Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/page-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
11 * 10Apr2002 Andrew Morton
12 * Initial version
13 */
14
15#include <linux/kernel.h>
16#include <linux/export.h>
17#include <linux/spinlock.h>
18#include <linux/fs.h>
19#include <linux/mm.h>
20#include <linux/swap.h>
21#include <linux/slab.h>
22#include <linux/pagemap.h>
23#include <linux/writeback.h>
24#include <linux/init.h>
25#include <linux/backing-dev.h>
26#include <linux/task_io_accounting_ops.h>
27#include <linux/blkdev.h>
28#include <linux/mpage.h>
29#include <linux/rmap.h>
30#include <linux/percpu.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/pagevec.h>
36#include <linux/timer.h>
37#include <linux/sched/rt.h>
38#include <linux/sched/signal.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 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
120struct wb_domain global_wb_domain;
121
122/* consolidated parameters for balance_dirty_pages() and its subroutines */
123struct dirty_throttle_control {
124#ifdef CONFIG_CGROUP_WRITEBACK
125 struct wb_domain *dom;
126 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
127#endif
128 struct bdi_writeback *wb;
129 struct fprop_local_percpu *wb_completions;
130
131 unsigned long avail; /* dirtyable */
132 unsigned long dirty; /* file_dirty + write + nfs */
133 unsigned long thresh; /* dirty threshold */
134 unsigned long bg_thresh; /* dirty background threshold */
135
136 unsigned long wb_dirty; /* per-wb counterparts */
137 unsigned long wb_thresh;
138 unsigned long wb_bg_thresh;
139
140 unsigned long pos_ratio;
141};
142
143/*
144 * Length of period for aging writeout fractions of bdis. This is an
145 * arbitrarily chosen number. The longer the period, the slower fractions will
146 * reflect changes in current writeout rate.
147 */
148#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
149
150#ifdef CONFIG_CGROUP_WRITEBACK
151
152#define GDTC_INIT(__wb) .wb = (__wb), \
153 .dom = &global_wb_domain, \
154 .wb_completions = &(__wb)->completions
155
156#define GDTC_INIT_NO_WB .dom = &global_wb_domain
157
158#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
159 .dom = mem_cgroup_wb_domain(__wb), \
160 .wb_completions = &(__wb)->memcg_completions, \
161 .gdtc = __gdtc
162
163static bool mdtc_valid(struct dirty_throttle_control *dtc)
164{
165 return dtc->dom;
166}
167
168static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
172
173static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
174{
175 return mdtc->gdtc;
176}
177
178static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
179{
180 return &wb->memcg_completions;
181}
182
183static void wb_min_max_ratio(struct bdi_writeback *wb,
184 unsigned long *minp, unsigned long *maxp)
185{
186 unsigned long this_bw = wb->avg_write_bandwidth;
187 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
188 unsigned long long min = wb->bdi->min_ratio;
189 unsigned long long max = wb->bdi->max_ratio;
190
191 /*
192 * @wb may already be clean by the time control reaches here and
193 * the total may not include its bw.
194 */
195 if (this_bw < tot_bw) {
196 if (min) {
197 min *= this_bw;
198 min = div64_ul(min, tot_bw);
199 }
200 if (max < 100) {
201 max *= this_bw;
202 max = div64_ul(max, tot_bw);
203 }
204 }
205
206 *minp = min;
207 *maxp = max;
208}
209
210#else /* CONFIG_CGROUP_WRITEBACK */
211
212#define GDTC_INIT(__wb) .wb = (__wb), \
213 .wb_completions = &(__wb)->completions
214#define GDTC_INIT_NO_WB
215#define MDTC_INIT(__wb, __gdtc)
216
217static bool mdtc_valid(struct dirty_throttle_control *dtc)
218{
219 return false;
220}
221
222static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
223{
224 return &global_wb_domain;
225}
226
227static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
228{
229 return NULL;
230}
231
232static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
233{
234 return NULL;
235}
236
237static void wb_min_max_ratio(struct bdi_writeback *wb,
238 unsigned long *minp, unsigned long *maxp)
239{
240 *minp = wb->bdi->min_ratio;
241 *maxp = wb->bdi->max_ratio;
242}
243
244#endif /* CONFIG_CGROUP_WRITEBACK */
245
246/*
247 * In a memory zone, there is a certain amount of pages we consider
248 * available for the page cache, which is essentially the number of
249 * free and reclaimable pages, minus some zone reserves to protect
250 * lowmem and the ability to uphold the zone's watermarks without
251 * requiring writeback.
252 *
253 * This number of dirtyable pages is the base value of which the
254 * user-configurable dirty ratio is the effective number of pages that
255 * are allowed to be actually dirtied. Per individual zone, or
256 * globally by using the sum of dirtyable pages over all zones.
257 *
258 * Because the user is allowed to specify the dirty limit globally as
259 * absolute number of bytes, calculating the per-zone dirty limit can
260 * require translating the configured limit into a percentage of
261 * global dirtyable memory first.
262 */
263
264/**
265 * node_dirtyable_memory - number of dirtyable pages in a node
266 * @pgdat: the node
267 *
268 * Return: the node's number of pages potentially available for dirty
269 * page cache. This is the base value for the per-node dirty limits.
270 */
271static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
272{
273 unsigned long nr_pages = 0;
274 int z;
275
276 for (z = 0; z < MAX_NR_ZONES; z++) {
277 struct zone *zone = pgdat->node_zones + z;
278
279 if (!populated_zone(zone))
280 continue;
281
282 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
283 }
284
285 /*
286 * Pages reserved for the kernel should not be considered
287 * dirtyable, to prevent a situation where reclaim has to
288 * clean pages in order to balance the zones.
289 */
290 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
291
292 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
293 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
294
295 return nr_pages;
296}
297
298static unsigned long highmem_dirtyable_memory(unsigned long total)
299{
300#ifdef CONFIG_HIGHMEM
301 int node;
302 unsigned long x = 0;
303 int i;
304
305 for_each_node_state(node, N_HIGH_MEMORY) {
306 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
307 struct zone *z;
308 unsigned long nr_pages;
309
310 if (!is_highmem_idx(i))
311 continue;
312
313 z = &NODE_DATA(node)->node_zones[i];
314 if (!populated_zone(z))
315 continue;
316
317 nr_pages = zone_page_state(z, NR_FREE_PAGES);
318 /* watch for underflows */
319 nr_pages -= min(nr_pages, high_wmark_pages(z));
320 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
321 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
322 x += nr_pages;
323 }
324 }
325
326 /*
327 * Unreclaimable memory (kernel memory or anonymous memory
328 * without swap) can bring down the dirtyable pages below
329 * the zone's dirty balance reserve and the above calculation
330 * will underflow. However we still want to add in nodes
331 * which are below threshold (negative values) to get a more
332 * accurate calculation but make sure that the total never
333 * underflows.
334 */
335 if ((long)x < 0)
336 x = 0;
337
338 /*
339 * Make sure that the number of highmem pages is never larger
340 * than the number of the total dirtyable memory. This can only
341 * occur in very strange VM situations but we want to make sure
342 * that this does not occur.
343 */
344 return min(x, total);
345#else
346 return 0;
347#endif
348}
349
350/**
351 * global_dirtyable_memory - number of globally dirtyable pages
352 *
353 * Return: the global number of pages potentially available for dirty
354 * page cache. This is the base value for the global dirty limits.
355 */
356static unsigned long global_dirtyable_memory(void)
357{
358 unsigned long x;
359
360 x = global_zone_page_state(NR_FREE_PAGES);
361 /*
362 * Pages reserved for the kernel should not be considered
363 * dirtyable, to prevent a situation where reclaim has to
364 * clean pages in order to balance the zones.
365 */
366 x -= min(x, totalreserve_pages);
367
368 x += global_node_page_state(NR_INACTIVE_FILE);
369 x += global_node_page_state(NR_ACTIVE_FILE);
370
371 if (!vm_highmem_is_dirtyable)
372 x -= highmem_dirtyable_memory(x);
373
374 return x + 1; /* Ensure that we never return 0 */
375}
376
377/**
378 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
379 * @dtc: dirty_throttle_control of interest
380 *
381 * Calculate @dtc->thresh and ->bg_thresh considering
382 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
383 * must ensure that @dtc->avail is set before calling this function. The
384 * dirty limits will be lifted by 1/4 for real-time tasks.
385 */
386static void domain_dirty_limits(struct dirty_throttle_control *dtc)
387{
388 const unsigned long available_memory = dtc->avail;
389 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
390 unsigned long bytes = vm_dirty_bytes;
391 unsigned long bg_bytes = dirty_background_bytes;
392 /* convert ratios to per-PAGE_SIZE for higher precision */
393 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
394 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
395 unsigned long thresh;
396 unsigned long bg_thresh;
397 struct task_struct *tsk;
398
399 /* gdtc is !NULL iff @dtc is for memcg domain */
400 if (gdtc) {
401 unsigned long global_avail = gdtc->avail;
402
403 /*
404 * The byte settings can't be applied directly to memcg
405 * domains. Convert them to ratios by scaling against
406 * globally available memory. As the ratios are in
407 * per-PAGE_SIZE, they can be obtained by dividing bytes by
408 * number of pages.
409 */
410 if (bytes)
411 ratio = min(DIV_ROUND_UP(bytes, global_avail),
412 PAGE_SIZE);
413 if (bg_bytes)
414 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
415 PAGE_SIZE);
416 bytes = bg_bytes = 0;
417 }
418
419 if (bytes)
420 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
421 else
422 thresh = (ratio * available_memory) / PAGE_SIZE;
423
424 if (bg_bytes)
425 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
426 else
427 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
428
429 if (bg_thresh >= thresh)
430 bg_thresh = thresh / 2;
431 tsk = current;
432 if (rt_task(tsk)) {
433 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
434 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
435 }
436 dtc->thresh = thresh;
437 dtc->bg_thresh = bg_thresh;
438
439 /* we should eventually report the domain in the TP */
440 if (!gdtc)
441 trace_global_dirty_state(bg_thresh, thresh);
442}
443
444/**
445 * global_dirty_limits - background-writeback and dirty-throttling thresholds
446 * @pbackground: out parameter for bg_thresh
447 * @pdirty: out parameter for thresh
448 *
449 * Calculate bg_thresh and thresh for global_wb_domain. See
450 * domain_dirty_limits() for details.
451 */
452void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
453{
454 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
455
456 gdtc.avail = global_dirtyable_memory();
457 domain_dirty_limits(&gdtc);
458
459 *pbackground = gdtc.bg_thresh;
460 *pdirty = gdtc.thresh;
461}
462
463/**
464 * node_dirty_limit - maximum number of dirty pages allowed in a node
465 * @pgdat: the node
466 *
467 * Return: the maximum number of dirty pages allowed in a node, based
468 * on the node's dirtyable memory.
469 */
470static unsigned long node_dirty_limit(struct pglist_data *pgdat)
471{
472 unsigned long node_memory = node_dirtyable_memory(pgdat);
473 struct task_struct *tsk = current;
474 unsigned long dirty;
475
476 if (vm_dirty_bytes)
477 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
478 node_memory / global_dirtyable_memory();
479 else
480 dirty = vm_dirty_ratio * node_memory / 100;
481
482 if (rt_task(tsk))
483 dirty += dirty / 4;
484
485 return dirty;
486}
487
488/**
489 * node_dirty_ok - tells whether a node is within its dirty limits
490 * @pgdat: the node to check
491 *
492 * Return: %true when the dirty pages in @pgdat are within the node's
493 * dirty limit, %false if the limit is exceeded.
494 */
495bool node_dirty_ok(struct pglist_data *pgdat)
496{
497 unsigned long limit = node_dirty_limit(pgdat);
498 unsigned long nr_pages = 0;
499
500 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
501 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
502
503 return nr_pages <= limit;
504}
505
506int dirty_background_ratio_handler(struct ctl_table *table, int write,
507 void *buffer, size_t *lenp, loff_t *ppos)
508{
509 int ret;
510
511 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
512 if (ret == 0 && write)
513 dirty_background_bytes = 0;
514 return ret;
515}
516
517int dirty_background_bytes_handler(struct ctl_table *table, int write,
518 void *buffer, size_t *lenp, loff_t *ppos)
519{
520 int ret;
521
522 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
523 if (ret == 0 && write)
524 dirty_background_ratio = 0;
525 return ret;
526}
527
528int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
529 size_t *lenp, loff_t *ppos)
530{
531 int old_ratio = vm_dirty_ratio;
532 int ret;
533
534 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
535 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
536 writeback_set_ratelimit();
537 vm_dirty_bytes = 0;
538 }
539 return ret;
540}
541
542int dirty_bytes_handler(struct ctl_table *table, int write,
543 void *buffer, size_t *lenp, loff_t *ppos)
544{
545 unsigned long old_bytes = vm_dirty_bytes;
546 int ret;
547
548 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
549 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
550 writeback_set_ratelimit();
551 vm_dirty_ratio = 0;
552 }
553 return ret;
554}
555
556static unsigned long wp_next_time(unsigned long cur_time)
557{
558 cur_time += VM_COMPLETIONS_PERIOD_LEN;
559 /* 0 has a special meaning... */
560 if (!cur_time)
561 return 1;
562 return cur_time;
563}
564
565static void wb_domain_writeout_inc(struct wb_domain *dom,
566 struct fprop_local_percpu *completions,
567 unsigned int max_prop_frac)
568{
569 __fprop_inc_percpu_max(&dom->completions, completions,
570 max_prop_frac);
571 /* First event after period switching was turned off? */
572 if (unlikely(!dom->period_time)) {
573 /*
574 * We can race with other __bdi_writeout_inc calls here but
575 * it does not cause any harm since the resulting time when
576 * timer will fire and what is in writeout_period_time will be
577 * roughly the same.
578 */
579 dom->period_time = wp_next_time(jiffies);
580 mod_timer(&dom->period_timer, dom->period_time);
581 }
582}
583
584/*
585 * Increment @wb's writeout completion count and the global writeout
586 * completion count. Called from test_clear_page_writeback().
587 */
588static inline void __wb_writeout_inc(struct bdi_writeback *wb)
589{
590 struct wb_domain *cgdom;
591
592 inc_wb_stat(wb, WB_WRITTEN);
593 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
594 wb->bdi->max_prop_frac);
595
596 cgdom = mem_cgroup_wb_domain(wb);
597 if (cgdom)
598 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
599 wb->bdi->max_prop_frac);
600}
601
602void wb_writeout_inc(struct bdi_writeback *wb)
603{
604 unsigned long flags;
605
606 local_irq_save(flags);
607 __wb_writeout_inc(wb);
608 local_irq_restore(flags);
609}
610EXPORT_SYMBOL_GPL(wb_writeout_inc);
611
612/*
613 * On idle system, we can be called long after we scheduled because we use
614 * deferred timers so count with missed periods.
615 */
616static void writeout_period(struct timer_list *t)
617{
618 struct wb_domain *dom = from_timer(dom, t, period_timer);
619 int miss_periods = (jiffies - dom->period_time) /
620 VM_COMPLETIONS_PERIOD_LEN;
621
622 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
623 dom->period_time = wp_next_time(dom->period_time +
624 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
625 mod_timer(&dom->period_timer, dom->period_time);
626 } else {
627 /*
628 * Aging has zeroed all fractions. Stop wasting CPU on period
629 * updates.
630 */
631 dom->period_time = 0;
632 }
633}
634
635int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
636{
637 memset(dom, 0, sizeof(*dom));
638
639 spin_lock_init(&dom->lock);
640
641 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
642
643 dom->dirty_limit_tstamp = jiffies;
644
645 return fprop_global_init(&dom->completions, gfp);
646}
647
648#ifdef CONFIG_CGROUP_WRITEBACK
649void wb_domain_exit(struct wb_domain *dom)
650{
651 del_timer_sync(&dom->period_timer);
652 fprop_global_destroy(&dom->completions);
653}
654#endif
655
656/*
657 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
658 * registered backing devices, which, for obvious reasons, can not
659 * exceed 100%.
660 */
661static unsigned int bdi_min_ratio;
662
663int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
664{
665 int ret = 0;
666
667 spin_lock_bh(&bdi_lock);
668 if (min_ratio > bdi->max_ratio) {
669 ret = -EINVAL;
670 } else {
671 min_ratio -= bdi->min_ratio;
672 if (bdi_min_ratio + min_ratio < 100) {
673 bdi_min_ratio += min_ratio;
674 bdi->min_ratio += min_ratio;
675 } else {
676 ret = -EINVAL;
677 }
678 }
679 spin_unlock_bh(&bdi_lock);
680
681 return ret;
682}
683
684int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
685{
686 int ret = 0;
687
688 if (max_ratio > 100)
689 return -EINVAL;
690
691 spin_lock_bh(&bdi_lock);
692 if (bdi->min_ratio > max_ratio) {
693 ret = -EINVAL;
694 } else {
695 bdi->max_ratio = max_ratio;
696 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
697 }
698 spin_unlock_bh(&bdi_lock);
699
700 return ret;
701}
702EXPORT_SYMBOL(bdi_set_max_ratio);
703
704static unsigned long dirty_freerun_ceiling(unsigned long thresh,
705 unsigned long bg_thresh)
706{
707 return (thresh + bg_thresh) / 2;
708}
709
710static unsigned long hard_dirty_limit(struct wb_domain *dom,
711 unsigned long thresh)
712{
713 return max(thresh, dom->dirty_limit);
714}
715
716/*
717 * Memory which can be further allocated to a memcg domain is capped by
718 * system-wide clean memory excluding the amount being used in the domain.
719 */
720static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
721 unsigned long filepages, unsigned long headroom)
722{
723 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
724 unsigned long clean = filepages - min(filepages, mdtc->dirty);
725 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
726 unsigned long other_clean = global_clean - min(global_clean, clean);
727
728 mdtc->avail = filepages + min(headroom, other_clean);
729}
730
731/**
732 * __wb_calc_thresh - @wb's share of dirty throttling threshold
733 * @dtc: dirty_throttle_context of interest
734 *
735 * Note that balance_dirty_pages() will only seriously take it as a hard limit
736 * when sleeping max_pause per page is not enough to keep the dirty pages under
737 * control. For example, when the device is completely stalled due to some error
738 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
739 * In the other normal situations, it acts more gently by throttling the tasks
740 * more (rather than completely block them) when the wb dirty pages go high.
741 *
742 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
743 * - starving fast devices
744 * - piling up dirty pages (that will take long time to sync) on slow devices
745 *
746 * The wb's share of dirty limit will be adapting to its throughput and
747 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
748 *
749 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
750 * dirty balancing includes all PG_dirty and PG_writeback pages.
751 */
752static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
753{
754 struct wb_domain *dom = dtc_dom(dtc);
755 unsigned long thresh = dtc->thresh;
756 u64 wb_thresh;
757 unsigned long numerator, denominator;
758 unsigned long wb_min_ratio, wb_max_ratio;
759
760 /*
761 * Calculate this BDI's share of the thresh ratio.
762 */
763 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
764 &numerator, &denominator);
765
766 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
767 wb_thresh *= numerator;
768 wb_thresh = div64_ul(wb_thresh, denominator);
769
770 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
771
772 wb_thresh += (thresh * wb_min_ratio) / 100;
773 if (wb_thresh > (thresh * wb_max_ratio) / 100)
774 wb_thresh = thresh * wb_max_ratio / 100;
775
776 return wb_thresh;
777}
778
779unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
780{
781 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
782 .thresh = thresh };
783 return __wb_calc_thresh(&gdtc);
784}
785
786/*
787 * setpoint - dirty 3
788 * f(dirty) := 1.0 + (----------------)
789 * limit - setpoint
790 *
791 * it's a 3rd order polynomial that subjects to
792 *
793 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
794 * (2) f(setpoint) = 1.0 => the balance point
795 * (3) f(limit) = 0 => the hard limit
796 * (4) df/dx <= 0 => negative feedback control
797 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
798 * => fast response on large errors; small oscillation near setpoint
799 */
800static long long pos_ratio_polynom(unsigned long setpoint,
801 unsigned long dirty,
802 unsigned long limit)
803{
804 long long pos_ratio;
805 long x;
806
807 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
808 (limit - setpoint) | 1);
809 pos_ratio = x;
810 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
811 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
812 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
813
814 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
815}
816
817/*
818 * Dirty position control.
819 *
820 * (o) global/bdi setpoints
821 *
822 * We want the dirty pages be balanced around the global/wb setpoints.
823 * When the number of dirty pages is higher/lower than the setpoint, the
824 * dirty position control ratio (and hence task dirty ratelimit) will be
825 * decreased/increased to bring the dirty pages back to the setpoint.
826 *
827 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
828 *
829 * if (dirty < setpoint) scale up pos_ratio
830 * if (dirty > setpoint) scale down pos_ratio
831 *
832 * if (wb_dirty < wb_setpoint) scale up pos_ratio
833 * if (wb_dirty > wb_setpoint) scale down pos_ratio
834 *
835 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
836 *
837 * (o) global control line
838 *
839 * ^ pos_ratio
840 * |
841 * | |<===== global dirty control scope ======>|
842 * 2.0 * * * * * * *
843 * | .*
844 * | . *
845 * | . *
846 * | . *
847 * | . *
848 * | . *
849 * 1.0 ................................*
850 * | . . *
851 * | . . *
852 * | . . *
853 * | . . *
854 * | . . *
855 * 0 +------------.------------------.----------------------*------------->
856 * freerun^ setpoint^ limit^ dirty pages
857 *
858 * (o) wb control line
859 *
860 * ^ pos_ratio
861 * |
862 * | *
863 * | *
864 * | *
865 * | *
866 * | * |<=========== span ============>|
867 * 1.0 .......................*
868 * | . *
869 * | . *
870 * | . *
871 * | . *
872 * | . *
873 * | . *
874 * | . *
875 * | . *
876 * | . *
877 * | . *
878 * | . *
879 * 1/4 ...............................................* * * * * * * * * * * *
880 * | . .
881 * | . .
882 * | . .
883 * 0 +----------------------.-------------------------------.------------->
884 * wb_setpoint^ x_intercept^
885 *
886 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
887 * be smoothly throttled down to normal if it starts high in situations like
888 * - start writing to a slow SD card and a fast disk at the same time. The SD
889 * card's wb_dirty may rush to many times higher than wb_setpoint.
890 * - the wb dirty thresh drops quickly due to change of JBOD workload
891 */
892static void wb_position_ratio(struct dirty_throttle_control *dtc)
893{
894 struct bdi_writeback *wb = dtc->wb;
895 unsigned long write_bw = wb->avg_write_bandwidth;
896 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
897 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
898 unsigned long wb_thresh = dtc->wb_thresh;
899 unsigned long x_intercept;
900 unsigned long setpoint; /* dirty pages' target balance point */
901 unsigned long wb_setpoint;
902 unsigned long span;
903 long long pos_ratio; /* for scaling up/down the rate limit */
904 long x;
905
906 dtc->pos_ratio = 0;
907
908 if (unlikely(dtc->dirty >= limit))
909 return;
910
911 /*
912 * global setpoint
913 *
914 * See comment for pos_ratio_polynom().
915 */
916 setpoint = (freerun + limit) / 2;
917 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
918
919 /*
920 * The strictlimit feature is a tool preventing mistrusted filesystems
921 * from growing a large number of dirty pages before throttling. For
922 * such filesystems balance_dirty_pages always checks wb counters
923 * against wb limits. Even if global "nr_dirty" is under "freerun".
924 * This is especially important for fuse which sets bdi->max_ratio to
925 * 1% by default. Without strictlimit feature, fuse writeback may
926 * consume arbitrary amount of RAM because it is accounted in
927 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
928 *
929 * Here, in wb_position_ratio(), we calculate pos_ratio based on
930 * two values: wb_dirty and wb_thresh. Let's consider an example:
931 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
932 * limits are set by default to 10% and 20% (background and throttle).
933 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
934 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
935 * about ~6K pages (as the average of background and throttle wb
936 * limits). The 3rd order polynomial will provide positive feedback if
937 * wb_dirty is under wb_setpoint and vice versa.
938 *
939 * Note, that we cannot use global counters in these calculations
940 * because we want to throttle process writing to a strictlimit wb
941 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
942 * in the example above).
943 */
944 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
945 long long wb_pos_ratio;
946
947 if (dtc->wb_dirty < 8) {
948 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
949 2 << RATELIMIT_CALC_SHIFT);
950 return;
951 }
952
953 if (dtc->wb_dirty >= wb_thresh)
954 return;
955
956 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
957 dtc->wb_bg_thresh);
958
959 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
960 return;
961
962 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
963 wb_thresh);
964
965 /*
966 * Typically, for strictlimit case, wb_setpoint << setpoint
967 * and pos_ratio >> wb_pos_ratio. In the other words global
968 * state ("dirty") is not limiting factor and we have to
969 * make decision based on wb counters. But there is an
970 * important case when global pos_ratio should get precedence:
971 * global limits are exceeded (e.g. due to activities on other
972 * wb's) while given strictlimit wb is below limit.
973 *
974 * "pos_ratio * wb_pos_ratio" would work for the case above,
975 * but it would look too non-natural for the case of all
976 * activity in the system coming from a single strictlimit wb
977 * with bdi->max_ratio == 100%.
978 *
979 * Note that min() below somewhat changes the dynamics of the
980 * control system. Normally, pos_ratio value can be well over 3
981 * (when globally we are at freerun and wb is well below wb
982 * setpoint). Now the maximum pos_ratio in the same situation
983 * is 2. We might want to tweak this if we observe the control
984 * system is too slow to adapt.
985 */
986 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
987 return;
988 }
989
990 /*
991 * We have computed basic pos_ratio above based on global situation. If
992 * the wb is over/under its share of dirty pages, we want to scale
993 * pos_ratio further down/up. That is done by the following mechanism.
994 */
995
996 /*
997 * wb setpoint
998 *
999 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1000 *
1001 * x_intercept - wb_dirty
1002 * := --------------------------
1003 * x_intercept - wb_setpoint
1004 *
1005 * The main wb control line is a linear function that subjects to
1006 *
1007 * (1) f(wb_setpoint) = 1.0
1008 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1009 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1010 *
1011 * For single wb case, the dirty pages are observed to fluctuate
1012 * regularly within range
1013 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1014 * for various filesystems, where (2) can yield in a reasonable 12.5%
1015 * fluctuation range for pos_ratio.
1016 *
1017 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1018 * own size, so move the slope over accordingly and choose a slope that
1019 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1020 */
1021 if (unlikely(wb_thresh > dtc->thresh))
1022 wb_thresh = dtc->thresh;
1023 /*
1024 * It's very possible that wb_thresh is close to 0 not because the
1025 * device is slow, but that it has remained inactive for long time.
1026 * Honour such devices a reasonable good (hopefully IO efficient)
1027 * threshold, so that the occasional writes won't be blocked and active
1028 * writes can rampup the threshold quickly.
1029 */
1030 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1031 /*
1032 * scale global setpoint to wb's:
1033 * wb_setpoint = setpoint * wb_thresh / thresh
1034 */
1035 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1036 wb_setpoint = setpoint * (u64)x >> 16;
1037 /*
1038 * Use span=(8*write_bw) in single wb case as indicated by
1039 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1040 *
1041 * wb_thresh thresh - wb_thresh
1042 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1043 * thresh thresh
1044 */
1045 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1046 x_intercept = wb_setpoint + span;
1047
1048 if (dtc->wb_dirty < x_intercept - span / 4) {
1049 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1050 (x_intercept - wb_setpoint) | 1);
1051 } else
1052 pos_ratio /= 4;
1053
1054 /*
1055 * wb reserve area, safeguard against dirty pool underrun and disk idle
1056 * It may push the desired control point of global dirty pages higher
1057 * than setpoint.
1058 */
1059 x_intercept = wb_thresh / 2;
1060 if (dtc->wb_dirty < x_intercept) {
1061 if (dtc->wb_dirty > x_intercept / 8)
1062 pos_ratio = div_u64(pos_ratio * x_intercept,
1063 dtc->wb_dirty);
1064 else
1065 pos_ratio *= 8;
1066 }
1067
1068 dtc->pos_ratio = pos_ratio;
1069}
1070
1071static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1072 unsigned long elapsed,
1073 unsigned long written)
1074{
1075 const unsigned long period = roundup_pow_of_two(3 * HZ);
1076 unsigned long avg = wb->avg_write_bandwidth;
1077 unsigned long old = wb->write_bandwidth;
1078 u64 bw;
1079
1080 /*
1081 * bw = written * HZ / elapsed
1082 *
1083 * bw * elapsed + write_bandwidth * (period - elapsed)
1084 * write_bandwidth = ---------------------------------------------------
1085 * period
1086 *
1087 * @written may have decreased due to account_page_redirty().
1088 * Avoid underflowing @bw calculation.
1089 */
1090 bw = written - min(written, wb->written_stamp);
1091 bw *= HZ;
1092 if (unlikely(elapsed > period)) {
1093 bw = div64_ul(bw, elapsed);
1094 avg = bw;
1095 goto out;
1096 }
1097 bw += (u64)wb->write_bandwidth * (period - elapsed);
1098 bw >>= ilog2(period);
1099
1100 /*
1101 * one more level of smoothing, for filtering out sudden spikes
1102 */
1103 if (avg > old && old >= (unsigned long)bw)
1104 avg -= (avg - old) >> 3;
1105
1106 if (avg < old && old <= (unsigned long)bw)
1107 avg += (old - avg) >> 3;
1108
1109out:
1110 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1111 avg = max(avg, 1LU);
1112 if (wb_has_dirty_io(wb)) {
1113 long delta = avg - wb->avg_write_bandwidth;
1114 WARN_ON_ONCE(atomic_long_add_return(delta,
1115 &wb->bdi->tot_write_bandwidth) <= 0);
1116 }
1117 wb->write_bandwidth = bw;
1118 wb->avg_write_bandwidth = avg;
1119}
1120
1121static void update_dirty_limit(struct dirty_throttle_control *dtc)
1122{
1123 struct wb_domain *dom = dtc_dom(dtc);
1124 unsigned long thresh = dtc->thresh;
1125 unsigned long limit = dom->dirty_limit;
1126
1127 /*
1128 * Follow up in one step.
1129 */
1130 if (limit < thresh) {
1131 limit = thresh;
1132 goto update;
1133 }
1134
1135 /*
1136 * Follow down slowly. Use the higher one as the target, because thresh
1137 * may drop below dirty. This is exactly the reason to introduce
1138 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1139 */
1140 thresh = max(thresh, dtc->dirty);
1141 if (limit > thresh) {
1142 limit -= (limit - thresh) >> 5;
1143 goto update;
1144 }
1145 return;
1146update:
1147 dom->dirty_limit = limit;
1148}
1149
1150static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1151 unsigned long now)
1152{
1153 struct wb_domain *dom = dtc_dom(dtc);
1154
1155 /*
1156 * check locklessly first to optimize away locking for the most time
1157 */
1158 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1159 return;
1160
1161 spin_lock(&dom->lock);
1162 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1163 update_dirty_limit(dtc);
1164 dom->dirty_limit_tstamp = now;
1165 }
1166 spin_unlock(&dom->lock);
1167}
1168
1169/*
1170 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1171 *
1172 * Normal wb tasks will be curbed at or below it in long term.
1173 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1174 */
1175static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1176 unsigned long dirtied,
1177 unsigned long elapsed)
1178{
1179 struct bdi_writeback *wb = dtc->wb;
1180 unsigned long dirty = dtc->dirty;
1181 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1182 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1183 unsigned long setpoint = (freerun + limit) / 2;
1184 unsigned long write_bw = wb->avg_write_bandwidth;
1185 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1186 unsigned long dirty_rate;
1187 unsigned long task_ratelimit;
1188 unsigned long balanced_dirty_ratelimit;
1189 unsigned long step;
1190 unsigned long x;
1191 unsigned long shift;
1192
1193 /*
1194 * The dirty rate will match the writeout rate in long term, except
1195 * when dirty pages are truncated by userspace or re-dirtied by FS.
1196 */
1197 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1198
1199 /*
1200 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1201 */
1202 task_ratelimit = (u64)dirty_ratelimit *
1203 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1204 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1205
1206 /*
1207 * A linear estimation of the "balanced" throttle rate. The theory is,
1208 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1209 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1210 * formula will yield the balanced rate limit (write_bw / N).
1211 *
1212 * Note that the expanded form is not a pure rate feedback:
1213 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1214 * but also takes pos_ratio into account:
1215 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1216 *
1217 * (1) is not realistic because pos_ratio also takes part in balancing
1218 * the dirty rate. Consider the state
1219 * pos_ratio = 0.5 (3)
1220 * rate = 2 * (write_bw / N) (4)
1221 * If (1) is used, it will stuck in that state! Because each dd will
1222 * be throttled at
1223 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1224 * yielding
1225 * dirty_rate = N * task_ratelimit = write_bw (6)
1226 * put (6) into (1) we get
1227 * rate_(i+1) = rate_(i) (7)
1228 *
1229 * So we end up using (2) to always keep
1230 * rate_(i+1) ~= (write_bw / N) (8)
1231 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1232 * pos_ratio is able to drive itself to 1.0, which is not only where
1233 * the dirty count meet the setpoint, but also where the slope of
1234 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1235 */
1236 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1237 dirty_rate | 1);
1238 /*
1239 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1240 */
1241 if (unlikely(balanced_dirty_ratelimit > write_bw))
1242 balanced_dirty_ratelimit = write_bw;
1243
1244 /*
1245 * We could safely do this and return immediately:
1246 *
1247 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1248 *
1249 * However to get a more stable dirty_ratelimit, the below elaborated
1250 * code makes use of task_ratelimit to filter out singular points and
1251 * limit the step size.
1252 *
1253 * The below code essentially only uses the relative value of
1254 *
1255 * task_ratelimit - dirty_ratelimit
1256 * = (pos_ratio - 1) * dirty_ratelimit
1257 *
1258 * which reflects the direction and size of dirty position error.
1259 */
1260
1261 /*
1262 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1263 * task_ratelimit is on the same side of dirty_ratelimit, too.
1264 * For example, when
1265 * - dirty_ratelimit > balanced_dirty_ratelimit
1266 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1267 * lowering dirty_ratelimit will help meet both the position and rate
1268 * control targets. Otherwise, don't update dirty_ratelimit if it will
1269 * only help meet the rate target. After all, what the users ultimately
1270 * feel and care are stable dirty rate and small position error.
1271 *
1272 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1273 * and filter out the singular points of balanced_dirty_ratelimit. Which
1274 * keeps jumping around randomly and can even leap far away at times
1275 * due to the small 200ms estimation period of dirty_rate (we want to
1276 * keep that period small to reduce time lags).
1277 */
1278 step = 0;
1279
1280 /*
1281 * For strictlimit case, calculations above were based on wb counters
1282 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1283 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1284 * Hence, to calculate "step" properly, we have to use wb_dirty as
1285 * "dirty" and wb_setpoint as "setpoint".
1286 *
1287 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1288 * it's possible that wb_thresh is close to zero due to inactivity
1289 * of backing device.
1290 */
1291 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1292 dirty = dtc->wb_dirty;
1293 if (dtc->wb_dirty < 8)
1294 setpoint = dtc->wb_dirty + 1;
1295 else
1296 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1297 }
1298
1299 if (dirty < setpoint) {
1300 x = min3(wb->balanced_dirty_ratelimit,
1301 balanced_dirty_ratelimit, task_ratelimit);
1302 if (dirty_ratelimit < x)
1303 step = x - dirty_ratelimit;
1304 } else {
1305 x = max3(wb->balanced_dirty_ratelimit,
1306 balanced_dirty_ratelimit, task_ratelimit);
1307 if (dirty_ratelimit > x)
1308 step = dirty_ratelimit - x;
1309 }
1310
1311 /*
1312 * Don't pursue 100% rate matching. It's impossible since the balanced
1313 * rate itself is constantly fluctuating. So decrease the track speed
1314 * when it gets close to the target. Helps eliminate pointless tremors.
1315 */
1316 shift = dirty_ratelimit / (2 * step + 1);
1317 if (shift < BITS_PER_LONG)
1318 step = DIV_ROUND_UP(step >> shift, 8);
1319 else
1320 step = 0;
1321
1322 if (dirty_ratelimit < balanced_dirty_ratelimit)
1323 dirty_ratelimit += step;
1324 else
1325 dirty_ratelimit -= step;
1326
1327 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1328 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1329
1330 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1331}
1332
1333static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1334 struct dirty_throttle_control *mdtc,
1335 unsigned long start_time,
1336 bool update_ratelimit)
1337{
1338 struct bdi_writeback *wb = gdtc->wb;
1339 unsigned long now = jiffies;
1340 unsigned long elapsed = now - wb->bw_time_stamp;
1341 unsigned long dirtied;
1342 unsigned long written;
1343
1344 lockdep_assert_held(&wb->list_lock);
1345
1346 /*
1347 * rate-limit, only update once every 200ms.
1348 */
1349 if (elapsed < BANDWIDTH_INTERVAL)
1350 return;
1351
1352 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1353 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1354
1355 /*
1356 * Skip quiet periods when disk bandwidth is under-utilized.
1357 * (at least 1s idle time between two flusher runs)
1358 */
1359 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1360 goto snapshot;
1361
1362 if (update_ratelimit) {
1363 domain_update_bandwidth(gdtc, now);
1364 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1365
1366 /*
1367 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1368 * compiler has no way to figure that out. Help it.
1369 */
1370 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1371 domain_update_bandwidth(mdtc, now);
1372 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1373 }
1374 }
1375 wb_update_write_bandwidth(wb, elapsed, written);
1376
1377snapshot:
1378 wb->dirtied_stamp = dirtied;
1379 wb->written_stamp = written;
1380 wb->bw_time_stamp = now;
1381}
1382
1383void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1384{
1385 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1386
1387 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1388}
1389
1390/*
1391 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1392 * will look to see if it needs to start dirty throttling.
1393 *
1394 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1395 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1396 * (the number of pages we may dirty without exceeding the dirty limits).
1397 */
1398static unsigned long dirty_poll_interval(unsigned long dirty,
1399 unsigned long thresh)
1400{
1401 if (thresh > dirty)
1402 return 1UL << (ilog2(thresh - dirty) >> 1);
1403
1404 return 1;
1405}
1406
1407static unsigned long wb_max_pause(struct bdi_writeback *wb,
1408 unsigned long wb_dirty)
1409{
1410 unsigned long bw = wb->avg_write_bandwidth;
1411 unsigned long t;
1412
1413 /*
1414 * Limit pause time for small memory systems. If sleeping for too long
1415 * time, a small pool of dirty/writeback pages may go empty and disk go
1416 * idle.
1417 *
1418 * 8 serves as the safety ratio.
1419 */
1420 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1421 t++;
1422
1423 return min_t(unsigned long, t, MAX_PAUSE);
1424}
1425
1426static long wb_min_pause(struct bdi_writeback *wb,
1427 long max_pause,
1428 unsigned long task_ratelimit,
1429 unsigned long dirty_ratelimit,
1430 int *nr_dirtied_pause)
1431{
1432 long hi = ilog2(wb->avg_write_bandwidth);
1433 long lo = ilog2(wb->dirty_ratelimit);
1434 long t; /* target pause */
1435 long pause; /* estimated next pause */
1436 int pages; /* target nr_dirtied_pause */
1437
1438 /* target for 10ms pause on 1-dd case */
1439 t = max(1, HZ / 100);
1440
1441 /*
1442 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1443 * overheads.
1444 *
1445 * (N * 10ms) on 2^N concurrent tasks.
1446 */
1447 if (hi > lo)
1448 t += (hi - lo) * (10 * HZ) / 1024;
1449
1450 /*
1451 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1452 * on the much more stable dirty_ratelimit. However the next pause time
1453 * will be computed based on task_ratelimit and the two rate limits may
1454 * depart considerably at some time. Especially if task_ratelimit goes
1455 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1456 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1457 * result task_ratelimit won't be executed faithfully, which could
1458 * eventually bring down dirty_ratelimit.
1459 *
1460 * We apply two rules to fix it up:
1461 * 1) try to estimate the next pause time and if necessary, use a lower
1462 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1463 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1464 * 2) limit the target pause time to max_pause/2, so that the normal
1465 * small fluctuations of task_ratelimit won't trigger rule (1) and
1466 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1467 */
1468 t = min(t, 1 + max_pause / 2);
1469 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1470
1471 /*
1472 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1473 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1474 * When the 16 consecutive reads are often interrupted by some dirty
1475 * throttling pause during the async writes, cfq will go into idles
1476 * (deadline is fine). So push nr_dirtied_pause as high as possible
1477 * until reaches DIRTY_POLL_THRESH=32 pages.
1478 */
1479 if (pages < DIRTY_POLL_THRESH) {
1480 t = max_pause;
1481 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482 if (pages > DIRTY_POLL_THRESH) {
1483 pages = DIRTY_POLL_THRESH;
1484 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1485 }
1486 }
1487
1488 pause = HZ * pages / (task_ratelimit + 1);
1489 if (pause > max_pause) {
1490 t = max_pause;
1491 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1492 }
1493
1494 *nr_dirtied_pause = pages;
1495 /*
1496 * The minimal pause time will normally be half the target pause time.
1497 */
1498 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1499}
1500
1501static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1502{
1503 struct bdi_writeback *wb = dtc->wb;
1504 unsigned long wb_reclaimable;
1505
1506 /*
1507 * wb_thresh is not treated as some limiting factor as
1508 * dirty_thresh, due to reasons
1509 * - in JBOD setup, wb_thresh can fluctuate a lot
1510 * - in a system with HDD and USB key, the USB key may somehow
1511 * go into state (wb_dirty >> wb_thresh) either because
1512 * wb_dirty starts high, or because wb_thresh drops low.
1513 * In this case we don't want to hard throttle the USB key
1514 * dirtiers for 100 seconds until wb_dirty drops under
1515 * wb_thresh. Instead the auxiliary wb control line in
1516 * wb_position_ratio() will let the dirtier task progress
1517 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1518 */
1519 dtc->wb_thresh = __wb_calc_thresh(dtc);
1520 dtc->wb_bg_thresh = dtc->thresh ?
1521 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1522
1523 /*
1524 * In order to avoid the stacked BDI deadlock we need
1525 * to ensure we accurately count the 'dirty' pages when
1526 * the threshold is low.
1527 *
1528 * Otherwise it would be possible to get thresh+n pages
1529 * reported dirty, even though there are thresh-m pages
1530 * actually dirty; with m+n sitting in the percpu
1531 * deltas.
1532 */
1533 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1534 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1535 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1536 } else {
1537 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1538 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1539 }
1540}
1541
1542/*
1543 * balance_dirty_pages() must be called by processes which are generating dirty
1544 * data. It looks at the number of dirty pages in the machine and will force
1545 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1546 * If we're over `background_thresh' then the writeback threads are woken to
1547 * perform some writeout.
1548 */
1549static void balance_dirty_pages(struct bdi_writeback *wb,
1550 unsigned long pages_dirtied)
1551{
1552 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1553 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1554 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1555 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1556 &mdtc_stor : NULL;
1557 struct dirty_throttle_control *sdtc;
1558 unsigned long nr_reclaimable; /* = file_dirty */
1559 long period;
1560 long pause;
1561 long max_pause;
1562 long min_pause;
1563 int nr_dirtied_pause;
1564 bool dirty_exceeded = false;
1565 unsigned long task_ratelimit;
1566 unsigned long dirty_ratelimit;
1567 struct backing_dev_info *bdi = wb->bdi;
1568 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1569 unsigned long start_time = jiffies;
1570
1571 for (;;) {
1572 unsigned long now = jiffies;
1573 unsigned long dirty, thresh, bg_thresh;
1574 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1575 unsigned long m_thresh = 0;
1576 unsigned long m_bg_thresh = 0;
1577
1578 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1579 gdtc->avail = global_dirtyable_memory();
1580 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1581
1582 domain_dirty_limits(gdtc);
1583
1584 if (unlikely(strictlimit)) {
1585 wb_dirty_limits(gdtc);
1586
1587 dirty = gdtc->wb_dirty;
1588 thresh = gdtc->wb_thresh;
1589 bg_thresh = gdtc->wb_bg_thresh;
1590 } else {
1591 dirty = gdtc->dirty;
1592 thresh = gdtc->thresh;
1593 bg_thresh = gdtc->bg_thresh;
1594 }
1595
1596 if (mdtc) {
1597 unsigned long filepages, headroom, writeback;
1598
1599 /*
1600 * If @wb belongs to !root memcg, repeat the same
1601 * basic calculations for the memcg domain.
1602 */
1603 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1604 &mdtc->dirty, &writeback);
1605 mdtc->dirty += writeback;
1606 mdtc_calc_avail(mdtc, filepages, headroom);
1607
1608 domain_dirty_limits(mdtc);
1609
1610 if (unlikely(strictlimit)) {
1611 wb_dirty_limits(mdtc);
1612 m_dirty = mdtc->wb_dirty;
1613 m_thresh = mdtc->wb_thresh;
1614 m_bg_thresh = mdtc->wb_bg_thresh;
1615 } else {
1616 m_dirty = mdtc->dirty;
1617 m_thresh = mdtc->thresh;
1618 m_bg_thresh = mdtc->bg_thresh;
1619 }
1620 }
1621
1622 /*
1623 * Throttle it only when the background writeback cannot
1624 * catch-up. This avoids (excessively) small writeouts
1625 * when the wb limits are ramping up in case of !strictlimit.
1626 *
1627 * In strictlimit case make decision based on the wb counters
1628 * and limits. Small writeouts when the wb limits are ramping
1629 * up are the price we consciously pay for strictlimit-ing.
1630 *
1631 * If memcg domain is in effect, @dirty should be under
1632 * both global and memcg freerun ceilings.
1633 */
1634 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1635 (!mdtc ||
1636 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1637 unsigned long intv;
1638 unsigned long m_intv;
1639
1640free_running:
1641 intv = dirty_poll_interval(dirty, thresh);
1642 m_intv = ULONG_MAX;
1643
1644 current->dirty_paused_when = now;
1645 current->nr_dirtied = 0;
1646 if (mdtc)
1647 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1648 current->nr_dirtied_pause = min(intv, m_intv);
1649 break;
1650 }
1651
1652 if (unlikely(!writeback_in_progress(wb)))
1653 wb_start_background_writeback(wb);
1654
1655 mem_cgroup_flush_foreign(wb);
1656
1657 /*
1658 * Calculate global domain's pos_ratio and select the
1659 * global dtc by default.
1660 */
1661 if (!strictlimit) {
1662 wb_dirty_limits(gdtc);
1663
1664 if ((current->flags & PF_LOCAL_THROTTLE) &&
1665 gdtc->wb_dirty <
1666 dirty_freerun_ceiling(gdtc->wb_thresh,
1667 gdtc->wb_bg_thresh))
1668 /*
1669 * LOCAL_THROTTLE tasks must not be throttled
1670 * when below the per-wb freerun ceiling.
1671 */
1672 goto free_running;
1673 }
1674
1675 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1676 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1677
1678 wb_position_ratio(gdtc);
1679 sdtc = gdtc;
1680
1681 if (mdtc) {
1682 /*
1683 * If memcg domain is in effect, calculate its
1684 * pos_ratio. @wb should satisfy constraints from
1685 * both global and memcg domains. Choose the one
1686 * w/ lower pos_ratio.
1687 */
1688 if (!strictlimit) {
1689 wb_dirty_limits(mdtc);
1690
1691 if ((current->flags & PF_LOCAL_THROTTLE) &&
1692 mdtc->wb_dirty <
1693 dirty_freerun_ceiling(mdtc->wb_thresh,
1694 mdtc->wb_bg_thresh))
1695 /*
1696 * LOCAL_THROTTLE tasks must not be
1697 * throttled when below the per-wb
1698 * freerun ceiling.
1699 */
1700 goto free_running;
1701 }
1702 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1703 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1704
1705 wb_position_ratio(mdtc);
1706 if (mdtc->pos_ratio < gdtc->pos_ratio)
1707 sdtc = mdtc;
1708 }
1709
1710 if (dirty_exceeded && !wb->dirty_exceeded)
1711 wb->dirty_exceeded = 1;
1712
1713 if (time_is_before_jiffies(wb->bw_time_stamp +
1714 BANDWIDTH_INTERVAL)) {
1715 spin_lock(&wb->list_lock);
1716 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1717 spin_unlock(&wb->list_lock);
1718 }
1719
1720 /* throttle according to the chosen dtc */
1721 dirty_ratelimit = wb->dirty_ratelimit;
1722 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1723 RATELIMIT_CALC_SHIFT;
1724 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1725 min_pause = wb_min_pause(wb, max_pause,
1726 task_ratelimit, dirty_ratelimit,
1727 &nr_dirtied_pause);
1728
1729 if (unlikely(task_ratelimit == 0)) {
1730 period = max_pause;
1731 pause = max_pause;
1732 goto pause;
1733 }
1734 period = HZ * pages_dirtied / task_ratelimit;
1735 pause = period;
1736 if (current->dirty_paused_when)
1737 pause -= now - current->dirty_paused_when;
1738 /*
1739 * For less than 1s think time (ext3/4 may block the dirtier
1740 * for up to 800ms from time to time on 1-HDD; so does xfs,
1741 * however at much less frequency), try to compensate it in
1742 * future periods by updating the virtual time; otherwise just
1743 * do a reset, as it may be a light dirtier.
1744 */
1745 if (pause < min_pause) {
1746 trace_balance_dirty_pages(wb,
1747 sdtc->thresh,
1748 sdtc->bg_thresh,
1749 sdtc->dirty,
1750 sdtc->wb_thresh,
1751 sdtc->wb_dirty,
1752 dirty_ratelimit,
1753 task_ratelimit,
1754 pages_dirtied,
1755 period,
1756 min(pause, 0L),
1757 start_time);
1758 if (pause < -HZ) {
1759 current->dirty_paused_when = now;
1760 current->nr_dirtied = 0;
1761 } else if (period) {
1762 current->dirty_paused_when += period;
1763 current->nr_dirtied = 0;
1764 } else if (current->nr_dirtied_pause <= pages_dirtied)
1765 current->nr_dirtied_pause += pages_dirtied;
1766 break;
1767 }
1768 if (unlikely(pause > max_pause)) {
1769 /* for occasional dropped task_ratelimit */
1770 now += min(pause - max_pause, max_pause);
1771 pause = max_pause;
1772 }
1773
1774pause:
1775 trace_balance_dirty_pages(wb,
1776 sdtc->thresh,
1777 sdtc->bg_thresh,
1778 sdtc->dirty,
1779 sdtc->wb_thresh,
1780 sdtc->wb_dirty,
1781 dirty_ratelimit,
1782 task_ratelimit,
1783 pages_dirtied,
1784 period,
1785 pause,
1786 start_time);
1787 __set_current_state(TASK_KILLABLE);
1788 wb->dirty_sleep = now;
1789 io_schedule_timeout(pause);
1790
1791 current->dirty_paused_when = now + pause;
1792 current->nr_dirtied = 0;
1793 current->nr_dirtied_pause = nr_dirtied_pause;
1794
1795 /*
1796 * This is typically equal to (dirty < thresh) and can also
1797 * keep "1000+ dd on a slow USB stick" under control.
1798 */
1799 if (task_ratelimit)
1800 break;
1801
1802 /*
1803 * In the case of an unresponsive NFS server and the NFS dirty
1804 * pages exceeds dirty_thresh, give the other good wb's a pipe
1805 * to go through, so that tasks on them still remain responsive.
1806 *
1807 * In theory 1 page is enough to keep the consumer-producer
1808 * pipe going: the flusher cleans 1 page => the task dirties 1
1809 * more page. However wb_dirty has accounting errors. So use
1810 * the larger and more IO friendly wb_stat_error.
1811 */
1812 if (sdtc->wb_dirty <= wb_stat_error())
1813 break;
1814
1815 if (fatal_signal_pending(current))
1816 break;
1817 }
1818
1819 if (!dirty_exceeded && wb->dirty_exceeded)
1820 wb->dirty_exceeded = 0;
1821
1822 if (writeback_in_progress(wb))
1823 return;
1824
1825 /*
1826 * In laptop mode, we wait until hitting the higher threshold before
1827 * starting background writeout, and then write out all the way down
1828 * to the lower threshold. So slow writers cause minimal disk activity.
1829 *
1830 * In normal mode, we start background writeout at the lower
1831 * background_thresh, to keep the amount of dirty memory low.
1832 */
1833 if (laptop_mode)
1834 return;
1835
1836 if (nr_reclaimable > gdtc->bg_thresh)
1837 wb_start_background_writeback(wb);
1838}
1839
1840static DEFINE_PER_CPU(int, bdp_ratelimits);
1841
1842/*
1843 * Normal tasks are throttled by
1844 * loop {
1845 * dirty tsk->nr_dirtied_pause pages;
1846 * take a snap in balance_dirty_pages();
1847 * }
1848 * However there is a worst case. If every task exit immediately when dirtied
1849 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1850 * called to throttle the page dirties. The solution is to save the not yet
1851 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1852 * randomly into the running tasks. This works well for the above worst case,
1853 * as the new task will pick up and accumulate the old task's leaked dirty
1854 * count and eventually get throttled.
1855 */
1856DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1857
1858/**
1859 * balance_dirty_pages_ratelimited - balance dirty memory state
1860 * @mapping: address_space which was dirtied
1861 *
1862 * Processes which are dirtying memory should call in here once for each page
1863 * which was newly dirtied. The function will periodically check the system's
1864 * dirty state and will initiate writeback if needed.
1865 *
1866 * Once we're over the dirty memory limit we decrease the ratelimiting
1867 * by a lot, to prevent individual processes from overshooting the limit
1868 * by (ratelimit_pages) each.
1869 */
1870void balance_dirty_pages_ratelimited(struct address_space *mapping)
1871{
1872 struct inode *inode = mapping->host;
1873 struct backing_dev_info *bdi = inode_to_bdi(inode);
1874 struct bdi_writeback *wb = NULL;
1875 int ratelimit;
1876 int *p;
1877
1878 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1879 return;
1880
1881 if (inode_cgwb_enabled(inode))
1882 wb = wb_get_create_current(bdi, GFP_KERNEL);
1883 if (!wb)
1884 wb = &bdi->wb;
1885
1886 ratelimit = current->nr_dirtied_pause;
1887 if (wb->dirty_exceeded)
1888 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1889
1890 preempt_disable();
1891 /*
1892 * This prevents one CPU to accumulate too many dirtied pages without
1893 * calling into balance_dirty_pages(), which can happen when there are
1894 * 1000+ tasks, all of them start dirtying pages at exactly the same
1895 * time, hence all honoured too large initial task->nr_dirtied_pause.
1896 */
1897 p = this_cpu_ptr(&bdp_ratelimits);
1898 if (unlikely(current->nr_dirtied >= ratelimit))
1899 *p = 0;
1900 else if (unlikely(*p >= ratelimit_pages)) {
1901 *p = 0;
1902 ratelimit = 0;
1903 }
1904 /*
1905 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1906 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1907 * the dirty throttling and livelock other long-run dirtiers.
1908 */
1909 p = this_cpu_ptr(&dirty_throttle_leaks);
1910 if (*p > 0 && current->nr_dirtied < ratelimit) {
1911 unsigned long nr_pages_dirtied;
1912 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1913 *p -= nr_pages_dirtied;
1914 current->nr_dirtied += nr_pages_dirtied;
1915 }
1916 preempt_enable();
1917
1918 if (unlikely(current->nr_dirtied >= ratelimit))
1919 balance_dirty_pages(wb, current->nr_dirtied);
1920
1921 wb_put(wb);
1922}
1923EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1924
1925/**
1926 * wb_over_bg_thresh - does @wb need to be written back?
1927 * @wb: bdi_writeback of interest
1928 *
1929 * Determines whether background writeback should keep writing @wb or it's
1930 * clean enough.
1931 *
1932 * Return: %true if writeback should continue.
1933 */
1934bool wb_over_bg_thresh(struct bdi_writeback *wb)
1935{
1936 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1937 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1938 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1939 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1940 &mdtc_stor : NULL;
1941 unsigned long reclaimable;
1942 unsigned long thresh;
1943
1944 /*
1945 * Similar to balance_dirty_pages() but ignores pages being written
1946 * as we're trying to decide whether to put more under writeback.
1947 */
1948 gdtc->avail = global_dirtyable_memory();
1949 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1950 domain_dirty_limits(gdtc);
1951
1952 if (gdtc->dirty > gdtc->bg_thresh)
1953 return true;
1954
1955 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1956 if (thresh < 2 * wb_stat_error())
1957 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1958 else
1959 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1960
1961 if (reclaimable > thresh)
1962 return true;
1963
1964 if (mdtc) {
1965 unsigned long filepages, headroom, writeback;
1966
1967 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1968 &writeback);
1969 mdtc_calc_avail(mdtc, filepages, headroom);
1970 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1971
1972 if (mdtc->dirty > mdtc->bg_thresh)
1973 return true;
1974
1975 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1976 if (thresh < 2 * wb_stat_error())
1977 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1978 else
1979 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1980
1981 if (reclaimable > thresh)
1982 return true;
1983 }
1984
1985 return false;
1986}
1987
1988/*
1989 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1990 */
1991int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1992 void *buffer, size_t *length, loff_t *ppos)
1993{
1994 unsigned int old_interval = dirty_writeback_interval;
1995 int ret;
1996
1997 ret = proc_dointvec(table, write, buffer, length, ppos);
1998
1999 /*
2000 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2001 * and a different non-zero value will wakeup the writeback threads.
2002 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2003 * iterate over all bdis and wbs.
2004 * The reason we do this is to make the change take effect immediately.
2005 */
2006 if (!ret && write && dirty_writeback_interval &&
2007 dirty_writeback_interval != old_interval)
2008 wakeup_flusher_threads(WB_REASON_PERIODIC);
2009
2010 return ret;
2011}
2012
2013#ifdef CONFIG_BLOCK
2014void laptop_mode_timer_fn(struct timer_list *t)
2015{
2016 struct backing_dev_info *backing_dev_info =
2017 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2018
2019 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2020}
2021
2022/*
2023 * We've spun up the disk and we're in laptop mode: schedule writeback
2024 * of all dirty data a few seconds from now. If the flush is already scheduled
2025 * then push it back - the user is still using the disk.
2026 */
2027void laptop_io_completion(struct backing_dev_info *info)
2028{
2029 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2030}
2031
2032/*
2033 * We're in laptop mode and we've just synced. The sync's writes will have
2034 * caused another writeback to be scheduled by laptop_io_completion.
2035 * Nothing needs to be written back anymore, so we unschedule the writeback.
2036 */
2037void laptop_sync_completion(void)
2038{
2039 struct backing_dev_info *bdi;
2040
2041 rcu_read_lock();
2042
2043 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2044 del_timer(&bdi->laptop_mode_wb_timer);
2045
2046 rcu_read_unlock();
2047}
2048#endif
2049
2050/*
2051 * If ratelimit_pages is too high then we can get into dirty-data overload
2052 * if a large number of processes all perform writes at the same time.
2053 *
2054 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2055 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2056 * thresholds.
2057 */
2058
2059void writeback_set_ratelimit(void)
2060{
2061 struct wb_domain *dom = &global_wb_domain;
2062 unsigned long background_thresh;
2063 unsigned long dirty_thresh;
2064
2065 global_dirty_limits(&background_thresh, &dirty_thresh);
2066 dom->dirty_limit = dirty_thresh;
2067 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2068 if (ratelimit_pages < 16)
2069 ratelimit_pages = 16;
2070}
2071
2072static int page_writeback_cpu_online(unsigned int cpu)
2073{
2074 writeback_set_ratelimit();
2075 return 0;
2076}
2077
2078/*
2079 * Called early on to tune the page writeback dirty limits.
2080 *
2081 * We used to scale dirty pages according to how total memory
2082 * related to pages that could be allocated for buffers.
2083 *
2084 * However, that was when we used "dirty_ratio" to scale with
2085 * all memory, and we don't do that any more. "dirty_ratio"
2086 * is now applied to total non-HIGHPAGE memory, and as such we can't
2087 * get into the old insane situation any more where we had
2088 * large amounts of dirty pages compared to a small amount of
2089 * non-HIGHMEM memory.
2090 *
2091 * But we might still want to scale the dirty_ratio by how
2092 * much memory the box has..
2093 */
2094void __init page_writeback_init(void)
2095{
2096 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2097
2098 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2099 page_writeback_cpu_online, NULL);
2100 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2101 page_writeback_cpu_online);
2102}
2103
2104/**
2105 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2106 * @mapping: address space structure to write
2107 * @start: starting page index
2108 * @end: ending page index (inclusive)
2109 *
2110 * This function scans the page range from @start to @end (inclusive) and tags
2111 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2112 * that write_cache_pages (or whoever calls this function) will then use
2113 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2114 * used to avoid livelocking of writeback by a process steadily creating new
2115 * dirty pages in the file (thus it is important for this function to be quick
2116 * so that it can tag pages faster than a dirtying process can create them).
2117 */
2118void tag_pages_for_writeback(struct address_space *mapping,
2119 pgoff_t start, pgoff_t end)
2120{
2121 XA_STATE(xas, &mapping->i_pages, start);
2122 unsigned int tagged = 0;
2123 void *page;
2124
2125 xas_lock_irq(&xas);
2126 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2127 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2128 if (++tagged % XA_CHECK_SCHED)
2129 continue;
2130
2131 xas_pause(&xas);
2132 xas_unlock_irq(&xas);
2133 cond_resched();
2134 xas_lock_irq(&xas);
2135 }
2136 xas_unlock_irq(&xas);
2137}
2138EXPORT_SYMBOL(tag_pages_for_writeback);
2139
2140/**
2141 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2142 * @mapping: address space structure to write
2143 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2144 * @writepage: function called for each page
2145 * @data: data passed to writepage function
2146 *
2147 * If a page is already under I/O, write_cache_pages() skips it, even
2148 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2149 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2150 * and msync() need to guarantee that all the data which was dirty at the time
2151 * the call was made get new I/O started against them. If wbc->sync_mode is
2152 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2153 * existing IO to complete.
2154 *
2155 * To avoid livelocks (when other process dirties new pages), we first tag
2156 * pages which should be written back with TOWRITE tag and only then start
2157 * writing them. For data-integrity sync we have to be careful so that we do
2158 * not miss some pages (e.g., because some other process has cleared TOWRITE
2159 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2160 * by the process clearing the DIRTY tag (and submitting the page for IO).
2161 *
2162 * To avoid deadlocks between range_cyclic writeback and callers that hold
2163 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2164 * we do not loop back to the start of the file. Doing so causes a page
2165 * lock/page writeback access order inversion - we should only ever lock
2166 * multiple pages in ascending page->index order, and looping back to the start
2167 * of the file violates that rule and causes deadlocks.
2168 *
2169 * Return: %0 on success, negative error code otherwise
2170 */
2171int write_cache_pages(struct address_space *mapping,
2172 struct writeback_control *wbc, writepage_t writepage,
2173 void *data)
2174{
2175 int ret = 0;
2176 int done = 0;
2177 int error;
2178 struct pagevec pvec;
2179 int nr_pages;
2180 pgoff_t index;
2181 pgoff_t end; /* Inclusive */
2182 pgoff_t done_index;
2183 int range_whole = 0;
2184 xa_mark_t tag;
2185
2186 pagevec_init(&pvec);
2187 if (wbc->range_cyclic) {
2188 index = mapping->writeback_index; /* prev offset */
2189 end = -1;
2190 } else {
2191 index = wbc->range_start >> PAGE_SHIFT;
2192 end = wbc->range_end >> PAGE_SHIFT;
2193 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2194 range_whole = 1;
2195 }
2196 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2197 tag_pages_for_writeback(mapping, index, end);
2198 tag = PAGECACHE_TAG_TOWRITE;
2199 } else {
2200 tag = PAGECACHE_TAG_DIRTY;
2201 }
2202 done_index = index;
2203 while (!done && (index <= end)) {
2204 int i;
2205
2206 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2207 tag);
2208 if (nr_pages == 0)
2209 break;
2210
2211 for (i = 0; i < nr_pages; i++) {
2212 struct page *page = pvec.pages[i];
2213
2214 done_index = page->index;
2215
2216 lock_page(page);
2217
2218 /*
2219 * Page truncated or invalidated. We can freely skip it
2220 * then, even for data integrity operations: the page
2221 * has disappeared concurrently, so there could be no
2222 * real expectation of this data integrity operation
2223 * even if there is now a new, dirty page at the same
2224 * pagecache address.
2225 */
2226 if (unlikely(page->mapping != mapping)) {
2227continue_unlock:
2228 unlock_page(page);
2229 continue;
2230 }
2231
2232 if (!PageDirty(page)) {
2233 /* someone wrote it for us */
2234 goto continue_unlock;
2235 }
2236
2237 if (PageWriteback(page)) {
2238 if (wbc->sync_mode != WB_SYNC_NONE)
2239 wait_on_page_writeback(page);
2240 else
2241 goto continue_unlock;
2242 }
2243
2244 BUG_ON(PageWriteback(page));
2245 if (!clear_page_dirty_for_io(page))
2246 goto continue_unlock;
2247
2248 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2249 error = (*writepage)(page, wbc, data);
2250 if (unlikely(error)) {
2251 /*
2252 * Handle errors according to the type of
2253 * writeback. There's no need to continue for
2254 * background writeback. Just push done_index
2255 * past this page so media errors won't choke
2256 * writeout for the entire file. For integrity
2257 * writeback, we must process the entire dirty
2258 * set regardless of errors because the fs may
2259 * still have state to clear for each page. In
2260 * that case we continue processing and return
2261 * the first error.
2262 */
2263 if (error == AOP_WRITEPAGE_ACTIVATE) {
2264 unlock_page(page);
2265 error = 0;
2266 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2267 ret = error;
2268 done_index = page->index + 1;
2269 done = 1;
2270 break;
2271 }
2272 if (!ret)
2273 ret = error;
2274 }
2275
2276 /*
2277 * We stop writing back only if we are not doing
2278 * integrity sync. In case of integrity sync we have to
2279 * keep going until we have written all the pages
2280 * we tagged for writeback prior to entering this loop.
2281 */
2282 if (--wbc->nr_to_write <= 0 &&
2283 wbc->sync_mode == WB_SYNC_NONE) {
2284 done = 1;
2285 break;
2286 }
2287 }
2288 pagevec_release(&pvec);
2289 cond_resched();
2290 }
2291
2292 /*
2293 * If we hit the last page and there is more work to be done: wrap
2294 * back the index back to the start of the file for the next
2295 * time we are called.
2296 */
2297 if (wbc->range_cyclic && !done)
2298 done_index = 0;
2299 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2300 mapping->writeback_index = done_index;
2301
2302 return ret;
2303}
2304EXPORT_SYMBOL(write_cache_pages);
2305
2306/*
2307 * Function used by generic_writepages to call the real writepage
2308 * function and set the mapping flags on error
2309 */
2310static int __writepage(struct page *page, struct writeback_control *wbc,
2311 void *data)
2312{
2313 struct address_space *mapping = data;
2314 int ret = mapping->a_ops->writepage(page, wbc);
2315 mapping_set_error(mapping, ret);
2316 return ret;
2317}
2318
2319/**
2320 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2321 * @mapping: address space structure to write
2322 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2323 *
2324 * This is a library function, which implements the writepages()
2325 * address_space_operation.
2326 *
2327 * Return: %0 on success, negative error code otherwise
2328 */
2329int generic_writepages(struct address_space *mapping,
2330 struct writeback_control *wbc)
2331{
2332 struct blk_plug plug;
2333 int ret;
2334
2335 /* deal with chardevs and other special file */
2336 if (!mapping->a_ops->writepage)
2337 return 0;
2338
2339 blk_start_plug(&plug);
2340 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2341 blk_finish_plug(&plug);
2342 return ret;
2343}
2344
2345EXPORT_SYMBOL(generic_writepages);
2346
2347int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2348{
2349 int ret;
2350
2351 if (wbc->nr_to_write <= 0)
2352 return 0;
2353 while (1) {
2354 if (mapping->a_ops->writepages)
2355 ret = mapping->a_ops->writepages(mapping, wbc);
2356 else
2357 ret = generic_writepages(mapping, wbc);
2358 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2359 break;
2360 cond_resched();
2361 congestion_wait(BLK_RW_ASYNC, HZ/50);
2362 }
2363 return ret;
2364}
2365
2366/**
2367 * write_one_page - write out a single page and wait on I/O
2368 * @page: the page to write
2369 *
2370 * The page must be locked by the caller and will be unlocked upon return.
2371 *
2372 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2373 * function returns.
2374 *
2375 * Return: %0 on success, negative error code otherwise
2376 */
2377int write_one_page(struct page *page)
2378{
2379 struct address_space *mapping = page->mapping;
2380 int ret = 0;
2381 struct writeback_control wbc = {
2382 .sync_mode = WB_SYNC_ALL,
2383 .nr_to_write = 1,
2384 };
2385
2386 BUG_ON(!PageLocked(page));
2387
2388 wait_on_page_writeback(page);
2389
2390 if (clear_page_dirty_for_io(page)) {
2391 get_page(page);
2392 ret = mapping->a_ops->writepage(page, &wbc);
2393 if (ret == 0)
2394 wait_on_page_writeback(page);
2395 put_page(page);
2396 } else {
2397 unlock_page(page);
2398 }
2399
2400 if (!ret)
2401 ret = filemap_check_errors(mapping);
2402 return ret;
2403}
2404EXPORT_SYMBOL(write_one_page);
2405
2406/*
2407 * For address_spaces which do not use buffers nor write back.
2408 */
2409int __set_page_dirty_no_writeback(struct page *page)
2410{
2411 if (!PageDirty(page))
2412 return !TestSetPageDirty(page);
2413 return 0;
2414}
2415EXPORT_SYMBOL(__set_page_dirty_no_writeback);
2416
2417/*
2418 * Helper function for set_page_dirty family.
2419 *
2420 * Caller must hold lock_page_memcg().
2421 *
2422 * NOTE: This relies on being atomic wrt interrupts.
2423 */
2424static void account_page_dirtied(struct page *page,
2425 struct address_space *mapping)
2426{
2427 struct inode *inode = mapping->host;
2428
2429 trace_writeback_dirty_page(page, mapping);
2430
2431 if (mapping_can_writeback(mapping)) {
2432 struct bdi_writeback *wb;
2433
2434 inode_attach_wb(inode, page);
2435 wb = inode_to_wb(inode);
2436
2437 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2438 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2439 __inc_node_page_state(page, NR_DIRTIED);
2440 inc_wb_stat(wb, WB_RECLAIMABLE);
2441 inc_wb_stat(wb, WB_DIRTIED);
2442 task_io_account_write(PAGE_SIZE);
2443 current->nr_dirtied++;
2444 __this_cpu_inc(bdp_ratelimits);
2445
2446 mem_cgroup_track_foreign_dirty(page, wb);
2447 }
2448}
2449
2450/*
2451 * Helper function for deaccounting dirty page without writeback.
2452 *
2453 * Caller must hold lock_page_memcg().
2454 */
2455void account_page_cleaned(struct page *page, struct address_space *mapping,
2456 struct bdi_writeback *wb)
2457{
2458 if (mapping_can_writeback(mapping)) {
2459 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2460 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2461 dec_wb_stat(wb, WB_RECLAIMABLE);
2462 task_io_account_cancelled_write(PAGE_SIZE);
2463 }
2464}
2465
2466/*
2467 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
2468 * dirty.
2469 *
2470 * If warn is true, then emit a warning if the page is not uptodate and has
2471 * not been truncated.
2472 *
2473 * The caller must hold lock_page_memcg().
2474 */
2475void __set_page_dirty(struct page *page, struct address_space *mapping,
2476 int warn)
2477{
2478 unsigned long flags;
2479
2480 xa_lock_irqsave(&mapping->i_pages, flags);
2481 if (page->mapping) { /* Race with truncate? */
2482 WARN_ON_ONCE(warn && !PageUptodate(page));
2483 account_page_dirtied(page, mapping);
2484 __xa_set_mark(&mapping->i_pages, page_index(page),
2485 PAGECACHE_TAG_DIRTY);
2486 }
2487 xa_unlock_irqrestore(&mapping->i_pages, flags);
2488}
2489
2490/*
2491 * For address_spaces which do not use buffers. Just tag the page as dirty in
2492 * the xarray.
2493 *
2494 * This is also used when a single buffer is being dirtied: we want to set the
2495 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2496 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2497 *
2498 * The caller must ensure this doesn't race with truncation. Most will simply
2499 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2500 * the pte lock held, which also locks out truncation.
2501 */
2502int __set_page_dirty_nobuffers(struct page *page)
2503{
2504 lock_page_memcg(page);
2505 if (!TestSetPageDirty(page)) {
2506 struct address_space *mapping = page_mapping(page);
2507
2508 if (!mapping) {
2509 unlock_page_memcg(page);
2510 return 1;
2511 }
2512 __set_page_dirty(page, mapping, !PagePrivate(page));
2513 unlock_page_memcg(page);
2514
2515 if (mapping->host) {
2516 /* !PageAnon && !swapper_space */
2517 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2518 }
2519 return 1;
2520 }
2521 unlock_page_memcg(page);
2522 return 0;
2523}
2524EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2525
2526/*
2527 * Call this whenever redirtying a page, to de-account the dirty counters
2528 * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2529 * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2530 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2531 * control.
2532 */
2533void account_page_redirty(struct page *page)
2534{
2535 struct address_space *mapping = page->mapping;
2536
2537 if (mapping && mapping_can_writeback(mapping)) {
2538 struct inode *inode = mapping->host;
2539 struct bdi_writeback *wb;
2540 struct wb_lock_cookie cookie = {};
2541
2542 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2543 current->nr_dirtied--;
2544 dec_node_page_state(page, NR_DIRTIED);
2545 dec_wb_stat(wb, WB_DIRTIED);
2546 unlocked_inode_to_wb_end(inode, &cookie);
2547 }
2548}
2549EXPORT_SYMBOL(account_page_redirty);
2550
2551/*
2552 * When a writepage implementation decides that it doesn't want to write this
2553 * page for some reason, it should redirty the locked page via
2554 * redirty_page_for_writepage() and it should then unlock the page and return 0
2555 */
2556int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2557{
2558 int ret;
2559
2560 wbc->pages_skipped++;
2561 ret = __set_page_dirty_nobuffers(page);
2562 account_page_redirty(page);
2563 return ret;
2564}
2565EXPORT_SYMBOL(redirty_page_for_writepage);
2566
2567/*
2568 * Dirty a page.
2569 *
2570 * For pages with a mapping this should be done under the page lock for the
2571 * benefit of asynchronous memory errors who prefer a consistent dirty state.
2572 * This rule can be broken in some special cases, but should be better not to.
2573 */
2574int set_page_dirty(struct page *page)
2575{
2576 struct address_space *mapping = page_mapping(page);
2577
2578 page = compound_head(page);
2579 if (likely(mapping)) {
2580 /*
2581 * readahead/lru_deactivate_page could remain
2582 * PG_readahead/PG_reclaim due to race with end_page_writeback
2583 * About readahead, if the page is written, the flags would be
2584 * reset. So no problem.
2585 * About lru_deactivate_page, if the page is redirty, the flag
2586 * will be reset. So no problem. but if the page is used by readahead
2587 * it will confuse readahead and make it restart the size rampup
2588 * process. But it's a trivial problem.
2589 */
2590 if (PageReclaim(page))
2591 ClearPageReclaim(page);
2592 return mapping->a_ops->set_page_dirty(page);
2593 }
2594 if (!PageDirty(page)) {
2595 if (!TestSetPageDirty(page))
2596 return 1;
2597 }
2598 return 0;
2599}
2600EXPORT_SYMBOL(set_page_dirty);
2601
2602/*
2603 * set_page_dirty() is racy if the caller has no reference against
2604 * page->mapping->host, and if the page is unlocked. This is because another
2605 * CPU could truncate the page off the mapping and then free the mapping.
2606 *
2607 * Usually, the page _is_ locked, or the caller is a user-space process which
2608 * holds a reference on the inode by having an open file.
2609 *
2610 * In other cases, the page should be locked before running set_page_dirty().
2611 */
2612int set_page_dirty_lock(struct page *page)
2613{
2614 int ret;
2615
2616 lock_page(page);
2617 ret = set_page_dirty(page);
2618 unlock_page(page);
2619 return ret;
2620}
2621EXPORT_SYMBOL(set_page_dirty_lock);
2622
2623/*
2624 * This cancels just the dirty bit on the kernel page itself, it does NOT
2625 * actually remove dirty bits on any mmap's that may be around. It also
2626 * leaves the page tagged dirty, so any sync activity will still find it on
2627 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2628 * look at the dirty bits in the VM.
2629 *
2630 * Doing this should *normally* only ever be done when a page is truncated,
2631 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2632 * this when it notices that somebody has cleaned out all the buffers on a
2633 * page without actually doing it through the VM. Can you say "ext3 is
2634 * horribly ugly"? Thought you could.
2635 */
2636void __cancel_dirty_page(struct page *page)
2637{
2638 struct address_space *mapping = page_mapping(page);
2639
2640 if (mapping_can_writeback(mapping)) {
2641 struct inode *inode = mapping->host;
2642 struct bdi_writeback *wb;
2643 struct wb_lock_cookie cookie = {};
2644
2645 lock_page_memcg(page);
2646 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2647
2648 if (TestClearPageDirty(page))
2649 account_page_cleaned(page, mapping, wb);
2650
2651 unlocked_inode_to_wb_end(inode, &cookie);
2652 unlock_page_memcg(page);
2653 } else {
2654 ClearPageDirty(page);
2655 }
2656}
2657EXPORT_SYMBOL(__cancel_dirty_page);
2658
2659/*
2660 * Clear a page's dirty flag, while caring for dirty memory accounting.
2661 * Returns true if the page was previously dirty.
2662 *
2663 * This is for preparing to put the page under writeout. We leave the page
2664 * tagged as dirty in the xarray so that a concurrent write-for-sync
2665 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2666 * implementation will run either set_page_writeback() or set_page_dirty(),
2667 * at which stage we bring the page's dirty flag and xarray dirty tag
2668 * back into sync.
2669 *
2670 * This incoherency between the page's dirty flag and xarray tag is
2671 * unfortunate, but it only exists while the page is locked.
2672 */
2673int clear_page_dirty_for_io(struct page *page)
2674{
2675 struct address_space *mapping = page_mapping(page);
2676 int ret = 0;
2677
2678 VM_BUG_ON_PAGE(!PageLocked(page), page);
2679
2680 if (mapping && mapping_can_writeback(mapping)) {
2681 struct inode *inode = mapping->host;
2682 struct bdi_writeback *wb;
2683 struct wb_lock_cookie cookie = {};
2684
2685 /*
2686 * Yes, Virginia, this is indeed insane.
2687 *
2688 * We use this sequence to make sure that
2689 * (a) we account for dirty stats properly
2690 * (b) we tell the low-level filesystem to
2691 * mark the whole page dirty if it was
2692 * dirty in a pagetable. Only to then
2693 * (c) clean the page again and return 1 to
2694 * cause the writeback.
2695 *
2696 * This way we avoid all nasty races with the
2697 * dirty bit in multiple places and clearing
2698 * them concurrently from different threads.
2699 *
2700 * Note! Normally the "set_page_dirty(page)"
2701 * has no effect on the actual dirty bit - since
2702 * that will already usually be set. But we
2703 * need the side effects, and it can help us
2704 * avoid races.
2705 *
2706 * We basically use the page "master dirty bit"
2707 * as a serialization point for all the different
2708 * threads doing their things.
2709 */
2710 if (page_mkclean(page))
2711 set_page_dirty(page);
2712 /*
2713 * We carefully synchronise fault handlers against
2714 * installing a dirty pte and marking the page dirty
2715 * at this point. We do this by having them hold the
2716 * page lock while dirtying the page, and pages are
2717 * always locked coming in here, so we get the desired
2718 * exclusion.
2719 */
2720 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2721 if (TestClearPageDirty(page)) {
2722 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2723 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2724 dec_wb_stat(wb, WB_RECLAIMABLE);
2725 ret = 1;
2726 }
2727 unlocked_inode_to_wb_end(inode, &cookie);
2728 return ret;
2729 }
2730 return TestClearPageDirty(page);
2731}
2732EXPORT_SYMBOL(clear_page_dirty_for_io);
2733
2734int test_clear_page_writeback(struct page *page)
2735{
2736 struct address_space *mapping = page_mapping(page);
2737 int ret;
2738
2739 lock_page_memcg(page);
2740 if (mapping && mapping_use_writeback_tags(mapping)) {
2741 struct inode *inode = mapping->host;
2742 struct backing_dev_info *bdi = inode_to_bdi(inode);
2743 unsigned long flags;
2744
2745 xa_lock_irqsave(&mapping->i_pages, flags);
2746 ret = TestClearPageWriteback(page);
2747 if (ret) {
2748 __xa_clear_mark(&mapping->i_pages, page_index(page),
2749 PAGECACHE_TAG_WRITEBACK);
2750 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2751 struct bdi_writeback *wb = inode_to_wb(inode);
2752
2753 dec_wb_stat(wb, WB_WRITEBACK);
2754 __wb_writeout_inc(wb);
2755 }
2756 }
2757
2758 if (mapping->host && !mapping_tagged(mapping,
2759 PAGECACHE_TAG_WRITEBACK))
2760 sb_clear_inode_writeback(mapping->host);
2761
2762 xa_unlock_irqrestore(&mapping->i_pages, flags);
2763 } else {
2764 ret = TestClearPageWriteback(page);
2765 }
2766 if (ret) {
2767 dec_lruvec_page_state(page, NR_WRITEBACK);
2768 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2769 inc_node_page_state(page, NR_WRITTEN);
2770 }
2771 unlock_page_memcg(page);
2772 return ret;
2773}
2774
2775int __test_set_page_writeback(struct page *page, bool keep_write)
2776{
2777 struct address_space *mapping = page_mapping(page);
2778 int ret, access_ret;
2779
2780 lock_page_memcg(page);
2781 if (mapping && mapping_use_writeback_tags(mapping)) {
2782 XA_STATE(xas, &mapping->i_pages, page_index(page));
2783 struct inode *inode = mapping->host;
2784 struct backing_dev_info *bdi = inode_to_bdi(inode);
2785 unsigned long flags;
2786
2787 xas_lock_irqsave(&xas, flags);
2788 xas_load(&xas);
2789 ret = TestSetPageWriteback(page);
2790 if (!ret) {
2791 bool on_wblist;
2792
2793 on_wblist = mapping_tagged(mapping,
2794 PAGECACHE_TAG_WRITEBACK);
2795
2796 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2797 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT)
2798 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2799
2800 /*
2801 * We can come through here when swapping anonymous
2802 * pages, so we don't necessarily have an inode to track
2803 * for sync.
2804 */
2805 if (mapping->host && !on_wblist)
2806 sb_mark_inode_writeback(mapping->host);
2807 }
2808 if (!PageDirty(page))
2809 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2810 if (!keep_write)
2811 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2812 xas_unlock_irqrestore(&xas, flags);
2813 } else {
2814 ret = TestSetPageWriteback(page);
2815 }
2816 if (!ret) {
2817 inc_lruvec_page_state(page, NR_WRITEBACK);
2818 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2819 }
2820 unlock_page_memcg(page);
2821 access_ret = arch_make_page_accessible(page);
2822 /*
2823 * If writeback has been triggered on a page that cannot be made
2824 * accessible, it is too late to recover here.
2825 */
2826 VM_BUG_ON_PAGE(access_ret != 0, page);
2827
2828 return ret;
2829
2830}
2831EXPORT_SYMBOL(__test_set_page_writeback);
2832
2833/*
2834 * Wait for a page to complete writeback
2835 */
2836void wait_on_page_writeback(struct page *page)
2837{
2838 while (PageWriteback(page)) {
2839 trace_wait_on_page_writeback(page, page_mapping(page));
2840 wait_on_page_bit(page, PG_writeback);
2841 }
2842}
2843EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2844
2845/*
2846 * Wait for a page to complete writeback. Returns -EINTR if we get a
2847 * fatal signal while waiting.
2848 */
2849int wait_on_page_writeback_killable(struct page *page)
2850{
2851 while (PageWriteback(page)) {
2852 trace_wait_on_page_writeback(page, page_mapping(page));
2853 if (wait_on_page_bit_killable(page, PG_writeback))
2854 return -EINTR;
2855 }
2856
2857 return 0;
2858}
2859EXPORT_SYMBOL_GPL(wait_on_page_writeback_killable);
2860
2861/**
2862 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2863 * @page: The page to wait on.
2864 *
2865 * This function determines if the given page is related to a backing device
2866 * that requires page contents to be held stable during writeback. If so, then
2867 * it will wait for any pending writeback to complete.
2868 */
2869void wait_for_stable_page(struct page *page)
2870{
2871 page = thp_head(page);
2872 if (page->mapping->host->i_sb->s_iflags & SB_I_STABLE_WRITES)
2873 wait_on_page_writeback(page);
2874}
2875EXPORT_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);