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/module.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>
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
;
131 static struct prop_descriptor vm_dirties
;
134 * couple the period to the dirty_ratio:
136 * period/2 ~ roundup_pow_of_two(dirty limit)
138 static int calc_period_shift(void)
140 unsigned long dirty_total
;
143 dirty_total
= vm_dirty_bytes
/ PAGE_SIZE
;
145 dirty_total
= (vm_dirty_ratio
* determine_dirtyable_memory()) /
147 return 2 + ilog2(dirty_total
- 1);
151 * update the period when the dirty threshold changes.
153 static void update_completion_period(void)
155 int shift
= calc_period_shift();
156 prop_change_shift(&vm_completions
, shift
);
157 prop_change_shift(&vm_dirties
, shift
);
159 writeback_set_ratelimit();
162 int dirty_background_ratio_handler(struct ctl_table
*table
, int write
,
163 void __user
*buffer
, size_t *lenp
,
168 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
169 if (ret
== 0 && write
)
170 dirty_background_bytes
= 0;
174 int dirty_background_bytes_handler(struct ctl_table
*table
, int write
,
175 void __user
*buffer
, size_t *lenp
,
180 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
181 if (ret
== 0 && write
)
182 dirty_background_ratio
= 0;
186 int dirty_ratio_handler(struct ctl_table
*table
, int write
,
187 void __user
*buffer
, size_t *lenp
,
190 int old_ratio
= vm_dirty_ratio
;
193 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
194 if (ret
== 0 && write
&& vm_dirty_ratio
!= old_ratio
) {
195 update_completion_period();
202 int dirty_bytes_handler(struct ctl_table
*table
, int write
,
203 void __user
*buffer
, size_t *lenp
,
206 unsigned long old_bytes
= vm_dirty_bytes
;
209 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
210 if (ret
== 0 && write
&& vm_dirty_bytes
!= old_bytes
) {
211 update_completion_period();
218 * Increment the BDI's writeout completion count and the global writeout
219 * completion count. Called from test_clear_page_writeback().
221 static inline void __bdi_writeout_inc(struct backing_dev_info
*bdi
)
223 __inc_bdi_stat(bdi
, BDI_WRITTEN
);
224 __prop_inc_percpu_max(&vm_completions
, &bdi
->completions
,
228 void bdi_writeout_inc(struct backing_dev_info
*bdi
)
232 local_irq_save(flags
);
233 __bdi_writeout_inc(bdi
);
234 local_irq_restore(flags
);
236 EXPORT_SYMBOL_GPL(bdi_writeout_inc
);
238 void task_dirty_inc(struct task_struct
*tsk
)
240 prop_inc_single(&vm_dirties
, &tsk
->dirties
);
244 * Obtain an accurate fraction of the BDI's portion.
246 static void bdi_writeout_fraction(struct backing_dev_info
*bdi
,
247 long *numerator
, long *denominator
)
249 prop_fraction_percpu(&vm_completions
, &bdi
->completions
,
250 numerator
, denominator
);
256 static unsigned int bdi_min_ratio
;
258 int bdi_set_min_ratio(struct backing_dev_info
*bdi
, unsigned int min_ratio
)
262 spin_lock_bh(&bdi_lock
);
263 if (min_ratio
> bdi
->max_ratio
) {
266 min_ratio
-= bdi
->min_ratio
;
267 if (bdi_min_ratio
+ min_ratio
< 100) {
268 bdi_min_ratio
+= min_ratio
;
269 bdi
->min_ratio
+= min_ratio
;
274 spin_unlock_bh(&bdi_lock
);
279 int bdi_set_max_ratio(struct backing_dev_info
*bdi
, unsigned max_ratio
)
286 spin_lock_bh(&bdi_lock
);
287 if (bdi
->min_ratio
> max_ratio
) {
290 bdi
->max_ratio
= max_ratio
;
291 bdi
->max_prop_frac
= (PROP_FRAC_BASE
* max_ratio
) / 100;
293 spin_unlock_bh(&bdi_lock
);
297 EXPORT_SYMBOL(bdi_set_max_ratio
);
300 * Work out the current dirty-memory clamping and background writeout
303 * The main aim here is to lower them aggressively if there is a lot of mapped
304 * memory around. To avoid stressing page reclaim with lots of unreclaimable
305 * pages. It is better to clamp down on writers than to start swapping, and
306 * performing lots of scanning.
308 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
310 * We don't permit the clamping level to fall below 5% - that is getting rather
313 * We make sure that the background writeout level is below the adjusted
317 static unsigned long highmem_dirtyable_memory(unsigned long total
)
319 #ifdef CONFIG_HIGHMEM
323 for_each_node_state(node
, N_HIGH_MEMORY
) {
325 &NODE_DATA(node
)->node_zones
[ZONE_HIGHMEM
];
327 x
+= zone_page_state(z
, NR_FREE_PAGES
) +
328 zone_reclaimable_pages(z
);
331 * Make sure that the number of highmem pages is never larger
332 * than the number of the total dirtyable memory. This can only
333 * occur in very strange VM situations but we want to make sure
334 * that this does not occur.
336 return min(x
, total
);
343 * determine_dirtyable_memory - amount of memory that may be used
345 * Returns the numebr of pages that can currently be freed and used
346 * by the kernel for direct mappings.
348 unsigned long determine_dirtyable_memory(void)
352 x
= global_page_state(NR_FREE_PAGES
) + global_reclaimable_pages();
354 if (!vm_highmem_is_dirtyable
)
355 x
-= highmem_dirtyable_memory(x
);
357 return x
+ 1; /* Ensure that we never return 0 */
360 static unsigned long dirty_freerun_ceiling(unsigned long thresh
,
361 unsigned long bg_thresh
)
363 return (thresh
+ bg_thresh
) / 2;
366 static unsigned long hard_dirty_limit(unsigned long thresh
)
368 return max(thresh
, global_dirty_limit
);
372 * global_dirty_limits - background-writeback and dirty-throttling thresholds
374 * Calculate the dirty thresholds based on sysctl parameters
375 * - vm.dirty_background_ratio or vm.dirty_background_bytes
376 * - vm.dirty_ratio or vm.dirty_bytes
377 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
380 void global_dirty_limits(unsigned long *pbackground
, unsigned long *pdirty
)
382 unsigned long background
;
384 unsigned long uninitialized_var(available_memory
);
385 struct task_struct
*tsk
;
387 if (!vm_dirty_bytes
|| !dirty_background_bytes
)
388 available_memory
= determine_dirtyable_memory();
391 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
);
393 dirty
= (vm_dirty_ratio
* available_memory
) / 100;
395 if (dirty_background_bytes
)
396 background
= DIV_ROUND_UP(dirty_background_bytes
, PAGE_SIZE
);
398 background
= (dirty_background_ratio
* available_memory
) / 100;
400 if (background
>= dirty
)
401 background
= dirty
/ 2;
403 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
404 background
+= background
/ 4;
407 *pbackground
= background
;
409 trace_global_dirty_state(background
, dirty
);
413 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
414 * @bdi: the backing_dev_info to query
415 * @dirty: global dirty limit in pages
417 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
418 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
419 * And the "limit" in the name is not seriously taken as hard limit in
420 * balance_dirty_pages().
422 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
423 * - starving fast devices
424 * - piling up dirty pages (that will take long time to sync) on slow devices
426 * The bdi's share of dirty limit will be adapting to its throughput and
427 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
429 unsigned long bdi_dirty_limit(struct backing_dev_info
*bdi
, unsigned long dirty
)
432 long numerator
, denominator
;
435 * Calculate this BDI's share of the dirty ratio.
437 bdi_writeout_fraction(bdi
, &numerator
, &denominator
);
439 bdi_dirty
= (dirty
* (100 - bdi_min_ratio
)) / 100;
440 bdi_dirty
*= numerator
;
441 do_div(bdi_dirty
, denominator
);
443 bdi_dirty
+= (dirty
* bdi
->min_ratio
) / 100;
444 if (bdi_dirty
> (dirty
* bdi
->max_ratio
) / 100)
445 bdi_dirty
= dirty
* bdi
->max_ratio
/ 100;
451 * Dirty position control.
453 * (o) global/bdi setpoints
455 * We want the dirty pages be balanced around the global/bdi setpoints.
456 * When the number of dirty pages is higher/lower than the setpoint, the
457 * dirty position control ratio (and hence task dirty ratelimit) will be
458 * decreased/increased to bring the dirty pages back to the setpoint.
460 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
462 * if (dirty < setpoint) scale up pos_ratio
463 * if (dirty > setpoint) scale down pos_ratio
465 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
466 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
468 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
470 * (o) global control line
474 * | |<===== global dirty control scope ======>|
482 * 1.0 ................................*
488 * 0 +------------.------------------.----------------------*------------->
489 * freerun^ setpoint^ limit^ dirty pages
491 * (o) bdi control line
499 * | * |<=========== span ============>|
500 * 1.0 .......................*
512 * 1/4 ...............................................* * * * * * * * * * * *
516 * 0 +----------------------.-------------------------------.------------->
517 * bdi_setpoint^ x_intercept^
519 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
520 * be smoothly throttled down to normal if it starts high in situations like
521 * - start writing to a slow SD card and a fast disk at the same time. The SD
522 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
523 * - the bdi dirty thresh drops quickly due to change of JBOD workload
525 static unsigned long bdi_position_ratio(struct backing_dev_info
*bdi
,
526 unsigned long thresh
,
527 unsigned long bg_thresh
,
529 unsigned long bdi_thresh
,
530 unsigned long bdi_dirty
)
532 unsigned long write_bw
= bdi
->avg_write_bandwidth
;
533 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
534 unsigned long limit
= hard_dirty_limit(thresh
);
535 unsigned long x_intercept
;
536 unsigned long setpoint
; /* dirty pages' target balance point */
537 unsigned long bdi_setpoint
;
539 long long pos_ratio
; /* for scaling up/down the rate limit */
542 if (unlikely(dirty
>= limit
))
549 * f(dirty) := 1.0 + (----------------)
552 * it's a 3rd order polynomial that subjects to
554 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
555 * (2) f(setpoint) = 1.0 => the balance point
556 * (3) f(limit) = 0 => the hard limit
557 * (4) df/dx <= 0 => negative feedback control
558 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
559 * => fast response on large errors; small oscillation near setpoint
561 setpoint
= (freerun
+ limit
) / 2;
562 x
= div_s64((setpoint
- dirty
) << RATELIMIT_CALC_SHIFT
,
563 limit
- setpoint
+ 1);
565 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
566 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
567 pos_ratio
+= 1 << RATELIMIT_CALC_SHIFT
;
570 * We have computed basic pos_ratio above based on global situation. If
571 * the bdi is over/under its share of dirty pages, we want to scale
572 * pos_ratio further down/up. That is done by the following mechanism.
578 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
580 * x_intercept - bdi_dirty
581 * := --------------------------
582 * x_intercept - bdi_setpoint
584 * The main bdi control line is a linear function that subjects to
586 * (1) f(bdi_setpoint) = 1.0
587 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
588 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
590 * For single bdi case, the dirty pages are observed to fluctuate
591 * regularly within range
592 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
593 * for various filesystems, where (2) can yield in a reasonable 12.5%
594 * fluctuation range for pos_ratio.
596 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
597 * own size, so move the slope over accordingly and choose a slope that
598 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
600 if (unlikely(bdi_thresh
> thresh
))
603 * scale global setpoint to bdi's:
604 * bdi_setpoint = setpoint * bdi_thresh / thresh
606 x
= div_u64((u64
)bdi_thresh
<< 16, thresh
+ 1);
607 bdi_setpoint
= setpoint
* (u64
)x
>> 16;
609 * Use span=(8*write_bw) in single bdi case as indicated by
610 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
612 * bdi_thresh thresh - bdi_thresh
613 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
616 span
= (thresh
- bdi_thresh
+ 8 * write_bw
) * (u64
)x
>> 16;
617 x_intercept
= bdi_setpoint
+ span
;
619 if (bdi_dirty
< x_intercept
- span
/ 4) {
620 pos_ratio
*= x_intercept
- bdi_dirty
;
621 do_div(pos_ratio
, x_intercept
- bdi_setpoint
+ 1);
628 static void bdi_update_write_bandwidth(struct backing_dev_info
*bdi
,
629 unsigned long elapsed
,
630 unsigned long written
)
632 const unsigned long period
= roundup_pow_of_two(3 * HZ
);
633 unsigned long avg
= bdi
->avg_write_bandwidth
;
634 unsigned long old
= bdi
->write_bandwidth
;
638 * bw = written * HZ / elapsed
640 * bw * elapsed + write_bandwidth * (period - elapsed)
641 * write_bandwidth = ---------------------------------------------------
644 bw
= written
- bdi
->written_stamp
;
646 if (unlikely(elapsed
> period
)) {
651 bw
+= (u64
)bdi
->write_bandwidth
* (period
- elapsed
);
652 bw
>>= ilog2(period
);
655 * one more level of smoothing, for filtering out sudden spikes
657 if (avg
> old
&& old
>= (unsigned long)bw
)
658 avg
-= (avg
- old
) >> 3;
660 if (avg
< old
&& old
<= (unsigned long)bw
)
661 avg
+= (old
- avg
) >> 3;
664 bdi
->write_bandwidth
= bw
;
665 bdi
->avg_write_bandwidth
= avg
;
669 * The global dirtyable memory and dirty threshold could be suddenly knocked
670 * down by a large amount (eg. on the startup of KVM in a swapless system).
671 * This may throw the system into deep dirty exceeded state and throttle
672 * heavy/light dirtiers alike. To retain good responsiveness, maintain
673 * global_dirty_limit for tracking slowly down to the knocked down dirty
676 static void update_dirty_limit(unsigned long thresh
, unsigned long dirty
)
678 unsigned long limit
= global_dirty_limit
;
681 * Follow up in one step.
683 if (limit
< thresh
) {
689 * Follow down slowly. Use the higher one as the target, because thresh
690 * may drop below dirty. This is exactly the reason to introduce
691 * global_dirty_limit which is guaranteed to lie above the dirty pages.
693 thresh
= max(thresh
, dirty
);
694 if (limit
> thresh
) {
695 limit
-= (limit
- thresh
) >> 5;
700 global_dirty_limit
= limit
;
703 static void global_update_bandwidth(unsigned long thresh
,
707 static DEFINE_SPINLOCK(dirty_lock
);
708 static unsigned long update_time
;
711 * check locklessly first to optimize away locking for the most time
713 if (time_before(now
, update_time
+ BANDWIDTH_INTERVAL
))
716 spin_lock(&dirty_lock
);
717 if (time_after_eq(now
, update_time
+ BANDWIDTH_INTERVAL
)) {
718 update_dirty_limit(thresh
, dirty
);
721 spin_unlock(&dirty_lock
);
725 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
727 * Normal bdi tasks will be curbed at or below it in long term.
728 * Obviously it should be around (write_bw / N) when there are N dd tasks.
730 static void bdi_update_dirty_ratelimit(struct backing_dev_info
*bdi
,
731 unsigned long thresh
,
732 unsigned long bg_thresh
,
734 unsigned long bdi_thresh
,
735 unsigned long bdi_dirty
,
736 unsigned long dirtied
,
737 unsigned long elapsed
)
739 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
740 unsigned long limit
= hard_dirty_limit(thresh
);
741 unsigned long setpoint
= (freerun
+ limit
) / 2;
742 unsigned long write_bw
= bdi
->avg_write_bandwidth
;
743 unsigned long dirty_ratelimit
= bdi
->dirty_ratelimit
;
744 unsigned long dirty_rate
;
745 unsigned long task_ratelimit
;
746 unsigned long balanced_dirty_ratelimit
;
747 unsigned long pos_ratio
;
752 * The dirty rate will match the writeout rate in long term, except
753 * when dirty pages are truncated by userspace or re-dirtied by FS.
755 dirty_rate
= (dirtied
- bdi
->dirtied_stamp
) * HZ
/ elapsed
;
757 pos_ratio
= bdi_position_ratio(bdi
, thresh
, bg_thresh
, dirty
,
758 bdi_thresh
, bdi_dirty
);
760 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
762 task_ratelimit
= (u64
)dirty_ratelimit
*
763 pos_ratio
>> RATELIMIT_CALC_SHIFT
;
764 task_ratelimit
++; /* it helps rampup dirty_ratelimit from tiny values */
767 * A linear estimation of the "balanced" throttle rate. The theory is,
768 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
769 * dirty_rate will be measured to be (N * task_ratelimit). So the below
770 * formula will yield the balanced rate limit (write_bw / N).
772 * Note that the expanded form is not a pure rate feedback:
773 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
774 * but also takes pos_ratio into account:
775 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
777 * (1) is not realistic because pos_ratio also takes part in balancing
778 * the dirty rate. Consider the state
779 * pos_ratio = 0.5 (3)
780 * rate = 2 * (write_bw / N) (4)
781 * If (1) is used, it will stuck in that state! Because each dd will
783 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
785 * dirty_rate = N * task_ratelimit = write_bw (6)
786 * put (6) into (1) we get
787 * rate_(i+1) = rate_(i) (7)
789 * So we end up using (2) to always keep
790 * rate_(i+1) ~= (write_bw / N) (8)
791 * regardless of the value of pos_ratio. As long as (8) is satisfied,
792 * pos_ratio is able to drive itself to 1.0, which is not only where
793 * the dirty count meet the setpoint, but also where the slope of
794 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
796 balanced_dirty_ratelimit
= div_u64((u64
)task_ratelimit
* write_bw
,
800 * We could safely do this and return immediately:
802 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
804 * However to get a more stable dirty_ratelimit, the below elaborated
805 * code makes use of task_ratelimit to filter out sigular points and
806 * limit the step size.
808 * The below code essentially only uses the relative value of
810 * task_ratelimit - dirty_ratelimit
811 * = (pos_ratio - 1) * dirty_ratelimit
813 * which reflects the direction and size of dirty position error.
817 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
818 * task_ratelimit is on the same side of dirty_ratelimit, too.
820 * - dirty_ratelimit > balanced_dirty_ratelimit
821 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
822 * lowering dirty_ratelimit will help meet both the position and rate
823 * control targets. Otherwise, don't update dirty_ratelimit if it will
824 * only help meet the rate target. After all, what the users ultimately
825 * feel and care are stable dirty rate and small position error.
827 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
828 * and filter out the sigular points of balanced_dirty_ratelimit. Which
829 * keeps jumping around randomly and can even leap far away at times
830 * due to the small 200ms estimation period of dirty_rate (we want to
831 * keep that period small to reduce time lags).
834 if (dirty
< setpoint
) {
835 x
= min(bdi
->balanced_dirty_ratelimit
,
836 min(balanced_dirty_ratelimit
, task_ratelimit
));
837 if (dirty_ratelimit
< x
)
838 step
= x
- dirty_ratelimit
;
840 x
= max(bdi
->balanced_dirty_ratelimit
,
841 max(balanced_dirty_ratelimit
, task_ratelimit
));
842 if (dirty_ratelimit
> x
)
843 step
= dirty_ratelimit
- x
;
847 * Don't pursue 100% rate matching. It's impossible since the balanced
848 * rate itself is constantly fluctuating. So decrease the track speed
849 * when it gets close to the target. Helps eliminate pointless tremors.
851 step
>>= dirty_ratelimit
/ (2 * step
+ 1);
853 * Limit the tracking speed to avoid overshooting.
855 step
= (step
+ 7) / 8;
857 if (dirty_ratelimit
< balanced_dirty_ratelimit
)
858 dirty_ratelimit
+= step
;
860 dirty_ratelimit
-= step
;
862 bdi
->dirty_ratelimit
= max(dirty_ratelimit
, 1UL);
863 bdi
->balanced_dirty_ratelimit
= balanced_dirty_ratelimit
;
866 void __bdi_update_bandwidth(struct backing_dev_info
*bdi
,
867 unsigned long thresh
,
868 unsigned long bg_thresh
,
870 unsigned long bdi_thresh
,
871 unsigned long bdi_dirty
,
872 unsigned long start_time
)
874 unsigned long now
= jiffies
;
875 unsigned long elapsed
= now
- bdi
->bw_time_stamp
;
876 unsigned long dirtied
;
877 unsigned long written
;
880 * rate-limit, only update once every 200ms.
882 if (elapsed
< BANDWIDTH_INTERVAL
)
885 dirtied
= percpu_counter_read(&bdi
->bdi_stat
[BDI_DIRTIED
]);
886 written
= percpu_counter_read(&bdi
->bdi_stat
[BDI_WRITTEN
]);
889 * Skip quiet periods when disk bandwidth is under-utilized.
890 * (at least 1s idle time between two flusher runs)
892 if (elapsed
> HZ
&& time_before(bdi
->bw_time_stamp
, start_time
))
896 global_update_bandwidth(thresh
, dirty
, now
);
897 bdi_update_dirty_ratelimit(bdi
, thresh
, bg_thresh
, dirty
,
898 bdi_thresh
, bdi_dirty
,
901 bdi_update_write_bandwidth(bdi
, elapsed
, written
);
904 bdi
->dirtied_stamp
= dirtied
;
905 bdi
->written_stamp
= written
;
906 bdi
->bw_time_stamp
= now
;
909 static void bdi_update_bandwidth(struct backing_dev_info
*bdi
,
910 unsigned long thresh
,
911 unsigned long bg_thresh
,
913 unsigned long bdi_thresh
,
914 unsigned long bdi_dirty
,
915 unsigned long start_time
)
917 if (time_is_after_eq_jiffies(bdi
->bw_time_stamp
+ BANDWIDTH_INTERVAL
))
919 spin_lock(&bdi
->wb
.list_lock
);
920 __bdi_update_bandwidth(bdi
, thresh
, bg_thresh
, dirty
,
921 bdi_thresh
, bdi_dirty
, start_time
);
922 spin_unlock(&bdi
->wb
.list_lock
);
926 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
927 * will look to see if it needs to start dirty throttling.
929 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
930 * global_page_state() too often. So scale it near-sqrt to the safety margin
931 * (the number of pages we may dirty without exceeding the dirty limits).
933 static unsigned long dirty_poll_interval(unsigned long dirty
,
934 unsigned long thresh
)
937 return 1UL << (ilog2(thresh
- dirty
) >> 1);
942 static unsigned long bdi_max_pause(struct backing_dev_info
*bdi
,
943 unsigned long bdi_dirty
)
945 unsigned long bw
= bdi
->avg_write_bandwidth
;
946 unsigned long hi
= ilog2(bw
);
947 unsigned long lo
= ilog2(bdi
->dirty_ratelimit
);
950 /* target for 20ms max pause on 1-dd case */
954 * Scale up pause time for concurrent dirtiers in order to reduce CPU
957 * (N * 20ms) on 2^N concurrent tasks.
960 t
+= (hi
- lo
) * (20 * HZ
) / 1024;
963 * Limit pause time for small memory systems. If sleeping for too long
964 * time, a small pool of dirty/writeback pages may go empty and disk go
967 * 8 serves as the safety ratio.
970 t
= min(t
, bdi_dirty
* HZ
/ (8 * bw
+ 1));
973 * The pause time will be settled within range (max_pause/4, max_pause).
974 * Apply a minimal value of 4 to get a non-zero max_pause/4.
976 return clamp_val(t
, 4, MAX_PAUSE
);
980 * balance_dirty_pages() must be called by processes which are generating dirty
981 * data. It looks at the number of dirty pages in the machine and will force
982 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
983 * If we're over `background_thresh' then the writeback threads are woken to
984 * perform some writeout.
986 static void balance_dirty_pages(struct address_space
*mapping
,
987 unsigned long pages_dirtied
)
989 unsigned long nr_reclaimable
; /* = file_dirty + unstable_nfs */
990 unsigned long bdi_reclaimable
;
991 unsigned long nr_dirty
; /* = file_dirty + writeback + unstable_nfs */
992 unsigned long bdi_dirty
;
993 unsigned long freerun
;
994 unsigned long background_thresh
;
995 unsigned long dirty_thresh
;
996 unsigned long bdi_thresh
;
999 bool dirty_exceeded
= false;
1000 unsigned long task_ratelimit
;
1001 unsigned long dirty_ratelimit
;
1002 unsigned long pos_ratio
;
1003 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1004 unsigned long start_time
= jiffies
;
1008 * Unstable writes are a feature of certain networked
1009 * filesystems (i.e. NFS) in which data may have been
1010 * written to the server's write cache, but has not yet
1011 * been flushed to permanent storage.
1013 nr_reclaimable
= global_page_state(NR_FILE_DIRTY
) +
1014 global_page_state(NR_UNSTABLE_NFS
);
1015 nr_dirty
= nr_reclaimable
+ global_page_state(NR_WRITEBACK
);
1017 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1020 * Throttle it only when the background writeback cannot
1021 * catch-up. This avoids (excessively) small writeouts
1022 * when the bdi limits are ramping up.
1024 freerun
= dirty_freerun_ceiling(dirty_thresh
,
1026 if (nr_dirty
<= freerun
)
1029 if (unlikely(!writeback_in_progress(bdi
)))
1030 bdi_start_background_writeback(bdi
);
1033 * bdi_thresh is not treated as some limiting factor as
1034 * dirty_thresh, due to reasons
1035 * - in JBOD setup, bdi_thresh can fluctuate a lot
1036 * - in a system with HDD and USB key, the USB key may somehow
1037 * go into state (bdi_dirty >> bdi_thresh) either because
1038 * bdi_dirty starts high, or because bdi_thresh drops low.
1039 * In this case we don't want to hard throttle the USB key
1040 * dirtiers for 100 seconds until bdi_dirty drops under
1041 * bdi_thresh. Instead the auxiliary bdi control line in
1042 * bdi_position_ratio() will let the dirtier task progress
1043 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1045 bdi_thresh
= bdi_dirty_limit(bdi
, dirty_thresh
);
1048 * In order to avoid the stacked BDI deadlock we need
1049 * to ensure we accurately count the 'dirty' pages when
1050 * the threshold is low.
1052 * Otherwise it would be possible to get thresh+n pages
1053 * reported dirty, even though there are thresh-m pages
1054 * actually dirty; with m+n sitting in the percpu
1057 if (bdi_thresh
< 2 * bdi_stat_error(bdi
)) {
1058 bdi_reclaimable
= bdi_stat_sum(bdi
, BDI_RECLAIMABLE
);
1059 bdi_dirty
= bdi_reclaimable
+
1060 bdi_stat_sum(bdi
, BDI_WRITEBACK
);
1062 bdi_reclaimable
= bdi_stat(bdi
, BDI_RECLAIMABLE
);
1063 bdi_dirty
= bdi_reclaimable
+
1064 bdi_stat(bdi
, BDI_WRITEBACK
);
1067 dirty_exceeded
= (bdi_dirty
> bdi_thresh
) ||
1068 (nr_dirty
> dirty_thresh
);
1069 if (dirty_exceeded
&& !bdi
->dirty_exceeded
)
1070 bdi
->dirty_exceeded
= 1;
1072 bdi_update_bandwidth(bdi
, dirty_thresh
, background_thresh
,
1073 nr_dirty
, bdi_thresh
, bdi_dirty
,
1076 max_pause
= bdi_max_pause(bdi
, bdi_dirty
);
1078 dirty_ratelimit
= bdi
->dirty_ratelimit
;
1079 pos_ratio
= bdi_position_ratio(bdi
, dirty_thresh
,
1080 background_thresh
, nr_dirty
,
1081 bdi_thresh
, bdi_dirty
);
1082 if (unlikely(pos_ratio
== 0)) {
1086 task_ratelimit
= (u64
)dirty_ratelimit
*
1087 pos_ratio
>> RATELIMIT_CALC_SHIFT
;
1088 pause
= (HZ
* pages_dirtied
) / (task_ratelimit
| 1);
1089 if (unlikely(pause
<= 0)) {
1090 pause
= 1; /* avoid resetting nr_dirtied_pause below */
1093 pause
= min(pause
, max_pause
);
1096 __set_current_state(TASK_UNINTERRUPTIBLE
);
1097 io_schedule_timeout(pause
);
1099 dirty_thresh
= hard_dirty_limit(dirty_thresh
);
1101 * max-pause area. If dirty exceeded but still within this
1102 * area, no need to sleep for more than 200ms: (a) 8 pages per
1103 * 200ms is typically more than enough to curb heavy dirtiers;
1104 * (b) the pause time limit makes the dirtiers more responsive.
1106 if (nr_dirty
< dirty_thresh
)
1110 if (!dirty_exceeded
&& bdi
->dirty_exceeded
)
1111 bdi
->dirty_exceeded
= 0;
1113 current
->nr_dirtied
= 0;
1114 if (pause
== 0) { /* in freerun area */
1115 current
->nr_dirtied_pause
=
1116 dirty_poll_interval(nr_dirty
, dirty_thresh
);
1117 } else if (pause
<= max_pause
/ 4 &&
1118 pages_dirtied
>= current
->nr_dirtied_pause
) {
1119 current
->nr_dirtied_pause
= clamp_val(
1120 dirty_ratelimit
* (max_pause
/ 2) / HZ
,
1121 pages_dirtied
+ pages_dirtied
/ 8,
1123 } else if (pause
>= max_pause
) {
1124 current
->nr_dirtied_pause
= 1 | clamp_val(
1125 dirty_ratelimit
* (max_pause
/ 2) / HZ
,
1127 pages_dirtied
- pages_dirtied
/ 8);
1130 if (writeback_in_progress(bdi
))
1134 * In laptop mode, we wait until hitting the higher threshold before
1135 * starting background writeout, and then write out all the way down
1136 * to the lower threshold. So slow writers cause minimal disk activity.
1138 * In normal mode, we start background writeout at the lower
1139 * background_thresh, to keep the amount of dirty memory low.
1144 if (nr_reclaimable
> background_thresh
)
1145 bdi_start_background_writeback(bdi
);
1148 void set_page_dirty_balance(struct page
*page
, int page_mkwrite
)
1150 if (set_page_dirty(page
) || page_mkwrite
) {
1151 struct address_space
*mapping
= page_mapping(page
);
1154 balance_dirty_pages_ratelimited(mapping
);
1158 static DEFINE_PER_CPU(int, bdp_ratelimits
);
1161 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1162 * @mapping: address_space which was dirtied
1163 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1165 * Processes which are dirtying memory should call in here once for each page
1166 * which was newly dirtied. The function will periodically check the system's
1167 * dirty state and will initiate writeback if needed.
1169 * On really big machines, get_writeback_state is expensive, so try to avoid
1170 * calling it too often (ratelimiting). But once we're over the dirty memory
1171 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1172 * from overshooting the limit by (ratelimit_pages) each.
1174 void balance_dirty_pages_ratelimited_nr(struct address_space
*mapping
,
1175 unsigned long nr_pages_dirtied
)
1177 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1181 if (!bdi_cap_account_dirty(bdi
))
1184 ratelimit
= current
->nr_dirtied_pause
;
1185 if (bdi
->dirty_exceeded
)
1186 ratelimit
= min(ratelimit
, 32 >> (PAGE_SHIFT
- 10));
1188 current
->nr_dirtied
+= nr_pages_dirtied
;
1192 * This prevents one CPU to accumulate too many dirtied pages without
1193 * calling into balance_dirty_pages(), which can happen when there are
1194 * 1000+ tasks, all of them start dirtying pages at exactly the same
1195 * time, hence all honoured too large initial task->nr_dirtied_pause.
1197 p
= &__get_cpu_var(bdp_ratelimits
);
1198 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1201 *p
+= nr_pages_dirtied
;
1202 if (unlikely(*p
>= ratelimit_pages
)) {
1209 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1210 balance_dirty_pages(mapping
, current
->nr_dirtied
);
1212 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr
);
1214 void throttle_vm_writeout(gfp_t gfp_mask
)
1216 unsigned long background_thresh
;
1217 unsigned long dirty_thresh
;
1220 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1223 * Boost the allowable dirty threshold a bit for page
1224 * allocators so they don't get DoS'ed by heavy writers
1226 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
1228 if (global_page_state(NR_UNSTABLE_NFS
) +
1229 global_page_state(NR_WRITEBACK
) <= dirty_thresh
)
1231 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1234 * The caller might hold locks which can prevent IO completion
1235 * or progress in the filesystem. So we cannot just sit here
1236 * waiting for IO to complete.
1238 if ((gfp_mask
& (__GFP_FS
|__GFP_IO
)) != (__GFP_FS
|__GFP_IO
))
1244 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1246 int dirty_writeback_centisecs_handler(ctl_table
*table
, int write
,
1247 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1249 proc_dointvec(table
, write
, buffer
, length
, ppos
);
1250 bdi_arm_supers_timer();
1255 void laptop_mode_timer_fn(unsigned long data
)
1257 struct request_queue
*q
= (struct request_queue
*)data
;
1258 int nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1259 global_page_state(NR_UNSTABLE_NFS
);
1262 * We want to write everything out, not just down to the dirty
1265 if (bdi_has_dirty_io(&q
->backing_dev_info
))
1266 bdi_start_writeback(&q
->backing_dev_info
, nr_pages
);
1270 * We've spun up the disk and we're in laptop mode: schedule writeback
1271 * of all dirty data a few seconds from now. If the flush is already scheduled
1272 * then push it back - the user is still using the disk.
1274 void laptop_io_completion(struct backing_dev_info
*info
)
1276 mod_timer(&info
->laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
1280 * We're in laptop mode and we've just synced. The sync's writes will have
1281 * caused another writeback to be scheduled by laptop_io_completion.
1282 * Nothing needs to be written back anymore, so we unschedule the writeback.
1284 void laptop_sync_completion(void)
1286 struct backing_dev_info
*bdi
;
1290 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
1291 del_timer(&bdi
->laptop_mode_wb_timer
);
1298 * If ratelimit_pages is too high then we can get into dirty-data overload
1299 * if a large number of processes all perform writes at the same time.
1300 * If it is too low then SMP machines will call the (expensive)
1301 * get_writeback_state too often.
1303 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1304 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1308 void writeback_set_ratelimit(void)
1310 unsigned long background_thresh
;
1311 unsigned long dirty_thresh
;
1312 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1313 ratelimit_pages
= dirty_thresh
/ (num_online_cpus() * 32);
1314 if (ratelimit_pages
< 16)
1315 ratelimit_pages
= 16;
1318 static int __cpuinit
1319 ratelimit_handler(struct notifier_block
*self
, unsigned long u
, void *v
)
1321 writeback_set_ratelimit();
1325 static struct notifier_block __cpuinitdata ratelimit_nb
= {
1326 .notifier_call
= ratelimit_handler
,
1331 * Called early on to tune the page writeback dirty limits.
1333 * We used to scale dirty pages according to how total memory
1334 * related to pages that could be allocated for buffers (by
1335 * comparing nr_free_buffer_pages() to vm_total_pages.
1337 * However, that was when we used "dirty_ratio" to scale with
1338 * all memory, and we don't do that any more. "dirty_ratio"
1339 * is now applied to total non-HIGHPAGE memory (by subtracting
1340 * totalhigh_pages from vm_total_pages), and as such we can't
1341 * get into the old insane situation any more where we had
1342 * large amounts of dirty pages compared to a small amount of
1343 * non-HIGHMEM memory.
1345 * But we might still want to scale the dirty_ratio by how
1346 * much memory the box has..
1348 void __init
page_writeback_init(void)
1352 writeback_set_ratelimit();
1353 register_cpu_notifier(&ratelimit_nb
);
1355 shift
= calc_period_shift();
1356 prop_descriptor_init(&vm_completions
, shift
);
1357 prop_descriptor_init(&vm_dirties
, shift
);
1361 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1362 * @mapping: address space structure to write
1363 * @start: starting page index
1364 * @end: ending page index (inclusive)
1366 * This function scans the page range from @start to @end (inclusive) and tags
1367 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1368 * that write_cache_pages (or whoever calls this function) will then use
1369 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1370 * used to avoid livelocking of writeback by a process steadily creating new
1371 * dirty pages in the file (thus it is important for this function to be quick
1372 * so that it can tag pages faster than a dirtying process can create them).
1375 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1377 void tag_pages_for_writeback(struct address_space
*mapping
,
1378 pgoff_t start
, pgoff_t end
)
1380 #define WRITEBACK_TAG_BATCH 4096
1381 unsigned long tagged
;
1384 spin_lock_irq(&mapping
->tree_lock
);
1385 tagged
= radix_tree_range_tag_if_tagged(&mapping
->page_tree
,
1386 &start
, end
, WRITEBACK_TAG_BATCH
,
1387 PAGECACHE_TAG_DIRTY
, PAGECACHE_TAG_TOWRITE
);
1388 spin_unlock_irq(&mapping
->tree_lock
);
1389 WARN_ON_ONCE(tagged
> WRITEBACK_TAG_BATCH
);
1391 /* We check 'start' to handle wrapping when end == ~0UL */
1392 } while (tagged
>= WRITEBACK_TAG_BATCH
&& start
);
1394 EXPORT_SYMBOL(tag_pages_for_writeback
);
1397 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1398 * @mapping: address space structure to write
1399 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1400 * @writepage: function called for each page
1401 * @data: data passed to writepage function
1403 * If a page is already under I/O, write_cache_pages() skips it, even
1404 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1405 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1406 * and msync() need to guarantee that all the data which was dirty at the time
1407 * the call was made get new I/O started against them. If wbc->sync_mode is
1408 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1409 * existing IO to complete.
1411 * To avoid livelocks (when other process dirties new pages), we first tag
1412 * pages which should be written back with TOWRITE tag and only then start
1413 * writing them. For data-integrity sync we have to be careful so that we do
1414 * not miss some pages (e.g., because some other process has cleared TOWRITE
1415 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1416 * by the process clearing the DIRTY tag (and submitting the page for IO).
1418 int write_cache_pages(struct address_space
*mapping
,
1419 struct writeback_control
*wbc
, writepage_t writepage
,
1424 struct pagevec pvec
;
1426 pgoff_t
uninitialized_var(writeback_index
);
1428 pgoff_t end
; /* Inclusive */
1431 int range_whole
= 0;
1434 pagevec_init(&pvec
, 0);
1435 if (wbc
->range_cyclic
) {
1436 writeback_index
= mapping
->writeback_index
; /* prev offset */
1437 index
= writeback_index
;
1444 index
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1445 end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1446 if (wbc
->range_start
== 0 && wbc
->range_end
== LLONG_MAX
)
1448 cycled
= 1; /* ignore range_cyclic tests */
1450 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1451 tag
= PAGECACHE_TAG_TOWRITE
;
1453 tag
= PAGECACHE_TAG_DIRTY
;
1455 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1456 tag_pages_for_writeback(mapping
, index
, end
);
1458 while (!done
&& (index
<= end
)) {
1461 nr_pages
= pagevec_lookup_tag(&pvec
, mapping
, &index
, tag
,
1462 min(end
- index
, (pgoff_t
)PAGEVEC_SIZE
-1) + 1);
1466 for (i
= 0; i
< nr_pages
; i
++) {
1467 struct page
*page
= pvec
.pages
[i
];
1470 * At this point, the page may be truncated or
1471 * invalidated (changing page->mapping to NULL), or
1472 * even swizzled back from swapper_space to tmpfs file
1473 * mapping. However, page->index will not change
1474 * because we have a reference on the page.
1476 if (page
->index
> end
) {
1478 * can't be range_cyclic (1st pass) because
1479 * end == -1 in that case.
1485 done_index
= page
->index
;
1490 * Page truncated or invalidated. We can freely skip it
1491 * then, even for data integrity operations: the page
1492 * has disappeared concurrently, so there could be no
1493 * real expectation of this data interity operation
1494 * even if there is now a new, dirty page at the same
1495 * pagecache address.
1497 if (unlikely(page
->mapping
!= mapping
)) {
1503 if (!PageDirty(page
)) {
1504 /* someone wrote it for us */
1505 goto continue_unlock
;
1508 if (PageWriteback(page
)) {
1509 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
1510 wait_on_page_writeback(page
);
1512 goto continue_unlock
;
1515 BUG_ON(PageWriteback(page
));
1516 if (!clear_page_dirty_for_io(page
))
1517 goto continue_unlock
;
1519 trace_wbc_writepage(wbc
, mapping
->backing_dev_info
);
1520 ret
= (*writepage
)(page
, wbc
, data
);
1521 if (unlikely(ret
)) {
1522 if (ret
== AOP_WRITEPAGE_ACTIVATE
) {
1527 * done_index is set past this page,
1528 * so media errors will not choke
1529 * background writeout for the entire
1530 * file. This has consequences for
1531 * range_cyclic semantics (ie. it may
1532 * not be suitable for data integrity
1535 done_index
= page
->index
+ 1;
1542 * We stop writing back only if we are not doing
1543 * integrity sync. In case of integrity sync we have to
1544 * keep going until we have written all the pages
1545 * we tagged for writeback prior to entering this loop.
1547 if (--wbc
->nr_to_write
<= 0 &&
1548 wbc
->sync_mode
== WB_SYNC_NONE
) {
1553 pagevec_release(&pvec
);
1556 if (!cycled
&& !done
) {
1559 * We hit the last page and there is more work to be done: wrap
1560 * back to the start of the file
1564 end
= writeback_index
- 1;
1567 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0))
1568 mapping
->writeback_index
= done_index
;
1572 EXPORT_SYMBOL(write_cache_pages
);
1575 * Function used by generic_writepages to call the real writepage
1576 * function and set the mapping flags on error
1578 static int __writepage(struct page
*page
, struct writeback_control
*wbc
,
1581 struct address_space
*mapping
= data
;
1582 int ret
= mapping
->a_ops
->writepage(page
, wbc
);
1583 mapping_set_error(mapping
, ret
);
1588 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1589 * @mapping: address space structure to write
1590 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1592 * This is a library function, which implements the writepages()
1593 * address_space_operation.
1595 int generic_writepages(struct address_space
*mapping
,
1596 struct writeback_control
*wbc
)
1598 struct blk_plug plug
;
1601 /* deal with chardevs and other special file */
1602 if (!mapping
->a_ops
->writepage
)
1605 blk_start_plug(&plug
);
1606 ret
= write_cache_pages(mapping
, wbc
, __writepage
, mapping
);
1607 blk_finish_plug(&plug
);
1611 EXPORT_SYMBOL(generic_writepages
);
1613 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
1617 if (wbc
->nr_to_write
<= 0)
1619 if (mapping
->a_ops
->writepages
)
1620 ret
= mapping
->a_ops
->writepages(mapping
, wbc
);
1622 ret
= generic_writepages(mapping
, wbc
);
1627 * write_one_page - write out a single page and optionally wait on I/O
1628 * @page: the page to write
1629 * @wait: if true, wait on writeout
1631 * The page must be locked by the caller and will be unlocked upon return.
1633 * write_one_page() returns a negative error code if I/O failed.
1635 int write_one_page(struct page
*page
, int wait
)
1637 struct address_space
*mapping
= page
->mapping
;
1639 struct writeback_control wbc
= {
1640 .sync_mode
= WB_SYNC_ALL
,
1644 BUG_ON(!PageLocked(page
));
1647 wait_on_page_writeback(page
);
1649 if (clear_page_dirty_for_io(page
)) {
1650 page_cache_get(page
);
1651 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
1652 if (ret
== 0 && wait
) {
1653 wait_on_page_writeback(page
);
1654 if (PageError(page
))
1657 page_cache_release(page
);
1663 EXPORT_SYMBOL(write_one_page
);
1666 * For address_spaces which do not use buffers nor write back.
1668 int __set_page_dirty_no_writeback(struct page
*page
)
1670 if (!PageDirty(page
))
1671 return !TestSetPageDirty(page
);
1676 * Helper function for set_page_dirty family.
1677 * NOTE: This relies on being atomic wrt interrupts.
1679 void account_page_dirtied(struct page
*page
, struct address_space
*mapping
)
1681 if (mapping_cap_account_dirty(mapping
)) {
1682 __inc_zone_page_state(page
, NR_FILE_DIRTY
);
1683 __inc_zone_page_state(page
, NR_DIRTIED
);
1684 __inc_bdi_stat(mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
1685 __inc_bdi_stat(mapping
->backing_dev_info
, BDI_DIRTIED
);
1686 task_dirty_inc(current
);
1687 task_io_account_write(PAGE_CACHE_SIZE
);
1690 EXPORT_SYMBOL(account_page_dirtied
);
1693 * Helper function for set_page_writeback family.
1694 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1697 void account_page_writeback(struct page
*page
)
1699 inc_zone_page_state(page
, NR_WRITEBACK
);
1701 EXPORT_SYMBOL(account_page_writeback
);
1704 * For address_spaces which do not use buffers. Just tag the page as dirty in
1707 * This is also used when a single buffer is being dirtied: we want to set the
1708 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1709 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1711 * Most callers have locked the page, which pins the address_space in memory.
1712 * But zap_pte_range() does not lock the page, however in that case the
1713 * mapping is pinned by the vma's ->vm_file reference.
1715 * We take care to handle the case where the page was truncated from the
1716 * mapping by re-checking page_mapping() inside tree_lock.
1718 int __set_page_dirty_nobuffers(struct page
*page
)
1720 if (!TestSetPageDirty(page
)) {
1721 struct address_space
*mapping
= page_mapping(page
);
1722 struct address_space
*mapping2
;
1727 spin_lock_irq(&mapping
->tree_lock
);
1728 mapping2
= page_mapping(page
);
1729 if (mapping2
) { /* Race with truncate? */
1730 BUG_ON(mapping2
!= mapping
);
1731 WARN_ON_ONCE(!PagePrivate(page
) && !PageUptodate(page
));
1732 account_page_dirtied(page
, mapping
);
1733 radix_tree_tag_set(&mapping
->page_tree
,
1734 page_index(page
), PAGECACHE_TAG_DIRTY
);
1736 spin_unlock_irq(&mapping
->tree_lock
);
1737 if (mapping
->host
) {
1738 /* !PageAnon && !swapper_space */
1739 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
1745 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
1748 * When a writepage implementation decides that it doesn't want to write this
1749 * page for some reason, it should redirty the locked page via
1750 * redirty_page_for_writepage() and it should then unlock the page and return 0
1752 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
1754 wbc
->pages_skipped
++;
1755 return __set_page_dirty_nobuffers(page
);
1757 EXPORT_SYMBOL(redirty_page_for_writepage
);
1762 * For pages with a mapping this should be done under the page lock
1763 * for the benefit of asynchronous memory errors who prefer a consistent
1764 * dirty state. This rule can be broken in some special cases,
1765 * but should be better not to.
1767 * If the mapping doesn't provide a set_page_dirty a_op, then
1768 * just fall through and assume that it wants buffer_heads.
1770 int set_page_dirty(struct page
*page
)
1772 struct address_space
*mapping
= page_mapping(page
);
1774 if (likely(mapping
)) {
1775 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
1777 * readahead/lru_deactivate_page could remain
1778 * PG_readahead/PG_reclaim due to race with end_page_writeback
1779 * About readahead, if the page is written, the flags would be
1780 * reset. So no problem.
1781 * About lru_deactivate_page, if the page is redirty, the flag
1782 * will be reset. So no problem. but if the page is used by readahead
1783 * it will confuse readahead and make it restart the size rampup
1784 * process. But it's a trivial problem.
1786 ClearPageReclaim(page
);
1789 spd
= __set_page_dirty_buffers
;
1791 return (*spd
)(page
);
1793 if (!PageDirty(page
)) {
1794 if (!TestSetPageDirty(page
))
1799 EXPORT_SYMBOL(set_page_dirty
);
1802 * set_page_dirty() is racy if the caller has no reference against
1803 * page->mapping->host, and if the page is unlocked. This is because another
1804 * CPU could truncate the page off the mapping and then free the mapping.
1806 * Usually, the page _is_ locked, or the caller is a user-space process which
1807 * holds a reference on the inode by having an open file.
1809 * In other cases, the page should be locked before running set_page_dirty().
1811 int set_page_dirty_lock(struct page
*page
)
1816 ret
= set_page_dirty(page
);
1820 EXPORT_SYMBOL(set_page_dirty_lock
);
1823 * Clear a page's dirty flag, while caring for dirty memory accounting.
1824 * Returns true if the page was previously dirty.
1826 * This is for preparing to put the page under writeout. We leave the page
1827 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1828 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1829 * implementation will run either set_page_writeback() or set_page_dirty(),
1830 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1833 * This incoherency between the page's dirty flag and radix-tree tag is
1834 * unfortunate, but it only exists while the page is locked.
1836 int clear_page_dirty_for_io(struct page
*page
)
1838 struct address_space
*mapping
= page_mapping(page
);
1840 BUG_ON(!PageLocked(page
));
1842 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
1844 * Yes, Virginia, this is indeed insane.
1846 * We use this sequence to make sure that
1847 * (a) we account for dirty stats properly
1848 * (b) we tell the low-level filesystem to
1849 * mark the whole page dirty if it was
1850 * dirty in a pagetable. Only to then
1851 * (c) clean the page again and return 1 to
1852 * cause the writeback.
1854 * This way we avoid all nasty races with the
1855 * dirty bit in multiple places and clearing
1856 * them concurrently from different threads.
1858 * Note! Normally the "set_page_dirty(page)"
1859 * has no effect on the actual dirty bit - since
1860 * that will already usually be set. But we
1861 * need the side effects, and it can help us
1864 * We basically use the page "master dirty bit"
1865 * as a serialization point for all the different
1866 * threads doing their things.
1868 if (page_mkclean(page
))
1869 set_page_dirty(page
);
1871 * We carefully synchronise fault handlers against
1872 * installing a dirty pte and marking the page dirty
1873 * at this point. We do this by having them hold the
1874 * page lock at some point after installing their
1875 * pte, but before marking the page dirty.
1876 * Pages are always locked coming in here, so we get
1877 * the desired exclusion. See mm/memory.c:do_wp_page()
1878 * for more comments.
1880 if (TestClearPageDirty(page
)) {
1881 dec_zone_page_state(page
, NR_FILE_DIRTY
);
1882 dec_bdi_stat(mapping
->backing_dev_info
,
1888 return TestClearPageDirty(page
);
1890 EXPORT_SYMBOL(clear_page_dirty_for_io
);
1892 int test_clear_page_writeback(struct page
*page
)
1894 struct address_space
*mapping
= page_mapping(page
);
1898 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1899 unsigned long flags
;
1901 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
1902 ret
= TestClearPageWriteback(page
);
1904 radix_tree_tag_clear(&mapping
->page_tree
,
1906 PAGECACHE_TAG_WRITEBACK
);
1907 if (bdi_cap_account_writeback(bdi
)) {
1908 __dec_bdi_stat(bdi
, BDI_WRITEBACK
);
1909 __bdi_writeout_inc(bdi
);
1912 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
1914 ret
= TestClearPageWriteback(page
);
1917 dec_zone_page_state(page
, NR_WRITEBACK
);
1918 inc_zone_page_state(page
, NR_WRITTEN
);
1923 int test_set_page_writeback(struct page
*page
)
1925 struct address_space
*mapping
= page_mapping(page
);
1929 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1930 unsigned long flags
;
1932 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
1933 ret
= TestSetPageWriteback(page
);
1935 radix_tree_tag_set(&mapping
->page_tree
,
1937 PAGECACHE_TAG_WRITEBACK
);
1938 if (bdi_cap_account_writeback(bdi
))
1939 __inc_bdi_stat(bdi
, BDI_WRITEBACK
);
1941 if (!PageDirty(page
))
1942 radix_tree_tag_clear(&mapping
->page_tree
,
1944 PAGECACHE_TAG_DIRTY
);
1945 radix_tree_tag_clear(&mapping
->page_tree
,
1947 PAGECACHE_TAG_TOWRITE
);
1948 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
1950 ret
= TestSetPageWriteback(page
);
1953 account_page_writeback(page
);
1957 EXPORT_SYMBOL(test_set_page_writeback
);
1960 * Return true if any of the pages in the mapping are marked with the
1963 int mapping_tagged(struct address_space
*mapping
, int tag
)
1965 return radix_tree_tagged(&mapping
->page_tree
, tag
);
1967 EXPORT_SYMBOL(mapping_tagged
);