Loading...
1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/export.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
25#include <linux/task_io_accounting_ops.h>
26#include <linux/blkdev.h>
27#include <linux/mpage.h>
28#include <linux/rmap.h>
29#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/mm_inline.h>
40#include <trace/events/writeback.h>
41
42#include "internal.h"
43
44/*
45 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
56 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
60#define RATELIMIT_CALC_SHIFT 10
61
62/*
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
68/* The following parameters are exported via /proc/sys/vm */
69
70/*
71 * Start background writeback (via writeback threads) at this percentage
72 */
73int dirty_background_ratio = 10;
74
75/*
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
88 * The generator of dirty data starts writeback at this percentage
89 */
90int vm_dirty_ratio = 20;
91
92/*
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
99 * The interval between `kupdate'-style writebacks
100 */
101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102
103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
105/*
106 * The longest time for which data is allowed to remain dirty
107 */
108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int block_dump;
114
115/*
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
118 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
125struct wb_domain global_wb_domain;
126
127/* consolidated parameters for balance_dirty_pages() and its subroutines */
128struct dirty_throttle_control {
129#ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain *dom;
131 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
132#endif
133 struct bdi_writeback *wb;
134 struct fprop_local_percpu *wb_completions;
135
136 unsigned long avail; /* dirtyable */
137 unsigned long dirty; /* file_dirty + write + nfs */
138 unsigned long thresh; /* dirty threshold */
139 unsigned long bg_thresh; /* dirty background threshold */
140
141 unsigned long wb_dirty; /* per-wb counterparts */
142 unsigned long wb_thresh;
143 unsigned long wb_bg_thresh;
144
145 unsigned long pos_ratio;
146};
147
148/*
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
152 */
153#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
154
155#ifdef CONFIG_CGROUP_WRITEBACK
156
157#define GDTC_INIT(__wb) .wb = (__wb), \
158 .dom = &global_wb_domain, \
159 .wb_completions = &(__wb)->completions
160
161#define GDTC_INIT_NO_WB .dom = &global_wb_domain
162
163#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
164 .dom = mem_cgroup_wb_domain(__wb), \
165 .wb_completions = &(__wb)->memcg_completions, \
166 .gdtc = __gdtc
167
168static bool mdtc_valid(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
172
173static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174{
175 return dtc->dom;
176}
177
178static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179{
180 return mdtc->gdtc;
181}
182
183static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184{
185 return &wb->memcg_completions;
186}
187
188static void wb_min_max_ratio(struct bdi_writeback *wb,
189 unsigned long *minp, unsigned long *maxp)
190{
191 unsigned long this_bw = wb->avg_write_bandwidth;
192 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193 unsigned long long min = wb->bdi->min_ratio;
194 unsigned long long max = wb->bdi->max_ratio;
195
196 /*
197 * @wb may already be clean by the time control reaches here and
198 * the total may not include its bw.
199 */
200 if (this_bw < tot_bw) {
201 if (min) {
202 min *= this_bw;
203 do_div(min, tot_bw);
204 }
205 if (max < 100) {
206 max *= this_bw;
207 do_div(max, tot_bw);
208 }
209 }
210
211 *minp = min;
212 *maxp = max;
213}
214
215#else /* CONFIG_CGROUP_WRITEBACK */
216
217#define GDTC_INIT(__wb) .wb = (__wb), \
218 .wb_completions = &(__wb)->completions
219#define GDTC_INIT_NO_WB
220#define MDTC_INIT(__wb, __gdtc)
221
222static bool mdtc_valid(struct dirty_throttle_control *dtc)
223{
224 return false;
225}
226
227static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228{
229 return &global_wb_domain;
230}
231
232static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233{
234 return NULL;
235}
236
237static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238{
239 return NULL;
240}
241
242static void wb_min_max_ratio(struct bdi_writeback *wb,
243 unsigned long *minp, unsigned long *maxp)
244{
245 *minp = wb->bdi->min_ratio;
246 *maxp = wb->bdi->max_ratio;
247}
248
249#endif /* CONFIG_CGROUP_WRITEBACK */
250
251/*
252 * In a memory zone, there is a certain amount of pages we consider
253 * available for the page cache, which is essentially the number of
254 * free and reclaimable pages, minus some zone reserves to protect
255 * lowmem and the ability to uphold the zone's watermarks without
256 * requiring writeback.
257 *
258 * This number of dirtyable pages is the base value of which the
259 * user-configurable dirty ratio is the effictive number of pages that
260 * are allowed to be actually dirtied. Per individual zone, or
261 * globally by using the sum of dirtyable pages over all zones.
262 *
263 * Because the user is allowed to specify the dirty limit globally as
264 * absolute number of bytes, calculating the per-zone dirty limit can
265 * require translating the configured limit into a percentage of
266 * global dirtyable memory first.
267 */
268
269/**
270 * zone_dirtyable_memory - number of dirtyable pages in a zone
271 * @zone: the zone
272 *
273 * Returns the zone's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-zone dirty limits.
275 */
276static unsigned long zone_dirtyable_memory(struct zone *zone)
277{
278 unsigned long nr_pages;
279
280 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
281 /*
282 * Pages reserved for the kernel should not be considered
283 * dirtyable, to prevent a situation where reclaim has to
284 * clean pages in order to balance the zones.
285 */
286 nr_pages -= min(nr_pages, zone->totalreserve_pages);
287
288 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
289 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
290
291 return nr_pages;
292}
293
294static unsigned long highmem_dirtyable_memory(unsigned long total)
295{
296#ifdef CONFIG_HIGHMEM
297 int node;
298 unsigned long x = 0;
299
300 for_each_node_state(node, N_HIGH_MEMORY) {
301 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
302
303 x += zone_dirtyable_memory(z);
304 }
305 /*
306 * Unreclaimable memory (kernel memory or anonymous memory
307 * without swap) can bring down the dirtyable pages below
308 * the zone's dirty balance reserve and the above calculation
309 * will underflow. However we still want to add in nodes
310 * which are below threshold (negative values) to get a more
311 * accurate calculation but make sure that the total never
312 * underflows.
313 */
314 if ((long)x < 0)
315 x = 0;
316
317 /*
318 * Make sure that the number of highmem pages is never larger
319 * than the number of the total dirtyable memory. This can only
320 * occur in very strange VM situations but we want to make sure
321 * that this does not occur.
322 */
323 return min(x, total);
324#else
325 return 0;
326#endif
327}
328
329/**
330 * global_dirtyable_memory - number of globally dirtyable pages
331 *
332 * Returns the global number of pages potentially available for dirty
333 * page cache. This is the base value for the global dirty limits.
334 */
335static unsigned long global_dirtyable_memory(void)
336{
337 unsigned long x;
338
339 x = global_page_state(NR_FREE_PAGES);
340 /*
341 * Pages reserved for the kernel should not be considered
342 * dirtyable, to prevent a situation where reclaim has to
343 * clean pages in order to balance the zones.
344 */
345 x -= min(x, totalreserve_pages);
346
347 x += global_page_state(NR_INACTIVE_FILE);
348 x += global_page_state(NR_ACTIVE_FILE);
349
350 if (!vm_highmem_is_dirtyable)
351 x -= highmem_dirtyable_memory(x);
352
353 return x + 1; /* Ensure that we never return 0 */
354}
355
356/**
357 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
358 * @dtc: dirty_throttle_control of interest
359 *
360 * Calculate @dtc->thresh and ->bg_thresh considering
361 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
362 * must ensure that @dtc->avail is set before calling this function. The
363 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
364 * real-time tasks.
365 */
366static void domain_dirty_limits(struct dirty_throttle_control *dtc)
367{
368 const unsigned long available_memory = dtc->avail;
369 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
370 unsigned long bytes = vm_dirty_bytes;
371 unsigned long bg_bytes = dirty_background_bytes;
372 unsigned long ratio = vm_dirty_ratio;
373 unsigned long bg_ratio = dirty_background_ratio;
374 unsigned long thresh;
375 unsigned long bg_thresh;
376 struct task_struct *tsk;
377
378 /* gdtc is !NULL iff @dtc is for memcg domain */
379 if (gdtc) {
380 unsigned long global_avail = gdtc->avail;
381
382 /*
383 * The byte settings can't be applied directly to memcg
384 * domains. Convert them to ratios by scaling against
385 * globally available memory.
386 */
387 if (bytes)
388 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
389 global_avail, 100UL);
390 if (bg_bytes)
391 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
392 global_avail, 100UL);
393 bytes = bg_bytes = 0;
394 }
395
396 if (bytes)
397 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
398 else
399 thresh = (ratio * available_memory) / 100;
400
401 if (bg_bytes)
402 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
403 else
404 bg_thresh = (bg_ratio * available_memory) / 100;
405
406 if (bg_thresh >= thresh)
407 bg_thresh = thresh / 2;
408 tsk = current;
409 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
410 bg_thresh += bg_thresh / 4;
411 thresh += thresh / 4;
412 }
413 dtc->thresh = thresh;
414 dtc->bg_thresh = bg_thresh;
415
416 /* we should eventually report the domain in the TP */
417 if (!gdtc)
418 trace_global_dirty_state(bg_thresh, thresh);
419}
420
421/**
422 * global_dirty_limits - background-writeback and dirty-throttling thresholds
423 * @pbackground: out parameter for bg_thresh
424 * @pdirty: out parameter for thresh
425 *
426 * Calculate bg_thresh and thresh for global_wb_domain. See
427 * domain_dirty_limits() for details.
428 */
429void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
430{
431 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
432
433 gdtc.avail = global_dirtyable_memory();
434 domain_dirty_limits(&gdtc);
435
436 *pbackground = gdtc.bg_thresh;
437 *pdirty = gdtc.thresh;
438}
439
440/**
441 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
442 * @zone: the zone
443 *
444 * Returns the maximum number of dirty pages allowed in a zone, based
445 * on the zone's dirtyable memory.
446 */
447static unsigned long zone_dirty_limit(struct zone *zone)
448{
449 unsigned long zone_memory = zone_dirtyable_memory(zone);
450 struct task_struct *tsk = current;
451 unsigned long dirty;
452
453 if (vm_dirty_bytes)
454 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
455 zone_memory / global_dirtyable_memory();
456 else
457 dirty = vm_dirty_ratio * zone_memory / 100;
458
459 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
460 dirty += dirty / 4;
461
462 return dirty;
463}
464
465/**
466 * zone_dirty_ok - tells whether a zone is within its dirty limits
467 * @zone: the zone to check
468 *
469 * Returns %true when the dirty pages in @zone are within the zone's
470 * dirty limit, %false if the limit is exceeded.
471 */
472bool zone_dirty_ok(struct zone *zone)
473{
474 unsigned long limit = zone_dirty_limit(zone);
475
476 return zone_page_state(zone, NR_FILE_DIRTY) +
477 zone_page_state(zone, NR_UNSTABLE_NFS) +
478 zone_page_state(zone, NR_WRITEBACK) <= limit;
479}
480
481int dirty_background_ratio_handler(struct ctl_table *table, int write,
482 void __user *buffer, size_t *lenp,
483 loff_t *ppos)
484{
485 int ret;
486
487 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
488 if (ret == 0 && write)
489 dirty_background_bytes = 0;
490 return ret;
491}
492
493int dirty_background_bytes_handler(struct ctl_table *table, int write,
494 void __user *buffer, size_t *lenp,
495 loff_t *ppos)
496{
497 int ret;
498
499 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
500 if (ret == 0 && write)
501 dirty_background_ratio = 0;
502 return ret;
503}
504
505int dirty_ratio_handler(struct ctl_table *table, int write,
506 void __user *buffer, size_t *lenp,
507 loff_t *ppos)
508{
509 int old_ratio = vm_dirty_ratio;
510 int ret;
511
512 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
513 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
514 writeback_set_ratelimit();
515 vm_dirty_bytes = 0;
516 }
517 return ret;
518}
519
520int dirty_bytes_handler(struct ctl_table *table, int write,
521 void __user *buffer, size_t *lenp,
522 loff_t *ppos)
523{
524 unsigned long old_bytes = vm_dirty_bytes;
525 int ret;
526
527 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
528 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
529 writeback_set_ratelimit();
530 vm_dirty_ratio = 0;
531 }
532 return ret;
533}
534
535static unsigned long wp_next_time(unsigned long cur_time)
536{
537 cur_time += VM_COMPLETIONS_PERIOD_LEN;
538 /* 0 has a special meaning... */
539 if (!cur_time)
540 return 1;
541 return cur_time;
542}
543
544static void wb_domain_writeout_inc(struct wb_domain *dom,
545 struct fprop_local_percpu *completions,
546 unsigned int max_prop_frac)
547{
548 __fprop_inc_percpu_max(&dom->completions, completions,
549 max_prop_frac);
550 /* First event after period switching was turned off? */
551 if (!unlikely(dom->period_time)) {
552 /*
553 * We can race with other __bdi_writeout_inc calls here but
554 * it does not cause any harm since the resulting time when
555 * timer will fire and what is in writeout_period_time will be
556 * roughly the same.
557 */
558 dom->period_time = wp_next_time(jiffies);
559 mod_timer(&dom->period_timer, dom->period_time);
560 }
561}
562
563/*
564 * Increment @wb's writeout completion count and the global writeout
565 * completion count. Called from test_clear_page_writeback().
566 */
567static inline void __wb_writeout_inc(struct bdi_writeback *wb)
568{
569 struct wb_domain *cgdom;
570
571 __inc_wb_stat(wb, WB_WRITTEN);
572 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
573 wb->bdi->max_prop_frac);
574
575 cgdom = mem_cgroup_wb_domain(wb);
576 if (cgdom)
577 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
578 wb->bdi->max_prop_frac);
579}
580
581void wb_writeout_inc(struct bdi_writeback *wb)
582{
583 unsigned long flags;
584
585 local_irq_save(flags);
586 __wb_writeout_inc(wb);
587 local_irq_restore(flags);
588}
589EXPORT_SYMBOL_GPL(wb_writeout_inc);
590
591/*
592 * On idle system, we can be called long after we scheduled because we use
593 * deferred timers so count with missed periods.
594 */
595static void writeout_period(unsigned long t)
596{
597 struct wb_domain *dom = (void *)t;
598 int miss_periods = (jiffies - dom->period_time) /
599 VM_COMPLETIONS_PERIOD_LEN;
600
601 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
602 dom->period_time = wp_next_time(dom->period_time +
603 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
604 mod_timer(&dom->period_timer, dom->period_time);
605 } else {
606 /*
607 * Aging has zeroed all fractions. Stop wasting CPU on period
608 * updates.
609 */
610 dom->period_time = 0;
611 }
612}
613
614int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
615{
616 memset(dom, 0, sizeof(*dom));
617
618 spin_lock_init(&dom->lock);
619
620 init_timer_deferrable(&dom->period_timer);
621 dom->period_timer.function = writeout_period;
622 dom->period_timer.data = (unsigned long)dom;
623
624 dom->dirty_limit_tstamp = jiffies;
625
626 return fprop_global_init(&dom->completions, gfp);
627}
628
629#ifdef CONFIG_CGROUP_WRITEBACK
630void wb_domain_exit(struct wb_domain *dom)
631{
632 del_timer_sync(&dom->period_timer);
633 fprop_global_destroy(&dom->completions);
634}
635#endif
636
637/*
638 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
639 * registered backing devices, which, for obvious reasons, can not
640 * exceed 100%.
641 */
642static unsigned int bdi_min_ratio;
643
644int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
645{
646 int ret = 0;
647
648 spin_lock_bh(&bdi_lock);
649 if (min_ratio > bdi->max_ratio) {
650 ret = -EINVAL;
651 } else {
652 min_ratio -= bdi->min_ratio;
653 if (bdi_min_ratio + min_ratio < 100) {
654 bdi_min_ratio += min_ratio;
655 bdi->min_ratio += min_ratio;
656 } else {
657 ret = -EINVAL;
658 }
659 }
660 spin_unlock_bh(&bdi_lock);
661
662 return ret;
663}
664
665int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
666{
667 int ret = 0;
668
669 if (max_ratio > 100)
670 return -EINVAL;
671
672 spin_lock_bh(&bdi_lock);
673 if (bdi->min_ratio > max_ratio) {
674 ret = -EINVAL;
675 } else {
676 bdi->max_ratio = max_ratio;
677 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
678 }
679 spin_unlock_bh(&bdi_lock);
680
681 return ret;
682}
683EXPORT_SYMBOL(bdi_set_max_ratio);
684
685static unsigned long dirty_freerun_ceiling(unsigned long thresh,
686 unsigned long bg_thresh)
687{
688 return (thresh + bg_thresh) / 2;
689}
690
691static unsigned long hard_dirty_limit(struct wb_domain *dom,
692 unsigned long thresh)
693{
694 return max(thresh, dom->dirty_limit);
695}
696
697/*
698 * Memory which can be further allocated to a memcg domain is capped by
699 * system-wide clean memory excluding the amount being used in the domain.
700 */
701static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
702 unsigned long filepages, unsigned long headroom)
703{
704 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
705 unsigned long clean = filepages - min(filepages, mdtc->dirty);
706 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
707 unsigned long other_clean = global_clean - min(global_clean, clean);
708
709 mdtc->avail = filepages + min(headroom, other_clean);
710}
711
712/**
713 * __wb_calc_thresh - @wb's share of dirty throttling threshold
714 * @dtc: dirty_throttle_context of interest
715 *
716 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
717 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
718 *
719 * Note that balance_dirty_pages() will only seriously take it as a hard limit
720 * when sleeping max_pause per page is not enough to keep the dirty pages under
721 * control. For example, when the device is completely stalled due to some error
722 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
723 * In the other normal situations, it acts more gently by throttling the tasks
724 * more (rather than completely block them) when the wb dirty pages go high.
725 *
726 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
727 * - starving fast devices
728 * - piling up dirty pages (that will take long time to sync) on slow devices
729 *
730 * The wb's share of dirty limit will be adapting to its throughput and
731 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
732 */
733static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
734{
735 struct wb_domain *dom = dtc_dom(dtc);
736 unsigned long thresh = dtc->thresh;
737 u64 wb_thresh;
738 long numerator, denominator;
739 unsigned long wb_min_ratio, wb_max_ratio;
740
741 /*
742 * Calculate this BDI's share of the thresh ratio.
743 */
744 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
745 &numerator, &denominator);
746
747 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
748 wb_thresh *= numerator;
749 do_div(wb_thresh, denominator);
750
751 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
752
753 wb_thresh += (thresh * wb_min_ratio) / 100;
754 if (wb_thresh > (thresh * wb_max_ratio) / 100)
755 wb_thresh = thresh * wb_max_ratio / 100;
756
757 return wb_thresh;
758}
759
760unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
761{
762 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
763 .thresh = thresh };
764 return __wb_calc_thresh(&gdtc);
765}
766
767/*
768 * setpoint - dirty 3
769 * f(dirty) := 1.0 + (----------------)
770 * limit - setpoint
771 *
772 * it's a 3rd order polynomial that subjects to
773 *
774 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
775 * (2) f(setpoint) = 1.0 => the balance point
776 * (3) f(limit) = 0 => the hard limit
777 * (4) df/dx <= 0 => negative feedback control
778 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
779 * => fast response on large errors; small oscillation near setpoint
780 */
781static long long pos_ratio_polynom(unsigned long setpoint,
782 unsigned long dirty,
783 unsigned long limit)
784{
785 long long pos_ratio;
786 long x;
787
788 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
789 (limit - setpoint) | 1);
790 pos_ratio = x;
791 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
792 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
793 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
794
795 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
796}
797
798/*
799 * Dirty position control.
800 *
801 * (o) global/bdi setpoints
802 *
803 * We want the dirty pages be balanced around the global/wb setpoints.
804 * When the number of dirty pages is higher/lower than the setpoint, the
805 * dirty position control ratio (and hence task dirty ratelimit) will be
806 * decreased/increased to bring the dirty pages back to the setpoint.
807 *
808 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
809 *
810 * if (dirty < setpoint) scale up pos_ratio
811 * if (dirty > setpoint) scale down pos_ratio
812 *
813 * if (wb_dirty < wb_setpoint) scale up pos_ratio
814 * if (wb_dirty > wb_setpoint) scale down pos_ratio
815 *
816 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
817 *
818 * (o) global control line
819 *
820 * ^ pos_ratio
821 * |
822 * | |<===== global dirty control scope ======>|
823 * 2.0 .............*
824 * | .*
825 * | . *
826 * | . *
827 * | . *
828 * | . *
829 * | . *
830 * 1.0 ................................*
831 * | . . *
832 * | . . *
833 * | . . *
834 * | . . *
835 * | . . *
836 * 0 +------------.------------------.----------------------*------------->
837 * freerun^ setpoint^ limit^ dirty pages
838 *
839 * (o) wb control line
840 *
841 * ^ pos_ratio
842 * |
843 * | *
844 * | *
845 * | *
846 * | *
847 * | * |<=========== span ============>|
848 * 1.0 .......................*
849 * | . *
850 * | . *
851 * | . *
852 * | . *
853 * | . *
854 * | . *
855 * | . *
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * 1/4 ...............................................* * * * * * * * * * * *
861 * | . .
862 * | . .
863 * | . .
864 * 0 +----------------------.-------------------------------.------------->
865 * wb_setpoint^ x_intercept^
866 *
867 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
868 * be smoothly throttled down to normal if it starts high in situations like
869 * - start writing to a slow SD card and a fast disk at the same time. The SD
870 * card's wb_dirty may rush to many times higher than wb_setpoint.
871 * - the wb dirty thresh drops quickly due to change of JBOD workload
872 */
873static void wb_position_ratio(struct dirty_throttle_control *dtc)
874{
875 struct bdi_writeback *wb = dtc->wb;
876 unsigned long write_bw = wb->avg_write_bandwidth;
877 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
878 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
879 unsigned long wb_thresh = dtc->wb_thresh;
880 unsigned long x_intercept;
881 unsigned long setpoint; /* dirty pages' target balance point */
882 unsigned long wb_setpoint;
883 unsigned long span;
884 long long pos_ratio; /* for scaling up/down the rate limit */
885 long x;
886
887 dtc->pos_ratio = 0;
888
889 if (unlikely(dtc->dirty >= limit))
890 return;
891
892 /*
893 * global setpoint
894 *
895 * See comment for pos_ratio_polynom().
896 */
897 setpoint = (freerun + limit) / 2;
898 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
899
900 /*
901 * The strictlimit feature is a tool preventing mistrusted filesystems
902 * from growing a large number of dirty pages before throttling. For
903 * such filesystems balance_dirty_pages always checks wb counters
904 * against wb limits. Even if global "nr_dirty" is under "freerun".
905 * This is especially important for fuse which sets bdi->max_ratio to
906 * 1% by default. Without strictlimit feature, fuse writeback may
907 * consume arbitrary amount of RAM because it is accounted in
908 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
909 *
910 * Here, in wb_position_ratio(), we calculate pos_ratio based on
911 * two values: wb_dirty and wb_thresh. Let's consider an example:
912 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
913 * limits are set by default to 10% and 20% (background and throttle).
914 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
915 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
916 * about ~6K pages (as the average of background and throttle wb
917 * limits). The 3rd order polynomial will provide positive feedback if
918 * wb_dirty is under wb_setpoint and vice versa.
919 *
920 * Note, that we cannot use global counters in these calculations
921 * because we want to throttle process writing to a strictlimit wb
922 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
923 * in the example above).
924 */
925 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
926 long long wb_pos_ratio;
927
928 if (dtc->wb_dirty < 8) {
929 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
930 2 << RATELIMIT_CALC_SHIFT);
931 return;
932 }
933
934 if (dtc->wb_dirty >= wb_thresh)
935 return;
936
937 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
938 dtc->wb_bg_thresh);
939
940 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
941 return;
942
943 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
944 wb_thresh);
945
946 /*
947 * Typically, for strictlimit case, wb_setpoint << setpoint
948 * and pos_ratio >> wb_pos_ratio. In the other words global
949 * state ("dirty") is not limiting factor and we have to
950 * make decision based on wb counters. But there is an
951 * important case when global pos_ratio should get precedence:
952 * global limits are exceeded (e.g. due to activities on other
953 * wb's) while given strictlimit wb is below limit.
954 *
955 * "pos_ratio * wb_pos_ratio" would work for the case above,
956 * but it would look too non-natural for the case of all
957 * activity in the system coming from a single strictlimit wb
958 * with bdi->max_ratio == 100%.
959 *
960 * Note that min() below somewhat changes the dynamics of the
961 * control system. Normally, pos_ratio value can be well over 3
962 * (when globally we are at freerun and wb is well below wb
963 * setpoint). Now the maximum pos_ratio in the same situation
964 * is 2. We might want to tweak this if we observe the control
965 * system is too slow to adapt.
966 */
967 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
968 return;
969 }
970
971 /*
972 * We have computed basic pos_ratio above based on global situation. If
973 * the wb is over/under its share of dirty pages, we want to scale
974 * pos_ratio further down/up. That is done by the following mechanism.
975 */
976
977 /*
978 * wb setpoint
979 *
980 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
981 *
982 * x_intercept - wb_dirty
983 * := --------------------------
984 * x_intercept - wb_setpoint
985 *
986 * The main wb control line is a linear function that subjects to
987 *
988 * (1) f(wb_setpoint) = 1.0
989 * (2) k = - 1 / (8 * write_bw) (in single wb case)
990 * or equally: x_intercept = wb_setpoint + 8 * write_bw
991 *
992 * For single wb case, the dirty pages are observed to fluctuate
993 * regularly within range
994 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
995 * for various filesystems, where (2) can yield in a reasonable 12.5%
996 * fluctuation range for pos_ratio.
997 *
998 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
999 * own size, so move the slope over accordingly and choose a slope that
1000 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1001 */
1002 if (unlikely(wb_thresh > dtc->thresh))
1003 wb_thresh = dtc->thresh;
1004 /*
1005 * It's very possible that wb_thresh is close to 0 not because the
1006 * device is slow, but that it has remained inactive for long time.
1007 * Honour such devices a reasonable good (hopefully IO efficient)
1008 * threshold, so that the occasional writes won't be blocked and active
1009 * writes can rampup the threshold quickly.
1010 */
1011 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1012 /*
1013 * scale global setpoint to wb's:
1014 * wb_setpoint = setpoint * wb_thresh / thresh
1015 */
1016 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1017 wb_setpoint = setpoint * (u64)x >> 16;
1018 /*
1019 * Use span=(8*write_bw) in single wb case as indicated by
1020 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1021 *
1022 * wb_thresh thresh - wb_thresh
1023 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1024 * thresh thresh
1025 */
1026 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1027 x_intercept = wb_setpoint + span;
1028
1029 if (dtc->wb_dirty < x_intercept - span / 4) {
1030 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1031 (x_intercept - wb_setpoint) | 1);
1032 } else
1033 pos_ratio /= 4;
1034
1035 /*
1036 * wb reserve area, safeguard against dirty pool underrun and disk idle
1037 * It may push the desired control point of global dirty pages higher
1038 * than setpoint.
1039 */
1040 x_intercept = wb_thresh / 2;
1041 if (dtc->wb_dirty < x_intercept) {
1042 if (dtc->wb_dirty > x_intercept / 8)
1043 pos_ratio = div_u64(pos_ratio * x_intercept,
1044 dtc->wb_dirty);
1045 else
1046 pos_ratio *= 8;
1047 }
1048
1049 dtc->pos_ratio = pos_ratio;
1050}
1051
1052static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1053 unsigned long elapsed,
1054 unsigned long written)
1055{
1056 const unsigned long period = roundup_pow_of_two(3 * HZ);
1057 unsigned long avg = wb->avg_write_bandwidth;
1058 unsigned long old = wb->write_bandwidth;
1059 u64 bw;
1060
1061 /*
1062 * bw = written * HZ / elapsed
1063 *
1064 * bw * elapsed + write_bandwidth * (period - elapsed)
1065 * write_bandwidth = ---------------------------------------------------
1066 * period
1067 *
1068 * @written may have decreased due to account_page_redirty().
1069 * Avoid underflowing @bw calculation.
1070 */
1071 bw = written - min(written, wb->written_stamp);
1072 bw *= HZ;
1073 if (unlikely(elapsed > period)) {
1074 do_div(bw, elapsed);
1075 avg = bw;
1076 goto out;
1077 }
1078 bw += (u64)wb->write_bandwidth * (period - elapsed);
1079 bw >>= ilog2(period);
1080
1081 /*
1082 * one more level of smoothing, for filtering out sudden spikes
1083 */
1084 if (avg > old && old >= (unsigned long)bw)
1085 avg -= (avg - old) >> 3;
1086
1087 if (avg < old && old <= (unsigned long)bw)
1088 avg += (old - avg) >> 3;
1089
1090out:
1091 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1092 avg = max(avg, 1LU);
1093 if (wb_has_dirty_io(wb)) {
1094 long delta = avg - wb->avg_write_bandwidth;
1095 WARN_ON_ONCE(atomic_long_add_return(delta,
1096 &wb->bdi->tot_write_bandwidth) <= 0);
1097 }
1098 wb->write_bandwidth = bw;
1099 wb->avg_write_bandwidth = avg;
1100}
1101
1102static void update_dirty_limit(struct dirty_throttle_control *dtc)
1103{
1104 struct wb_domain *dom = dtc_dom(dtc);
1105 unsigned long thresh = dtc->thresh;
1106 unsigned long limit = dom->dirty_limit;
1107
1108 /*
1109 * Follow up in one step.
1110 */
1111 if (limit < thresh) {
1112 limit = thresh;
1113 goto update;
1114 }
1115
1116 /*
1117 * Follow down slowly. Use the higher one as the target, because thresh
1118 * may drop below dirty. This is exactly the reason to introduce
1119 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1120 */
1121 thresh = max(thresh, dtc->dirty);
1122 if (limit > thresh) {
1123 limit -= (limit - thresh) >> 5;
1124 goto update;
1125 }
1126 return;
1127update:
1128 dom->dirty_limit = limit;
1129}
1130
1131static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1132 unsigned long now)
1133{
1134 struct wb_domain *dom = dtc_dom(dtc);
1135
1136 /*
1137 * check locklessly first to optimize away locking for the most time
1138 */
1139 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1140 return;
1141
1142 spin_lock(&dom->lock);
1143 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1144 update_dirty_limit(dtc);
1145 dom->dirty_limit_tstamp = now;
1146 }
1147 spin_unlock(&dom->lock);
1148}
1149
1150/*
1151 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1152 *
1153 * Normal wb tasks will be curbed at or below it in long term.
1154 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1155 */
1156static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1157 unsigned long dirtied,
1158 unsigned long elapsed)
1159{
1160 struct bdi_writeback *wb = dtc->wb;
1161 unsigned long dirty = dtc->dirty;
1162 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1163 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1164 unsigned long setpoint = (freerun + limit) / 2;
1165 unsigned long write_bw = wb->avg_write_bandwidth;
1166 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1167 unsigned long dirty_rate;
1168 unsigned long task_ratelimit;
1169 unsigned long balanced_dirty_ratelimit;
1170 unsigned long step;
1171 unsigned long x;
1172 unsigned long shift;
1173
1174 /*
1175 * The dirty rate will match the writeout rate in long term, except
1176 * when dirty pages are truncated by userspace or re-dirtied by FS.
1177 */
1178 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1179
1180 /*
1181 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1182 */
1183 task_ratelimit = (u64)dirty_ratelimit *
1184 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1185 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1186
1187 /*
1188 * A linear estimation of the "balanced" throttle rate. The theory is,
1189 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1190 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1191 * formula will yield the balanced rate limit (write_bw / N).
1192 *
1193 * Note that the expanded form is not a pure rate feedback:
1194 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1195 * but also takes pos_ratio into account:
1196 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1197 *
1198 * (1) is not realistic because pos_ratio also takes part in balancing
1199 * the dirty rate. Consider the state
1200 * pos_ratio = 0.5 (3)
1201 * rate = 2 * (write_bw / N) (4)
1202 * If (1) is used, it will stuck in that state! Because each dd will
1203 * be throttled at
1204 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1205 * yielding
1206 * dirty_rate = N * task_ratelimit = write_bw (6)
1207 * put (6) into (1) we get
1208 * rate_(i+1) = rate_(i) (7)
1209 *
1210 * So we end up using (2) to always keep
1211 * rate_(i+1) ~= (write_bw / N) (8)
1212 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1213 * pos_ratio is able to drive itself to 1.0, which is not only where
1214 * the dirty count meet the setpoint, but also where the slope of
1215 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1216 */
1217 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1218 dirty_rate | 1);
1219 /*
1220 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1221 */
1222 if (unlikely(balanced_dirty_ratelimit > write_bw))
1223 balanced_dirty_ratelimit = write_bw;
1224
1225 /*
1226 * We could safely do this and return immediately:
1227 *
1228 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1229 *
1230 * However to get a more stable dirty_ratelimit, the below elaborated
1231 * code makes use of task_ratelimit to filter out singular points and
1232 * limit the step size.
1233 *
1234 * The below code essentially only uses the relative value of
1235 *
1236 * task_ratelimit - dirty_ratelimit
1237 * = (pos_ratio - 1) * dirty_ratelimit
1238 *
1239 * which reflects the direction and size of dirty position error.
1240 */
1241
1242 /*
1243 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1244 * task_ratelimit is on the same side of dirty_ratelimit, too.
1245 * For example, when
1246 * - dirty_ratelimit > balanced_dirty_ratelimit
1247 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1248 * lowering dirty_ratelimit will help meet both the position and rate
1249 * control targets. Otherwise, don't update dirty_ratelimit if it will
1250 * only help meet the rate target. After all, what the users ultimately
1251 * feel and care are stable dirty rate and small position error.
1252 *
1253 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1254 * and filter out the singular points of balanced_dirty_ratelimit. Which
1255 * keeps jumping around randomly and can even leap far away at times
1256 * due to the small 200ms estimation period of dirty_rate (we want to
1257 * keep that period small to reduce time lags).
1258 */
1259 step = 0;
1260
1261 /*
1262 * For strictlimit case, calculations above were based on wb counters
1263 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1264 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1265 * Hence, to calculate "step" properly, we have to use wb_dirty as
1266 * "dirty" and wb_setpoint as "setpoint".
1267 *
1268 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1269 * it's possible that wb_thresh is close to zero due to inactivity
1270 * of backing device.
1271 */
1272 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1273 dirty = dtc->wb_dirty;
1274 if (dtc->wb_dirty < 8)
1275 setpoint = dtc->wb_dirty + 1;
1276 else
1277 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1278 }
1279
1280 if (dirty < setpoint) {
1281 x = min3(wb->balanced_dirty_ratelimit,
1282 balanced_dirty_ratelimit, task_ratelimit);
1283 if (dirty_ratelimit < x)
1284 step = x - dirty_ratelimit;
1285 } else {
1286 x = max3(wb->balanced_dirty_ratelimit,
1287 balanced_dirty_ratelimit, task_ratelimit);
1288 if (dirty_ratelimit > x)
1289 step = dirty_ratelimit - x;
1290 }
1291
1292 /*
1293 * Don't pursue 100% rate matching. It's impossible since the balanced
1294 * rate itself is constantly fluctuating. So decrease the track speed
1295 * when it gets close to the target. Helps eliminate pointless tremors.
1296 */
1297 shift = dirty_ratelimit / (2 * step + 1);
1298 if (shift < BITS_PER_LONG)
1299 step = DIV_ROUND_UP(step >> shift, 8);
1300 else
1301 step = 0;
1302
1303 if (dirty_ratelimit < balanced_dirty_ratelimit)
1304 dirty_ratelimit += step;
1305 else
1306 dirty_ratelimit -= step;
1307
1308 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1309 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1310
1311 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1312}
1313
1314static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1315 struct dirty_throttle_control *mdtc,
1316 unsigned long start_time,
1317 bool update_ratelimit)
1318{
1319 struct bdi_writeback *wb = gdtc->wb;
1320 unsigned long now = jiffies;
1321 unsigned long elapsed = now - wb->bw_time_stamp;
1322 unsigned long dirtied;
1323 unsigned long written;
1324
1325 lockdep_assert_held(&wb->list_lock);
1326
1327 /*
1328 * rate-limit, only update once every 200ms.
1329 */
1330 if (elapsed < BANDWIDTH_INTERVAL)
1331 return;
1332
1333 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1334 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1335
1336 /*
1337 * Skip quiet periods when disk bandwidth is under-utilized.
1338 * (at least 1s idle time between two flusher runs)
1339 */
1340 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1341 goto snapshot;
1342
1343 if (update_ratelimit) {
1344 domain_update_bandwidth(gdtc, now);
1345 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1346
1347 /*
1348 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1349 * compiler has no way to figure that out. Help it.
1350 */
1351 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1352 domain_update_bandwidth(mdtc, now);
1353 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1354 }
1355 }
1356 wb_update_write_bandwidth(wb, elapsed, written);
1357
1358snapshot:
1359 wb->dirtied_stamp = dirtied;
1360 wb->written_stamp = written;
1361 wb->bw_time_stamp = now;
1362}
1363
1364void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1365{
1366 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1367
1368 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1369}
1370
1371/*
1372 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1373 * will look to see if it needs to start dirty throttling.
1374 *
1375 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1376 * global_page_state() too often. So scale it near-sqrt to the safety margin
1377 * (the number of pages we may dirty without exceeding the dirty limits).
1378 */
1379static unsigned long dirty_poll_interval(unsigned long dirty,
1380 unsigned long thresh)
1381{
1382 if (thresh > dirty)
1383 return 1UL << (ilog2(thresh - dirty) >> 1);
1384
1385 return 1;
1386}
1387
1388static unsigned long wb_max_pause(struct bdi_writeback *wb,
1389 unsigned long wb_dirty)
1390{
1391 unsigned long bw = wb->avg_write_bandwidth;
1392 unsigned long t;
1393
1394 /*
1395 * Limit pause time for small memory systems. If sleeping for too long
1396 * time, a small pool of dirty/writeback pages may go empty and disk go
1397 * idle.
1398 *
1399 * 8 serves as the safety ratio.
1400 */
1401 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1402 t++;
1403
1404 return min_t(unsigned long, t, MAX_PAUSE);
1405}
1406
1407static long wb_min_pause(struct bdi_writeback *wb,
1408 long max_pause,
1409 unsigned long task_ratelimit,
1410 unsigned long dirty_ratelimit,
1411 int *nr_dirtied_pause)
1412{
1413 long hi = ilog2(wb->avg_write_bandwidth);
1414 long lo = ilog2(wb->dirty_ratelimit);
1415 long t; /* target pause */
1416 long pause; /* estimated next pause */
1417 int pages; /* target nr_dirtied_pause */
1418
1419 /* target for 10ms pause on 1-dd case */
1420 t = max(1, HZ / 100);
1421
1422 /*
1423 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1424 * overheads.
1425 *
1426 * (N * 10ms) on 2^N concurrent tasks.
1427 */
1428 if (hi > lo)
1429 t += (hi - lo) * (10 * HZ) / 1024;
1430
1431 /*
1432 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1433 * on the much more stable dirty_ratelimit. However the next pause time
1434 * will be computed based on task_ratelimit and the two rate limits may
1435 * depart considerably at some time. Especially if task_ratelimit goes
1436 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1437 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1438 * result task_ratelimit won't be executed faithfully, which could
1439 * eventually bring down dirty_ratelimit.
1440 *
1441 * We apply two rules to fix it up:
1442 * 1) try to estimate the next pause time and if necessary, use a lower
1443 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1444 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1445 * 2) limit the target pause time to max_pause/2, so that the normal
1446 * small fluctuations of task_ratelimit won't trigger rule (1) and
1447 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1448 */
1449 t = min(t, 1 + max_pause / 2);
1450 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1451
1452 /*
1453 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1454 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1455 * When the 16 consecutive reads are often interrupted by some dirty
1456 * throttling pause during the async writes, cfq will go into idles
1457 * (deadline is fine). So push nr_dirtied_pause as high as possible
1458 * until reaches DIRTY_POLL_THRESH=32 pages.
1459 */
1460 if (pages < DIRTY_POLL_THRESH) {
1461 t = max_pause;
1462 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1463 if (pages > DIRTY_POLL_THRESH) {
1464 pages = DIRTY_POLL_THRESH;
1465 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1466 }
1467 }
1468
1469 pause = HZ * pages / (task_ratelimit + 1);
1470 if (pause > max_pause) {
1471 t = max_pause;
1472 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1473 }
1474
1475 *nr_dirtied_pause = pages;
1476 /*
1477 * The minimal pause time will normally be half the target pause time.
1478 */
1479 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1480}
1481
1482static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1483{
1484 struct bdi_writeback *wb = dtc->wb;
1485 unsigned long wb_reclaimable;
1486
1487 /*
1488 * wb_thresh is not treated as some limiting factor as
1489 * dirty_thresh, due to reasons
1490 * - in JBOD setup, wb_thresh can fluctuate a lot
1491 * - in a system with HDD and USB key, the USB key may somehow
1492 * go into state (wb_dirty >> wb_thresh) either because
1493 * wb_dirty starts high, or because wb_thresh drops low.
1494 * In this case we don't want to hard throttle the USB key
1495 * dirtiers for 100 seconds until wb_dirty drops under
1496 * wb_thresh. Instead the auxiliary wb control line in
1497 * wb_position_ratio() will let the dirtier task progress
1498 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1499 */
1500 dtc->wb_thresh = __wb_calc_thresh(dtc);
1501 dtc->wb_bg_thresh = dtc->thresh ?
1502 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1503
1504 /*
1505 * In order to avoid the stacked BDI deadlock we need
1506 * to ensure we accurately count the 'dirty' pages when
1507 * the threshold is low.
1508 *
1509 * Otherwise it would be possible to get thresh+n pages
1510 * reported dirty, even though there are thresh-m pages
1511 * actually dirty; with m+n sitting in the percpu
1512 * deltas.
1513 */
1514 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1515 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1516 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1517 } else {
1518 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1519 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1520 }
1521}
1522
1523/*
1524 * balance_dirty_pages() must be called by processes which are generating dirty
1525 * data. It looks at the number of dirty pages in the machine and will force
1526 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1527 * If we're over `background_thresh' then the writeback threads are woken to
1528 * perform some writeout.
1529 */
1530static void balance_dirty_pages(struct address_space *mapping,
1531 struct bdi_writeback *wb,
1532 unsigned long pages_dirtied)
1533{
1534 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1535 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1536 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1537 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1538 &mdtc_stor : NULL;
1539 struct dirty_throttle_control *sdtc;
1540 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1541 long period;
1542 long pause;
1543 long max_pause;
1544 long min_pause;
1545 int nr_dirtied_pause;
1546 bool dirty_exceeded = false;
1547 unsigned long task_ratelimit;
1548 unsigned long dirty_ratelimit;
1549 struct backing_dev_info *bdi = wb->bdi;
1550 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1551 unsigned long start_time = jiffies;
1552
1553 for (;;) {
1554 unsigned long now = jiffies;
1555 unsigned long dirty, thresh, bg_thresh;
1556 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1557 unsigned long m_thresh = 0;
1558 unsigned long m_bg_thresh = 0;
1559
1560 /*
1561 * Unstable writes are a feature of certain networked
1562 * filesystems (i.e. NFS) in which data may have been
1563 * written to the server's write cache, but has not yet
1564 * been flushed to permanent storage.
1565 */
1566 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1567 global_page_state(NR_UNSTABLE_NFS);
1568 gdtc->avail = global_dirtyable_memory();
1569 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1570
1571 domain_dirty_limits(gdtc);
1572
1573 if (unlikely(strictlimit)) {
1574 wb_dirty_limits(gdtc);
1575
1576 dirty = gdtc->wb_dirty;
1577 thresh = gdtc->wb_thresh;
1578 bg_thresh = gdtc->wb_bg_thresh;
1579 } else {
1580 dirty = gdtc->dirty;
1581 thresh = gdtc->thresh;
1582 bg_thresh = gdtc->bg_thresh;
1583 }
1584
1585 if (mdtc) {
1586 unsigned long filepages, headroom, writeback;
1587
1588 /*
1589 * If @wb belongs to !root memcg, repeat the same
1590 * basic calculations for the memcg domain.
1591 */
1592 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1593 &mdtc->dirty, &writeback);
1594 mdtc->dirty += writeback;
1595 mdtc_calc_avail(mdtc, filepages, headroom);
1596
1597 domain_dirty_limits(mdtc);
1598
1599 if (unlikely(strictlimit)) {
1600 wb_dirty_limits(mdtc);
1601 m_dirty = mdtc->wb_dirty;
1602 m_thresh = mdtc->wb_thresh;
1603 m_bg_thresh = mdtc->wb_bg_thresh;
1604 } else {
1605 m_dirty = mdtc->dirty;
1606 m_thresh = mdtc->thresh;
1607 m_bg_thresh = mdtc->bg_thresh;
1608 }
1609 }
1610
1611 /*
1612 * Throttle it only when the background writeback cannot
1613 * catch-up. This avoids (excessively) small writeouts
1614 * when the wb limits are ramping up in case of !strictlimit.
1615 *
1616 * In strictlimit case make decision based on the wb counters
1617 * and limits. Small writeouts when the wb limits are ramping
1618 * up are the price we consciously pay for strictlimit-ing.
1619 *
1620 * If memcg domain is in effect, @dirty should be under
1621 * both global and memcg freerun ceilings.
1622 */
1623 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1624 (!mdtc ||
1625 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1626 unsigned long intv = dirty_poll_interval(dirty, thresh);
1627 unsigned long m_intv = ULONG_MAX;
1628
1629 current->dirty_paused_when = now;
1630 current->nr_dirtied = 0;
1631 if (mdtc)
1632 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1633 current->nr_dirtied_pause = min(intv, m_intv);
1634 break;
1635 }
1636
1637 if (unlikely(!writeback_in_progress(wb)))
1638 wb_start_background_writeback(wb);
1639
1640 /*
1641 * Calculate global domain's pos_ratio and select the
1642 * global dtc by default.
1643 */
1644 if (!strictlimit)
1645 wb_dirty_limits(gdtc);
1646
1647 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1648 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1649
1650 wb_position_ratio(gdtc);
1651 sdtc = gdtc;
1652
1653 if (mdtc) {
1654 /*
1655 * If memcg domain is in effect, calculate its
1656 * pos_ratio. @wb should satisfy constraints from
1657 * both global and memcg domains. Choose the one
1658 * w/ lower pos_ratio.
1659 */
1660 if (!strictlimit)
1661 wb_dirty_limits(mdtc);
1662
1663 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1664 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1665
1666 wb_position_ratio(mdtc);
1667 if (mdtc->pos_ratio < gdtc->pos_ratio)
1668 sdtc = mdtc;
1669 }
1670
1671 if (dirty_exceeded && !wb->dirty_exceeded)
1672 wb->dirty_exceeded = 1;
1673
1674 if (time_is_before_jiffies(wb->bw_time_stamp +
1675 BANDWIDTH_INTERVAL)) {
1676 spin_lock(&wb->list_lock);
1677 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1678 spin_unlock(&wb->list_lock);
1679 }
1680
1681 /* throttle according to the chosen dtc */
1682 dirty_ratelimit = wb->dirty_ratelimit;
1683 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1684 RATELIMIT_CALC_SHIFT;
1685 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1686 min_pause = wb_min_pause(wb, max_pause,
1687 task_ratelimit, dirty_ratelimit,
1688 &nr_dirtied_pause);
1689
1690 if (unlikely(task_ratelimit == 0)) {
1691 period = max_pause;
1692 pause = max_pause;
1693 goto pause;
1694 }
1695 period = HZ * pages_dirtied / task_ratelimit;
1696 pause = period;
1697 if (current->dirty_paused_when)
1698 pause -= now - current->dirty_paused_when;
1699 /*
1700 * For less than 1s think time (ext3/4 may block the dirtier
1701 * for up to 800ms from time to time on 1-HDD; so does xfs,
1702 * however at much less frequency), try to compensate it in
1703 * future periods by updating the virtual time; otherwise just
1704 * do a reset, as it may be a light dirtier.
1705 */
1706 if (pause < min_pause) {
1707 trace_balance_dirty_pages(wb,
1708 sdtc->thresh,
1709 sdtc->bg_thresh,
1710 sdtc->dirty,
1711 sdtc->wb_thresh,
1712 sdtc->wb_dirty,
1713 dirty_ratelimit,
1714 task_ratelimit,
1715 pages_dirtied,
1716 period,
1717 min(pause, 0L),
1718 start_time);
1719 if (pause < -HZ) {
1720 current->dirty_paused_when = now;
1721 current->nr_dirtied = 0;
1722 } else if (period) {
1723 current->dirty_paused_when += period;
1724 current->nr_dirtied = 0;
1725 } else if (current->nr_dirtied_pause <= pages_dirtied)
1726 current->nr_dirtied_pause += pages_dirtied;
1727 break;
1728 }
1729 if (unlikely(pause > max_pause)) {
1730 /* for occasional dropped task_ratelimit */
1731 now += min(pause - max_pause, max_pause);
1732 pause = max_pause;
1733 }
1734
1735pause:
1736 trace_balance_dirty_pages(wb,
1737 sdtc->thresh,
1738 sdtc->bg_thresh,
1739 sdtc->dirty,
1740 sdtc->wb_thresh,
1741 sdtc->wb_dirty,
1742 dirty_ratelimit,
1743 task_ratelimit,
1744 pages_dirtied,
1745 period,
1746 pause,
1747 start_time);
1748 __set_current_state(TASK_KILLABLE);
1749 io_schedule_timeout(pause);
1750
1751 current->dirty_paused_when = now + pause;
1752 current->nr_dirtied = 0;
1753 current->nr_dirtied_pause = nr_dirtied_pause;
1754
1755 /*
1756 * This is typically equal to (dirty < thresh) and can also
1757 * keep "1000+ dd on a slow USB stick" under control.
1758 */
1759 if (task_ratelimit)
1760 break;
1761
1762 /*
1763 * In the case of an unresponding NFS server and the NFS dirty
1764 * pages exceeds dirty_thresh, give the other good wb's a pipe
1765 * to go through, so that tasks on them still remain responsive.
1766 *
1767 * In theory 1 page is enough to keep the comsumer-producer
1768 * pipe going: the flusher cleans 1 page => the task dirties 1
1769 * more page. However wb_dirty has accounting errors. So use
1770 * the larger and more IO friendly wb_stat_error.
1771 */
1772 if (sdtc->wb_dirty <= wb_stat_error(wb))
1773 break;
1774
1775 if (fatal_signal_pending(current))
1776 break;
1777 }
1778
1779 if (!dirty_exceeded && wb->dirty_exceeded)
1780 wb->dirty_exceeded = 0;
1781
1782 if (writeback_in_progress(wb))
1783 return;
1784
1785 /*
1786 * In laptop mode, we wait until hitting the higher threshold before
1787 * starting background writeout, and then write out all the way down
1788 * to the lower threshold. So slow writers cause minimal disk activity.
1789 *
1790 * In normal mode, we start background writeout at the lower
1791 * background_thresh, to keep the amount of dirty memory low.
1792 */
1793 if (laptop_mode)
1794 return;
1795
1796 if (nr_reclaimable > gdtc->bg_thresh)
1797 wb_start_background_writeback(wb);
1798}
1799
1800static DEFINE_PER_CPU(int, bdp_ratelimits);
1801
1802/*
1803 * Normal tasks are throttled by
1804 * loop {
1805 * dirty tsk->nr_dirtied_pause pages;
1806 * take a snap in balance_dirty_pages();
1807 * }
1808 * However there is a worst case. If every task exit immediately when dirtied
1809 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1810 * called to throttle the page dirties. The solution is to save the not yet
1811 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1812 * randomly into the running tasks. This works well for the above worst case,
1813 * as the new task will pick up and accumulate the old task's leaked dirty
1814 * count and eventually get throttled.
1815 */
1816DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1817
1818/**
1819 * balance_dirty_pages_ratelimited - balance dirty memory state
1820 * @mapping: address_space which was dirtied
1821 *
1822 * Processes which are dirtying memory should call in here once for each page
1823 * which was newly dirtied. The function will periodically check the system's
1824 * dirty state and will initiate writeback if needed.
1825 *
1826 * On really big machines, get_writeback_state is expensive, so try to avoid
1827 * calling it too often (ratelimiting). But once we're over the dirty memory
1828 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1829 * from overshooting the limit by (ratelimit_pages) each.
1830 */
1831void balance_dirty_pages_ratelimited(struct address_space *mapping)
1832{
1833 struct inode *inode = mapping->host;
1834 struct backing_dev_info *bdi = inode_to_bdi(inode);
1835 struct bdi_writeback *wb = NULL;
1836 int ratelimit;
1837 int *p;
1838
1839 if (!bdi_cap_account_dirty(bdi))
1840 return;
1841
1842 if (inode_cgwb_enabled(inode))
1843 wb = wb_get_create_current(bdi, GFP_KERNEL);
1844 if (!wb)
1845 wb = &bdi->wb;
1846
1847 ratelimit = current->nr_dirtied_pause;
1848 if (wb->dirty_exceeded)
1849 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1850
1851 preempt_disable();
1852 /*
1853 * This prevents one CPU to accumulate too many dirtied pages without
1854 * calling into balance_dirty_pages(), which can happen when there are
1855 * 1000+ tasks, all of them start dirtying pages at exactly the same
1856 * time, hence all honoured too large initial task->nr_dirtied_pause.
1857 */
1858 p = this_cpu_ptr(&bdp_ratelimits);
1859 if (unlikely(current->nr_dirtied >= ratelimit))
1860 *p = 0;
1861 else if (unlikely(*p >= ratelimit_pages)) {
1862 *p = 0;
1863 ratelimit = 0;
1864 }
1865 /*
1866 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1867 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1868 * the dirty throttling and livelock other long-run dirtiers.
1869 */
1870 p = this_cpu_ptr(&dirty_throttle_leaks);
1871 if (*p > 0 && current->nr_dirtied < ratelimit) {
1872 unsigned long nr_pages_dirtied;
1873 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1874 *p -= nr_pages_dirtied;
1875 current->nr_dirtied += nr_pages_dirtied;
1876 }
1877 preempt_enable();
1878
1879 if (unlikely(current->nr_dirtied >= ratelimit))
1880 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1881
1882 wb_put(wb);
1883}
1884EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1885
1886/**
1887 * wb_over_bg_thresh - does @wb need to be written back?
1888 * @wb: bdi_writeback of interest
1889 *
1890 * Determines whether background writeback should keep writing @wb or it's
1891 * clean enough. Returns %true if writeback should continue.
1892 */
1893bool wb_over_bg_thresh(struct bdi_writeback *wb)
1894{
1895 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1896 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1897 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1898 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1899 &mdtc_stor : NULL;
1900
1901 /*
1902 * Similar to balance_dirty_pages() but ignores pages being written
1903 * as we're trying to decide whether to put more under writeback.
1904 */
1905 gdtc->avail = global_dirtyable_memory();
1906 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1907 global_page_state(NR_UNSTABLE_NFS);
1908 domain_dirty_limits(gdtc);
1909
1910 if (gdtc->dirty > gdtc->bg_thresh)
1911 return true;
1912
1913 if (wb_stat(wb, WB_RECLAIMABLE) >
1914 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1915 return true;
1916
1917 if (mdtc) {
1918 unsigned long filepages, headroom, writeback;
1919
1920 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1921 &writeback);
1922 mdtc_calc_avail(mdtc, filepages, headroom);
1923 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1924
1925 if (mdtc->dirty > mdtc->bg_thresh)
1926 return true;
1927
1928 if (wb_stat(wb, WB_RECLAIMABLE) >
1929 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1930 return true;
1931 }
1932
1933 return false;
1934}
1935
1936void throttle_vm_writeout(gfp_t gfp_mask)
1937{
1938 unsigned long background_thresh;
1939 unsigned long dirty_thresh;
1940
1941 for ( ; ; ) {
1942 global_dirty_limits(&background_thresh, &dirty_thresh);
1943 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1944
1945 /*
1946 * Boost the allowable dirty threshold a bit for page
1947 * allocators so they don't get DoS'ed by heavy writers
1948 */
1949 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1950
1951 if (global_page_state(NR_UNSTABLE_NFS) +
1952 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1953 break;
1954 congestion_wait(BLK_RW_ASYNC, HZ/10);
1955
1956 /*
1957 * The caller might hold locks which can prevent IO completion
1958 * or progress in the filesystem. So we cannot just sit here
1959 * waiting for IO to complete.
1960 */
1961 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1962 break;
1963 }
1964}
1965
1966/*
1967 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1968 */
1969int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1970 void __user *buffer, size_t *length, loff_t *ppos)
1971{
1972 proc_dointvec(table, write, buffer, length, ppos);
1973 return 0;
1974}
1975
1976#ifdef CONFIG_BLOCK
1977void laptop_mode_timer_fn(unsigned long data)
1978{
1979 struct request_queue *q = (struct request_queue *)data;
1980 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1981 global_page_state(NR_UNSTABLE_NFS);
1982 struct bdi_writeback *wb;
1983
1984 /*
1985 * We want to write everything out, not just down to the dirty
1986 * threshold
1987 */
1988 if (!bdi_has_dirty_io(&q->backing_dev_info))
1989 return;
1990
1991 rcu_read_lock();
1992 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
1993 if (wb_has_dirty_io(wb))
1994 wb_start_writeback(wb, nr_pages, true,
1995 WB_REASON_LAPTOP_TIMER);
1996 rcu_read_unlock();
1997}
1998
1999/*
2000 * We've spun up the disk and we're in laptop mode: schedule writeback
2001 * of all dirty data a few seconds from now. If the flush is already scheduled
2002 * then push it back - the user is still using the disk.
2003 */
2004void laptop_io_completion(struct backing_dev_info *info)
2005{
2006 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2007}
2008
2009/*
2010 * We're in laptop mode and we've just synced. The sync's writes will have
2011 * caused another writeback to be scheduled by laptop_io_completion.
2012 * Nothing needs to be written back anymore, so we unschedule the writeback.
2013 */
2014void laptop_sync_completion(void)
2015{
2016 struct backing_dev_info *bdi;
2017
2018 rcu_read_lock();
2019
2020 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2021 del_timer(&bdi->laptop_mode_wb_timer);
2022
2023 rcu_read_unlock();
2024}
2025#endif
2026
2027/*
2028 * If ratelimit_pages is too high then we can get into dirty-data overload
2029 * if a large number of processes all perform writes at the same time.
2030 * If it is too low then SMP machines will call the (expensive)
2031 * get_writeback_state too often.
2032 *
2033 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2034 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2035 * thresholds.
2036 */
2037
2038void writeback_set_ratelimit(void)
2039{
2040 struct wb_domain *dom = &global_wb_domain;
2041 unsigned long background_thresh;
2042 unsigned long dirty_thresh;
2043
2044 global_dirty_limits(&background_thresh, &dirty_thresh);
2045 dom->dirty_limit = dirty_thresh;
2046 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2047 if (ratelimit_pages < 16)
2048 ratelimit_pages = 16;
2049}
2050
2051static int
2052ratelimit_handler(struct notifier_block *self, unsigned long action,
2053 void *hcpu)
2054{
2055
2056 switch (action & ~CPU_TASKS_FROZEN) {
2057 case CPU_ONLINE:
2058 case CPU_DEAD:
2059 writeback_set_ratelimit();
2060 return NOTIFY_OK;
2061 default:
2062 return NOTIFY_DONE;
2063 }
2064}
2065
2066static struct notifier_block ratelimit_nb = {
2067 .notifier_call = ratelimit_handler,
2068 .next = NULL,
2069};
2070
2071/*
2072 * Called early on to tune the page writeback dirty limits.
2073 *
2074 * We used to scale dirty pages according to how total memory
2075 * related to pages that could be allocated for buffers (by
2076 * comparing nr_free_buffer_pages() to vm_total_pages.
2077 *
2078 * However, that was when we used "dirty_ratio" to scale with
2079 * all memory, and we don't do that any more. "dirty_ratio"
2080 * is now applied to total non-HIGHPAGE memory (by subtracting
2081 * totalhigh_pages from vm_total_pages), and as such we can't
2082 * get into the old insane situation any more where we had
2083 * large amounts of dirty pages compared to a small amount of
2084 * non-HIGHMEM memory.
2085 *
2086 * But we might still want to scale the dirty_ratio by how
2087 * much memory the box has..
2088 */
2089void __init page_writeback_init(void)
2090{
2091 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2092
2093 writeback_set_ratelimit();
2094 register_cpu_notifier(&ratelimit_nb);
2095}
2096
2097/**
2098 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2099 * @mapping: address space structure to write
2100 * @start: starting page index
2101 * @end: ending page index (inclusive)
2102 *
2103 * This function scans the page range from @start to @end (inclusive) and tags
2104 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2105 * that write_cache_pages (or whoever calls this function) will then use
2106 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2107 * used to avoid livelocking of writeback by a process steadily creating new
2108 * dirty pages in the file (thus it is important for this function to be quick
2109 * so that it can tag pages faster than a dirtying process can create them).
2110 */
2111/*
2112 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2113 */
2114void tag_pages_for_writeback(struct address_space *mapping,
2115 pgoff_t start, pgoff_t end)
2116{
2117#define WRITEBACK_TAG_BATCH 4096
2118 unsigned long tagged;
2119
2120 do {
2121 spin_lock_irq(&mapping->tree_lock);
2122 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2123 &start, end, WRITEBACK_TAG_BATCH,
2124 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2125 spin_unlock_irq(&mapping->tree_lock);
2126 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2127 cond_resched();
2128 /* We check 'start' to handle wrapping when end == ~0UL */
2129 } while (tagged >= WRITEBACK_TAG_BATCH && start);
2130}
2131EXPORT_SYMBOL(tag_pages_for_writeback);
2132
2133/**
2134 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2135 * @mapping: address space structure to write
2136 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2137 * @writepage: function called for each page
2138 * @data: data passed to writepage function
2139 *
2140 * If a page is already under I/O, write_cache_pages() skips it, even
2141 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2142 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2143 * and msync() need to guarantee that all the data which was dirty at the time
2144 * the call was made get new I/O started against them. If wbc->sync_mode is
2145 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2146 * existing IO to complete.
2147 *
2148 * To avoid livelocks (when other process dirties new pages), we first tag
2149 * pages which should be written back with TOWRITE tag and only then start
2150 * writing them. For data-integrity sync we have to be careful so that we do
2151 * not miss some pages (e.g., because some other process has cleared TOWRITE
2152 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2153 * by the process clearing the DIRTY tag (and submitting the page for IO).
2154 */
2155int write_cache_pages(struct address_space *mapping,
2156 struct writeback_control *wbc, writepage_t writepage,
2157 void *data)
2158{
2159 int ret = 0;
2160 int done = 0;
2161 struct pagevec pvec;
2162 int nr_pages;
2163 pgoff_t uninitialized_var(writeback_index);
2164 pgoff_t index;
2165 pgoff_t end; /* Inclusive */
2166 pgoff_t done_index;
2167 int cycled;
2168 int range_whole = 0;
2169 int tag;
2170
2171 pagevec_init(&pvec, 0);
2172 if (wbc->range_cyclic) {
2173 writeback_index = mapping->writeback_index; /* prev offset */
2174 index = writeback_index;
2175 if (index == 0)
2176 cycled = 1;
2177 else
2178 cycled = 0;
2179 end = -1;
2180 } else {
2181 index = wbc->range_start >> PAGE_SHIFT;
2182 end = wbc->range_end >> PAGE_SHIFT;
2183 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2184 range_whole = 1;
2185 cycled = 1; /* ignore range_cyclic tests */
2186 }
2187 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2188 tag = PAGECACHE_TAG_TOWRITE;
2189 else
2190 tag = PAGECACHE_TAG_DIRTY;
2191retry:
2192 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2193 tag_pages_for_writeback(mapping, index, end);
2194 done_index = index;
2195 while (!done && (index <= end)) {
2196 int i;
2197
2198 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2199 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2200 if (nr_pages == 0)
2201 break;
2202
2203 for (i = 0; i < nr_pages; i++) {
2204 struct page *page = pvec.pages[i];
2205
2206 /*
2207 * At this point, the page may be truncated or
2208 * invalidated (changing page->mapping to NULL), or
2209 * even swizzled back from swapper_space to tmpfs file
2210 * mapping. However, page->index will not change
2211 * because we have a reference on the page.
2212 */
2213 if (page->index > end) {
2214 /*
2215 * can't be range_cyclic (1st pass) because
2216 * end == -1 in that case.
2217 */
2218 done = 1;
2219 break;
2220 }
2221
2222 done_index = page->index;
2223
2224 lock_page(page);
2225
2226 /*
2227 * Page truncated or invalidated. We can freely skip it
2228 * then, even for data integrity operations: the page
2229 * has disappeared concurrently, so there could be no
2230 * real expectation of this data interity operation
2231 * even if there is now a new, dirty page at the same
2232 * pagecache address.
2233 */
2234 if (unlikely(page->mapping != mapping)) {
2235continue_unlock:
2236 unlock_page(page);
2237 continue;
2238 }
2239
2240 if (!PageDirty(page)) {
2241 /* someone wrote it for us */
2242 goto continue_unlock;
2243 }
2244
2245 if (PageWriteback(page)) {
2246 if (wbc->sync_mode != WB_SYNC_NONE)
2247 wait_on_page_writeback(page);
2248 else
2249 goto continue_unlock;
2250 }
2251
2252 BUG_ON(PageWriteback(page));
2253 if (!clear_page_dirty_for_io(page))
2254 goto continue_unlock;
2255
2256 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2257 ret = (*writepage)(page, wbc, data);
2258 if (unlikely(ret)) {
2259 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2260 unlock_page(page);
2261 ret = 0;
2262 } else {
2263 /*
2264 * done_index is set past this page,
2265 * so media errors will not choke
2266 * background writeout for the entire
2267 * file. This has consequences for
2268 * range_cyclic semantics (ie. it may
2269 * not be suitable for data integrity
2270 * writeout).
2271 */
2272 done_index = page->index + 1;
2273 done = 1;
2274 break;
2275 }
2276 }
2277
2278 /*
2279 * We stop writing back only if we are not doing
2280 * integrity sync. In case of integrity sync we have to
2281 * keep going until we have written all the pages
2282 * we tagged for writeback prior to entering this loop.
2283 */
2284 if (--wbc->nr_to_write <= 0 &&
2285 wbc->sync_mode == WB_SYNC_NONE) {
2286 done = 1;
2287 break;
2288 }
2289 }
2290 pagevec_release(&pvec);
2291 cond_resched();
2292 }
2293 if (!cycled && !done) {
2294 /*
2295 * range_cyclic:
2296 * We hit the last page and there is more work to be done: wrap
2297 * back to the start of the file
2298 */
2299 cycled = 1;
2300 index = 0;
2301 end = writeback_index - 1;
2302 goto retry;
2303 }
2304 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2305 mapping->writeback_index = done_index;
2306
2307 return ret;
2308}
2309EXPORT_SYMBOL(write_cache_pages);
2310
2311/*
2312 * Function used by generic_writepages to call the real writepage
2313 * function and set the mapping flags on error
2314 */
2315static int __writepage(struct page *page, struct writeback_control *wbc,
2316 void *data)
2317{
2318 struct address_space *mapping = data;
2319 int ret = mapping->a_ops->writepage(page, wbc);
2320 mapping_set_error(mapping, ret);
2321 return ret;
2322}
2323
2324/**
2325 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2326 * @mapping: address space structure to write
2327 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2328 *
2329 * This is a library function, which implements the writepages()
2330 * address_space_operation.
2331 */
2332int generic_writepages(struct address_space *mapping,
2333 struct writeback_control *wbc)
2334{
2335 struct blk_plug plug;
2336 int ret;
2337
2338 /* deal with chardevs and other special file */
2339 if (!mapping->a_ops->writepage)
2340 return 0;
2341
2342 blk_start_plug(&plug);
2343 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2344 blk_finish_plug(&plug);
2345 return ret;
2346}
2347
2348EXPORT_SYMBOL(generic_writepages);
2349
2350int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2351{
2352 int ret;
2353
2354 if (wbc->nr_to_write <= 0)
2355 return 0;
2356 if (mapping->a_ops->writepages)
2357 ret = mapping->a_ops->writepages(mapping, wbc);
2358 else
2359 ret = generic_writepages(mapping, wbc);
2360 return ret;
2361}
2362
2363/**
2364 * write_one_page - write out a single page and optionally wait on I/O
2365 * @page: the page to write
2366 * @wait: if true, wait on writeout
2367 *
2368 * The page must be locked by the caller and will be unlocked upon return.
2369 *
2370 * write_one_page() returns a negative error code if I/O failed.
2371 */
2372int write_one_page(struct page *page, int wait)
2373{
2374 struct address_space *mapping = page->mapping;
2375 int ret = 0;
2376 struct writeback_control wbc = {
2377 .sync_mode = WB_SYNC_ALL,
2378 .nr_to_write = 1,
2379 };
2380
2381 BUG_ON(!PageLocked(page));
2382
2383 if (wait)
2384 wait_on_page_writeback(page);
2385
2386 if (clear_page_dirty_for_io(page)) {
2387 get_page(page);
2388 ret = mapping->a_ops->writepage(page, &wbc);
2389 if (ret == 0 && wait) {
2390 wait_on_page_writeback(page);
2391 if (PageError(page))
2392 ret = -EIO;
2393 }
2394 put_page(page);
2395 } else {
2396 unlock_page(page);
2397 }
2398 return ret;
2399}
2400EXPORT_SYMBOL(write_one_page);
2401
2402/*
2403 * For address_spaces which do not use buffers nor write back.
2404 */
2405int __set_page_dirty_no_writeback(struct page *page)
2406{
2407 if (!PageDirty(page))
2408 return !TestSetPageDirty(page);
2409 return 0;
2410}
2411
2412/*
2413 * Helper function for set_page_dirty family.
2414 *
2415 * Caller must hold lock_page_memcg().
2416 *
2417 * NOTE: This relies on being atomic wrt interrupts.
2418 */
2419void account_page_dirtied(struct page *page, struct address_space *mapping)
2420{
2421 struct inode *inode = mapping->host;
2422
2423 trace_writeback_dirty_page(page, mapping);
2424
2425 if (mapping_cap_account_dirty(mapping)) {
2426 struct bdi_writeback *wb;
2427
2428 inode_attach_wb(inode, page);
2429 wb = inode_to_wb(inode);
2430
2431 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2432 __inc_zone_page_state(page, NR_FILE_DIRTY);
2433 __inc_zone_page_state(page, NR_DIRTIED);
2434 __inc_wb_stat(wb, WB_RECLAIMABLE);
2435 __inc_wb_stat(wb, WB_DIRTIED);
2436 task_io_account_write(PAGE_SIZE);
2437 current->nr_dirtied++;
2438 this_cpu_inc(bdp_ratelimits);
2439 }
2440}
2441EXPORT_SYMBOL(account_page_dirtied);
2442
2443/*
2444 * Helper function for deaccounting dirty page without writeback.
2445 *
2446 * Caller must hold lock_page_memcg().
2447 */
2448void account_page_cleaned(struct page *page, struct address_space *mapping,
2449 struct bdi_writeback *wb)
2450{
2451 if (mapping_cap_account_dirty(mapping)) {
2452 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2453 dec_zone_page_state(page, NR_FILE_DIRTY);
2454 dec_wb_stat(wb, WB_RECLAIMABLE);
2455 task_io_account_cancelled_write(PAGE_SIZE);
2456 }
2457}
2458
2459/*
2460 * For address_spaces which do not use buffers. Just tag the page as dirty in
2461 * its radix tree.
2462 *
2463 * This is also used when a single buffer is being dirtied: we want to set the
2464 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2465 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2466 *
2467 * The caller must ensure this doesn't race with truncation. Most will simply
2468 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2469 * the pte lock held, which also locks out truncation.
2470 */
2471int __set_page_dirty_nobuffers(struct page *page)
2472{
2473 lock_page_memcg(page);
2474 if (!TestSetPageDirty(page)) {
2475 struct address_space *mapping = page_mapping(page);
2476 unsigned long flags;
2477
2478 if (!mapping) {
2479 unlock_page_memcg(page);
2480 return 1;
2481 }
2482
2483 spin_lock_irqsave(&mapping->tree_lock, flags);
2484 BUG_ON(page_mapping(page) != mapping);
2485 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2486 account_page_dirtied(page, mapping);
2487 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2488 PAGECACHE_TAG_DIRTY);
2489 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2490 unlock_page_memcg(page);
2491
2492 if (mapping->host) {
2493 /* !PageAnon && !swapper_space */
2494 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2495 }
2496 return 1;
2497 }
2498 unlock_page_memcg(page);
2499 return 0;
2500}
2501EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2502
2503/*
2504 * Call this whenever redirtying a page, to de-account the dirty counters
2505 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2506 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2507 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2508 * control.
2509 */
2510void account_page_redirty(struct page *page)
2511{
2512 struct address_space *mapping = page->mapping;
2513
2514 if (mapping && mapping_cap_account_dirty(mapping)) {
2515 struct inode *inode = mapping->host;
2516 struct bdi_writeback *wb;
2517 bool locked;
2518
2519 wb = unlocked_inode_to_wb_begin(inode, &locked);
2520 current->nr_dirtied--;
2521 dec_zone_page_state(page, NR_DIRTIED);
2522 dec_wb_stat(wb, WB_DIRTIED);
2523 unlocked_inode_to_wb_end(inode, locked);
2524 }
2525}
2526EXPORT_SYMBOL(account_page_redirty);
2527
2528/*
2529 * When a writepage implementation decides that it doesn't want to write this
2530 * page for some reason, it should redirty the locked page via
2531 * redirty_page_for_writepage() and it should then unlock the page and return 0
2532 */
2533int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2534{
2535 int ret;
2536
2537 wbc->pages_skipped++;
2538 ret = __set_page_dirty_nobuffers(page);
2539 account_page_redirty(page);
2540 return ret;
2541}
2542EXPORT_SYMBOL(redirty_page_for_writepage);
2543
2544/*
2545 * Dirty a page.
2546 *
2547 * For pages with a mapping this should be done under the page lock
2548 * for the benefit of asynchronous memory errors who prefer a consistent
2549 * dirty state. This rule can be broken in some special cases,
2550 * but should be better not to.
2551 *
2552 * If the mapping doesn't provide a set_page_dirty a_op, then
2553 * just fall through and assume that it wants buffer_heads.
2554 */
2555int set_page_dirty(struct page *page)
2556{
2557 struct address_space *mapping = page_mapping(page);
2558
2559 if (likely(mapping)) {
2560 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2561 /*
2562 * readahead/lru_deactivate_page could remain
2563 * PG_readahead/PG_reclaim due to race with end_page_writeback
2564 * About readahead, if the page is written, the flags would be
2565 * reset. So no problem.
2566 * About lru_deactivate_page, if the page is redirty, the flag
2567 * will be reset. So no problem. but if the page is used by readahead
2568 * it will confuse readahead and make it restart the size rampup
2569 * process. But it's a trivial problem.
2570 */
2571 if (PageReclaim(page))
2572 ClearPageReclaim(page);
2573#ifdef CONFIG_BLOCK
2574 if (!spd)
2575 spd = __set_page_dirty_buffers;
2576#endif
2577 return (*spd)(page);
2578 }
2579 if (!PageDirty(page)) {
2580 if (!TestSetPageDirty(page))
2581 return 1;
2582 }
2583 return 0;
2584}
2585EXPORT_SYMBOL(set_page_dirty);
2586
2587/*
2588 * set_page_dirty() is racy if the caller has no reference against
2589 * page->mapping->host, and if the page is unlocked. This is because another
2590 * CPU could truncate the page off the mapping and then free the mapping.
2591 *
2592 * Usually, the page _is_ locked, or the caller is a user-space process which
2593 * holds a reference on the inode by having an open file.
2594 *
2595 * In other cases, the page should be locked before running set_page_dirty().
2596 */
2597int set_page_dirty_lock(struct page *page)
2598{
2599 int ret;
2600
2601 lock_page(page);
2602 ret = set_page_dirty(page);
2603 unlock_page(page);
2604 return ret;
2605}
2606EXPORT_SYMBOL(set_page_dirty_lock);
2607
2608/*
2609 * This cancels just the dirty bit on the kernel page itself, it does NOT
2610 * actually remove dirty bits on any mmap's that may be around. It also
2611 * leaves the page tagged dirty, so any sync activity will still find it on
2612 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2613 * look at the dirty bits in the VM.
2614 *
2615 * Doing this should *normally* only ever be done when a page is truncated,
2616 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2617 * this when it notices that somebody has cleaned out all the buffers on a
2618 * page without actually doing it through the VM. Can you say "ext3 is
2619 * horribly ugly"? Thought you could.
2620 */
2621void cancel_dirty_page(struct page *page)
2622{
2623 struct address_space *mapping = page_mapping(page);
2624
2625 if (mapping_cap_account_dirty(mapping)) {
2626 struct inode *inode = mapping->host;
2627 struct bdi_writeback *wb;
2628 bool locked;
2629
2630 lock_page_memcg(page);
2631 wb = unlocked_inode_to_wb_begin(inode, &locked);
2632
2633 if (TestClearPageDirty(page))
2634 account_page_cleaned(page, mapping, wb);
2635
2636 unlocked_inode_to_wb_end(inode, locked);
2637 unlock_page_memcg(page);
2638 } else {
2639 ClearPageDirty(page);
2640 }
2641}
2642EXPORT_SYMBOL(cancel_dirty_page);
2643
2644/*
2645 * Clear a page's dirty flag, while caring for dirty memory accounting.
2646 * Returns true if the page was previously dirty.
2647 *
2648 * This is for preparing to put the page under writeout. We leave the page
2649 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2650 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2651 * implementation will run either set_page_writeback() or set_page_dirty(),
2652 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2653 * back into sync.
2654 *
2655 * This incoherency between the page's dirty flag and radix-tree tag is
2656 * unfortunate, but it only exists while the page is locked.
2657 */
2658int clear_page_dirty_for_io(struct page *page)
2659{
2660 struct address_space *mapping = page_mapping(page);
2661 int ret = 0;
2662
2663 BUG_ON(!PageLocked(page));
2664
2665 if (mapping && mapping_cap_account_dirty(mapping)) {
2666 struct inode *inode = mapping->host;
2667 struct bdi_writeback *wb;
2668 bool locked;
2669
2670 /*
2671 * Yes, Virginia, this is indeed insane.
2672 *
2673 * We use this sequence to make sure that
2674 * (a) we account for dirty stats properly
2675 * (b) we tell the low-level filesystem to
2676 * mark the whole page dirty if it was
2677 * dirty in a pagetable. Only to then
2678 * (c) clean the page again and return 1 to
2679 * cause the writeback.
2680 *
2681 * This way we avoid all nasty races with the
2682 * dirty bit in multiple places and clearing
2683 * them concurrently from different threads.
2684 *
2685 * Note! Normally the "set_page_dirty(page)"
2686 * has no effect on the actual dirty bit - since
2687 * that will already usually be set. But we
2688 * need the side effects, and it can help us
2689 * avoid races.
2690 *
2691 * We basically use the page "master dirty bit"
2692 * as a serialization point for all the different
2693 * threads doing their things.
2694 */
2695 if (page_mkclean(page))
2696 set_page_dirty(page);
2697 /*
2698 * We carefully synchronise fault handlers against
2699 * installing a dirty pte and marking the page dirty
2700 * at this point. We do this by having them hold the
2701 * page lock while dirtying the page, and pages are
2702 * always locked coming in here, so we get the desired
2703 * exclusion.
2704 */
2705 wb = unlocked_inode_to_wb_begin(inode, &locked);
2706 if (TestClearPageDirty(page)) {
2707 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2708 dec_zone_page_state(page, NR_FILE_DIRTY);
2709 dec_wb_stat(wb, WB_RECLAIMABLE);
2710 ret = 1;
2711 }
2712 unlocked_inode_to_wb_end(inode, locked);
2713 return ret;
2714 }
2715 return TestClearPageDirty(page);
2716}
2717EXPORT_SYMBOL(clear_page_dirty_for_io);
2718
2719int test_clear_page_writeback(struct page *page)
2720{
2721 struct address_space *mapping = page_mapping(page);
2722 int ret;
2723
2724 lock_page_memcg(page);
2725 if (mapping) {
2726 struct inode *inode = mapping->host;
2727 struct backing_dev_info *bdi = inode_to_bdi(inode);
2728 unsigned long flags;
2729
2730 spin_lock_irqsave(&mapping->tree_lock, flags);
2731 ret = TestClearPageWriteback(page);
2732 if (ret) {
2733 radix_tree_tag_clear(&mapping->page_tree,
2734 page_index(page),
2735 PAGECACHE_TAG_WRITEBACK);
2736 if (bdi_cap_account_writeback(bdi)) {
2737 struct bdi_writeback *wb = inode_to_wb(inode);
2738
2739 __dec_wb_stat(wb, WB_WRITEBACK);
2740 __wb_writeout_inc(wb);
2741 }
2742 }
2743 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2744 } else {
2745 ret = TestClearPageWriteback(page);
2746 }
2747 if (ret) {
2748 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2749 dec_zone_page_state(page, NR_WRITEBACK);
2750 inc_zone_page_state(page, NR_WRITTEN);
2751 }
2752 unlock_page_memcg(page);
2753 return ret;
2754}
2755
2756int __test_set_page_writeback(struct page *page, bool keep_write)
2757{
2758 struct address_space *mapping = page_mapping(page);
2759 int ret;
2760
2761 lock_page_memcg(page);
2762 if (mapping) {
2763 struct inode *inode = mapping->host;
2764 struct backing_dev_info *bdi = inode_to_bdi(inode);
2765 unsigned long flags;
2766
2767 spin_lock_irqsave(&mapping->tree_lock, flags);
2768 ret = TestSetPageWriteback(page);
2769 if (!ret) {
2770 radix_tree_tag_set(&mapping->page_tree,
2771 page_index(page),
2772 PAGECACHE_TAG_WRITEBACK);
2773 if (bdi_cap_account_writeback(bdi))
2774 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2775 }
2776 if (!PageDirty(page))
2777 radix_tree_tag_clear(&mapping->page_tree,
2778 page_index(page),
2779 PAGECACHE_TAG_DIRTY);
2780 if (!keep_write)
2781 radix_tree_tag_clear(&mapping->page_tree,
2782 page_index(page),
2783 PAGECACHE_TAG_TOWRITE);
2784 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2785 } else {
2786 ret = TestSetPageWriteback(page);
2787 }
2788 if (!ret) {
2789 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2790 inc_zone_page_state(page, NR_WRITEBACK);
2791 }
2792 unlock_page_memcg(page);
2793 return ret;
2794
2795}
2796EXPORT_SYMBOL(__test_set_page_writeback);
2797
2798/*
2799 * Return true if any of the pages in the mapping are marked with the
2800 * passed tag.
2801 */
2802int mapping_tagged(struct address_space *mapping, int tag)
2803{
2804 return radix_tree_tagged(&mapping->page_tree, tag);
2805}
2806EXPORT_SYMBOL(mapping_tagged);
2807
2808/**
2809 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2810 * @page: The page to wait on.
2811 *
2812 * This function determines if the given page is related to a backing device
2813 * that requires page contents to be held stable during writeback. If so, then
2814 * it will wait for any pending writeback to complete.
2815 */
2816void wait_for_stable_page(struct page *page)
2817{
2818 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2819 wait_on_page_writeback(page);
2820}
2821EXPORT_SYMBOL_GPL(wait_for_stable_page);
1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/export.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.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>
38
39/*
40 * Sleep at most 200ms at a time in balance_dirty_pages().
41 */
42#define MAX_PAUSE max(HZ/5, 1)
43
44/*
45 * Try to keep balance_dirty_pages() call intervals higher than this many pages
46 * by raising pause time to max_pause when falls below it.
47 */
48#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
49
50/*
51 * Estimate write bandwidth at 200ms intervals.
52 */
53#define BANDWIDTH_INTERVAL max(HZ/5, 1)
54
55#define RATELIMIT_CALC_SHIFT 10
56
57/*
58 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
59 * will look to see if it needs to force writeback or throttling.
60 */
61static long ratelimit_pages = 32;
62
63/* The following parameters are exported via /proc/sys/vm */
64
65/*
66 * Start background writeback (via writeback threads) at this percentage
67 */
68int dirty_background_ratio = 10;
69
70/*
71 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
72 * dirty_background_ratio * the amount of dirtyable memory
73 */
74unsigned long dirty_background_bytes;
75
76/*
77 * free highmem will not be subtracted from the total free memory
78 * for calculating free ratios if vm_highmem_is_dirtyable is true
79 */
80int vm_highmem_is_dirtyable;
81
82/*
83 * The generator of dirty data starts writeback at this percentage
84 */
85int vm_dirty_ratio = 20;
86
87/*
88 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
89 * vm_dirty_ratio * the amount of dirtyable memory
90 */
91unsigned long vm_dirty_bytes;
92
93/*
94 * The interval between `kupdate'-style writebacks
95 */
96unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
97
98EXPORT_SYMBOL_GPL(dirty_writeback_interval);
99
100/*
101 * The longest time for which data is allowed to remain dirty
102 */
103unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
104
105/*
106 * Flag that makes the machine dump writes/reads and block dirtyings.
107 */
108int block_dump;
109
110/*
111 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112 * a full sync is triggered after this time elapses without any disk activity.
113 */
114int laptop_mode;
115
116EXPORT_SYMBOL(laptop_mode);
117
118/* End of sysctl-exported parameters */
119
120unsigned long global_dirty_limit;
121
122/*
123 * Scale the writeback cache size proportional to the relative writeout speeds.
124 *
125 * We do this by keeping a floating proportion between BDIs, based on page
126 * writeback completions [end_page_writeback()]. Those devices that write out
127 * pages fastest will get the larger share, while the slower will get a smaller
128 * share.
129 *
130 * We use page writeout completions because we are interested in getting rid of
131 * dirty pages. Having them written out is the primary goal.
132 *
133 * We introduce a concept of time, a period over which we measure these events,
134 * because demand can/will vary over time. The length of this period itself is
135 * measured in page writeback completions.
136 *
137 */
138static struct prop_descriptor vm_completions;
139
140/*
141 * Work out the current dirty-memory clamping and background writeout
142 * thresholds.
143 *
144 * The main aim here is to lower them aggressively if there is a lot of mapped
145 * memory around. To avoid stressing page reclaim with lots of unreclaimable
146 * pages. It is better to clamp down on writers than to start swapping, and
147 * performing lots of scanning.
148 *
149 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
150 *
151 * We don't permit the clamping level to fall below 5% - that is getting rather
152 * excessive.
153 *
154 * We make sure that the background writeout level is below the adjusted
155 * clamping level.
156 */
157
158/*
159 * In a memory zone, there is a certain amount of pages we consider
160 * available for the page cache, which is essentially the number of
161 * free and reclaimable pages, minus some zone reserves to protect
162 * lowmem and the ability to uphold the zone's watermarks without
163 * requiring writeback.
164 *
165 * This number of dirtyable pages is the base value of which the
166 * user-configurable dirty ratio is the effictive number of pages that
167 * are allowed to be actually dirtied. Per individual zone, or
168 * globally by using the sum of dirtyable pages over all zones.
169 *
170 * Because the user is allowed to specify the dirty limit globally as
171 * absolute number of bytes, calculating the per-zone dirty limit can
172 * require translating the configured limit into a percentage of
173 * global dirtyable memory first.
174 */
175
176static unsigned long highmem_dirtyable_memory(unsigned long total)
177{
178#ifdef CONFIG_HIGHMEM
179 int node;
180 unsigned long x = 0;
181
182 for_each_node_state(node, N_HIGH_MEMORY) {
183 struct zone *z =
184 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
185
186 x += zone_page_state(z, NR_FREE_PAGES) +
187 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
188 }
189 /*
190 * Make sure that the number of highmem pages is never larger
191 * than the number of the total dirtyable memory. This can only
192 * occur in very strange VM situations but we want to make sure
193 * that this does not occur.
194 */
195 return min(x, total);
196#else
197 return 0;
198#endif
199}
200
201/**
202 * global_dirtyable_memory - number of globally dirtyable pages
203 *
204 * Returns the global number of pages potentially available for dirty
205 * page cache. This is the base value for the global dirty limits.
206 */
207static unsigned long global_dirtyable_memory(void)
208{
209 unsigned long x;
210
211 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
212 dirty_balance_reserve;
213
214 if (!vm_highmem_is_dirtyable)
215 x -= highmem_dirtyable_memory(x);
216
217 return x + 1; /* Ensure that we never return 0 */
218}
219
220/*
221 * global_dirty_limits - background-writeback and dirty-throttling thresholds
222 *
223 * Calculate the dirty thresholds based on sysctl parameters
224 * - vm.dirty_background_ratio or vm.dirty_background_bytes
225 * - vm.dirty_ratio or vm.dirty_bytes
226 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
227 * real-time tasks.
228 */
229void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
230{
231 unsigned long background;
232 unsigned long dirty;
233 unsigned long uninitialized_var(available_memory);
234 struct task_struct *tsk;
235
236 if (!vm_dirty_bytes || !dirty_background_bytes)
237 available_memory = global_dirtyable_memory();
238
239 if (vm_dirty_bytes)
240 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
241 else
242 dirty = (vm_dirty_ratio * available_memory) / 100;
243
244 if (dirty_background_bytes)
245 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
246 else
247 background = (dirty_background_ratio * available_memory) / 100;
248
249 if (background >= dirty)
250 background = dirty / 2;
251 tsk = current;
252 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
253 background += background / 4;
254 dirty += dirty / 4;
255 }
256 *pbackground = background;
257 *pdirty = dirty;
258 trace_global_dirty_state(background, dirty);
259}
260
261/**
262 * zone_dirtyable_memory - number of dirtyable pages in a zone
263 * @zone: the zone
264 *
265 * Returns the zone's number of pages potentially available for dirty
266 * page cache. This is the base value for the per-zone dirty limits.
267 */
268static unsigned long zone_dirtyable_memory(struct zone *zone)
269{
270 /*
271 * The effective global number of dirtyable pages may exclude
272 * highmem as a big-picture measure to keep the ratio between
273 * dirty memory and lowmem reasonable.
274 *
275 * But this function is purely about the individual zone and a
276 * highmem zone can hold its share of dirty pages, so we don't
277 * care about vm_highmem_is_dirtyable here.
278 */
279 return zone_page_state(zone, NR_FREE_PAGES) +
280 zone_reclaimable_pages(zone) -
281 zone->dirty_balance_reserve;
282}
283
284/**
285 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
286 * @zone: the zone
287 *
288 * Returns the maximum number of dirty pages allowed in a zone, based
289 * on the zone's dirtyable memory.
290 */
291static unsigned long zone_dirty_limit(struct zone *zone)
292{
293 unsigned long zone_memory = zone_dirtyable_memory(zone);
294 struct task_struct *tsk = current;
295 unsigned long dirty;
296
297 if (vm_dirty_bytes)
298 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
299 zone_memory / global_dirtyable_memory();
300 else
301 dirty = vm_dirty_ratio * zone_memory / 100;
302
303 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
304 dirty += dirty / 4;
305
306 return dirty;
307}
308
309/**
310 * zone_dirty_ok - tells whether a zone is within its dirty limits
311 * @zone: the zone to check
312 *
313 * Returns %true when the dirty pages in @zone are within the zone's
314 * dirty limit, %false if the limit is exceeded.
315 */
316bool zone_dirty_ok(struct zone *zone)
317{
318 unsigned long limit = zone_dirty_limit(zone);
319
320 return zone_page_state(zone, NR_FILE_DIRTY) +
321 zone_page_state(zone, NR_UNSTABLE_NFS) +
322 zone_page_state(zone, NR_WRITEBACK) <= limit;
323}
324
325/*
326 * couple the period to the dirty_ratio:
327 *
328 * period/2 ~ roundup_pow_of_two(dirty limit)
329 */
330static int calc_period_shift(void)
331{
332 unsigned long dirty_total;
333
334 if (vm_dirty_bytes)
335 dirty_total = vm_dirty_bytes / PAGE_SIZE;
336 else
337 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
338 100;
339 return 2 + ilog2(dirty_total - 1);
340}
341
342/*
343 * update the period when the dirty threshold changes.
344 */
345static void update_completion_period(void)
346{
347 int shift = calc_period_shift();
348 prop_change_shift(&vm_completions, shift);
349
350 writeback_set_ratelimit();
351}
352
353int dirty_background_ratio_handler(struct ctl_table *table, int write,
354 void __user *buffer, size_t *lenp,
355 loff_t *ppos)
356{
357 int ret;
358
359 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
360 if (ret == 0 && write)
361 dirty_background_bytes = 0;
362 return ret;
363}
364
365int dirty_background_bytes_handler(struct ctl_table *table, int write,
366 void __user *buffer, size_t *lenp,
367 loff_t *ppos)
368{
369 int ret;
370
371 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
372 if (ret == 0 && write)
373 dirty_background_ratio = 0;
374 return ret;
375}
376
377int dirty_ratio_handler(struct ctl_table *table, int write,
378 void __user *buffer, size_t *lenp,
379 loff_t *ppos)
380{
381 int old_ratio = vm_dirty_ratio;
382 int ret;
383
384 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
385 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
386 update_completion_period();
387 vm_dirty_bytes = 0;
388 }
389 return ret;
390}
391
392int dirty_bytes_handler(struct ctl_table *table, int write,
393 void __user *buffer, size_t *lenp,
394 loff_t *ppos)
395{
396 unsigned long old_bytes = vm_dirty_bytes;
397 int ret;
398
399 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
400 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
401 update_completion_period();
402 vm_dirty_ratio = 0;
403 }
404 return ret;
405}
406
407/*
408 * Increment the BDI's writeout completion count and the global writeout
409 * completion count. Called from test_clear_page_writeback().
410 */
411static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
412{
413 __inc_bdi_stat(bdi, BDI_WRITTEN);
414 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
415 bdi->max_prop_frac);
416}
417
418void bdi_writeout_inc(struct backing_dev_info *bdi)
419{
420 unsigned long flags;
421
422 local_irq_save(flags);
423 __bdi_writeout_inc(bdi);
424 local_irq_restore(flags);
425}
426EXPORT_SYMBOL_GPL(bdi_writeout_inc);
427
428/*
429 * Obtain an accurate fraction of the BDI's portion.
430 */
431static void bdi_writeout_fraction(struct backing_dev_info *bdi,
432 long *numerator, long *denominator)
433{
434 prop_fraction_percpu(&vm_completions, &bdi->completions,
435 numerator, denominator);
436}
437
438/*
439 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
440 * registered backing devices, which, for obvious reasons, can not
441 * exceed 100%.
442 */
443static unsigned int bdi_min_ratio;
444
445int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
446{
447 int ret = 0;
448
449 spin_lock_bh(&bdi_lock);
450 if (min_ratio > bdi->max_ratio) {
451 ret = -EINVAL;
452 } else {
453 min_ratio -= bdi->min_ratio;
454 if (bdi_min_ratio + min_ratio < 100) {
455 bdi_min_ratio += min_ratio;
456 bdi->min_ratio += min_ratio;
457 } else {
458 ret = -EINVAL;
459 }
460 }
461 spin_unlock_bh(&bdi_lock);
462
463 return ret;
464}
465
466int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
467{
468 int ret = 0;
469
470 if (max_ratio > 100)
471 return -EINVAL;
472
473 spin_lock_bh(&bdi_lock);
474 if (bdi->min_ratio > max_ratio) {
475 ret = -EINVAL;
476 } else {
477 bdi->max_ratio = max_ratio;
478 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
479 }
480 spin_unlock_bh(&bdi_lock);
481
482 return ret;
483}
484EXPORT_SYMBOL(bdi_set_max_ratio);
485
486static unsigned long dirty_freerun_ceiling(unsigned long thresh,
487 unsigned long bg_thresh)
488{
489 return (thresh + bg_thresh) / 2;
490}
491
492static unsigned long hard_dirty_limit(unsigned long thresh)
493{
494 return max(thresh, global_dirty_limit);
495}
496
497/**
498 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
499 * @bdi: the backing_dev_info to query
500 * @dirty: global dirty limit in pages
501 *
502 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
503 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
504 *
505 * Note that balance_dirty_pages() will only seriously take it as a hard limit
506 * when sleeping max_pause per page is not enough to keep the dirty pages under
507 * control. For example, when the device is completely stalled due to some error
508 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
509 * In the other normal situations, it acts more gently by throttling the tasks
510 * more (rather than completely block them) when the bdi dirty pages go high.
511 *
512 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
513 * - starving fast devices
514 * - piling up dirty pages (that will take long time to sync) on slow devices
515 *
516 * The bdi's share of dirty limit will be adapting to its throughput and
517 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
518 */
519unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
520{
521 u64 bdi_dirty;
522 long numerator, denominator;
523
524 /*
525 * Calculate this BDI's share of the dirty ratio.
526 */
527 bdi_writeout_fraction(bdi, &numerator, &denominator);
528
529 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
530 bdi_dirty *= numerator;
531 do_div(bdi_dirty, denominator);
532
533 bdi_dirty += (dirty * bdi->min_ratio) / 100;
534 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
535 bdi_dirty = dirty * bdi->max_ratio / 100;
536
537 return bdi_dirty;
538}
539
540/*
541 * Dirty position control.
542 *
543 * (o) global/bdi setpoints
544 *
545 * We want the dirty pages be balanced around the global/bdi setpoints.
546 * When the number of dirty pages is higher/lower than the setpoint, the
547 * dirty position control ratio (and hence task dirty ratelimit) will be
548 * decreased/increased to bring the dirty pages back to the setpoint.
549 *
550 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
551 *
552 * if (dirty < setpoint) scale up pos_ratio
553 * if (dirty > setpoint) scale down pos_ratio
554 *
555 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
556 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
557 *
558 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
559 *
560 * (o) global control line
561 *
562 * ^ pos_ratio
563 * |
564 * | |<===== global dirty control scope ======>|
565 * 2.0 .............*
566 * | .*
567 * | . *
568 * | . *
569 * | . *
570 * | . *
571 * | . *
572 * 1.0 ................................*
573 * | . . *
574 * | . . *
575 * | . . *
576 * | . . *
577 * | . . *
578 * 0 +------------.------------------.----------------------*------------->
579 * freerun^ setpoint^ limit^ dirty pages
580 *
581 * (o) bdi control line
582 *
583 * ^ pos_ratio
584 * |
585 * | *
586 * | *
587 * | *
588 * | *
589 * | * |<=========== span ============>|
590 * 1.0 .......................*
591 * | . *
592 * | . *
593 * | . *
594 * | . *
595 * | . *
596 * | . *
597 * | . *
598 * | . *
599 * | . *
600 * | . *
601 * | . *
602 * 1/4 ...............................................* * * * * * * * * * * *
603 * | . .
604 * | . .
605 * | . .
606 * 0 +----------------------.-------------------------------.------------->
607 * bdi_setpoint^ x_intercept^
608 *
609 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
610 * be smoothly throttled down to normal if it starts high in situations like
611 * - start writing to a slow SD card and a fast disk at the same time. The SD
612 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
613 * - the bdi dirty thresh drops quickly due to change of JBOD workload
614 */
615static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
616 unsigned long thresh,
617 unsigned long bg_thresh,
618 unsigned long dirty,
619 unsigned long bdi_thresh,
620 unsigned long bdi_dirty)
621{
622 unsigned long write_bw = bdi->avg_write_bandwidth;
623 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
624 unsigned long limit = hard_dirty_limit(thresh);
625 unsigned long x_intercept;
626 unsigned long setpoint; /* dirty pages' target balance point */
627 unsigned long bdi_setpoint;
628 unsigned long span;
629 long long pos_ratio; /* for scaling up/down the rate limit */
630 long x;
631
632 if (unlikely(dirty >= limit))
633 return 0;
634
635 /*
636 * global setpoint
637 *
638 * setpoint - dirty 3
639 * f(dirty) := 1.0 + (----------------)
640 * limit - setpoint
641 *
642 * it's a 3rd order polynomial that subjects to
643 *
644 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
645 * (2) f(setpoint) = 1.0 => the balance point
646 * (3) f(limit) = 0 => the hard limit
647 * (4) df/dx <= 0 => negative feedback control
648 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
649 * => fast response on large errors; small oscillation near setpoint
650 */
651 setpoint = (freerun + limit) / 2;
652 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
653 limit - setpoint + 1);
654 pos_ratio = x;
655 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
656 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
657 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
658
659 /*
660 * We have computed basic pos_ratio above based on global situation. If
661 * the bdi is over/under its share of dirty pages, we want to scale
662 * pos_ratio further down/up. That is done by the following mechanism.
663 */
664
665 /*
666 * bdi setpoint
667 *
668 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
669 *
670 * x_intercept - bdi_dirty
671 * := --------------------------
672 * x_intercept - bdi_setpoint
673 *
674 * The main bdi control line is a linear function that subjects to
675 *
676 * (1) f(bdi_setpoint) = 1.0
677 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
678 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
679 *
680 * For single bdi case, the dirty pages are observed to fluctuate
681 * regularly within range
682 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
683 * for various filesystems, where (2) can yield in a reasonable 12.5%
684 * fluctuation range for pos_ratio.
685 *
686 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
687 * own size, so move the slope over accordingly and choose a slope that
688 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
689 */
690 if (unlikely(bdi_thresh > thresh))
691 bdi_thresh = thresh;
692 /*
693 * It's very possible that bdi_thresh is close to 0 not because the
694 * device is slow, but that it has remained inactive for long time.
695 * Honour such devices a reasonable good (hopefully IO efficient)
696 * threshold, so that the occasional writes won't be blocked and active
697 * writes can rampup the threshold quickly.
698 */
699 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
700 /*
701 * scale global setpoint to bdi's:
702 * bdi_setpoint = setpoint * bdi_thresh / thresh
703 */
704 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
705 bdi_setpoint = setpoint * (u64)x >> 16;
706 /*
707 * Use span=(8*write_bw) in single bdi case as indicated by
708 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
709 *
710 * bdi_thresh thresh - bdi_thresh
711 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
712 * thresh thresh
713 */
714 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
715 x_intercept = bdi_setpoint + span;
716
717 if (bdi_dirty < x_intercept - span / 4) {
718 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
719 x_intercept - bdi_setpoint + 1);
720 } else
721 pos_ratio /= 4;
722
723 /*
724 * bdi reserve area, safeguard against dirty pool underrun and disk idle
725 * It may push the desired control point of global dirty pages higher
726 * than setpoint.
727 */
728 x_intercept = bdi_thresh / 2;
729 if (bdi_dirty < x_intercept) {
730 if (bdi_dirty > x_intercept / 8)
731 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
732 else
733 pos_ratio *= 8;
734 }
735
736 return pos_ratio;
737}
738
739static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
740 unsigned long elapsed,
741 unsigned long written)
742{
743 const unsigned long period = roundup_pow_of_two(3 * HZ);
744 unsigned long avg = bdi->avg_write_bandwidth;
745 unsigned long old = bdi->write_bandwidth;
746 u64 bw;
747
748 /*
749 * bw = written * HZ / elapsed
750 *
751 * bw * elapsed + write_bandwidth * (period - elapsed)
752 * write_bandwidth = ---------------------------------------------------
753 * period
754 */
755 bw = written - bdi->written_stamp;
756 bw *= HZ;
757 if (unlikely(elapsed > period)) {
758 do_div(bw, elapsed);
759 avg = bw;
760 goto out;
761 }
762 bw += (u64)bdi->write_bandwidth * (period - elapsed);
763 bw >>= ilog2(period);
764
765 /*
766 * one more level of smoothing, for filtering out sudden spikes
767 */
768 if (avg > old && old >= (unsigned long)bw)
769 avg -= (avg - old) >> 3;
770
771 if (avg < old && old <= (unsigned long)bw)
772 avg += (old - avg) >> 3;
773
774out:
775 bdi->write_bandwidth = bw;
776 bdi->avg_write_bandwidth = avg;
777}
778
779/*
780 * The global dirtyable memory and dirty threshold could be suddenly knocked
781 * down by a large amount (eg. on the startup of KVM in a swapless system).
782 * This may throw the system into deep dirty exceeded state and throttle
783 * heavy/light dirtiers alike. To retain good responsiveness, maintain
784 * global_dirty_limit for tracking slowly down to the knocked down dirty
785 * threshold.
786 */
787static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
788{
789 unsigned long limit = global_dirty_limit;
790
791 /*
792 * Follow up in one step.
793 */
794 if (limit < thresh) {
795 limit = thresh;
796 goto update;
797 }
798
799 /*
800 * Follow down slowly. Use the higher one as the target, because thresh
801 * may drop below dirty. This is exactly the reason to introduce
802 * global_dirty_limit which is guaranteed to lie above the dirty pages.
803 */
804 thresh = max(thresh, dirty);
805 if (limit > thresh) {
806 limit -= (limit - thresh) >> 5;
807 goto update;
808 }
809 return;
810update:
811 global_dirty_limit = limit;
812}
813
814static void global_update_bandwidth(unsigned long thresh,
815 unsigned long dirty,
816 unsigned long now)
817{
818 static DEFINE_SPINLOCK(dirty_lock);
819 static unsigned long update_time;
820
821 /*
822 * check locklessly first to optimize away locking for the most time
823 */
824 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
825 return;
826
827 spin_lock(&dirty_lock);
828 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
829 update_dirty_limit(thresh, dirty);
830 update_time = now;
831 }
832 spin_unlock(&dirty_lock);
833}
834
835/*
836 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
837 *
838 * Normal bdi tasks will be curbed at or below it in long term.
839 * Obviously it should be around (write_bw / N) when there are N dd tasks.
840 */
841static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
842 unsigned long thresh,
843 unsigned long bg_thresh,
844 unsigned long dirty,
845 unsigned long bdi_thresh,
846 unsigned long bdi_dirty,
847 unsigned long dirtied,
848 unsigned long elapsed)
849{
850 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
851 unsigned long limit = hard_dirty_limit(thresh);
852 unsigned long setpoint = (freerun + limit) / 2;
853 unsigned long write_bw = bdi->avg_write_bandwidth;
854 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
855 unsigned long dirty_rate;
856 unsigned long task_ratelimit;
857 unsigned long balanced_dirty_ratelimit;
858 unsigned long pos_ratio;
859 unsigned long step;
860 unsigned long x;
861
862 /*
863 * The dirty rate will match the writeout rate in long term, except
864 * when dirty pages are truncated by userspace or re-dirtied by FS.
865 */
866 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
867
868 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
869 bdi_thresh, bdi_dirty);
870 /*
871 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
872 */
873 task_ratelimit = (u64)dirty_ratelimit *
874 pos_ratio >> RATELIMIT_CALC_SHIFT;
875 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
876
877 /*
878 * A linear estimation of the "balanced" throttle rate. The theory is,
879 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
880 * dirty_rate will be measured to be (N * task_ratelimit). So the below
881 * formula will yield the balanced rate limit (write_bw / N).
882 *
883 * Note that the expanded form is not a pure rate feedback:
884 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
885 * but also takes pos_ratio into account:
886 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
887 *
888 * (1) is not realistic because pos_ratio also takes part in balancing
889 * the dirty rate. Consider the state
890 * pos_ratio = 0.5 (3)
891 * rate = 2 * (write_bw / N) (4)
892 * If (1) is used, it will stuck in that state! Because each dd will
893 * be throttled at
894 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
895 * yielding
896 * dirty_rate = N * task_ratelimit = write_bw (6)
897 * put (6) into (1) we get
898 * rate_(i+1) = rate_(i) (7)
899 *
900 * So we end up using (2) to always keep
901 * rate_(i+1) ~= (write_bw / N) (8)
902 * regardless of the value of pos_ratio. As long as (8) is satisfied,
903 * pos_ratio is able to drive itself to 1.0, which is not only where
904 * the dirty count meet the setpoint, but also where the slope of
905 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
906 */
907 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
908 dirty_rate | 1);
909 /*
910 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
911 */
912 if (unlikely(balanced_dirty_ratelimit > write_bw))
913 balanced_dirty_ratelimit = write_bw;
914
915 /*
916 * We could safely do this and return immediately:
917 *
918 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
919 *
920 * However to get a more stable dirty_ratelimit, the below elaborated
921 * code makes use of task_ratelimit to filter out sigular points and
922 * limit the step size.
923 *
924 * The below code essentially only uses the relative value of
925 *
926 * task_ratelimit - dirty_ratelimit
927 * = (pos_ratio - 1) * dirty_ratelimit
928 *
929 * which reflects the direction and size of dirty position error.
930 */
931
932 /*
933 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
934 * task_ratelimit is on the same side of dirty_ratelimit, too.
935 * For example, when
936 * - dirty_ratelimit > balanced_dirty_ratelimit
937 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
938 * lowering dirty_ratelimit will help meet both the position and rate
939 * control targets. Otherwise, don't update dirty_ratelimit if it will
940 * only help meet the rate target. After all, what the users ultimately
941 * feel and care are stable dirty rate and small position error.
942 *
943 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
944 * and filter out the sigular points of balanced_dirty_ratelimit. Which
945 * keeps jumping around randomly and can even leap far away at times
946 * due to the small 200ms estimation period of dirty_rate (we want to
947 * keep that period small to reduce time lags).
948 */
949 step = 0;
950 if (dirty < setpoint) {
951 x = min(bdi->balanced_dirty_ratelimit,
952 min(balanced_dirty_ratelimit, task_ratelimit));
953 if (dirty_ratelimit < x)
954 step = x - dirty_ratelimit;
955 } else {
956 x = max(bdi->balanced_dirty_ratelimit,
957 max(balanced_dirty_ratelimit, task_ratelimit));
958 if (dirty_ratelimit > x)
959 step = dirty_ratelimit - x;
960 }
961
962 /*
963 * Don't pursue 100% rate matching. It's impossible since the balanced
964 * rate itself is constantly fluctuating. So decrease the track speed
965 * when it gets close to the target. Helps eliminate pointless tremors.
966 */
967 step >>= dirty_ratelimit / (2 * step + 1);
968 /*
969 * Limit the tracking speed to avoid overshooting.
970 */
971 step = (step + 7) / 8;
972
973 if (dirty_ratelimit < balanced_dirty_ratelimit)
974 dirty_ratelimit += step;
975 else
976 dirty_ratelimit -= step;
977
978 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
979 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
980
981 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
982}
983
984void __bdi_update_bandwidth(struct backing_dev_info *bdi,
985 unsigned long thresh,
986 unsigned long bg_thresh,
987 unsigned long dirty,
988 unsigned long bdi_thresh,
989 unsigned long bdi_dirty,
990 unsigned long start_time)
991{
992 unsigned long now = jiffies;
993 unsigned long elapsed = now - bdi->bw_time_stamp;
994 unsigned long dirtied;
995 unsigned long written;
996
997 /*
998 * rate-limit, only update once every 200ms.
999 */
1000 if (elapsed < BANDWIDTH_INTERVAL)
1001 return;
1002
1003 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1004 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1005
1006 /*
1007 * Skip quiet periods when disk bandwidth is under-utilized.
1008 * (at least 1s idle time between two flusher runs)
1009 */
1010 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1011 goto snapshot;
1012
1013 if (thresh) {
1014 global_update_bandwidth(thresh, dirty, now);
1015 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1016 bdi_thresh, bdi_dirty,
1017 dirtied, elapsed);
1018 }
1019 bdi_update_write_bandwidth(bdi, elapsed, written);
1020
1021snapshot:
1022 bdi->dirtied_stamp = dirtied;
1023 bdi->written_stamp = written;
1024 bdi->bw_time_stamp = now;
1025}
1026
1027static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1028 unsigned long thresh,
1029 unsigned long bg_thresh,
1030 unsigned long dirty,
1031 unsigned long bdi_thresh,
1032 unsigned long bdi_dirty,
1033 unsigned long start_time)
1034{
1035 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1036 return;
1037 spin_lock(&bdi->wb.list_lock);
1038 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1039 bdi_thresh, bdi_dirty, start_time);
1040 spin_unlock(&bdi->wb.list_lock);
1041}
1042
1043/*
1044 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1045 * will look to see if it needs to start dirty throttling.
1046 *
1047 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1048 * global_page_state() too often. So scale it near-sqrt to the safety margin
1049 * (the number of pages we may dirty without exceeding the dirty limits).
1050 */
1051static unsigned long dirty_poll_interval(unsigned long dirty,
1052 unsigned long thresh)
1053{
1054 if (thresh > dirty)
1055 return 1UL << (ilog2(thresh - dirty) >> 1);
1056
1057 return 1;
1058}
1059
1060static long bdi_max_pause(struct backing_dev_info *bdi,
1061 unsigned long bdi_dirty)
1062{
1063 long bw = bdi->avg_write_bandwidth;
1064 long t;
1065
1066 /*
1067 * Limit pause time for small memory systems. If sleeping for too long
1068 * time, a small pool of dirty/writeback pages may go empty and disk go
1069 * idle.
1070 *
1071 * 8 serves as the safety ratio.
1072 */
1073 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1074 t++;
1075
1076 return min_t(long, t, MAX_PAUSE);
1077}
1078
1079static long bdi_min_pause(struct backing_dev_info *bdi,
1080 long max_pause,
1081 unsigned long task_ratelimit,
1082 unsigned long dirty_ratelimit,
1083 int *nr_dirtied_pause)
1084{
1085 long hi = ilog2(bdi->avg_write_bandwidth);
1086 long lo = ilog2(bdi->dirty_ratelimit);
1087 long t; /* target pause */
1088 long pause; /* estimated next pause */
1089 int pages; /* target nr_dirtied_pause */
1090
1091 /* target for 10ms pause on 1-dd case */
1092 t = max(1, HZ / 100);
1093
1094 /*
1095 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1096 * overheads.
1097 *
1098 * (N * 10ms) on 2^N concurrent tasks.
1099 */
1100 if (hi > lo)
1101 t += (hi - lo) * (10 * HZ) / 1024;
1102
1103 /*
1104 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1105 * on the much more stable dirty_ratelimit. However the next pause time
1106 * will be computed based on task_ratelimit and the two rate limits may
1107 * depart considerably at some time. Especially if task_ratelimit goes
1108 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1109 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1110 * result task_ratelimit won't be executed faithfully, which could
1111 * eventually bring down dirty_ratelimit.
1112 *
1113 * We apply two rules to fix it up:
1114 * 1) try to estimate the next pause time and if necessary, use a lower
1115 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1116 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1117 * 2) limit the target pause time to max_pause/2, so that the normal
1118 * small fluctuations of task_ratelimit won't trigger rule (1) and
1119 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1120 */
1121 t = min(t, 1 + max_pause / 2);
1122 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1123
1124 /*
1125 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1126 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1127 * When the 16 consecutive reads are often interrupted by some dirty
1128 * throttling pause during the async writes, cfq will go into idles
1129 * (deadline is fine). So push nr_dirtied_pause as high as possible
1130 * until reaches DIRTY_POLL_THRESH=32 pages.
1131 */
1132 if (pages < DIRTY_POLL_THRESH) {
1133 t = max_pause;
1134 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1135 if (pages > DIRTY_POLL_THRESH) {
1136 pages = DIRTY_POLL_THRESH;
1137 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1138 }
1139 }
1140
1141 pause = HZ * pages / (task_ratelimit + 1);
1142 if (pause > max_pause) {
1143 t = max_pause;
1144 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1145 }
1146
1147 *nr_dirtied_pause = pages;
1148 /*
1149 * The minimal pause time will normally be half the target pause time.
1150 */
1151 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1152}
1153
1154/*
1155 * balance_dirty_pages() must be called by processes which are generating dirty
1156 * data. It looks at the number of dirty pages in the machine and will force
1157 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1158 * If we're over `background_thresh' then the writeback threads are woken to
1159 * perform some writeout.
1160 */
1161static void balance_dirty_pages(struct address_space *mapping,
1162 unsigned long pages_dirtied)
1163{
1164 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1165 unsigned long bdi_reclaimable;
1166 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1167 unsigned long bdi_dirty;
1168 unsigned long freerun;
1169 unsigned long background_thresh;
1170 unsigned long dirty_thresh;
1171 unsigned long bdi_thresh;
1172 long period;
1173 long pause;
1174 long max_pause;
1175 long min_pause;
1176 int nr_dirtied_pause;
1177 bool dirty_exceeded = false;
1178 unsigned long task_ratelimit;
1179 unsigned long dirty_ratelimit;
1180 unsigned long pos_ratio;
1181 struct backing_dev_info *bdi = mapping->backing_dev_info;
1182 unsigned long start_time = jiffies;
1183
1184 for (;;) {
1185 unsigned long now = jiffies;
1186
1187 /*
1188 * Unstable writes are a feature of certain networked
1189 * filesystems (i.e. NFS) in which data may have been
1190 * written to the server's write cache, but has not yet
1191 * been flushed to permanent storage.
1192 */
1193 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1194 global_page_state(NR_UNSTABLE_NFS);
1195 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1196
1197 global_dirty_limits(&background_thresh, &dirty_thresh);
1198
1199 /*
1200 * Throttle it only when the background writeback cannot
1201 * catch-up. This avoids (excessively) small writeouts
1202 * when the bdi limits are ramping up.
1203 */
1204 freerun = dirty_freerun_ceiling(dirty_thresh,
1205 background_thresh);
1206 if (nr_dirty <= freerun) {
1207 current->dirty_paused_when = now;
1208 current->nr_dirtied = 0;
1209 current->nr_dirtied_pause =
1210 dirty_poll_interval(nr_dirty, dirty_thresh);
1211 break;
1212 }
1213
1214 if (unlikely(!writeback_in_progress(bdi)))
1215 bdi_start_background_writeback(bdi);
1216
1217 /*
1218 * bdi_thresh is not treated as some limiting factor as
1219 * dirty_thresh, due to reasons
1220 * - in JBOD setup, bdi_thresh can fluctuate a lot
1221 * - in a system with HDD and USB key, the USB key may somehow
1222 * go into state (bdi_dirty >> bdi_thresh) either because
1223 * bdi_dirty starts high, or because bdi_thresh drops low.
1224 * In this case we don't want to hard throttle the USB key
1225 * dirtiers for 100 seconds until bdi_dirty drops under
1226 * bdi_thresh. Instead the auxiliary bdi control line in
1227 * bdi_position_ratio() will let the dirtier task progress
1228 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1229 */
1230 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1231
1232 /*
1233 * In order to avoid the stacked BDI deadlock we need
1234 * to ensure we accurately count the 'dirty' pages when
1235 * the threshold is low.
1236 *
1237 * Otherwise it would be possible to get thresh+n pages
1238 * reported dirty, even though there are thresh-m pages
1239 * actually dirty; with m+n sitting in the percpu
1240 * deltas.
1241 */
1242 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1243 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1244 bdi_dirty = bdi_reclaimable +
1245 bdi_stat_sum(bdi, BDI_WRITEBACK);
1246 } else {
1247 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1248 bdi_dirty = bdi_reclaimable +
1249 bdi_stat(bdi, BDI_WRITEBACK);
1250 }
1251
1252 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1253 (nr_dirty > dirty_thresh);
1254 if (dirty_exceeded && !bdi->dirty_exceeded)
1255 bdi->dirty_exceeded = 1;
1256
1257 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1258 nr_dirty, bdi_thresh, bdi_dirty,
1259 start_time);
1260
1261 dirty_ratelimit = bdi->dirty_ratelimit;
1262 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1263 background_thresh, nr_dirty,
1264 bdi_thresh, bdi_dirty);
1265 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1266 RATELIMIT_CALC_SHIFT;
1267 max_pause = bdi_max_pause(bdi, bdi_dirty);
1268 min_pause = bdi_min_pause(bdi, max_pause,
1269 task_ratelimit, dirty_ratelimit,
1270 &nr_dirtied_pause);
1271
1272 if (unlikely(task_ratelimit == 0)) {
1273 period = max_pause;
1274 pause = max_pause;
1275 goto pause;
1276 }
1277 period = HZ * pages_dirtied / task_ratelimit;
1278 pause = period;
1279 if (current->dirty_paused_when)
1280 pause -= now - current->dirty_paused_when;
1281 /*
1282 * For less than 1s think time (ext3/4 may block the dirtier
1283 * for up to 800ms from time to time on 1-HDD; so does xfs,
1284 * however at much less frequency), try to compensate it in
1285 * future periods by updating the virtual time; otherwise just
1286 * do a reset, as it may be a light dirtier.
1287 */
1288 if (pause < min_pause) {
1289 trace_balance_dirty_pages(bdi,
1290 dirty_thresh,
1291 background_thresh,
1292 nr_dirty,
1293 bdi_thresh,
1294 bdi_dirty,
1295 dirty_ratelimit,
1296 task_ratelimit,
1297 pages_dirtied,
1298 period,
1299 min(pause, 0L),
1300 start_time);
1301 if (pause < -HZ) {
1302 current->dirty_paused_when = now;
1303 current->nr_dirtied = 0;
1304 } else if (period) {
1305 current->dirty_paused_when += period;
1306 current->nr_dirtied = 0;
1307 } else if (current->nr_dirtied_pause <= pages_dirtied)
1308 current->nr_dirtied_pause += pages_dirtied;
1309 break;
1310 }
1311 if (unlikely(pause > max_pause)) {
1312 /* for occasional dropped task_ratelimit */
1313 now += min(pause - max_pause, max_pause);
1314 pause = max_pause;
1315 }
1316
1317pause:
1318 trace_balance_dirty_pages(bdi,
1319 dirty_thresh,
1320 background_thresh,
1321 nr_dirty,
1322 bdi_thresh,
1323 bdi_dirty,
1324 dirty_ratelimit,
1325 task_ratelimit,
1326 pages_dirtied,
1327 period,
1328 pause,
1329 start_time);
1330 __set_current_state(TASK_KILLABLE);
1331 io_schedule_timeout(pause);
1332
1333 current->dirty_paused_when = now + pause;
1334 current->nr_dirtied = 0;
1335 current->nr_dirtied_pause = nr_dirtied_pause;
1336
1337 /*
1338 * This is typically equal to (nr_dirty < dirty_thresh) and can
1339 * also keep "1000+ dd on a slow USB stick" under control.
1340 */
1341 if (task_ratelimit)
1342 break;
1343
1344 /*
1345 * In the case of an unresponding NFS server and the NFS dirty
1346 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1347 * to go through, so that tasks on them still remain responsive.
1348 *
1349 * In theory 1 page is enough to keep the comsumer-producer
1350 * pipe going: the flusher cleans 1 page => the task dirties 1
1351 * more page. However bdi_dirty has accounting errors. So use
1352 * the larger and more IO friendly bdi_stat_error.
1353 */
1354 if (bdi_dirty <= bdi_stat_error(bdi))
1355 break;
1356
1357 if (fatal_signal_pending(current))
1358 break;
1359 }
1360
1361 if (!dirty_exceeded && bdi->dirty_exceeded)
1362 bdi->dirty_exceeded = 0;
1363
1364 if (writeback_in_progress(bdi))
1365 return;
1366
1367 /*
1368 * In laptop mode, we wait until hitting the higher threshold before
1369 * starting background writeout, and then write out all the way down
1370 * to the lower threshold. So slow writers cause minimal disk activity.
1371 *
1372 * In normal mode, we start background writeout at the lower
1373 * background_thresh, to keep the amount of dirty memory low.
1374 */
1375 if (laptop_mode)
1376 return;
1377
1378 if (nr_reclaimable > background_thresh)
1379 bdi_start_background_writeback(bdi);
1380}
1381
1382void set_page_dirty_balance(struct page *page, int page_mkwrite)
1383{
1384 if (set_page_dirty(page) || page_mkwrite) {
1385 struct address_space *mapping = page_mapping(page);
1386
1387 if (mapping)
1388 balance_dirty_pages_ratelimited(mapping);
1389 }
1390}
1391
1392static DEFINE_PER_CPU(int, bdp_ratelimits);
1393
1394/*
1395 * Normal tasks are throttled by
1396 * loop {
1397 * dirty tsk->nr_dirtied_pause pages;
1398 * take a snap in balance_dirty_pages();
1399 * }
1400 * However there is a worst case. If every task exit immediately when dirtied
1401 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1402 * called to throttle the page dirties. The solution is to save the not yet
1403 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1404 * randomly into the running tasks. This works well for the above worst case,
1405 * as the new task will pick up and accumulate the old task's leaked dirty
1406 * count and eventually get throttled.
1407 */
1408DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1409
1410/**
1411 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1412 * @mapping: address_space which was dirtied
1413 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1414 *
1415 * Processes which are dirtying memory should call in here once for each page
1416 * which was newly dirtied. The function will periodically check the system's
1417 * dirty state and will initiate writeback if needed.
1418 *
1419 * On really big machines, get_writeback_state is expensive, so try to avoid
1420 * calling it too often (ratelimiting). But once we're over the dirty memory
1421 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1422 * from overshooting the limit by (ratelimit_pages) each.
1423 */
1424void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1425 unsigned long nr_pages_dirtied)
1426{
1427 struct backing_dev_info *bdi = mapping->backing_dev_info;
1428 int ratelimit;
1429 int *p;
1430
1431 if (!bdi_cap_account_dirty(bdi))
1432 return;
1433
1434 ratelimit = current->nr_dirtied_pause;
1435 if (bdi->dirty_exceeded)
1436 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1437
1438 preempt_disable();
1439 /*
1440 * This prevents one CPU to accumulate too many dirtied pages without
1441 * calling into balance_dirty_pages(), which can happen when there are
1442 * 1000+ tasks, all of them start dirtying pages at exactly the same
1443 * time, hence all honoured too large initial task->nr_dirtied_pause.
1444 */
1445 p = &__get_cpu_var(bdp_ratelimits);
1446 if (unlikely(current->nr_dirtied >= ratelimit))
1447 *p = 0;
1448 else if (unlikely(*p >= ratelimit_pages)) {
1449 *p = 0;
1450 ratelimit = 0;
1451 }
1452 /*
1453 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1454 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1455 * the dirty throttling and livelock other long-run dirtiers.
1456 */
1457 p = &__get_cpu_var(dirty_throttle_leaks);
1458 if (*p > 0 && current->nr_dirtied < ratelimit) {
1459 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1460 *p -= nr_pages_dirtied;
1461 current->nr_dirtied += nr_pages_dirtied;
1462 }
1463 preempt_enable();
1464
1465 if (unlikely(current->nr_dirtied >= ratelimit))
1466 balance_dirty_pages(mapping, current->nr_dirtied);
1467}
1468EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1469
1470void throttle_vm_writeout(gfp_t gfp_mask)
1471{
1472 unsigned long background_thresh;
1473 unsigned long dirty_thresh;
1474
1475 for ( ; ; ) {
1476 global_dirty_limits(&background_thresh, &dirty_thresh);
1477 dirty_thresh = hard_dirty_limit(dirty_thresh);
1478
1479 /*
1480 * Boost the allowable dirty threshold a bit for page
1481 * allocators so they don't get DoS'ed by heavy writers
1482 */
1483 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1484
1485 if (global_page_state(NR_UNSTABLE_NFS) +
1486 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1487 break;
1488 congestion_wait(BLK_RW_ASYNC, HZ/10);
1489
1490 /*
1491 * The caller might hold locks which can prevent IO completion
1492 * or progress in the filesystem. So we cannot just sit here
1493 * waiting for IO to complete.
1494 */
1495 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1496 break;
1497 }
1498}
1499
1500/*
1501 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1502 */
1503int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1504 void __user *buffer, size_t *length, loff_t *ppos)
1505{
1506 proc_dointvec(table, write, buffer, length, ppos);
1507 bdi_arm_supers_timer();
1508 return 0;
1509}
1510
1511#ifdef CONFIG_BLOCK
1512void laptop_mode_timer_fn(unsigned long data)
1513{
1514 struct request_queue *q = (struct request_queue *)data;
1515 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1516 global_page_state(NR_UNSTABLE_NFS);
1517
1518 /*
1519 * We want to write everything out, not just down to the dirty
1520 * threshold
1521 */
1522 if (bdi_has_dirty_io(&q->backing_dev_info))
1523 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1524 WB_REASON_LAPTOP_TIMER);
1525}
1526
1527/*
1528 * We've spun up the disk and we're in laptop mode: schedule writeback
1529 * of all dirty data a few seconds from now. If the flush is already scheduled
1530 * then push it back - the user is still using the disk.
1531 */
1532void laptop_io_completion(struct backing_dev_info *info)
1533{
1534 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1535}
1536
1537/*
1538 * We're in laptop mode and we've just synced. The sync's writes will have
1539 * caused another writeback to be scheduled by laptop_io_completion.
1540 * Nothing needs to be written back anymore, so we unschedule the writeback.
1541 */
1542void laptop_sync_completion(void)
1543{
1544 struct backing_dev_info *bdi;
1545
1546 rcu_read_lock();
1547
1548 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1549 del_timer(&bdi->laptop_mode_wb_timer);
1550
1551 rcu_read_unlock();
1552}
1553#endif
1554
1555/*
1556 * If ratelimit_pages is too high then we can get into dirty-data overload
1557 * if a large number of processes all perform writes at the same time.
1558 * If it is too low then SMP machines will call the (expensive)
1559 * get_writeback_state too often.
1560 *
1561 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1562 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1563 * thresholds.
1564 */
1565
1566void writeback_set_ratelimit(void)
1567{
1568 unsigned long background_thresh;
1569 unsigned long dirty_thresh;
1570 global_dirty_limits(&background_thresh, &dirty_thresh);
1571 global_dirty_limit = dirty_thresh;
1572 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1573 if (ratelimit_pages < 16)
1574 ratelimit_pages = 16;
1575}
1576
1577static int __cpuinit
1578ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1579{
1580 writeback_set_ratelimit();
1581 return NOTIFY_DONE;
1582}
1583
1584static struct notifier_block __cpuinitdata ratelimit_nb = {
1585 .notifier_call = ratelimit_handler,
1586 .next = NULL,
1587};
1588
1589/*
1590 * Called early on to tune the page writeback dirty limits.
1591 *
1592 * We used to scale dirty pages according to how total memory
1593 * related to pages that could be allocated for buffers (by
1594 * comparing nr_free_buffer_pages() to vm_total_pages.
1595 *
1596 * However, that was when we used "dirty_ratio" to scale with
1597 * all memory, and we don't do that any more. "dirty_ratio"
1598 * is now applied to total non-HIGHPAGE memory (by subtracting
1599 * totalhigh_pages from vm_total_pages), and as such we can't
1600 * get into the old insane situation any more where we had
1601 * large amounts of dirty pages compared to a small amount of
1602 * non-HIGHMEM memory.
1603 *
1604 * But we might still want to scale the dirty_ratio by how
1605 * much memory the box has..
1606 */
1607void __init page_writeback_init(void)
1608{
1609 int shift;
1610
1611 writeback_set_ratelimit();
1612 register_cpu_notifier(&ratelimit_nb);
1613
1614 shift = calc_period_shift();
1615 prop_descriptor_init(&vm_completions, shift);
1616}
1617
1618/**
1619 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1620 * @mapping: address space structure to write
1621 * @start: starting page index
1622 * @end: ending page index (inclusive)
1623 *
1624 * This function scans the page range from @start to @end (inclusive) and tags
1625 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1626 * that write_cache_pages (or whoever calls this function) will then use
1627 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1628 * used to avoid livelocking of writeback by a process steadily creating new
1629 * dirty pages in the file (thus it is important for this function to be quick
1630 * so that it can tag pages faster than a dirtying process can create them).
1631 */
1632/*
1633 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1634 */
1635void tag_pages_for_writeback(struct address_space *mapping,
1636 pgoff_t start, pgoff_t end)
1637{
1638#define WRITEBACK_TAG_BATCH 4096
1639 unsigned long tagged;
1640
1641 do {
1642 spin_lock_irq(&mapping->tree_lock);
1643 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1644 &start, end, WRITEBACK_TAG_BATCH,
1645 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1646 spin_unlock_irq(&mapping->tree_lock);
1647 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1648 cond_resched();
1649 /* We check 'start' to handle wrapping when end == ~0UL */
1650 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1651}
1652EXPORT_SYMBOL(tag_pages_for_writeback);
1653
1654/**
1655 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1656 * @mapping: address space structure to write
1657 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1658 * @writepage: function called for each page
1659 * @data: data passed to writepage function
1660 *
1661 * If a page is already under I/O, write_cache_pages() skips it, even
1662 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1663 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1664 * and msync() need to guarantee that all the data which was dirty at the time
1665 * the call was made get new I/O started against them. If wbc->sync_mode is
1666 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1667 * existing IO to complete.
1668 *
1669 * To avoid livelocks (when other process dirties new pages), we first tag
1670 * pages which should be written back with TOWRITE tag and only then start
1671 * writing them. For data-integrity sync we have to be careful so that we do
1672 * not miss some pages (e.g., because some other process has cleared TOWRITE
1673 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1674 * by the process clearing the DIRTY tag (and submitting the page for IO).
1675 */
1676int write_cache_pages(struct address_space *mapping,
1677 struct writeback_control *wbc, writepage_t writepage,
1678 void *data)
1679{
1680 int ret = 0;
1681 int done = 0;
1682 struct pagevec pvec;
1683 int nr_pages;
1684 pgoff_t uninitialized_var(writeback_index);
1685 pgoff_t index;
1686 pgoff_t end; /* Inclusive */
1687 pgoff_t done_index;
1688 int cycled;
1689 int range_whole = 0;
1690 int tag;
1691
1692 pagevec_init(&pvec, 0);
1693 if (wbc->range_cyclic) {
1694 writeback_index = mapping->writeback_index; /* prev offset */
1695 index = writeback_index;
1696 if (index == 0)
1697 cycled = 1;
1698 else
1699 cycled = 0;
1700 end = -1;
1701 } else {
1702 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1703 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1704 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1705 range_whole = 1;
1706 cycled = 1; /* ignore range_cyclic tests */
1707 }
1708 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1709 tag = PAGECACHE_TAG_TOWRITE;
1710 else
1711 tag = PAGECACHE_TAG_DIRTY;
1712retry:
1713 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1714 tag_pages_for_writeback(mapping, index, end);
1715 done_index = index;
1716 while (!done && (index <= end)) {
1717 int i;
1718
1719 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1720 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1721 if (nr_pages == 0)
1722 break;
1723
1724 for (i = 0; i < nr_pages; i++) {
1725 struct page *page = pvec.pages[i];
1726
1727 /*
1728 * At this point, the page may be truncated or
1729 * invalidated (changing page->mapping to NULL), or
1730 * even swizzled back from swapper_space to tmpfs file
1731 * mapping. However, page->index will not change
1732 * because we have a reference on the page.
1733 */
1734 if (page->index > end) {
1735 /*
1736 * can't be range_cyclic (1st pass) because
1737 * end == -1 in that case.
1738 */
1739 done = 1;
1740 break;
1741 }
1742
1743 done_index = page->index;
1744
1745 lock_page(page);
1746
1747 /*
1748 * Page truncated or invalidated. We can freely skip it
1749 * then, even for data integrity operations: the page
1750 * has disappeared concurrently, so there could be no
1751 * real expectation of this data interity operation
1752 * even if there is now a new, dirty page at the same
1753 * pagecache address.
1754 */
1755 if (unlikely(page->mapping != mapping)) {
1756continue_unlock:
1757 unlock_page(page);
1758 continue;
1759 }
1760
1761 if (!PageDirty(page)) {
1762 /* someone wrote it for us */
1763 goto continue_unlock;
1764 }
1765
1766 if (PageWriteback(page)) {
1767 if (wbc->sync_mode != WB_SYNC_NONE)
1768 wait_on_page_writeback(page);
1769 else
1770 goto continue_unlock;
1771 }
1772
1773 BUG_ON(PageWriteback(page));
1774 if (!clear_page_dirty_for_io(page))
1775 goto continue_unlock;
1776
1777 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1778 ret = (*writepage)(page, wbc, data);
1779 if (unlikely(ret)) {
1780 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1781 unlock_page(page);
1782 ret = 0;
1783 } else {
1784 /*
1785 * done_index is set past this page,
1786 * so media errors will not choke
1787 * background writeout for the entire
1788 * file. This has consequences for
1789 * range_cyclic semantics (ie. it may
1790 * not be suitable for data integrity
1791 * writeout).
1792 */
1793 done_index = page->index + 1;
1794 done = 1;
1795 break;
1796 }
1797 }
1798
1799 /*
1800 * We stop writing back only if we are not doing
1801 * integrity sync. In case of integrity sync we have to
1802 * keep going until we have written all the pages
1803 * we tagged for writeback prior to entering this loop.
1804 */
1805 if (--wbc->nr_to_write <= 0 &&
1806 wbc->sync_mode == WB_SYNC_NONE) {
1807 done = 1;
1808 break;
1809 }
1810 }
1811 pagevec_release(&pvec);
1812 cond_resched();
1813 }
1814 if (!cycled && !done) {
1815 /*
1816 * range_cyclic:
1817 * We hit the last page and there is more work to be done: wrap
1818 * back to the start of the file
1819 */
1820 cycled = 1;
1821 index = 0;
1822 end = writeback_index - 1;
1823 goto retry;
1824 }
1825 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1826 mapping->writeback_index = done_index;
1827
1828 return ret;
1829}
1830EXPORT_SYMBOL(write_cache_pages);
1831
1832/*
1833 * Function used by generic_writepages to call the real writepage
1834 * function and set the mapping flags on error
1835 */
1836static int __writepage(struct page *page, struct writeback_control *wbc,
1837 void *data)
1838{
1839 struct address_space *mapping = data;
1840 int ret = mapping->a_ops->writepage(page, wbc);
1841 mapping_set_error(mapping, ret);
1842 return ret;
1843}
1844
1845/**
1846 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1847 * @mapping: address space structure to write
1848 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1849 *
1850 * This is a library function, which implements the writepages()
1851 * address_space_operation.
1852 */
1853int generic_writepages(struct address_space *mapping,
1854 struct writeback_control *wbc)
1855{
1856 struct blk_plug plug;
1857 int ret;
1858
1859 /* deal with chardevs and other special file */
1860 if (!mapping->a_ops->writepage)
1861 return 0;
1862
1863 blk_start_plug(&plug);
1864 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1865 blk_finish_plug(&plug);
1866 return ret;
1867}
1868
1869EXPORT_SYMBOL(generic_writepages);
1870
1871int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1872{
1873 int ret;
1874
1875 if (wbc->nr_to_write <= 0)
1876 return 0;
1877 if (mapping->a_ops->writepages)
1878 ret = mapping->a_ops->writepages(mapping, wbc);
1879 else
1880 ret = generic_writepages(mapping, wbc);
1881 return ret;
1882}
1883
1884/**
1885 * write_one_page - write out a single page and optionally wait on I/O
1886 * @page: the page to write
1887 * @wait: if true, wait on writeout
1888 *
1889 * The page must be locked by the caller and will be unlocked upon return.
1890 *
1891 * write_one_page() returns a negative error code if I/O failed.
1892 */
1893int write_one_page(struct page *page, int wait)
1894{
1895 struct address_space *mapping = page->mapping;
1896 int ret = 0;
1897 struct writeback_control wbc = {
1898 .sync_mode = WB_SYNC_ALL,
1899 .nr_to_write = 1,
1900 };
1901
1902 BUG_ON(!PageLocked(page));
1903
1904 if (wait)
1905 wait_on_page_writeback(page);
1906
1907 if (clear_page_dirty_for_io(page)) {
1908 page_cache_get(page);
1909 ret = mapping->a_ops->writepage(page, &wbc);
1910 if (ret == 0 && wait) {
1911 wait_on_page_writeback(page);
1912 if (PageError(page))
1913 ret = -EIO;
1914 }
1915 page_cache_release(page);
1916 } else {
1917 unlock_page(page);
1918 }
1919 return ret;
1920}
1921EXPORT_SYMBOL(write_one_page);
1922
1923/*
1924 * For address_spaces which do not use buffers nor write back.
1925 */
1926int __set_page_dirty_no_writeback(struct page *page)
1927{
1928 if (!PageDirty(page))
1929 return !TestSetPageDirty(page);
1930 return 0;
1931}
1932
1933/*
1934 * Helper function for set_page_dirty family.
1935 * NOTE: This relies on being atomic wrt interrupts.
1936 */
1937void account_page_dirtied(struct page *page, struct address_space *mapping)
1938{
1939 if (mapping_cap_account_dirty(mapping)) {
1940 __inc_zone_page_state(page, NR_FILE_DIRTY);
1941 __inc_zone_page_state(page, NR_DIRTIED);
1942 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1943 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1944 task_io_account_write(PAGE_CACHE_SIZE);
1945 current->nr_dirtied++;
1946 this_cpu_inc(bdp_ratelimits);
1947 }
1948}
1949EXPORT_SYMBOL(account_page_dirtied);
1950
1951/*
1952 * Helper function for set_page_writeback family.
1953 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1954 * wrt interrupts.
1955 */
1956void account_page_writeback(struct page *page)
1957{
1958 inc_zone_page_state(page, NR_WRITEBACK);
1959}
1960EXPORT_SYMBOL(account_page_writeback);
1961
1962/*
1963 * For address_spaces which do not use buffers. Just tag the page as dirty in
1964 * its radix tree.
1965 *
1966 * This is also used when a single buffer is being dirtied: we want to set the
1967 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1968 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1969 *
1970 * Most callers have locked the page, which pins the address_space in memory.
1971 * But zap_pte_range() does not lock the page, however in that case the
1972 * mapping is pinned by the vma's ->vm_file reference.
1973 *
1974 * We take care to handle the case where the page was truncated from the
1975 * mapping by re-checking page_mapping() inside tree_lock.
1976 */
1977int __set_page_dirty_nobuffers(struct page *page)
1978{
1979 if (!TestSetPageDirty(page)) {
1980 struct address_space *mapping = page_mapping(page);
1981 struct address_space *mapping2;
1982
1983 if (!mapping)
1984 return 1;
1985
1986 spin_lock_irq(&mapping->tree_lock);
1987 mapping2 = page_mapping(page);
1988 if (mapping2) { /* Race with truncate? */
1989 BUG_ON(mapping2 != mapping);
1990 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1991 account_page_dirtied(page, mapping);
1992 radix_tree_tag_set(&mapping->page_tree,
1993 page_index(page), PAGECACHE_TAG_DIRTY);
1994 }
1995 spin_unlock_irq(&mapping->tree_lock);
1996 if (mapping->host) {
1997 /* !PageAnon && !swapper_space */
1998 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1999 }
2000 return 1;
2001 }
2002 return 0;
2003}
2004EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2005
2006/*
2007 * Call this whenever redirtying a page, to de-account the dirty counters
2008 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2009 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2010 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2011 * control.
2012 */
2013void account_page_redirty(struct page *page)
2014{
2015 struct address_space *mapping = page->mapping;
2016 if (mapping && mapping_cap_account_dirty(mapping)) {
2017 current->nr_dirtied--;
2018 dec_zone_page_state(page, NR_DIRTIED);
2019 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2020 }
2021}
2022EXPORT_SYMBOL(account_page_redirty);
2023
2024/*
2025 * When a writepage implementation decides that it doesn't want to write this
2026 * page for some reason, it should redirty the locked page via
2027 * redirty_page_for_writepage() and it should then unlock the page and return 0
2028 */
2029int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2030{
2031 wbc->pages_skipped++;
2032 account_page_redirty(page);
2033 return __set_page_dirty_nobuffers(page);
2034}
2035EXPORT_SYMBOL(redirty_page_for_writepage);
2036
2037/*
2038 * Dirty a page.
2039 *
2040 * For pages with a mapping this should be done under the page lock
2041 * for the benefit of asynchronous memory errors who prefer a consistent
2042 * dirty state. This rule can be broken in some special cases,
2043 * but should be better not to.
2044 *
2045 * If the mapping doesn't provide a set_page_dirty a_op, then
2046 * just fall through and assume that it wants buffer_heads.
2047 */
2048int set_page_dirty(struct page *page)
2049{
2050 struct address_space *mapping = page_mapping(page);
2051
2052 if (likely(mapping)) {
2053 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2054 /*
2055 * readahead/lru_deactivate_page could remain
2056 * PG_readahead/PG_reclaim due to race with end_page_writeback
2057 * About readahead, if the page is written, the flags would be
2058 * reset. So no problem.
2059 * About lru_deactivate_page, if the page is redirty, the flag
2060 * will be reset. So no problem. but if the page is used by readahead
2061 * it will confuse readahead and make it restart the size rampup
2062 * process. But it's a trivial problem.
2063 */
2064 ClearPageReclaim(page);
2065#ifdef CONFIG_BLOCK
2066 if (!spd)
2067 spd = __set_page_dirty_buffers;
2068#endif
2069 return (*spd)(page);
2070 }
2071 if (!PageDirty(page)) {
2072 if (!TestSetPageDirty(page))
2073 return 1;
2074 }
2075 return 0;
2076}
2077EXPORT_SYMBOL(set_page_dirty);
2078
2079/*
2080 * set_page_dirty() is racy if the caller has no reference against
2081 * page->mapping->host, and if the page is unlocked. This is because another
2082 * CPU could truncate the page off the mapping and then free the mapping.
2083 *
2084 * Usually, the page _is_ locked, or the caller is a user-space process which
2085 * holds a reference on the inode by having an open file.
2086 *
2087 * In other cases, the page should be locked before running set_page_dirty().
2088 */
2089int set_page_dirty_lock(struct page *page)
2090{
2091 int ret;
2092
2093 lock_page(page);
2094 ret = set_page_dirty(page);
2095 unlock_page(page);
2096 return ret;
2097}
2098EXPORT_SYMBOL(set_page_dirty_lock);
2099
2100/*
2101 * Clear a page's dirty flag, while caring for dirty memory accounting.
2102 * Returns true if the page was previously dirty.
2103 *
2104 * This is for preparing to put the page under writeout. We leave the page
2105 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2106 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2107 * implementation will run either set_page_writeback() or set_page_dirty(),
2108 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2109 * back into sync.
2110 *
2111 * This incoherency between the page's dirty flag and radix-tree tag is
2112 * unfortunate, but it only exists while the page is locked.
2113 */
2114int clear_page_dirty_for_io(struct page *page)
2115{
2116 struct address_space *mapping = page_mapping(page);
2117
2118 BUG_ON(!PageLocked(page));
2119
2120 if (mapping && mapping_cap_account_dirty(mapping)) {
2121 /*
2122 * Yes, Virginia, this is indeed insane.
2123 *
2124 * We use this sequence to make sure that
2125 * (a) we account for dirty stats properly
2126 * (b) we tell the low-level filesystem to
2127 * mark the whole page dirty if it was
2128 * dirty in a pagetable. Only to then
2129 * (c) clean the page again and return 1 to
2130 * cause the writeback.
2131 *
2132 * This way we avoid all nasty races with the
2133 * dirty bit in multiple places and clearing
2134 * them concurrently from different threads.
2135 *
2136 * Note! Normally the "set_page_dirty(page)"
2137 * has no effect on the actual dirty bit - since
2138 * that will already usually be set. But we
2139 * need the side effects, and it can help us
2140 * avoid races.
2141 *
2142 * We basically use the page "master dirty bit"
2143 * as a serialization point for all the different
2144 * threads doing their things.
2145 */
2146 if (page_mkclean(page))
2147 set_page_dirty(page);
2148 /*
2149 * We carefully synchronise fault handlers against
2150 * installing a dirty pte and marking the page dirty
2151 * at this point. We do this by having them hold the
2152 * page lock at some point after installing their
2153 * pte, but before marking the page dirty.
2154 * Pages are always locked coming in here, so we get
2155 * the desired exclusion. See mm/memory.c:do_wp_page()
2156 * for more comments.
2157 */
2158 if (TestClearPageDirty(page)) {
2159 dec_zone_page_state(page, NR_FILE_DIRTY);
2160 dec_bdi_stat(mapping->backing_dev_info,
2161 BDI_RECLAIMABLE);
2162 return 1;
2163 }
2164 return 0;
2165 }
2166 return TestClearPageDirty(page);
2167}
2168EXPORT_SYMBOL(clear_page_dirty_for_io);
2169
2170int test_clear_page_writeback(struct page *page)
2171{
2172 struct address_space *mapping = page_mapping(page);
2173 int ret;
2174
2175 if (mapping) {
2176 struct backing_dev_info *bdi = mapping->backing_dev_info;
2177 unsigned long flags;
2178
2179 spin_lock_irqsave(&mapping->tree_lock, flags);
2180 ret = TestClearPageWriteback(page);
2181 if (ret) {
2182 radix_tree_tag_clear(&mapping->page_tree,
2183 page_index(page),
2184 PAGECACHE_TAG_WRITEBACK);
2185 if (bdi_cap_account_writeback(bdi)) {
2186 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2187 __bdi_writeout_inc(bdi);
2188 }
2189 }
2190 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2191 } else {
2192 ret = TestClearPageWriteback(page);
2193 }
2194 if (ret) {
2195 dec_zone_page_state(page, NR_WRITEBACK);
2196 inc_zone_page_state(page, NR_WRITTEN);
2197 }
2198 return ret;
2199}
2200
2201int test_set_page_writeback(struct page *page)
2202{
2203 struct address_space *mapping = page_mapping(page);
2204 int ret;
2205
2206 if (mapping) {
2207 struct backing_dev_info *bdi = mapping->backing_dev_info;
2208 unsigned long flags;
2209
2210 spin_lock_irqsave(&mapping->tree_lock, flags);
2211 ret = TestSetPageWriteback(page);
2212 if (!ret) {
2213 radix_tree_tag_set(&mapping->page_tree,
2214 page_index(page),
2215 PAGECACHE_TAG_WRITEBACK);
2216 if (bdi_cap_account_writeback(bdi))
2217 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2218 }
2219 if (!PageDirty(page))
2220 radix_tree_tag_clear(&mapping->page_tree,
2221 page_index(page),
2222 PAGECACHE_TAG_DIRTY);
2223 radix_tree_tag_clear(&mapping->page_tree,
2224 page_index(page),
2225 PAGECACHE_TAG_TOWRITE);
2226 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2227 } else {
2228 ret = TestSetPageWriteback(page);
2229 }
2230 if (!ret)
2231 account_page_writeback(page);
2232 return ret;
2233
2234}
2235EXPORT_SYMBOL(test_set_page_writeback);
2236
2237/*
2238 * Return true if any of the pages in the mapping are marked with the
2239 * passed tag.
2240 */
2241int mapping_tagged(struct address_space *mapping, int tag)
2242{
2243 return radix_tree_tagged(&mapping->page_tree, tag);
2244}
2245EXPORT_SYMBOL(mapping_tagged);