Loading...
1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/export.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
25#include <linux/task_io_accounting_ops.h>
26#include <linux/blkdev.h>
27#include <linux/mpage.h>
28#include <linux/rmap.h>
29#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/sched/signal.h>
40#include <linux/mm_inline.h>
41#include <trace/events/writeback.h>
42
43#include "internal.h"
44
45/*
46 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48#define MAX_PAUSE max(HZ/5, 1)
49
50/*
51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56/*
57 * Estimate write bandwidth at 200ms intervals.
58 */
59#define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
61#define RATELIMIT_CALC_SHIFT 10
62
63/*
64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67static long ratelimit_pages = 32;
68
69/* The following parameters are exported via /proc/sys/vm */
70
71/*
72 * Start background writeback (via writeback threads) at this percentage
73 */
74int dirty_background_ratio = 10;
75
76/*
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
79 */
80unsigned long dirty_background_bytes;
81
82/*
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 */
86int vm_highmem_is_dirtyable;
87
88/*
89 * The generator of dirty data starts writeback at this percentage
90 */
91int vm_dirty_ratio = 20;
92
93/*
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
96 */
97unsigned long vm_dirty_bytes;
98
99/*
100 * The interval between `kupdate'-style writebacks
101 */
102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106/*
107 * The longest time for which data is allowed to remain dirty
108 */
109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111/*
112 * Flag that makes the machine dump writes/reads and block dirtyings.
113 */
114int block_dump;
115
116/*
117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
119 */
120int laptop_mode;
121
122EXPORT_SYMBOL(laptop_mode);
123
124/* End of sysctl-exported parameters */
125
126struct wb_domain global_wb_domain;
127
128/* consolidated parameters for balance_dirty_pages() and its subroutines */
129struct dirty_throttle_control {
130#ifdef CONFIG_CGROUP_WRITEBACK
131 struct wb_domain *dom;
132 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
133#endif
134 struct bdi_writeback *wb;
135 struct fprop_local_percpu *wb_completions;
136
137 unsigned long avail; /* dirtyable */
138 unsigned long dirty; /* file_dirty + write + nfs */
139 unsigned long thresh; /* dirty threshold */
140 unsigned long bg_thresh; /* dirty background threshold */
141
142 unsigned long wb_dirty; /* per-wb counterparts */
143 unsigned long wb_thresh;
144 unsigned long wb_bg_thresh;
145
146 unsigned long pos_ratio;
147};
148
149/*
150 * Length of period for aging writeout fractions of bdis. This is an
151 * arbitrarily chosen number. The longer the period, the slower fractions will
152 * reflect changes in current writeout rate.
153 */
154#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155
156#ifdef CONFIG_CGROUP_WRITEBACK
157
158#define GDTC_INIT(__wb) .wb = (__wb), \
159 .dom = &global_wb_domain, \
160 .wb_completions = &(__wb)->completions
161
162#define GDTC_INIT_NO_WB .dom = &global_wb_domain
163
164#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
165 .dom = mem_cgroup_wb_domain(__wb), \
166 .wb_completions = &(__wb)->memcg_completions, \
167 .gdtc = __gdtc
168
169static bool mdtc_valid(struct dirty_throttle_control *dtc)
170{
171 return dtc->dom;
172}
173
174static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175{
176 return dtc->dom;
177}
178
179static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180{
181 return mdtc->gdtc;
182}
183
184static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185{
186 return &wb->memcg_completions;
187}
188
189static void wb_min_max_ratio(struct bdi_writeback *wb,
190 unsigned long *minp, unsigned long *maxp)
191{
192 unsigned long this_bw = wb->avg_write_bandwidth;
193 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194 unsigned long long min = wb->bdi->min_ratio;
195 unsigned long long max = wb->bdi->max_ratio;
196
197 /*
198 * @wb may already be clean by the time control reaches here and
199 * the total may not include its bw.
200 */
201 if (this_bw < tot_bw) {
202 if (min) {
203 min *= this_bw;
204 do_div(min, tot_bw);
205 }
206 if (max < 100) {
207 max *= this_bw;
208 do_div(max, tot_bw);
209 }
210 }
211
212 *minp = min;
213 *maxp = max;
214}
215
216#else /* CONFIG_CGROUP_WRITEBACK */
217
218#define GDTC_INIT(__wb) .wb = (__wb), \
219 .wb_completions = &(__wb)->completions
220#define GDTC_INIT_NO_WB
221#define MDTC_INIT(__wb, __gdtc)
222
223static bool mdtc_valid(struct dirty_throttle_control *dtc)
224{
225 return false;
226}
227
228static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229{
230 return &global_wb_domain;
231}
232
233static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234{
235 return NULL;
236}
237
238static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239{
240 return NULL;
241}
242
243static void wb_min_max_ratio(struct bdi_writeback *wb,
244 unsigned long *minp, unsigned long *maxp)
245{
246 *minp = wb->bdi->min_ratio;
247 *maxp = wb->bdi->max_ratio;
248}
249
250#endif /* CONFIG_CGROUP_WRITEBACK */
251
252/*
253 * In a memory zone, there is a certain amount of pages we consider
254 * available for the page cache, which is essentially the number of
255 * free and reclaimable pages, minus some zone reserves to protect
256 * lowmem and the ability to uphold the zone's watermarks without
257 * requiring writeback.
258 *
259 * This number of dirtyable pages is the base value of which the
260 * user-configurable dirty ratio is the effictive number of pages that
261 * are allowed to be actually dirtied. Per individual zone, or
262 * globally by using the sum of dirtyable pages over all zones.
263 *
264 * Because the user is allowed to specify the dirty limit globally as
265 * absolute number of bytes, calculating the per-zone dirty limit can
266 * require translating the configured limit into a percentage of
267 * global dirtyable memory first.
268 */
269
270/**
271 * node_dirtyable_memory - number of dirtyable pages in a node
272 * @pgdat: the node
273 *
274 * Returns the node's number of pages potentially available for dirty
275 * page cache. This is the base value for the per-node dirty limits.
276 */
277static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278{
279 unsigned long nr_pages = 0;
280 int z;
281
282 for (z = 0; z < MAX_NR_ZONES; z++) {
283 struct zone *zone = pgdat->node_zones + z;
284
285 if (!populated_zone(zone))
286 continue;
287
288 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289 }
290
291 /*
292 * Pages reserved for the kernel should not be considered
293 * dirtyable, to prevent a situation where reclaim has to
294 * clean pages in order to balance the zones.
295 */
296 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297
298 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
300
301 return nr_pages;
302}
303
304static unsigned long highmem_dirtyable_memory(unsigned long total)
305{
306#ifdef CONFIG_HIGHMEM
307 int node;
308 unsigned long x = 0;
309 int i;
310
311 for_each_node_state(node, N_HIGH_MEMORY) {
312 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313 struct zone *z;
314 unsigned long nr_pages;
315
316 if (!is_highmem_idx(i))
317 continue;
318
319 z = &NODE_DATA(node)->node_zones[i];
320 if (!populated_zone(z))
321 continue;
322
323 nr_pages = zone_page_state(z, NR_FREE_PAGES);
324 /* watch for underflows */
325 nr_pages -= min(nr_pages, high_wmark_pages(z));
326 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328 x += nr_pages;
329 }
330 }
331
332 /*
333 * Unreclaimable memory (kernel memory or anonymous memory
334 * without swap) can bring down the dirtyable pages below
335 * the zone's dirty balance reserve and the above calculation
336 * will underflow. However we still want to add in nodes
337 * which are below threshold (negative values) to get a more
338 * accurate calculation but make sure that the total never
339 * underflows.
340 */
341 if ((long)x < 0)
342 x = 0;
343
344 /*
345 * Make sure that the number of highmem pages is never larger
346 * than the number of the total dirtyable memory. This can only
347 * occur in very strange VM situations but we want to make sure
348 * that this does not occur.
349 */
350 return min(x, total);
351#else
352 return 0;
353#endif
354}
355
356/**
357 * global_dirtyable_memory - number of globally dirtyable pages
358 *
359 * Returns the global number of pages potentially available for dirty
360 * page cache. This is the base value for the global dirty limits.
361 */
362static unsigned long global_dirtyable_memory(void)
363{
364 unsigned long x;
365
366 x = global_zone_page_state(NR_FREE_PAGES);
367 /*
368 * Pages reserved for the kernel should not be considered
369 * dirtyable, to prevent a situation where reclaim has to
370 * clean pages in order to balance the zones.
371 */
372 x -= min(x, totalreserve_pages);
373
374 x += global_node_page_state(NR_INACTIVE_FILE);
375 x += global_node_page_state(NR_ACTIVE_FILE);
376
377 if (!vm_highmem_is_dirtyable)
378 x -= highmem_dirtyable_memory(x);
379
380 return x + 1; /* Ensure that we never return 0 */
381}
382
383/**
384 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385 * @dtc: dirty_throttle_control of interest
386 *
387 * Calculate @dtc->thresh and ->bg_thresh considering
388 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
389 * must ensure that @dtc->avail is set before calling this function. The
390 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
391 * real-time tasks.
392 */
393static void domain_dirty_limits(struct dirty_throttle_control *dtc)
394{
395 const unsigned long available_memory = dtc->avail;
396 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
397 unsigned long bytes = vm_dirty_bytes;
398 unsigned long bg_bytes = dirty_background_bytes;
399 /* convert ratios to per-PAGE_SIZE for higher precision */
400 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
401 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
402 unsigned long thresh;
403 unsigned long bg_thresh;
404 struct task_struct *tsk;
405
406 /* gdtc is !NULL iff @dtc is for memcg domain */
407 if (gdtc) {
408 unsigned long global_avail = gdtc->avail;
409
410 /*
411 * The byte settings can't be applied directly to memcg
412 * domains. Convert them to ratios by scaling against
413 * globally available memory. As the ratios are in
414 * per-PAGE_SIZE, they can be obtained by dividing bytes by
415 * number of pages.
416 */
417 if (bytes)
418 ratio = min(DIV_ROUND_UP(bytes, global_avail),
419 PAGE_SIZE);
420 if (bg_bytes)
421 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
422 PAGE_SIZE);
423 bytes = bg_bytes = 0;
424 }
425
426 if (bytes)
427 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
428 else
429 thresh = (ratio * available_memory) / PAGE_SIZE;
430
431 if (bg_bytes)
432 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
433 else
434 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
435
436 if (bg_thresh >= thresh)
437 bg_thresh = thresh / 2;
438 tsk = current;
439 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
440 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
441 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
442 }
443 dtc->thresh = thresh;
444 dtc->bg_thresh = bg_thresh;
445
446 /* we should eventually report the domain in the TP */
447 if (!gdtc)
448 trace_global_dirty_state(bg_thresh, thresh);
449}
450
451/**
452 * global_dirty_limits - background-writeback and dirty-throttling thresholds
453 * @pbackground: out parameter for bg_thresh
454 * @pdirty: out parameter for thresh
455 *
456 * Calculate bg_thresh and thresh for global_wb_domain. See
457 * domain_dirty_limits() for details.
458 */
459void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
460{
461 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
462
463 gdtc.avail = global_dirtyable_memory();
464 domain_dirty_limits(&gdtc);
465
466 *pbackground = gdtc.bg_thresh;
467 *pdirty = gdtc.thresh;
468}
469
470/**
471 * node_dirty_limit - maximum number of dirty pages allowed in a node
472 * @pgdat: the node
473 *
474 * Returns the maximum number of dirty pages allowed in a node, based
475 * on the node's dirtyable memory.
476 */
477static unsigned long node_dirty_limit(struct pglist_data *pgdat)
478{
479 unsigned long node_memory = node_dirtyable_memory(pgdat);
480 struct task_struct *tsk = current;
481 unsigned long dirty;
482
483 if (vm_dirty_bytes)
484 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
485 node_memory / global_dirtyable_memory();
486 else
487 dirty = vm_dirty_ratio * node_memory / 100;
488
489 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
490 dirty += dirty / 4;
491
492 return dirty;
493}
494
495/**
496 * node_dirty_ok - tells whether a node is within its dirty limits
497 * @pgdat: the node to check
498 *
499 * Returns %true when the dirty pages in @pgdat are within the node's
500 * dirty limit, %false if the limit is exceeded.
501 */
502bool node_dirty_ok(struct pglist_data *pgdat)
503{
504 unsigned long limit = node_dirty_limit(pgdat);
505 unsigned long nr_pages = 0;
506
507 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
508 nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
509 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
510
511 return nr_pages <= limit;
512}
513
514int dirty_background_ratio_handler(struct ctl_table *table, int write,
515 void __user *buffer, size_t *lenp,
516 loff_t *ppos)
517{
518 int ret;
519
520 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
521 if (ret == 0 && write)
522 dirty_background_bytes = 0;
523 return ret;
524}
525
526int dirty_background_bytes_handler(struct ctl_table *table, int write,
527 void __user *buffer, size_t *lenp,
528 loff_t *ppos)
529{
530 int ret;
531
532 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
533 if (ret == 0 && write)
534 dirty_background_ratio = 0;
535 return ret;
536}
537
538int dirty_ratio_handler(struct ctl_table *table, int write,
539 void __user *buffer, size_t *lenp,
540 loff_t *ppos)
541{
542 int old_ratio = vm_dirty_ratio;
543 int ret;
544
545 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
546 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
547 writeback_set_ratelimit();
548 vm_dirty_bytes = 0;
549 }
550 return ret;
551}
552
553int dirty_bytes_handler(struct ctl_table *table, int write,
554 void __user *buffer, size_t *lenp,
555 loff_t *ppos)
556{
557 unsigned long old_bytes = vm_dirty_bytes;
558 int ret;
559
560 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
561 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
562 writeback_set_ratelimit();
563 vm_dirty_ratio = 0;
564 }
565 return ret;
566}
567
568static unsigned long wp_next_time(unsigned long cur_time)
569{
570 cur_time += VM_COMPLETIONS_PERIOD_LEN;
571 /* 0 has a special meaning... */
572 if (!cur_time)
573 return 1;
574 return cur_time;
575}
576
577static void wb_domain_writeout_inc(struct wb_domain *dom,
578 struct fprop_local_percpu *completions,
579 unsigned int max_prop_frac)
580{
581 __fprop_inc_percpu_max(&dom->completions, completions,
582 max_prop_frac);
583 /* First event after period switching was turned off? */
584 if (unlikely(!dom->period_time)) {
585 /*
586 * We can race with other __bdi_writeout_inc calls here but
587 * it does not cause any harm since the resulting time when
588 * timer will fire and what is in writeout_period_time will be
589 * roughly the same.
590 */
591 dom->period_time = wp_next_time(jiffies);
592 mod_timer(&dom->period_timer, dom->period_time);
593 }
594}
595
596/*
597 * Increment @wb's writeout completion count and the global writeout
598 * completion count. Called from test_clear_page_writeback().
599 */
600static inline void __wb_writeout_inc(struct bdi_writeback *wb)
601{
602 struct wb_domain *cgdom;
603
604 inc_wb_stat(wb, WB_WRITTEN);
605 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
606 wb->bdi->max_prop_frac);
607
608 cgdom = mem_cgroup_wb_domain(wb);
609 if (cgdom)
610 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
611 wb->bdi->max_prop_frac);
612}
613
614void wb_writeout_inc(struct bdi_writeback *wb)
615{
616 unsigned long flags;
617
618 local_irq_save(flags);
619 __wb_writeout_inc(wb);
620 local_irq_restore(flags);
621}
622EXPORT_SYMBOL_GPL(wb_writeout_inc);
623
624/*
625 * On idle system, we can be called long after we scheduled because we use
626 * deferred timers so count with missed periods.
627 */
628static void writeout_period(struct timer_list *t)
629{
630 struct wb_domain *dom = from_timer(dom, t, period_timer);
631 int miss_periods = (jiffies - dom->period_time) /
632 VM_COMPLETIONS_PERIOD_LEN;
633
634 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
635 dom->period_time = wp_next_time(dom->period_time +
636 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
637 mod_timer(&dom->period_timer, dom->period_time);
638 } else {
639 /*
640 * Aging has zeroed all fractions. Stop wasting CPU on period
641 * updates.
642 */
643 dom->period_time = 0;
644 }
645}
646
647int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
648{
649 memset(dom, 0, sizeof(*dom));
650
651 spin_lock_init(&dom->lock);
652
653 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
654
655 dom->dirty_limit_tstamp = jiffies;
656
657 return fprop_global_init(&dom->completions, gfp);
658}
659
660#ifdef CONFIG_CGROUP_WRITEBACK
661void wb_domain_exit(struct wb_domain *dom)
662{
663 del_timer_sync(&dom->period_timer);
664 fprop_global_destroy(&dom->completions);
665}
666#endif
667
668/*
669 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
670 * registered backing devices, which, for obvious reasons, can not
671 * exceed 100%.
672 */
673static unsigned int bdi_min_ratio;
674
675int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
676{
677 int ret = 0;
678
679 spin_lock_bh(&bdi_lock);
680 if (min_ratio > bdi->max_ratio) {
681 ret = -EINVAL;
682 } else {
683 min_ratio -= bdi->min_ratio;
684 if (bdi_min_ratio + min_ratio < 100) {
685 bdi_min_ratio += min_ratio;
686 bdi->min_ratio += min_ratio;
687 } else {
688 ret = -EINVAL;
689 }
690 }
691 spin_unlock_bh(&bdi_lock);
692
693 return ret;
694}
695
696int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
697{
698 int ret = 0;
699
700 if (max_ratio > 100)
701 return -EINVAL;
702
703 spin_lock_bh(&bdi_lock);
704 if (bdi->min_ratio > max_ratio) {
705 ret = -EINVAL;
706 } else {
707 bdi->max_ratio = max_ratio;
708 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
709 }
710 spin_unlock_bh(&bdi_lock);
711
712 return ret;
713}
714EXPORT_SYMBOL(bdi_set_max_ratio);
715
716static unsigned long dirty_freerun_ceiling(unsigned long thresh,
717 unsigned long bg_thresh)
718{
719 return (thresh + bg_thresh) / 2;
720}
721
722static unsigned long hard_dirty_limit(struct wb_domain *dom,
723 unsigned long thresh)
724{
725 return max(thresh, dom->dirty_limit);
726}
727
728/*
729 * Memory which can be further allocated to a memcg domain is capped by
730 * system-wide clean memory excluding the amount being used in the domain.
731 */
732static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
733 unsigned long filepages, unsigned long headroom)
734{
735 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
736 unsigned long clean = filepages - min(filepages, mdtc->dirty);
737 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
738 unsigned long other_clean = global_clean - min(global_clean, clean);
739
740 mdtc->avail = filepages + min(headroom, other_clean);
741}
742
743/**
744 * __wb_calc_thresh - @wb's share of dirty throttling threshold
745 * @dtc: dirty_throttle_context of interest
746 *
747 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
748 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
749 *
750 * Note that balance_dirty_pages() will only seriously take it as a hard limit
751 * when sleeping max_pause per page is not enough to keep the dirty pages under
752 * control. For example, when the device is completely stalled due to some error
753 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
754 * In the other normal situations, it acts more gently by throttling the tasks
755 * more (rather than completely block them) when the wb dirty pages go high.
756 *
757 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
758 * - starving fast devices
759 * - piling up dirty pages (that will take long time to sync) on slow devices
760 *
761 * The wb's share of dirty limit will be adapting to its throughput and
762 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
763 */
764static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
765{
766 struct wb_domain *dom = dtc_dom(dtc);
767 unsigned long thresh = dtc->thresh;
768 u64 wb_thresh;
769 long numerator, denominator;
770 unsigned long wb_min_ratio, wb_max_ratio;
771
772 /*
773 * Calculate this BDI's share of the thresh ratio.
774 */
775 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
776 &numerator, &denominator);
777
778 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
779 wb_thresh *= numerator;
780 do_div(wb_thresh, denominator);
781
782 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
783
784 wb_thresh += (thresh * wb_min_ratio) / 100;
785 if (wb_thresh > (thresh * wb_max_ratio) / 100)
786 wb_thresh = thresh * wb_max_ratio / 100;
787
788 return wb_thresh;
789}
790
791unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
792{
793 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
794 .thresh = thresh };
795 return __wb_calc_thresh(&gdtc);
796}
797
798/*
799 * setpoint - dirty 3
800 * f(dirty) := 1.0 + (----------------)
801 * limit - setpoint
802 *
803 * it's a 3rd order polynomial that subjects to
804 *
805 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
806 * (2) f(setpoint) = 1.0 => the balance point
807 * (3) f(limit) = 0 => the hard limit
808 * (4) df/dx <= 0 => negative feedback control
809 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
810 * => fast response on large errors; small oscillation near setpoint
811 */
812static long long pos_ratio_polynom(unsigned long setpoint,
813 unsigned long dirty,
814 unsigned long limit)
815{
816 long long pos_ratio;
817 long x;
818
819 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
820 (limit - setpoint) | 1);
821 pos_ratio = x;
822 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
823 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
824 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
825
826 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
827}
828
829/*
830 * Dirty position control.
831 *
832 * (o) global/bdi setpoints
833 *
834 * We want the dirty pages be balanced around the global/wb setpoints.
835 * When the number of dirty pages is higher/lower than the setpoint, the
836 * dirty position control ratio (and hence task dirty ratelimit) will be
837 * decreased/increased to bring the dirty pages back to the setpoint.
838 *
839 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
840 *
841 * if (dirty < setpoint) scale up pos_ratio
842 * if (dirty > setpoint) scale down pos_ratio
843 *
844 * if (wb_dirty < wb_setpoint) scale up pos_ratio
845 * if (wb_dirty > wb_setpoint) scale down pos_ratio
846 *
847 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
848 *
849 * (o) global control line
850 *
851 * ^ pos_ratio
852 * |
853 * | |<===== global dirty control scope ======>|
854 * 2.0 .............*
855 * | .*
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * | . *
861 * 1.0 ................................*
862 * | . . *
863 * | . . *
864 * | . . *
865 * | . . *
866 * | . . *
867 * 0 +------------.------------------.----------------------*------------->
868 * freerun^ setpoint^ limit^ dirty pages
869 *
870 * (o) wb control line
871 *
872 * ^ pos_ratio
873 * |
874 * | *
875 * | *
876 * | *
877 * | *
878 * | * |<=========== span ============>|
879 * 1.0 .......................*
880 * | . *
881 * | . *
882 * | . *
883 * | . *
884 * | . *
885 * | . *
886 * | . *
887 * | . *
888 * | . *
889 * | . *
890 * | . *
891 * 1/4 ...............................................* * * * * * * * * * * *
892 * | . .
893 * | . .
894 * | . .
895 * 0 +----------------------.-------------------------------.------------->
896 * wb_setpoint^ x_intercept^
897 *
898 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
899 * be smoothly throttled down to normal if it starts high in situations like
900 * - start writing to a slow SD card and a fast disk at the same time. The SD
901 * card's wb_dirty may rush to many times higher than wb_setpoint.
902 * - the wb dirty thresh drops quickly due to change of JBOD workload
903 */
904static void wb_position_ratio(struct dirty_throttle_control *dtc)
905{
906 struct bdi_writeback *wb = dtc->wb;
907 unsigned long write_bw = wb->avg_write_bandwidth;
908 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
909 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
910 unsigned long wb_thresh = dtc->wb_thresh;
911 unsigned long x_intercept;
912 unsigned long setpoint; /* dirty pages' target balance point */
913 unsigned long wb_setpoint;
914 unsigned long span;
915 long long pos_ratio; /* for scaling up/down the rate limit */
916 long x;
917
918 dtc->pos_ratio = 0;
919
920 if (unlikely(dtc->dirty >= limit))
921 return;
922
923 /*
924 * global setpoint
925 *
926 * See comment for pos_ratio_polynom().
927 */
928 setpoint = (freerun + limit) / 2;
929 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
930
931 /*
932 * The strictlimit feature is a tool preventing mistrusted filesystems
933 * from growing a large number of dirty pages before throttling. For
934 * such filesystems balance_dirty_pages always checks wb counters
935 * against wb limits. Even if global "nr_dirty" is under "freerun".
936 * This is especially important for fuse which sets bdi->max_ratio to
937 * 1% by default. Without strictlimit feature, fuse writeback may
938 * consume arbitrary amount of RAM because it is accounted in
939 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
940 *
941 * Here, in wb_position_ratio(), we calculate pos_ratio based on
942 * two values: wb_dirty and wb_thresh. Let's consider an example:
943 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
944 * limits are set by default to 10% and 20% (background and throttle).
945 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
946 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
947 * about ~6K pages (as the average of background and throttle wb
948 * limits). The 3rd order polynomial will provide positive feedback if
949 * wb_dirty is under wb_setpoint and vice versa.
950 *
951 * Note, that we cannot use global counters in these calculations
952 * because we want to throttle process writing to a strictlimit wb
953 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
954 * in the example above).
955 */
956 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
957 long long wb_pos_ratio;
958
959 if (dtc->wb_dirty < 8) {
960 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
961 2 << RATELIMIT_CALC_SHIFT);
962 return;
963 }
964
965 if (dtc->wb_dirty >= wb_thresh)
966 return;
967
968 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
969 dtc->wb_bg_thresh);
970
971 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
972 return;
973
974 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
975 wb_thresh);
976
977 /*
978 * Typically, for strictlimit case, wb_setpoint << setpoint
979 * and pos_ratio >> wb_pos_ratio. In the other words global
980 * state ("dirty") is not limiting factor and we have to
981 * make decision based on wb counters. But there is an
982 * important case when global pos_ratio should get precedence:
983 * global limits are exceeded (e.g. due to activities on other
984 * wb's) while given strictlimit wb is below limit.
985 *
986 * "pos_ratio * wb_pos_ratio" would work for the case above,
987 * but it would look too non-natural for the case of all
988 * activity in the system coming from a single strictlimit wb
989 * with bdi->max_ratio == 100%.
990 *
991 * Note that min() below somewhat changes the dynamics of the
992 * control system. Normally, pos_ratio value can be well over 3
993 * (when globally we are at freerun and wb is well below wb
994 * setpoint). Now the maximum pos_ratio in the same situation
995 * is 2. We might want to tweak this if we observe the control
996 * system is too slow to adapt.
997 */
998 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
999 return;
1000 }
1001
1002 /*
1003 * We have computed basic pos_ratio above based on global situation. If
1004 * the wb is over/under its share of dirty pages, we want to scale
1005 * pos_ratio further down/up. That is done by the following mechanism.
1006 */
1007
1008 /*
1009 * wb setpoint
1010 *
1011 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1012 *
1013 * x_intercept - wb_dirty
1014 * := --------------------------
1015 * x_intercept - wb_setpoint
1016 *
1017 * The main wb control line is a linear function that subjects to
1018 *
1019 * (1) f(wb_setpoint) = 1.0
1020 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1021 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1022 *
1023 * For single wb case, the dirty pages are observed to fluctuate
1024 * regularly within range
1025 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1026 * for various filesystems, where (2) can yield in a reasonable 12.5%
1027 * fluctuation range for pos_ratio.
1028 *
1029 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1030 * own size, so move the slope over accordingly and choose a slope that
1031 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1032 */
1033 if (unlikely(wb_thresh > dtc->thresh))
1034 wb_thresh = dtc->thresh;
1035 /*
1036 * It's very possible that wb_thresh is close to 0 not because the
1037 * device is slow, but that it has remained inactive for long time.
1038 * Honour such devices a reasonable good (hopefully IO efficient)
1039 * threshold, so that the occasional writes won't be blocked and active
1040 * writes can rampup the threshold quickly.
1041 */
1042 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1043 /*
1044 * scale global setpoint to wb's:
1045 * wb_setpoint = setpoint * wb_thresh / thresh
1046 */
1047 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1048 wb_setpoint = setpoint * (u64)x >> 16;
1049 /*
1050 * Use span=(8*write_bw) in single wb case as indicated by
1051 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1052 *
1053 * wb_thresh thresh - wb_thresh
1054 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1055 * thresh thresh
1056 */
1057 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1058 x_intercept = wb_setpoint + span;
1059
1060 if (dtc->wb_dirty < x_intercept - span / 4) {
1061 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1062 (x_intercept - wb_setpoint) | 1);
1063 } else
1064 pos_ratio /= 4;
1065
1066 /*
1067 * wb reserve area, safeguard against dirty pool underrun and disk idle
1068 * It may push the desired control point of global dirty pages higher
1069 * than setpoint.
1070 */
1071 x_intercept = wb_thresh / 2;
1072 if (dtc->wb_dirty < x_intercept) {
1073 if (dtc->wb_dirty > x_intercept / 8)
1074 pos_ratio = div_u64(pos_ratio * x_intercept,
1075 dtc->wb_dirty);
1076 else
1077 pos_ratio *= 8;
1078 }
1079
1080 dtc->pos_ratio = pos_ratio;
1081}
1082
1083static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1084 unsigned long elapsed,
1085 unsigned long written)
1086{
1087 const unsigned long period = roundup_pow_of_two(3 * HZ);
1088 unsigned long avg = wb->avg_write_bandwidth;
1089 unsigned long old = wb->write_bandwidth;
1090 u64 bw;
1091
1092 /*
1093 * bw = written * HZ / elapsed
1094 *
1095 * bw * elapsed + write_bandwidth * (period - elapsed)
1096 * write_bandwidth = ---------------------------------------------------
1097 * period
1098 *
1099 * @written may have decreased due to account_page_redirty().
1100 * Avoid underflowing @bw calculation.
1101 */
1102 bw = written - min(written, wb->written_stamp);
1103 bw *= HZ;
1104 if (unlikely(elapsed > period)) {
1105 do_div(bw, elapsed);
1106 avg = bw;
1107 goto out;
1108 }
1109 bw += (u64)wb->write_bandwidth * (period - elapsed);
1110 bw >>= ilog2(period);
1111
1112 /*
1113 * one more level of smoothing, for filtering out sudden spikes
1114 */
1115 if (avg > old && old >= (unsigned long)bw)
1116 avg -= (avg - old) >> 3;
1117
1118 if (avg < old && old <= (unsigned long)bw)
1119 avg += (old - avg) >> 3;
1120
1121out:
1122 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1123 avg = max(avg, 1LU);
1124 if (wb_has_dirty_io(wb)) {
1125 long delta = avg - wb->avg_write_bandwidth;
1126 WARN_ON_ONCE(atomic_long_add_return(delta,
1127 &wb->bdi->tot_write_bandwidth) <= 0);
1128 }
1129 wb->write_bandwidth = bw;
1130 wb->avg_write_bandwidth = avg;
1131}
1132
1133static void update_dirty_limit(struct dirty_throttle_control *dtc)
1134{
1135 struct wb_domain *dom = dtc_dom(dtc);
1136 unsigned long thresh = dtc->thresh;
1137 unsigned long limit = dom->dirty_limit;
1138
1139 /*
1140 * Follow up in one step.
1141 */
1142 if (limit < thresh) {
1143 limit = thresh;
1144 goto update;
1145 }
1146
1147 /*
1148 * Follow down slowly. Use the higher one as the target, because thresh
1149 * may drop below dirty. This is exactly the reason to introduce
1150 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1151 */
1152 thresh = max(thresh, dtc->dirty);
1153 if (limit > thresh) {
1154 limit -= (limit - thresh) >> 5;
1155 goto update;
1156 }
1157 return;
1158update:
1159 dom->dirty_limit = limit;
1160}
1161
1162static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1163 unsigned long now)
1164{
1165 struct wb_domain *dom = dtc_dom(dtc);
1166
1167 /*
1168 * check locklessly first to optimize away locking for the most time
1169 */
1170 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1171 return;
1172
1173 spin_lock(&dom->lock);
1174 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1175 update_dirty_limit(dtc);
1176 dom->dirty_limit_tstamp = now;
1177 }
1178 spin_unlock(&dom->lock);
1179}
1180
1181/*
1182 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1183 *
1184 * Normal wb tasks will be curbed at or below it in long term.
1185 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1186 */
1187static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1188 unsigned long dirtied,
1189 unsigned long elapsed)
1190{
1191 struct bdi_writeback *wb = dtc->wb;
1192 unsigned long dirty = dtc->dirty;
1193 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1194 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1195 unsigned long setpoint = (freerun + limit) / 2;
1196 unsigned long write_bw = wb->avg_write_bandwidth;
1197 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1198 unsigned long dirty_rate;
1199 unsigned long task_ratelimit;
1200 unsigned long balanced_dirty_ratelimit;
1201 unsigned long step;
1202 unsigned long x;
1203 unsigned long shift;
1204
1205 /*
1206 * The dirty rate will match the writeout rate in long term, except
1207 * when dirty pages are truncated by userspace or re-dirtied by FS.
1208 */
1209 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1210
1211 /*
1212 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1213 */
1214 task_ratelimit = (u64)dirty_ratelimit *
1215 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1216 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1217
1218 /*
1219 * A linear estimation of the "balanced" throttle rate. The theory is,
1220 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1221 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1222 * formula will yield the balanced rate limit (write_bw / N).
1223 *
1224 * Note that the expanded form is not a pure rate feedback:
1225 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1226 * but also takes pos_ratio into account:
1227 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1228 *
1229 * (1) is not realistic because pos_ratio also takes part in balancing
1230 * the dirty rate. Consider the state
1231 * pos_ratio = 0.5 (3)
1232 * rate = 2 * (write_bw / N) (4)
1233 * If (1) is used, it will stuck in that state! Because each dd will
1234 * be throttled at
1235 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1236 * yielding
1237 * dirty_rate = N * task_ratelimit = write_bw (6)
1238 * put (6) into (1) we get
1239 * rate_(i+1) = rate_(i) (7)
1240 *
1241 * So we end up using (2) to always keep
1242 * rate_(i+1) ~= (write_bw / N) (8)
1243 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1244 * pos_ratio is able to drive itself to 1.0, which is not only where
1245 * the dirty count meet the setpoint, but also where the slope of
1246 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1247 */
1248 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1249 dirty_rate | 1);
1250 /*
1251 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1252 */
1253 if (unlikely(balanced_dirty_ratelimit > write_bw))
1254 balanced_dirty_ratelimit = write_bw;
1255
1256 /*
1257 * We could safely do this and return immediately:
1258 *
1259 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1260 *
1261 * However to get a more stable dirty_ratelimit, the below elaborated
1262 * code makes use of task_ratelimit to filter out singular points and
1263 * limit the step size.
1264 *
1265 * The below code essentially only uses the relative value of
1266 *
1267 * task_ratelimit - dirty_ratelimit
1268 * = (pos_ratio - 1) * dirty_ratelimit
1269 *
1270 * which reflects the direction and size of dirty position error.
1271 */
1272
1273 /*
1274 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1275 * task_ratelimit is on the same side of dirty_ratelimit, too.
1276 * For example, when
1277 * - dirty_ratelimit > balanced_dirty_ratelimit
1278 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1279 * lowering dirty_ratelimit will help meet both the position and rate
1280 * control targets. Otherwise, don't update dirty_ratelimit if it will
1281 * only help meet the rate target. After all, what the users ultimately
1282 * feel and care are stable dirty rate and small position error.
1283 *
1284 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1285 * and filter out the singular points of balanced_dirty_ratelimit. Which
1286 * keeps jumping around randomly and can even leap far away at times
1287 * due to the small 200ms estimation period of dirty_rate (we want to
1288 * keep that period small to reduce time lags).
1289 */
1290 step = 0;
1291
1292 /*
1293 * For strictlimit case, calculations above were based on wb counters
1294 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1295 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1296 * Hence, to calculate "step" properly, we have to use wb_dirty as
1297 * "dirty" and wb_setpoint as "setpoint".
1298 *
1299 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1300 * it's possible that wb_thresh is close to zero due to inactivity
1301 * of backing device.
1302 */
1303 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1304 dirty = dtc->wb_dirty;
1305 if (dtc->wb_dirty < 8)
1306 setpoint = dtc->wb_dirty + 1;
1307 else
1308 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1309 }
1310
1311 if (dirty < setpoint) {
1312 x = min3(wb->balanced_dirty_ratelimit,
1313 balanced_dirty_ratelimit, task_ratelimit);
1314 if (dirty_ratelimit < x)
1315 step = x - dirty_ratelimit;
1316 } else {
1317 x = max3(wb->balanced_dirty_ratelimit,
1318 balanced_dirty_ratelimit, task_ratelimit);
1319 if (dirty_ratelimit > x)
1320 step = dirty_ratelimit - x;
1321 }
1322
1323 /*
1324 * Don't pursue 100% rate matching. It's impossible since the balanced
1325 * rate itself is constantly fluctuating. So decrease the track speed
1326 * when it gets close to the target. Helps eliminate pointless tremors.
1327 */
1328 shift = dirty_ratelimit / (2 * step + 1);
1329 if (shift < BITS_PER_LONG)
1330 step = DIV_ROUND_UP(step >> shift, 8);
1331 else
1332 step = 0;
1333
1334 if (dirty_ratelimit < balanced_dirty_ratelimit)
1335 dirty_ratelimit += step;
1336 else
1337 dirty_ratelimit -= step;
1338
1339 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1340 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1341
1342 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1343}
1344
1345static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1346 struct dirty_throttle_control *mdtc,
1347 unsigned long start_time,
1348 bool update_ratelimit)
1349{
1350 struct bdi_writeback *wb = gdtc->wb;
1351 unsigned long now = jiffies;
1352 unsigned long elapsed = now - wb->bw_time_stamp;
1353 unsigned long dirtied;
1354 unsigned long written;
1355
1356 lockdep_assert_held(&wb->list_lock);
1357
1358 /*
1359 * rate-limit, only update once every 200ms.
1360 */
1361 if (elapsed < BANDWIDTH_INTERVAL)
1362 return;
1363
1364 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1365 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1366
1367 /*
1368 * Skip quiet periods when disk bandwidth is under-utilized.
1369 * (at least 1s idle time between two flusher runs)
1370 */
1371 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1372 goto snapshot;
1373
1374 if (update_ratelimit) {
1375 domain_update_bandwidth(gdtc, now);
1376 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1377
1378 /*
1379 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1380 * compiler has no way to figure that out. Help it.
1381 */
1382 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1383 domain_update_bandwidth(mdtc, now);
1384 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1385 }
1386 }
1387 wb_update_write_bandwidth(wb, elapsed, written);
1388
1389snapshot:
1390 wb->dirtied_stamp = dirtied;
1391 wb->written_stamp = written;
1392 wb->bw_time_stamp = now;
1393}
1394
1395void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1396{
1397 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1398
1399 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1400}
1401
1402/*
1403 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1404 * will look to see if it needs to start dirty throttling.
1405 *
1406 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1407 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1408 * (the number of pages we may dirty without exceeding the dirty limits).
1409 */
1410static unsigned long dirty_poll_interval(unsigned long dirty,
1411 unsigned long thresh)
1412{
1413 if (thresh > dirty)
1414 return 1UL << (ilog2(thresh - dirty) >> 1);
1415
1416 return 1;
1417}
1418
1419static unsigned long wb_max_pause(struct bdi_writeback *wb,
1420 unsigned long wb_dirty)
1421{
1422 unsigned long bw = wb->avg_write_bandwidth;
1423 unsigned long t;
1424
1425 /*
1426 * Limit pause time for small memory systems. If sleeping for too long
1427 * time, a small pool of dirty/writeback pages may go empty and disk go
1428 * idle.
1429 *
1430 * 8 serves as the safety ratio.
1431 */
1432 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1433 t++;
1434
1435 return min_t(unsigned long, t, MAX_PAUSE);
1436}
1437
1438static long wb_min_pause(struct bdi_writeback *wb,
1439 long max_pause,
1440 unsigned long task_ratelimit,
1441 unsigned long dirty_ratelimit,
1442 int *nr_dirtied_pause)
1443{
1444 long hi = ilog2(wb->avg_write_bandwidth);
1445 long lo = ilog2(wb->dirty_ratelimit);
1446 long t; /* target pause */
1447 long pause; /* estimated next pause */
1448 int pages; /* target nr_dirtied_pause */
1449
1450 /* target for 10ms pause on 1-dd case */
1451 t = max(1, HZ / 100);
1452
1453 /*
1454 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1455 * overheads.
1456 *
1457 * (N * 10ms) on 2^N concurrent tasks.
1458 */
1459 if (hi > lo)
1460 t += (hi - lo) * (10 * HZ) / 1024;
1461
1462 /*
1463 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1464 * on the much more stable dirty_ratelimit. However the next pause time
1465 * will be computed based on task_ratelimit and the two rate limits may
1466 * depart considerably at some time. Especially if task_ratelimit goes
1467 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1468 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1469 * result task_ratelimit won't be executed faithfully, which could
1470 * eventually bring down dirty_ratelimit.
1471 *
1472 * We apply two rules to fix it up:
1473 * 1) try to estimate the next pause time and if necessary, use a lower
1474 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1475 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1476 * 2) limit the target pause time to max_pause/2, so that the normal
1477 * small fluctuations of task_ratelimit won't trigger rule (1) and
1478 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1479 */
1480 t = min(t, 1 + max_pause / 2);
1481 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482
1483 /*
1484 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1485 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1486 * When the 16 consecutive reads are often interrupted by some dirty
1487 * throttling pause during the async writes, cfq will go into idles
1488 * (deadline is fine). So push nr_dirtied_pause as high as possible
1489 * until reaches DIRTY_POLL_THRESH=32 pages.
1490 */
1491 if (pages < DIRTY_POLL_THRESH) {
1492 t = max_pause;
1493 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1494 if (pages > DIRTY_POLL_THRESH) {
1495 pages = DIRTY_POLL_THRESH;
1496 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1497 }
1498 }
1499
1500 pause = HZ * pages / (task_ratelimit + 1);
1501 if (pause > max_pause) {
1502 t = max_pause;
1503 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1504 }
1505
1506 *nr_dirtied_pause = pages;
1507 /*
1508 * The minimal pause time will normally be half the target pause time.
1509 */
1510 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1511}
1512
1513static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1514{
1515 struct bdi_writeback *wb = dtc->wb;
1516 unsigned long wb_reclaimable;
1517
1518 /*
1519 * wb_thresh is not treated as some limiting factor as
1520 * dirty_thresh, due to reasons
1521 * - in JBOD setup, wb_thresh can fluctuate a lot
1522 * - in a system with HDD and USB key, the USB key may somehow
1523 * go into state (wb_dirty >> wb_thresh) either because
1524 * wb_dirty starts high, or because wb_thresh drops low.
1525 * In this case we don't want to hard throttle the USB key
1526 * dirtiers for 100 seconds until wb_dirty drops under
1527 * wb_thresh. Instead the auxiliary wb control line in
1528 * wb_position_ratio() will let the dirtier task progress
1529 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1530 */
1531 dtc->wb_thresh = __wb_calc_thresh(dtc);
1532 dtc->wb_bg_thresh = dtc->thresh ?
1533 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1534
1535 /*
1536 * In order to avoid the stacked BDI deadlock we need
1537 * to ensure we accurately count the 'dirty' pages when
1538 * the threshold is low.
1539 *
1540 * Otherwise it would be possible to get thresh+n pages
1541 * reported dirty, even though there are thresh-m pages
1542 * actually dirty; with m+n sitting in the percpu
1543 * deltas.
1544 */
1545 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1546 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1547 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1548 } else {
1549 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1550 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1551 }
1552}
1553
1554/*
1555 * balance_dirty_pages() must be called by processes which are generating dirty
1556 * data. It looks at the number of dirty pages in the machine and will force
1557 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1558 * If we're over `background_thresh' then the writeback threads are woken to
1559 * perform some writeout.
1560 */
1561static void balance_dirty_pages(struct bdi_writeback *wb,
1562 unsigned long pages_dirtied)
1563{
1564 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1565 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1566 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1567 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1568 &mdtc_stor : NULL;
1569 struct dirty_throttle_control *sdtc;
1570 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1571 long period;
1572 long pause;
1573 long max_pause;
1574 long min_pause;
1575 int nr_dirtied_pause;
1576 bool dirty_exceeded = false;
1577 unsigned long task_ratelimit;
1578 unsigned long dirty_ratelimit;
1579 struct backing_dev_info *bdi = wb->bdi;
1580 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1581 unsigned long start_time = jiffies;
1582
1583 for (;;) {
1584 unsigned long now = jiffies;
1585 unsigned long dirty, thresh, bg_thresh;
1586 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1587 unsigned long m_thresh = 0;
1588 unsigned long m_bg_thresh = 0;
1589
1590 /*
1591 * Unstable writes are a feature of certain networked
1592 * filesystems (i.e. NFS) in which data may have been
1593 * written to the server's write cache, but has not yet
1594 * been flushed to permanent storage.
1595 */
1596 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1597 global_node_page_state(NR_UNSTABLE_NFS);
1598 gdtc->avail = global_dirtyable_memory();
1599 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1600
1601 domain_dirty_limits(gdtc);
1602
1603 if (unlikely(strictlimit)) {
1604 wb_dirty_limits(gdtc);
1605
1606 dirty = gdtc->wb_dirty;
1607 thresh = gdtc->wb_thresh;
1608 bg_thresh = gdtc->wb_bg_thresh;
1609 } else {
1610 dirty = gdtc->dirty;
1611 thresh = gdtc->thresh;
1612 bg_thresh = gdtc->bg_thresh;
1613 }
1614
1615 if (mdtc) {
1616 unsigned long filepages, headroom, writeback;
1617
1618 /*
1619 * If @wb belongs to !root memcg, repeat the same
1620 * basic calculations for the memcg domain.
1621 */
1622 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1623 &mdtc->dirty, &writeback);
1624 mdtc->dirty += writeback;
1625 mdtc_calc_avail(mdtc, filepages, headroom);
1626
1627 domain_dirty_limits(mdtc);
1628
1629 if (unlikely(strictlimit)) {
1630 wb_dirty_limits(mdtc);
1631 m_dirty = mdtc->wb_dirty;
1632 m_thresh = mdtc->wb_thresh;
1633 m_bg_thresh = mdtc->wb_bg_thresh;
1634 } else {
1635 m_dirty = mdtc->dirty;
1636 m_thresh = mdtc->thresh;
1637 m_bg_thresh = mdtc->bg_thresh;
1638 }
1639 }
1640
1641 /*
1642 * Throttle it only when the background writeback cannot
1643 * catch-up. This avoids (excessively) small writeouts
1644 * when the wb limits are ramping up in case of !strictlimit.
1645 *
1646 * In strictlimit case make decision based on the wb counters
1647 * and limits. Small writeouts when the wb limits are ramping
1648 * up are the price we consciously pay for strictlimit-ing.
1649 *
1650 * If memcg domain is in effect, @dirty should be under
1651 * both global and memcg freerun ceilings.
1652 */
1653 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1654 (!mdtc ||
1655 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1656 unsigned long intv = dirty_poll_interval(dirty, thresh);
1657 unsigned long m_intv = ULONG_MAX;
1658
1659 current->dirty_paused_when = now;
1660 current->nr_dirtied = 0;
1661 if (mdtc)
1662 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1663 current->nr_dirtied_pause = min(intv, m_intv);
1664 break;
1665 }
1666
1667 if (unlikely(!writeback_in_progress(wb)))
1668 wb_start_background_writeback(wb);
1669
1670 /*
1671 * Calculate global domain's pos_ratio and select the
1672 * global dtc by default.
1673 */
1674 if (!strictlimit)
1675 wb_dirty_limits(gdtc);
1676
1677 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1678 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1679
1680 wb_position_ratio(gdtc);
1681 sdtc = gdtc;
1682
1683 if (mdtc) {
1684 /*
1685 * If memcg domain is in effect, calculate its
1686 * pos_ratio. @wb should satisfy constraints from
1687 * both global and memcg domains. Choose the one
1688 * w/ lower pos_ratio.
1689 */
1690 if (!strictlimit)
1691 wb_dirty_limits(mdtc);
1692
1693 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1694 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1695
1696 wb_position_ratio(mdtc);
1697 if (mdtc->pos_ratio < gdtc->pos_ratio)
1698 sdtc = mdtc;
1699 }
1700
1701 if (dirty_exceeded && !wb->dirty_exceeded)
1702 wb->dirty_exceeded = 1;
1703
1704 if (time_is_before_jiffies(wb->bw_time_stamp +
1705 BANDWIDTH_INTERVAL)) {
1706 spin_lock(&wb->list_lock);
1707 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1708 spin_unlock(&wb->list_lock);
1709 }
1710
1711 /* throttle according to the chosen dtc */
1712 dirty_ratelimit = wb->dirty_ratelimit;
1713 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1714 RATELIMIT_CALC_SHIFT;
1715 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1716 min_pause = wb_min_pause(wb, max_pause,
1717 task_ratelimit, dirty_ratelimit,
1718 &nr_dirtied_pause);
1719
1720 if (unlikely(task_ratelimit == 0)) {
1721 period = max_pause;
1722 pause = max_pause;
1723 goto pause;
1724 }
1725 period = HZ * pages_dirtied / task_ratelimit;
1726 pause = period;
1727 if (current->dirty_paused_when)
1728 pause -= now - current->dirty_paused_when;
1729 /*
1730 * For less than 1s think time (ext3/4 may block the dirtier
1731 * for up to 800ms from time to time on 1-HDD; so does xfs,
1732 * however at much less frequency), try to compensate it in
1733 * future periods by updating the virtual time; otherwise just
1734 * do a reset, as it may be a light dirtier.
1735 */
1736 if (pause < min_pause) {
1737 trace_balance_dirty_pages(wb,
1738 sdtc->thresh,
1739 sdtc->bg_thresh,
1740 sdtc->dirty,
1741 sdtc->wb_thresh,
1742 sdtc->wb_dirty,
1743 dirty_ratelimit,
1744 task_ratelimit,
1745 pages_dirtied,
1746 period,
1747 min(pause, 0L),
1748 start_time);
1749 if (pause < -HZ) {
1750 current->dirty_paused_when = now;
1751 current->nr_dirtied = 0;
1752 } else if (period) {
1753 current->dirty_paused_when += period;
1754 current->nr_dirtied = 0;
1755 } else if (current->nr_dirtied_pause <= pages_dirtied)
1756 current->nr_dirtied_pause += pages_dirtied;
1757 break;
1758 }
1759 if (unlikely(pause > max_pause)) {
1760 /* for occasional dropped task_ratelimit */
1761 now += min(pause - max_pause, max_pause);
1762 pause = max_pause;
1763 }
1764
1765pause:
1766 trace_balance_dirty_pages(wb,
1767 sdtc->thresh,
1768 sdtc->bg_thresh,
1769 sdtc->dirty,
1770 sdtc->wb_thresh,
1771 sdtc->wb_dirty,
1772 dirty_ratelimit,
1773 task_ratelimit,
1774 pages_dirtied,
1775 period,
1776 pause,
1777 start_time);
1778 __set_current_state(TASK_KILLABLE);
1779 wb->dirty_sleep = now;
1780 io_schedule_timeout(pause);
1781
1782 current->dirty_paused_when = now + pause;
1783 current->nr_dirtied = 0;
1784 current->nr_dirtied_pause = nr_dirtied_pause;
1785
1786 /*
1787 * This is typically equal to (dirty < thresh) and can also
1788 * keep "1000+ dd on a slow USB stick" under control.
1789 */
1790 if (task_ratelimit)
1791 break;
1792
1793 /*
1794 * In the case of an unresponding NFS server and the NFS dirty
1795 * pages exceeds dirty_thresh, give the other good wb's a pipe
1796 * to go through, so that tasks on them still remain responsive.
1797 *
1798 * In theory 1 page is enough to keep the consumer-producer
1799 * pipe going: the flusher cleans 1 page => the task dirties 1
1800 * more page. However wb_dirty has accounting errors. So use
1801 * the larger and more IO friendly wb_stat_error.
1802 */
1803 if (sdtc->wb_dirty <= wb_stat_error())
1804 break;
1805
1806 if (fatal_signal_pending(current))
1807 break;
1808 }
1809
1810 if (!dirty_exceeded && wb->dirty_exceeded)
1811 wb->dirty_exceeded = 0;
1812
1813 if (writeback_in_progress(wb))
1814 return;
1815
1816 /*
1817 * In laptop mode, we wait until hitting the higher threshold before
1818 * starting background writeout, and then write out all the way down
1819 * to the lower threshold. So slow writers cause minimal disk activity.
1820 *
1821 * In normal mode, we start background writeout at the lower
1822 * background_thresh, to keep the amount of dirty memory low.
1823 */
1824 if (laptop_mode)
1825 return;
1826
1827 if (nr_reclaimable > gdtc->bg_thresh)
1828 wb_start_background_writeback(wb);
1829}
1830
1831static DEFINE_PER_CPU(int, bdp_ratelimits);
1832
1833/*
1834 * Normal tasks are throttled by
1835 * loop {
1836 * dirty tsk->nr_dirtied_pause pages;
1837 * take a snap in balance_dirty_pages();
1838 * }
1839 * However there is a worst case. If every task exit immediately when dirtied
1840 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1841 * called to throttle the page dirties. The solution is to save the not yet
1842 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1843 * randomly into the running tasks. This works well for the above worst case,
1844 * as the new task will pick up and accumulate the old task's leaked dirty
1845 * count and eventually get throttled.
1846 */
1847DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1848
1849/**
1850 * balance_dirty_pages_ratelimited - balance dirty memory state
1851 * @mapping: address_space which was dirtied
1852 *
1853 * Processes which are dirtying memory should call in here once for each page
1854 * which was newly dirtied. The function will periodically check the system's
1855 * dirty state and will initiate writeback if needed.
1856 *
1857 * On really big machines, get_writeback_state is expensive, so try to avoid
1858 * calling it too often (ratelimiting). But once we're over the dirty memory
1859 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1860 * from overshooting the limit by (ratelimit_pages) each.
1861 */
1862void balance_dirty_pages_ratelimited(struct address_space *mapping)
1863{
1864 struct inode *inode = mapping->host;
1865 struct backing_dev_info *bdi = inode_to_bdi(inode);
1866 struct bdi_writeback *wb = NULL;
1867 int ratelimit;
1868 int *p;
1869
1870 if (!bdi_cap_account_dirty(bdi))
1871 return;
1872
1873 if (inode_cgwb_enabled(inode))
1874 wb = wb_get_create_current(bdi, GFP_KERNEL);
1875 if (!wb)
1876 wb = &bdi->wb;
1877
1878 ratelimit = current->nr_dirtied_pause;
1879 if (wb->dirty_exceeded)
1880 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1881
1882 preempt_disable();
1883 /*
1884 * This prevents one CPU to accumulate too many dirtied pages without
1885 * calling into balance_dirty_pages(), which can happen when there are
1886 * 1000+ tasks, all of them start dirtying pages at exactly the same
1887 * time, hence all honoured too large initial task->nr_dirtied_pause.
1888 */
1889 p = this_cpu_ptr(&bdp_ratelimits);
1890 if (unlikely(current->nr_dirtied >= ratelimit))
1891 *p = 0;
1892 else if (unlikely(*p >= ratelimit_pages)) {
1893 *p = 0;
1894 ratelimit = 0;
1895 }
1896 /*
1897 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1898 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1899 * the dirty throttling and livelock other long-run dirtiers.
1900 */
1901 p = this_cpu_ptr(&dirty_throttle_leaks);
1902 if (*p > 0 && current->nr_dirtied < ratelimit) {
1903 unsigned long nr_pages_dirtied;
1904 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1905 *p -= nr_pages_dirtied;
1906 current->nr_dirtied += nr_pages_dirtied;
1907 }
1908 preempt_enable();
1909
1910 if (unlikely(current->nr_dirtied >= ratelimit))
1911 balance_dirty_pages(wb, current->nr_dirtied);
1912
1913 wb_put(wb);
1914}
1915EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1916
1917/**
1918 * wb_over_bg_thresh - does @wb need to be written back?
1919 * @wb: bdi_writeback of interest
1920 *
1921 * Determines whether background writeback should keep writing @wb or it's
1922 * clean enough. Returns %true if writeback should continue.
1923 */
1924bool wb_over_bg_thresh(struct bdi_writeback *wb)
1925{
1926 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1927 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1928 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1929 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1930 &mdtc_stor : NULL;
1931
1932 /*
1933 * Similar to balance_dirty_pages() but ignores pages being written
1934 * as we're trying to decide whether to put more under writeback.
1935 */
1936 gdtc->avail = global_dirtyable_memory();
1937 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1938 global_node_page_state(NR_UNSTABLE_NFS);
1939 domain_dirty_limits(gdtc);
1940
1941 if (gdtc->dirty > gdtc->bg_thresh)
1942 return true;
1943
1944 if (wb_stat(wb, WB_RECLAIMABLE) >
1945 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1946 return true;
1947
1948 if (mdtc) {
1949 unsigned long filepages, headroom, writeback;
1950
1951 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1952 &writeback);
1953 mdtc_calc_avail(mdtc, filepages, headroom);
1954 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1955
1956 if (mdtc->dirty > mdtc->bg_thresh)
1957 return true;
1958
1959 if (wb_stat(wb, WB_RECLAIMABLE) >
1960 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1961 return true;
1962 }
1963
1964 return false;
1965}
1966
1967/*
1968 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1969 */
1970int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1971 void __user *buffer, size_t *length, loff_t *ppos)
1972{
1973 unsigned int old_interval = dirty_writeback_interval;
1974 int ret;
1975
1976 ret = proc_dointvec(table, write, buffer, length, ppos);
1977
1978 /*
1979 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1980 * and a different non-zero value will wakeup the writeback threads.
1981 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1982 * iterate over all bdis and wbs.
1983 * The reason we do this is to make the change take effect immediately.
1984 */
1985 if (!ret && write && dirty_writeback_interval &&
1986 dirty_writeback_interval != old_interval)
1987 wakeup_flusher_threads(WB_REASON_PERIODIC);
1988
1989 return ret;
1990}
1991
1992#ifdef CONFIG_BLOCK
1993void laptop_mode_timer_fn(struct timer_list *t)
1994{
1995 struct backing_dev_info *backing_dev_info =
1996 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
1997
1998 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
1999}
2000
2001/*
2002 * We've spun up the disk and we're in laptop mode: schedule writeback
2003 * of all dirty data a few seconds from now. If the flush is already scheduled
2004 * then push it back - the user is still using the disk.
2005 */
2006void laptop_io_completion(struct backing_dev_info *info)
2007{
2008 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2009}
2010
2011/*
2012 * We're in laptop mode and we've just synced. The sync's writes will have
2013 * caused another writeback to be scheduled by laptop_io_completion.
2014 * Nothing needs to be written back anymore, so we unschedule the writeback.
2015 */
2016void laptop_sync_completion(void)
2017{
2018 struct backing_dev_info *bdi;
2019
2020 rcu_read_lock();
2021
2022 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2023 del_timer(&bdi->laptop_mode_wb_timer);
2024
2025 rcu_read_unlock();
2026}
2027#endif
2028
2029/*
2030 * If ratelimit_pages is too high then we can get into dirty-data overload
2031 * if a large number of processes all perform writes at the same time.
2032 * If it is too low then SMP machines will call the (expensive)
2033 * get_writeback_state too often.
2034 *
2035 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2036 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2037 * thresholds.
2038 */
2039
2040void writeback_set_ratelimit(void)
2041{
2042 struct wb_domain *dom = &global_wb_domain;
2043 unsigned long background_thresh;
2044 unsigned long dirty_thresh;
2045
2046 global_dirty_limits(&background_thresh, &dirty_thresh);
2047 dom->dirty_limit = dirty_thresh;
2048 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2049 if (ratelimit_pages < 16)
2050 ratelimit_pages = 16;
2051}
2052
2053static int page_writeback_cpu_online(unsigned int cpu)
2054{
2055 writeback_set_ratelimit();
2056 return 0;
2057}
2058
2059/*
2060 * Called early on to tune the page writeback dirty limits.
2061 *
2062 * We used to scale dirty pages according to how total memory
2063 * related to pages that could be allocated for buffers (by
2064 * comparing nr_free_buffer_pages() to vm_total_pages.
2065 *
2066 * However, that was when we used "dirty_ratio" to scale with
2067 * all memory, and we don't do that any more. "dirty_ratio"
2068 * is now applied to total non-HIGHPAGE memory (by subtracting
2069 * totalhigh_pages from vm_total_pages), and as such we can't
2070 * get into the old insane situation any more where we had
2071 * large amounts of dirty pages compared to a small amount of
2072 * non-HIGHMEM memory.
2073 *
2074 * But we might still want to scale the dirty_ratio by how
2075 * much memory the box has..
2076 */
2077void __init page_writeback_init(void)
2078{
2079 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2080
2081 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2082 page_writeback_cpu_online, NULL);
2083 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2084 page_writeback_cpu_online);
2085}
2086
2087/**
2088 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2089 * @mapping: address space structure to write
2090 * @start: starting page index
2091 * @end: ending page index (inclusive)
2092 *
2093 * This function scans the page range from @start to @end (inclusive) and tags
2094 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2095 * that write_cache_pages (or whoever calls this function) will then use
2096 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2097 * used to avoid livelocking of writeback by a process steadily creating new
2098 * dirty pages in the file (thus it is important for this function to be quick
2099 * so that it can tag pages faster than a dirtying process can create them).
2100 */
2101/*
2102 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce the i_pages lock
2103 * latency.
2104 */
2105void tag_pages_for_writeback(struct address_space *mapping,
2106 pgoff_t start, pgoff_t end)
2107{
2108#define WRITEBACK_TAG_BATCH 4096
2109 unsigned long tagged = 0;
2110 struct radix_tree_iter iter;
2111 void **slot;
2112
2113 xa_lock_irq(&mapping->i_pages);
2114 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start,
2115 PAGECACHE_TAG_DIRTY) {
2116 if (iter.index > end)
2117 break;
2118 radix_tree_iter_tag_set(&mapping->i_pages, &iter,
2119 PAGECACHE_TAG_TOWRITE);
2120 tagged++;
2121 if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2122 continue;
2123 slot = radix_tree_iter_resume(slot, &iter);
2124 xa_unlock_irq(&mapping->i_pages);
2125 cond_resched();
2126 xa_lock_irq(&mapping->i_pages);
2127 }
2128 xa_unlock_irq(&mapping->i_pages);
2129}
2130EXPORT_SYMBOL(tag_pages_for_writeback);
2131
2132/**
2133 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2134 * @mapping: address space structure to write
2135 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2136 * @writepage: function called for each page
2137 * @data: data passed to writepage function
2138 *
2139 * If a page is already under I/O, write_cache_pages() skips it, even
2140 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2141 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2142 * and msync() need to guarantee that all the data which was dirty at the time
2143 * the call was made get new I/O started against them. If wbc->sync_mode is
2144 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2145 * existing IO to complete.
2146 *
2147 * To avoid livelocks (when other process dirties new pages), we first tag
2148 * pages which should be written back with TOWRITE tag and only then start
2149 * writing them. For data-integrity sync we have to be careful so that we do
2150 * not miss some pages (e.g., because some other process has cleared TOWRITE
2151 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2152 * by the process clearing the DIRTY tag (and submitting the page for IO).
2153 */
2154int write_cache_pages(struct address_space *mapping,
2155 struct writeback_control *wbc, writepage_t writepage,
2156 void *data)
2157{
2158 int ret = 0;
2159 int done = 0;
2160 struct pagevec pvec;
2161 int nr_pages;
2162 pgoff_t uninitialized_var(writeback_index);
2163 pgoff_t index;
2164 pgoff_t end; /* Inclusive */
2165 pgoff_t done_index;
2166 int cycled;
2167 int range_whole = 0;
2168 int tag;
2169
2170 pagevec_init(&pvec);
2171 if (wbc->range_cyclic) {
2172 writeback_index = mapping->writeback_index; /* prev offset */
2173 index = writeback_index;
2174 if (index == 0)
2175 cycled = 1;
2176 else
2177 cycled = 0;
2178 end = -1;
2179 } else {
2180 index = wbc->range_start >> PAGE_SHIFT;
2181 end = wbc->range_end >> PAGE_SHIFT;
2182 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2183 range_whole = 1;
2184 cycled = 1; /* ignore range_cyclic tests */
2185 }
2186 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2187 tag = PAGECACHE_TAG_TOWRITE;
2188 else
2189 tag = PAGECACHE_TAG_DIRTY;
2190retry:
2191 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2192 tag_pages_for_writeback(mapping, index, end);
2193 done_index = index;
2194 while (!done && (index <= end)) {
2195 int i;
2196
2197 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2198 tag);
2199 if (nr_pages == 0)
2200 break;
2201
2202 for (i = 0; i < nr_pages; i++) {
2203 struct page *page = pvec.pages[i];
2204
2205 done_index = page->index;
2206
2207 lock_page(page);
2208
2209 /*
2210 * Page truncated or invalidated. We can freely skip it
2211 * then, even for data integrity operations: the page
2212 * has disappeared concurrently, so there could be no
2213 * real expectation of this data interity operation
2214 * even if there is now a new, dirty page at the same
2215 * pagecache address.
2216 */
2217 if (unlikely(page->mapping != mapping)) {
2218continue_unlock:
2219 unlock_page(page);
2220 continue;
2221 }
2222
2223 if (!PageDirty(page)) {
2224 /* someone wrote it for us */
2225 goto continue_unlock;
2226 }
2227
2228 if (PageWriteback(page)) {
2229 if (wbc->sync_mode != WB_SYNC_NONE)
2230 wait_on_page_writeback(page);
2231 else
2232 goto continue_unlock;
2233 }
2234
2235 BUG_ON(PageWriteback(page));
2236 if (!clear_page_dirty_for_io(page))
2237 goto continue_unlock;
2238
2239 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2240 ret = (*writepage)(page, wbc, data);
2241 if (unlikely(ret)) {
2242 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2243 unlock_page(page);
2244 ret = 0;
2245 } else {
2246 /*
2247 * done_index is set past this page,
2248 * so media errors will not choke
2249 * background writeout for the entire
2250 * file. This has consequences for
2251 * range_cyclic semantics (ie. it may
2252 * not be suitable for data integrity
2253 * writeout).
2254 */
2255 done_index = page->index + 1;
2256 done = 1;
2257 break;
2258 }
2259 }
2260
2261 /*
2262 * We stop writing back only if we are not doing
2263 * integrity sync. In case of integrity sync we have to
2264 * keep going until we have written all the pages
2265 * we tagged for writeback prior to entering this loop.
2266 */
2267 if (--wbc->nr_to_write <= 0 &&
2268 wbc->sync_mode == WB_SYNC_NONE) {
2269 done = 1;
2270 break;
2271 }
2272 }
2273 pagevec_release(&pvec);
2274 cond_resched();
2275 }
2276 if (!cycled && !done) {
2277 /*
2278 * range_cyclic:
2279 * We hit the last page and there is more work to be done: wrap
2280 * back to the start of the file
2281 */
2282 cycled = 1;
2283 index = 0;
2284 end = writeback_index - 1;
2285 goto retry;
2286 }
2287 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2288 mapping->writeback_index = done_index;
2289
2290 return ret;
2291}
2292EXPORT_SYMBOL(write_cache_pages);
2293
2294/*
2295 * Function used by generic_writepages to call the real writepage
2296 * function and set the mapping flags on error
2297 */
2298static int __writepage(struct page *page, struct writeback_control *wbc,
2299 void *data)
2300{
2301 struct address_space *mapping = data;
2302 int ret = mapping->a_ops->writepage(page, wbc);
2303 mapping_set_error(mapping, ret);
2304 return ret;
2305}
2306
2307/**
2308 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2309 * @mapping: address space structure to write
2310 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2311 *
2312 * This is a library function, which implements the writepages()
2313 * address_space_operation.
2314 */
2315int generic_writepages(struct address_space *mapping,
2316 struct writeback_control *wbc)
2317{
2318 struct blk_plug plug;
2319 int ret;
2320
2321 /* deal with chardevs and other special file */
2322 if (!mapping->a_ops->writepage)
2323 return 0;
2324
2325 blk_start_plug(&plug);
2326 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2327 blk_finish_plug(&plug);
2328 return ret;
2329}
2330
2331EXPORT_SYMBOL(generic_writepages);
2332
2333int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2334{
2335 int ret;
2336
2337 if (wbc->nr_to_write <= 0)
2338 return 0;
2339 while (1) {
2340 if (mapping->a_ops->writepages)
2341 ret = mapping->a_ops->writepages(mapping, wbc);
2342 else
2343 ret = generic_writepages(mapping, wbc);
2344 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2345 break;
2346 cond_resched();
2347 congestion_wait(BLK_RW_ASYNC, HZ/50);
2348 }
2349 return ret;
2350}
2351
2352/**
2353 * write_one_page - write out a single page and wait on I/O
2354 * @page: the page to write
2355 *
2356 * The page must be locked by the caller and will be unlocked upon return.
2357 *
2358 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2359 * function returns.
2360 */
2361int write_one_page(struct page *page)
2362{
2363 struct address_space *mapping = page->mapping;
2364 int ret = 0;
2365 struct writeback_control wbc = {
2366 .sync_mode = WB_SYNC_ALL,
2367 .nr_to_write = 1,
2368 };
2369
2370 BUG_ON(!PageLocked(page));
2371
2372 wait_on_page_writeback(page);
2373
2374 if (clear_page_dirty_for_io(page)) {
2375 get_page(page);
2376 ret = mapping->a_ops->writepage(page, &wbc);
2377 if (ret == 0)
2378 wait_on_page_writeback(page);
2379 put_page(page);
2380 } else {
2381 unlock_page(page);
2382 }
2383
2384 if (!ret)
2385 ret = filemap_check_errors(mapping);
2386 return ret;
2387}
2388EXPORT_SYMBOL(write_one_page);
2389
2390/*
2391 * For address_spaces which do not use buffers nor write back.
2392 */
2393int __set_page_dirty_no_writeback(struct page *page)
2394{
2395 if (!PageDirty(page))
2396 return !TestSetPageDirty(page);
2397 return 0;
2398}
2399
2400/*
2401 * Helper function for set_page_dirty family.
2402 *
2403 * Caller must hold lock_page_memcg().
2404 *
2405 * NOTE: This relies on being atomic wrt interrupts.
2406 */
2407void account_page_dirtied(struct page *page, struct address_space *mapping)
2408{
2409 struct inode *inode = mapping->host;
2410
2411 trace_writeback_dirty_page(page, mapping);
2412
2413 if (mapping_cap_account_dirty(mapping)) {
2414 struct bdi_writeback *wb;
2415
2416 inode_attach_wb(inode, page);
2417 wb = inode_to_wb(inode);
2418
2419 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2420 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2421 __inc_node_page_state(page, NR_DIRTIED);
2422 inc_wb_stat(wb, WB_RECLAIMABLE);
2423 inc_wb_stat(wb, WB_DIRTIED);
2424 task_io_account_write(PAGE_SIZE);
2425 current->nr_dirtied++;
2426 this_cpu_inc(bdp_ratelimits);
2427 }
2428}
2429EXPORT_SYMBOL(account_page_dirtied);
2430
2431/*
2432 * Helper function for deaccounting dirty page without writeback.
2433 *
2434 * Caller must hold lock_page_memcg().
2435 */
2436void account_page_cleaned(struct page *page, struct address_space *mapping,
2437 struct bdi_writeback *wb)
2438{
2439 if (mapping_cap_account_dirty(mapping)) {
2440 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2441 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2442 dec_wb_stat(wb, WB_RECLAIMABLE);
2443 task_io_account_cancelled_write(PAGE_SIZE);
2444 }
2445}
2446
2447/*
2448 * For address_spaces which do not use buffers. Just tag the page as dirty in
2449 * its radix tree.
2450 *
2451 * This is also used when a single buffer is being dirtied: we want to set the
2452 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2453 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454 *
2455 * The caller must ensure this doesn't race with truncation. Most will simply
2456 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457 * the pte lock held, which also locks out truncation.
2458 */
2459int __set_page_dirty_nobuffers(struct page *page)
2460{
2461 lock_page_memcg(page);
2462 if (!TestSetPageDirty(page)) {
2463 struct address_space *mapping = page_mapping(page);
2464 unsigned long flags;
2465
2466 if (!mapping) {
2467 unlock_page_memcg(page);
2468 return 1;
2469 }
2470
2471 xa_lock_irqsave(&mapping->i_pages, flags);
2472 BUG_ON(page_mapping(page) != mapping);
2473 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2474 account_page_dirtied(page, mapping);
2475 radix_tree_tag_set(&mapping->i_pages, page_index(page),
2476 PAGECACHE_TAG_DIRTY);
2477 xa_unlock_irqrestore(&mapping->i_pages, flags);
2478 unlock_page_memcg(page);
2479
2480 if (mapping->host) {
2481 /* !PageAnon && !swapper_space */
2482 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2483 }
2484 return 1;
2485 }
2486 unlock_page_memcg(page);
2487 return 0;
2488}
2489EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2490
2491/*
2492 * Call this whenever redirtying a page, to de-account the dirty counters
2493 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2494 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2495 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2496 * control.
2497 */
2498void account_page_redirty(struct page *page)
2499{
2500 struct address_space *mapping = page->mapping;
2501
2502 if (mapping && mapping_cap_account_dirty(mapping)) {
2503 struct inode *inode = mapping->host;
2504 struct bdi_writeback *wb;
2505 struct wb_lock_cookie cookie = {};
2506
2507 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2508 current->nr_dirtied--;
2509 dec_node_page_state(page, NR_DIRTIED);
2510 dec_wb_stat(wb, WB_DIRTIED);
2511 unlocked_inode_to_wb_end(inode, &cookie);
2512 }
2513}
2514EXPORT_SYMBOL(account_page_redirty);
2515
2516/*
2517 * When a writepage implementation decides that it doesn't want to write this
2518 * page for some reason, it should redirty the locked page via
2519 * redirty_page_for_writepage() and it should then unlock the page and return 0
2520 */
2521int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2522{
2523 int ret;
2524
2525 wbc->pages_skipped++;
2526 ret = __set_page_dirty_nobuffers(page);
2527 account_page_redirty(page);
2528 return ret;
2529}
2530EXPORT_SYMBOL(redirty_page_for_writepage);
2531
2532/*
2533 * Dirty a page.
2534 *
2535 * For pages with a mapping this should be done under the page lock
2536 * for the benefit of asynchronous memory errors who prefer a consistent
2537 * dirty state. This rule can be broken in some special cases,
2538 * but should be better not to.
2539 *
2540 * If the mapping doesn't provide a set_page_dirty a_op, then
2541 * just fall through and assume that it wants buffer_heads.
2542 */
2543int set_page_dirty(struct page *page)
2544{
2545 struct address_space *mapping = page_mapping(page);
2546
2547 page = compound_head(page);
2548 if (likely(mapping)) {
2549 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2550 /*
2551 * readahead/lru_deactivate_page could remain
2552 * PG_readahead/PG_reclaim due to race with end_page_writeback
2553 * About readahead, if the page is written, the flags would be
2554 * reset. So no problem.
2555 * About lru_deactivate_page, if the page is redirty, the flag
2556 * will be reset. So no problem. but if the page is used by readahead
2557 * it will confuse readahead and make it restart the size rampup
2558 * process. But it's a trivial problem.
2559 */
2560 if (PageReclaim(page))
2561 ClearPageReclaim(page);
2562#ifdef CONFIG_BLOCK
2563 if (!spd)
2564 spd = __set_page_dirty_buffers;
2565#endif
2566 return (*spd)(page);
2567 }
2568 if (!PageDirty(page)) {
2569 if (!TestSetPageDirty(page))
2570 return 1;
2571 }
2572 return 0;
2573}
2574EXPORT_SYMBOL(set_page_dirty);
2575
2576/*
2577 * set_page_dirty() is racy if the caller has no reference against
2578 * page->mapping->host, and if the page is unlocked. This is because another
2579 * CPU could truncate the page off the mapping and then free the mapping.
2580 *
2581 * Usually, the page _is_ locked, or the caller is a user-space process which
2582 * holds a reference on the inode by having an open file.
2583 *
2584 * In other cases, the page should be locked before running set_page_dirty().
2585 */
2586int set_page_dirty_lock(struct page *page)
2587{
2588 int ret;
2589
2590 lock_page(page);
2591 ret = set_page_dirty(page);
2592 unlock_page(page);
2593 return ret;
2594}
2595EXPORT_SYMBOL(set_page_dirty_lock);
2596
2597/*
2598 * This cancels just the dirty bit on the kernel page itself, it does NOT
2599 * actually remove dirty bits on any mmap's that may be around. It also
2600 * leaves the page tagged dirty, so any sync activity will still find it on
2601 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2602 * look at the dirty bits in the VM.
2603 *
2604 * Doing this should *normally* only ever be done when a page is truncated,
2605 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2606 * this when it notices that somebody has cleaned out all the buffers on a
2607 * page without actually doing it through the VM. Can you say "ext3 is
2608 * horribly ugly"? Thought you could.
2609 */
2610void __cancel_dirty_page(struct page *page)
2611{
2612 struct address_space *mapping = page_mapping(page);
2613
2614 if (mapping_cap_account_dirty(mapping)) {
2615 struct inode *inode = mapping->host;
2616 struct bdi_writeback *wb;
2617 struct wb_lock_cookie cookie = {};
2618
2619 lock_page_memcg(page);
2620 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2621
2622 if (TestClearPageDirty(page))
2623 account_page_cleaned(page, mapping, wb);
2624
2625 unlocked_inode_to_wb_end(inode, &cookie);
2626 unlock_page_memcg(page);
2627 } else {
2628 ClearPageDirty(page);
2629 }
2630}
2631EXPORT_SYMBOL(__cancel_dirty_page);
2632
2633/*
2634 * Clear a page's dirty flag, while caring for dirty memory accounting.
2635 * Returns true if the page was previously dirty.
2636 *
2637 * This is for preparing to put the page under writeout. We leave the page
2638 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2639 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2640 * implementation will run either set_page_writeback() or set_page_dirty(),
2641 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2642 * back into sync.
2643 *
2644 * This incoherency between the page's dirty flag and radix-tree tag is
2645 * unfortunate, but it only exists while the page is locked.
2646 */
2647int clear_page_dirty_for_io(struct page *page)
2648{
2649 struct address_space *mapping = page_mapping(page);
2650 int ret = 0;
2651
2652 BUG_ON(!PageLocked(page));
2653
2654 if (mapping && mapping_cap_account_dirty(mapping)) {
2655 struct inode *inode = mapping->host;
2656 struct bdi_writeback *wb;
2657 struct wb_lock_cookie cookie = {};
2658
2659 /*
2660 * Yes, Virginia, this is indeed insane.
2661 *
2662 * We use this sequence to make sure that
2663 * (a) we account for dirty stats properly
2664 * (b) we tell the low-level filesystem to
2665 * mark the whole page dirty if it was
2666 * dirty in a pagetable. Only to then
2667 * (c) clean the page again and return 1 to
2668 * cause the writeback.
2669 *
2670 * This way we avoid all nasty races with the
2671 * dirty bit in multiple places and clearing
2672 * them concurrently from different threads.
2673 *
2674 * Note! Normally the "set_page_dirty(page)"
2675 * has no effect on the actual dirty bit - since
2676 * that will already usually be set. But we
2677 * need the side effects, and it can help us
2678 * avoid races.
2679 *
2680 * We basically use the page "master dirty bit"
2681 * as a serialization point for all the different
2682 * threads doing their things.
2683 */
2684 if (page_mkclean(page))
2685 set_page_dirty(page);
2686 /*
2687 * We carefully synchronise fault handlers against
2688 * installing a dirty pte and marking the page dirty
2689 * at this point. We do this by having them hold the
2690 * page lock while dirtying the page, and pages are
2691 * always locked coming in here, so we get the desired
2692 * exclusion.
2693 */
2694 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2695 if (TestClearPageDirty(page)) {
2696 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2697 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2698 dec_wb_stat(wb, WB_RECLAIMABLE);
2699 ret = 1;
2700 }
2701 unlocked_inode_to_wb_end(inode, &cookie);
2702 return ret;
2703 }
2704 return TestClearPageDirty(page);
2705}
2706EXPORT_SYMBOL(clear_page_dirty_for_io);
2707
2708int test_clear_page_writeback(struct page *page)
2709{
2710 struct address_space *mapping = page_mapping(page);
2711 struct mem_cgroup *memcg;
2712 struct lruvec *lruvec;
2713 int ret;
2714
2715 memcg = lock_page_memcg(page);
2716 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2717 if (mapping && mapping_use_writeback_tags(mapping)) {
2718 struct inode *inode = mapping->host;
2719 struct backing_dev_info *bdi = inode_to_bdi(inode);
2720 unsigned long flags;
2721
2722 xa_lock_irqsave(&mapping->i_pages, flags);
2723 ret = TestClearPageWriteback(page);
2724 if (ret) {
2725 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2726 PAGECACHE_TAG_WRITEBACK);
2727 if (bdi_cap_account_writeback(bdi)) {
2728 struct bdi_writeback *wb = inode_to_wb(inode);
2729
2730 dec_wb_stat(wb, WB_WRITEBACK);
2731 __wb_writeout_inc(wb);
2732 }
2733 }
2734
2735 if (mapping->host && !mapping_tagged(mapping,
2736 PAGECACHE_TAG_WRITEBACK))
2737 sb_clear_inode_writeback(mapping->host);
2738
2739 xa_unlock_irqrestore(&mapping->i_pages, flags);
2740 } else {
2741 ret = TestClearPageWriteback(page);
2742 }
2743 /*
2744 * NOTE: Page might be free now! Writeback doesn't hold a page
2745 * reference on its own, it relies on truncation to wait for
2746 * the clearing of PG_writeback. The below can only access
2747 * page state that is static across allocation cycles.
2748 */
2749 if (ret) {
2750 dec_lruvec_state(lruvec, NR_WRITEBACK);
2751 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2752 inc_node_page_state(page, NR_WRITTEN);
2753 }
2754 __unlock_page_memcg(memcg);
2755 return ret;
2756}
2757
2758int __test_set_page_writeback(struct page *page, bool keep_write)
2759{
2760 struct address_space *mapping = page_mapping(page);
2761 int ret;
2762
2763 lock_page_memcg(page);
2764 if (mapping && mapping_use_writeback_tags(mapping)) {
2765 struct inode *inode = mapping->host;
2766 struct backing_dev_info *bdi = inode_to_bdi(inode);
2767 unsigned long flags;
2768
2769 xa_lock_irqsave(&mapping->i_pages, flags);
2770 ret = TestSetPageWriteback(page);
2771 if (!ret) {
2772 bool on_wblist;
2773
2774 on_wblist = mapping_tagged(mapping,
2775 PAGECACHE_TAG_WRITEBACK);
2776
2777 radix_tree_tag_set(&mapping->i_pages, page_index(page),
2778 PAGECACHE_TAG_WRITEBACK);
2779 if (bdi_cap_account_writeback(bdi))
2780 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2781
2782 /*
2783 * We can come through here when swapping anonymous
2784 * pages, so we don't necessarily have an inode to track
2785 * for sync.
2786 */
2787 if (mapping->host && !on_wblist)
2788 sb_mark_inode_writeback(mapping->host);
2789 }
2790 if (!PageDirty(page))
2791 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2792 PAGECACHE_TAG_DIRTY);
2793 if (!keep_write)
2794 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2795 PAGECACHE_TAG_TOWRITE);
2796 xa_unlock_irqrestore(&mapping->i_pages, flags);
2797 } else {
2798 ret = TestSetPageWriteback(page);
2799 }
2800 if (!ret) {
2801 inc_lruvec_page_state(page, NR_WRITEBACK);
2802 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2803 }
2804 unlock_page_memcg(page);
2805 return ret;
2806
2807}
2808EXPORT_SYMBOL(__test_set_page_writeback);
2809
2810/*
2811 * Return true if any of the pages in the mapping are marked with the
2812 * passed tag.
2813 */
2814int mapping_tagged(struct address_space *mapping, int tag)
2815{
2816 return radix_tree_tagged(&mapping->i_pages, tag);
2817}
2818EXPORT_SYMBOL(mapping_tagged);
2819
2820/**
2821 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2822 * @page: The page to wait on.
2823 *
2824 * This function determines if the given page is related to a backing device
2825 * that requires page contents to be held stable during writeback. If so, then
2826 * it will wait for any pending writeback to complete.
2827 */
2828void wait_for_stable_page(struct page *page)
2829{
2830 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2831 wait_on_page_writeback(page);
2832}
2833EXPORT_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 <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/mm_inline.h>
40#include <trace/events/writeback.h>
41
42#include "internal.h"
43
44/*
45 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
56 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
60#define RATELIMIT_CALC_SHIFT 10
61
62/*
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
68/* The following parameters are exported via /proc/sys/vm */
69
70/*
71 * Start background writeback (via writeback threads) at this percentage
72 */
73int dirty_background_ratio = 10;
74
75/*
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
88 * The generator of dirty data starts writeback at this percentage
89 */
90int vm_dirty_ratio = 20;
91
92/*
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
99 * The interval between `kupdate'-style writebacks
100 */
101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102
103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
105/*
106 * The longest time for which data is allowed to remain dirty
107 */
108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int block_dump;
114
115/*
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
118 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
125unsigned long global_dirty_limit;
126
127/*
128 * Scale the writeback cache size proportional to the relative writeout speeds.
129 *
130 * We do this by keeping a floating proportion between BDIs, based on page
131 * writeback completions [end_page_writeback()]. Those devices that write out
132 * pages fastest will get the larger share, while the slower will get a smaller
133 * share.
134 *
135 * We use page writeout completions because we are interested in getting rid of
136 * dirty pages. Having them written out is the primary goal.
137 *
138 * We introduce a concept of time, a period over which we measure these events,
139 * because demand can/will vary over time. The length of this period itself is
140 * measured in page writeback completions.
141 *
142 */
143static struct fprop_global writeout_completions;
144
145static void writeout_period(unsigned long t);
146/* Timer for aging of writeout_completions */
147static struct timer_list writeout_period_timer =
148 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
149static unsigned long writeout_period_time = 0;
150
151/*
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
155 */
156#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
157
158/*
159 * Work out the current dirty-memory clamping and background writeout
160 * thresholds.
161 *
162 * The main aim here is to lower them aggressively if there is a lot of mapped
163 * memory around. To avoid stressing page reclaim with lots of unreclaimable
164 * pages. It is better to clamp down on writers than to start swapping, and
165 * performing lots of scanning.
166 *
167 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
168 *
169 * We don't permit the clamping level to fall below 5% - that is getting rather
170 * excessive.
171 *
172 * We make sure that the background writeout level is below the adjusted
173 * clamping level.
174 */
175
176/*
177 * In a memory zone, there is a certain amount of pages we consider
178 * available for the page cache, which is essentially the number of
179 * free and reclaimable pages, minus some zone reserves to protect
180 * lowmem and the ability to uphold the zone's watermarks without
181 * requiring writeback.
182 *
183 * This number of dirtyable pages is the base value of which the
184 * user-configurable dirty ratio is the effictive number of pages that
185 * are allowed to be actually dirtied. Per individual zone, or
186 * globally by using the sum of dirtyable pages over all zones.
187 *
188 * Because the user is allowed to specify the dirty limit globally as
189 * absolute number of bytes, calculating the per-zone dirty limit can
190 * require translating the configured limit into a percentage of
191 * global dirtyable memory first.
192 */
193
194/**
195 * zone_dirtyable_memory - number of dirtyable pages in a zone
196 * @zone: the zone
197 *
198 * Returns the zone's number of pages potentially available for dirty
199 * page cache. This is the base value for the per-zone dirty limits.
200 */
201static unsigned long zone_dirtyable_memory(struct zone *zone)
202{
203 unsigned long nr_pages;
204
205 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
206 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
207
208 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
209 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
210
211 return nr_pages;
212}
213
214static unsigned long highmem_dirtyable_memory(unsigned long total)
215{
216#ifdef CONFIG_HIGHMEM
217 int node;
218 unsigned long x = 0;
219
220 for_each_node_state(node, N_HIGH_MEMORY) {
221 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
222
223 x += zone_dirtyable_memory(z);
224 }
225 /*
226 * Unreclaimable memory (kernel memory or anonymous memory
227 * without swap) can bring down the dirtyable pages below
228 * the zone's dirty balance reserve and the above calculation
229 * will underflow. However we still want to add in nodes
230 * which are below threshold (negative values) to get a more
231 * accurate calculation but make sure that the total never
232 * underflows.
233 */
234 if ((long)x < 0)
235 x = 0;
236
237 /*
238 * Make sure that the number of highmem pages is never larger
239 * than the number of the total dirtyable memory. This can only
240 * occur in very strange VM situations but we want to make sure
241 * that this does not occur.
242 */
243 return min(x, total);
244#else
245 return 0;
246#endif
247}
248
249/**
250 * global_dirtyable_memory - number of globally dirtyable pages
251 *
252 * Returns the global number of pages potentially available for dirty
253 * page cache. This is the base value for the global dirty limits.
254 */
255static unsigned long global_dirtyable_memory(void)
256{
257 unsigned long x;
258
259 x = global_page_state(NR_FREE_PAGES);
260 x -= min(x, dirty_balance_reserve);
261
262 x += global_page_state(NR_INACTIVE_FILE);
263 x += global_page_state(NR_ACTIVE_FILE);
264
265 if (!vm_highmem_is_dirtyable)
266 x -= highmem_dirtyable_memory(x);
267
268 return x + 1; /* Ensure that we never return 0 */
269}
270
271/*
272 * global_dirty_limits - background-writeback and dirty-throttling thresholds
273 *
274 * Calculate the dirty thresholds based on sysctl parameters
275 * - vm.dirty_background_ratio or vm.dirty_background_bytes
276 * - vm.dirty_ratio or vm.dirty_bytes
277 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
278 * real-time tasks.
279 */
280void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
281{
282 unsigned long background;
283 unsigned long dirty;
284 unsigned long uninitialized_var(available_memory);
285 struct task_struct *tsk;
286
287 if (!vm_dirty_bytes || !dirty_background_bytes)
288 available_memory = global_dirtyable_memory();
289
290 if (vm_dirty_bytes)
291 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
292 else
293 dirty = (vm_dirty_ratio * available_memory) / 100;
294
295 if (dirty_background_bytes)
296 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
297 else
298 background = (dirty_background_ratio * available_memory) / 100;
299
300 if (background >= dirty)
301 background = dirty / 2;
302 tsk = current;
303 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
304 background += background / 4;
305 dirty += dirty / 4;
306 }
307 *pbackground = background;
308 *pdirty = dirty;
309 trace_global_dirty_state(background, dirty);
310}
311
312/**
313 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
314 * @zone: the zone
315 *
316 * Returns the maximum number of dirty pages allowed in a zone, based
317 * on the zone's dirtyable memory.
318 */
319static unsigned long zone_dirty_limit(struct zone *zone)
320{
321 unsigned long zone_memory = zone_dirtyable_memory(zone);
322 struct task_struct *tsk = current;
323 unsigned long dirty;
324
325 if (vm_dirty_bytes)
326 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
327 zone_memory / global_dirtyable_memory();
328 else
329 dirty = vm_dirty_ratio * zone_memory / 100;
330
331 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
332 dirty += dirty / 4;
333
334 return dirty;
335}
336
337/**
338 * zone_dirty_ok - tells whether a zone is within its dirty limits
339 * @zone: the zone to check
340 *
341 * Returns %true when the dirty pages in @zone are within the zone's
342 * dirty limit, %false if the limit is exceeded.
343 */
344bool zone_dirty_ok(struct zone *zone)
345{
346 unsigned long limit = zone_dirty_limit(zone);
347
348 return zone_page_state(zone, NR_FILE_DIRTY) +
349 zone_page_state(zone, NR_UNSTABLE_NFS) +
350 zone_page_state(zone, NR_WRITEBACK) <= limit;
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 writeback_set_ratelimit();
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 writeback_set_ratelimit();
402 vm_dirty_ratio = 0;
403 }
404 return ret;
405}
406
407static unsigned long wp_next_time(unsigned long cur_time)
408{
409 cur_time += VM_COMPLETIONS_PERIOD_LEN;
410 /* 0 has a special meaning... */
411 if (!cur_time)
412 return 1;
413 return cur_time;
414}
415
416/*
417 * Increment the BDI's writeout completion count and the global writeout
418 * completion count. Called from test_clear_page_writeback().
419 */
420static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
421{
422 __inc_bdi_stat(bdi, BDI_WRITTEN);
423 __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
424 bdi->max_prop_frac);
425 /* First event after period switching was turned off? */
426 if (!unlikely(writeout_period_time)) {
427 /*
428 * We can race with other __bdi_writeout_inc calls here but
429 * it does not cause any harm since the resulting time when
430 * timer will fire and what is in writeout_period_time will be
431 * roughly the same.
432 */
433 writeout_period_time = wp_next_time(jiffies);
434 mod_timer(&writeout_period_timer, writeout_period_time);
435 }
436}
437
438void bdi_writeout_inc(struct backing_dev_info *bdi)
439{
440 unsigned long flags;
441
442 local_irq_save(flags);
443 __bdi_writeout_inc(bdi);
444 local_irq_restore(flags);
445}
446EXPORT_SYMBOL_GPL(bdi_writeout_inc);
447
448/*
449 * Obtain an accurate fraction of the BDI's portion.
450 */
451static void bdi_writeout_fraction(struct backing_dev_info *bdi,
452 long *numerator, long *denominator)
453{
454 fprop_fraction_percpu(&writeout_completions, &bdi->completions,
455 numerator, denominator);
456}
457
458/*
459 * On idle system, we can be called long after we scheduled because we use
460 * deferred timers so count with missed periods.
461 */
462static void writeout_period(unsigned long t)
463{
464 int miss_periods = (jiffies - writeout_period_time) /
465 VM_COMPLETIONS_PERIOD_LEN;
466
467 if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
468 writeout_period_time = wp_next_time(writeout_period_time +
469 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
470 mod_timer(&writeout_period_timer, writeout_period_time);
471 } else {
472 /*
473 * Aging has zeroed all fractions. Stop wasting CPU on period
474 * updates.
475 */
476 writeout_period_time = 0;
477 }
478}
479
480/*
481 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
482 * registered backing devices, which, for obvious reasons, can not
483 * exceed 100%.
484 */
485static unsigned int bdi_min_ratio;
486
487int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
488{
489 int ret = 0;
490
491 spin_lock_bh(&bdi_lock);
492 if (min_ratio > bdi->max_ratio) {
493 ret = -EINVAL;
494 } else {
495 min_ratio -= bdi->min_ratio;
496 if (bdi_min_ratio + min_ratio < 100) {
497 bdi_min_ratio += min_ratio;
498 bdi->min_ratio += min_ratio;
499 } else {
500 ret = -EINVAL;
501 }
502 }
503 spin_unlock_bh(&bdi_lock);
504
505 return ret;
506}
507
508int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
509{
510 int ret = 0;
511
512 if (max_ratio > 100)
513 return -EINVAL;
514
515 spin_lock_bh(&bdi_lock);
516 if (bdi->min_ratio > max_ratio) {
517 ret = -EINVAL;
518 } else {
519 bdi->max_ratio = max_ratio;
520 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
521 }
522 spin_unlock_bh(&bdi_lock);
523
524 return ret;
525}
526EXPORT_SYMBOL(bdi_set_max_ratio);
527
528static unsigned long dirty_freerun_ceiling(unsigned long thresh,
529 unsigned long bg_thresh)
530{
531 return (thresh + bg_thresh) / 2;
532}
533
534static unsigned long hard_dirty_limit(unsigned long thresh)
535{
536 return max(thresh, global_dirty_limit);
537}
538
539/**
540 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
541 * @bdi: the backing_dev_info to query
542 * @dirty: global dirty limit in pages
543 *
544 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
545 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
546 *
547 * Note that balance_dirty_pages() will only seriously take it as a hard limit
548 * when sleeping max_pause per page is not enough to keep the dirty pages under
549 * control. For example, when the device is completely stalled due to some error
550 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
551 * In the other normal situations, it acts more gently by throttling the tasks
552 * more (rather than completely block them) when the bdi dirty pages go high.
553 *
554 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
555 * - starving fast devices
556 * - piling up dirty pages (that will take long time to sync) on slow devices
557 *
558 * The bdi's share of dirty limit will be adapting to its throughput and
559 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
560 */
561unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
562{
563 u64 bdi_dirty;
564 long numerator, denominator;
565
566 /*
567 * Calculate this BDI's share of the dirty ratio.
568 */
569 bdi_writeout_fraction(bdi, &numerator, &denominator);
570
571 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
572 bdi_dirty *= numerator;
573 do_div(bdi_dirty, denominator);
574
575 bdi_dirty += (dirty * bdi->min_ratio) / 100;
576 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
577 bdi_dirty = dirty * bdi->max_ratio / 100;
578
579 return bdi_dirty;
580}
581
582/*
583 * setpoint - dirty 3
584 * f(dirty) := 1.0 + (----------------)
585 * limit - setpoint
586 *
587 * it's a 3rd order polynomial that subjects to
588 *
589 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
590 * (2) f(setpoint) = 1.0 => the balance point
591 * (3) f(limit) = 0 => the hard limit
592 * (4) df/dx <= 0 => negative feedback control
593 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
594 * => fast response on large errors; small oscillation near setpoint
595 */
596static long long pos_ratio_polynom(unsigned long setpoint,
597 unsigned long dirty,
598 unsigned long limit)
599{
600 long long pos_ratio;
601 long x;
602
603 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
604 limit - setpoint + 1);
605 pos_ratio = x;
606 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
607 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
608 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
609
610 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
611}
612
613/*
614 * Dirty position control.
615 *
616 * (o) global/bdi setpoints
617 *
618 * We want the dirty pages be balanced around the global/bdi setpoints.
619 * When the number of dirty pages is higher/lower than the setpoint, the
620 * dirty position control ratio (and hence task dirty ratelimit) will be
621 * decreased/increased to bring the dirty pages back to the setpoint.
622 *
623 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
624 *
625 * if (dirty < setpoint) scale up pos_ratio
626 * if (dirty > setpoint) scale down pos_ratio
627 *
628 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
629 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
630 *
631 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
632 *
633 * (o) global control line
634 *
635 * ^ pos_ratio
636 * |
637 * | |<===== global dirty control scope ======>|
638 * 2.0 .............*
639 * | .*
640 * | . *
641 * | . *
642 * | . *
643 * | . *
644 * | . *
645 * 1.0 ................................*
646 * | . . *
647 * | . . *
648 * | . . *
649 * | . . *
650 * | . . *
651 * 0 +------------.------------------.----------------------*------------->
652 * freerun^ setpoint^ limit^ dirty pages
653 *
654 * (o) bdi control line
655 *
656 * ^ pos_ratio
657 * |
658 * | *
659 * | *
660 * | *
661 * | *
662 * | * |<=========== span ============>|
663 * 1.0 .......................*
664 * | . *
665 * | . *
666 * | . *
667 * | . *
668 * | . *
669 * | . *
670 * | . *
671 * | . *
672 * | . *
673 * | . *
674 * | . *
675 * 1/4 ...............................................* * * * * * * * * * * *
676 * | . .
677 * | . .
678 * | . .
679 * 0 +----------------------.-------------------------------.------------->
680 * bdi_setpoint^ x_intercept^
681 *
682 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
683 * be smoothly throttled down to normal if it starts high in situations like
684 * - start writing to a slow SD card and a fast disk at the same time. The SD
685 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
686 * - the bdi dirty thresh drops quickly due to change of JBOD workload
687 */
688static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
689 unsigned long thresh,
690 unsigned long bg_thresh,
691 unsigned long dirty,
692 unsigned long bdi_thresh,
693 unsigned long bdi_dirty)
694{
695 unsigned long write_bw = bdi->avg_write_bandwidth;
696 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
697 unsigned long limit = hard_dirty_limit(thresh);
698 unsigned long x_intercept;
699 unsigned long setpoint; /* dirty pages' target balance point */
700 unsigned long bdi_setpoint;
701 unsigned long span;
702 long long pos_ratio; /* for scaling up/down the rate limit */
703 long x;
704
705 if (unlikely(dirty >= limit))
706 return 0;
707
708 /*
709 * global setpoint
710 *
711 * See comment for pos_ratio_polynom().
712 */
713 setpoint = (freerun + limit) / 2;
714 pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
715
716 /*
717 * The strictlimit feature is a tool preventing mistrusted filesystems
718 * from growing a large number of dirty pages before throttling. For
719 * such filesystems balance_dirty_pages always checks bdi counters
720 * against bdi limits. Even if global "nr_dirty" is under "freerun".
721 * This is especially important for fuse which sets bdi->max_ratio to
722 * 1% by default. Without strictlimit feature, fuse writeback may
723 * consume arbitrary amount of RAM because it is accounted in
724 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
725 *
726 * Here, in bdi_position_ratio(), we calculate pos_ratio based on
727 * two values: bdi_dirty and bdi_thresh. Let's consider an example:
728 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
729 * limits are set by default to 10% and 20% (background and throttle).
730 * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
731 * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
732 * about ~6K pages (as the average of background and throttle bdi
733 * limits). The 3rd order polynomial will provide positive feedback if
734 * bdi_dirty is under bdi_setpoint and vice versa.
735 *
736 * Note, that we cannot use global counters in these calculations
737 * because we want to throttle process writing to a strictlimit BDI
738 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
739 * in the example above).
740 */
741 if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
742 long long bdi_pos_ratio;
743 unsigned long bdi_bg_thresh;
744
745 if (bdi_dirty < 8)
746 return min_t(long long, pos_ratio * 2,
747 2 << RATELIMIT_CALC_SHIFT);
748
749 if (bdi_dirty >= bdi_thresh)
750 return 0;
751
752 bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
753 bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
754 bdi_bg_thresh);
755
756 if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
757 return 0;
758
759 bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
760 bdi_thresh);
761
762 /*
763 * Typically, for strictlimit case, bdi_setpoint << setpoint
764 * and pos_ratio >> bdi_pos_ratio. In the other words global
765 * state ("dirty") is not limiting factor and we have to
766 * make decision based on bdi counters. But there is an
767 * important case when global pos_ratio should get precedence:
768 * global limits are exceeded (e.g. due to activities on other
769 * BDIs) while given strictlimit BDI is below limit.
770 *
771 * "pos_ratio * bdi_pos_ratio" would work for the case above,
772 * but it would look too non-natural for the case of all
773 * activity in the system coming from a single strictlimit BDI
774 * with bdi->max_ratio == 100%.
775 *
776 * Note that min() below somewhat changes the dynamics of the
777 * control system. Normally, pos_ratio value can be well over 3
778 * (when globally we are at freerun and bdi is well below bdi
779 * setpoint). Now the maximum pos_ratio in the same situation
780 * is 2. We might want to tweak this if we observe the control
781 * system is too slow to adapt.
782 */
783 return min(pos_ratio, bdi_pos_ratio);
784 }
785
786 /*
787 * We have computed basic pos_ratio above based on global situation. If
788 * the bdi is over/under its share of dirty pages, we want to scale
789 * pos_ratio further down/up. That is done by the following mechanism.
790 */
791
792 /*
793 * bdi setpoint
794 *
795 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
796 *
797 * x_intercept - bdi_dirty
798 * := --------------------------
799 * x_intercept - bdi_setpoint
800 *
801 * The main bdi control line is a linear function that subjects to
802 *
803 * (1) f(bdi_setpoint) = 1.0
804 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
805 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
806 *
807 * For single bdi case, the dirty pages are observed to fluctuate
808 * regularly within range
809 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
810 * for various filesystems, where (2) can yield in a reasonable 12.5%
811 * fluctuation range for pos_ratio.
812 *
813 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
814 * own size, so move the slope over accordingly and choose a slope that
815 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
816 */
817 if (unlikely(bdi_thresh > thresh))
818 bdi_thresh = thresh;
819 /*
820 * It's very possible that bdi_thresh is close to 0 not because the
821 * device is slow, but that it has remained inactive for long time.
822 * Honour such devices a reasonable good (hopefully IO efficient)
823 * threshold, so that the occasional writes won't be blocked and active
824 * writes can rampup the threshold quickly.
825 */
826 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
827 /*
828 * scale global setpoint to bdi's:
829 * bdi_setpoint = setpoint * bdi_thresh / thresh
830 */
831 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
832 bdi_setpoint = setpoint * (u64)x >> 16;
833 /*
834 * Use span=(8*write_bw) in single bdi case as indicated by
835 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
836 *
837 * bdi_thresh thresh - bdi_thresh
838 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
839 * thresh thresh
840 */
841 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
842 x_intercept = bdi_setpoint + span;
843
844 if (bdi_dirty < x_intercept - span / 4) {
845 pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty),
846 x_intercept - bdi_setpoint + 1);
847 } else
848 pos_ratio /= 4;
849
850 /*
851 * bdi reserve area, safeguard against dirty pool underrun and disk idle
852 * It may push the desired control point of global dirty pages higher
853 * than setpoint.
854 */
855 x_intercept = bdi_thresh / 2;
856 if (bdi_dirty < x_intercept) {
857 if (bdi_dirty > x_intercept / 8)
858 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
859 else
860 pos_ratio *= 8;
861 }
862
863 return pos_ratio;
864}
865
866static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
867 unsigned long elapsed,
868 unsigned long written)
869{
870 const unsigned long period = roundup_pow_of_two(3 * HZ);
871 unsigned long avg = bdi->avg_write_bandwidth;
872 unsigned long old = bdi->write_bandwidth;
873 u64 bw;
874
875 /*
876 * bw = written * HZ / elapsed
877 *
878 * bw * elapsed + write_bandwidth * (period - elapsed)
879 * write_bandwidth = ---------------------------------------------------
880 * period
881 */
882 bw = written - bdi->written_stamp;
883 bw *= HZ;
884 if (unlikely(elapsed > period)) {
885 do_div(bw, elapsed);
886 avg = bw;
887 goto out;
888 }
889 bw += (u64)bdi->write_bandwidth * (period - elapsed);
890 bw >>= ilog2(period);
891
892 /*
893 * one more level of smoothing, for filtering out sudden spikes
894 */
895 if (avg > old && old >= (unsigned long)bw)
896 avg -= (avg - old) >> 3;
897
898 if (avg < old && old <= (unsigned long)bw)
899 avg += (old - avg) >> 3;
900
901out:
902 bdi->write_bandwidth = bw;
903 bdi->avg_write_bandwidth = avg;
904}
905
906/*
907 * The global dirtyable memory and dirty threshold could be suddenly knocked
908 * down by a large amount (eg. on the startup of KVM in a swapless system).
909 * This may throw the system into deep dirty exceeded state and throttle
910 * heavy/light dirtiers alike. To retain good responsiveness, maintain
911 * global_dirty_limit for tracking slowly down to the knocked down dirty
912 * threshold.
913 */
914static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
915{
916 unsigned long limit = global_dirty_limit;
917
918 /*
919 * Follow up in one step.
920 */
921 if (limit < thresh) {
922 limit = thresh;
923 goto update;
924 }
925
926 /*
927 * Follow down slowly. Use the higher one as the target, because thresh
928 * may drop below dirty. This is exactly the reason to introduce
929 * global_dirty_limit which is guaranteed to lie above the dirty pages.
930 */
931 thresh = max(thresh, dirty);
932 if (limit > thresh) {
933 limit -= (limit - thresh) >> 5;
934 goto update;
935 }
936 return;
937update:
938 global_dirty_limit = limit;
939}
940
941static void global_update_bandwidth(unsigned long thresh,
942 unsigned long dirty,
943 unsigned long now)
944{
945 static DEFINE_SPINLOCK(dirty_lock);
946 static unsigned long update_time;
947
948 /*
949 * check locklessly first to optimize away locking for the most time
950 */
951 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
952 return;
953
954 spin_lock(&dirty_lock);
955 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
956 update_dirty_limit(thresh, dirty);
957 update_time = now;
958 }
959 spin_unlock(&dirty_lock);
960}
961
962/*
963 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
964 *
965 * Normal bdi tasks will be curbed at or below it in long term.
966 * Obviously it should be around (write_bw / N) when there are N dd tasks.
967 */
968static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
969 unsigned long thresh,
970 unsigned long bg_thresh,
971 unsigned long dirty,
972 unsigned long bdi_thresh,
973 unsigned long bdi_dirty,
974 unsigned long dirtied,
975 unsigned long elapsed)
976{
977 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
978 unsigned long limit = hard_dirty_limit(thresh);
979 unsigned long setpoint = (freerun + limit) / 2;
980 unsigned long write_bw = bdi->avg_write_bandwidth;
981 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
982 unsigned long dirty_rate;
983 unsigned long task_ratelimit;
984 unsigned long balanced_dirty_ratelimit;
985 unsigned long pos_ratio;
986 unsigned long step;
987 unsigned long x;
988
989 /*
990 * The dirty rate will match the writeout rate in long term, except
991 * when dirty pages are truncated by userspace or re-dirtied by FS.
992 */
993 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
994
995 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
996 bdi_thresh, bdi_dirty);
997 /*
998 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
999 */
1000 task_ratelimit = (u64)dirty_ratelimit *
1001 pos_ratio >> RATELIMIT_CALC_SHIFT;
1002 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1003
1004 /*
1005 * A linear estimation of the "balanced" throttle rate. The theory is,
1006 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
1007 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1008 * formula will yield the balanced rate limit (write_bw / N).
1009 *
1010 * Note that the expanded form is not a pure rate feedback:
1011 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1012 * but also takes pos_ratio into account:
1013 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1014 *
1015 * (1) is not realistic because pos_ratio also takes part in balancing
1016 * the dirty rate. Consider the state
1017 * pos_ratio = 0.5 (3)
1018 * rate = 2 * (write_bw / N) (4)
1019 * If (1) is used, it will stuck in that state! Because each dd will
1020 * be throttled at
1021 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1022 * yielding
1023 * dirty_rate = N * task_ratelimit = write_bw (6)
1024 * put (6) into (1) we get
1025 * rate_(i+1) = rate_(i) (7)
1026 *
1027 * So we end up using (2) to always keep
1028 * rate_(i+1) ~= (write_bw / N) (8)
1029 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1030 * pos_ratio is able to drive itself to 1.0, which is not only where
1031 * the dirty count meet the setpoint, but also where the slope of
1032 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1033 */
1034 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1035 dirty_rate | 1);
1036 /*
1037 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1038 */
1039 if (unlikely(balanced_dirty_ratelimit > write_bw))
1040 balanced_dirty_ratelimit = write_bw;
1041
1042 /*
1043 * We could safely do this and return immediately:
1044 *
1045 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
1046 *
1047 * However to get a more stable dirty_ratelimit, the below elaborated
1048 * code makes use of task_ratelimit to filter out singular points and
1049 * limit the step size.
1050 *
1051 * The below code essentially only uses the relative value of
1052 *
1053 * task_ratelimit - dirty_ratelimit
1054 * = (pos_ratio - 1) * dirty_ratelimit
1055 *
1056 * which reflects the direction and size of dirty position error.
1057 */
1058
1059 /*
1060 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1061 * task_ratelimit is on the same side of dirty_ratelimit, too.
1062 * For example, when
1063 * - dirty_ratelimit > balanced_dirty_ratelimit
1064 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1065 * lowering dirty_ratelimit will help meet both the position and rate
1066 * control targets. Otherwise, don't update dirty_ratelimit if it will
1067 * only help meet the rate target. After all, what the users ultimately
1068 * feel and care are stable dirty rate and small position error.
1069 *
1070 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1071 * and filter out the singular points of balanced_dirty_ratelimit. Which
1072 * keeps jumping around randomly and can even leap far away at times
1073 * due to the small 200ms estimation period of dirty_rate (we want to
1074 * keep that period small to reduce time lags).
1075 */
1076 step = 0;
1077
1078 /*
1079 * For strictlimit case, calculations above were based on bdi counters
1080 * and limits (starting from pos_ratio = bdi_position_ratio() and up to
1081 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1082 * Hence, to calculate "step" properly, we have to use bdi_dirty as
1083 * "dirty" and bdi_setpoint as "setpoint".
1084 *
1085 * We rampup dirty_ratelimit forcibly if bdi_dirty is low because
1086 * it's possible that bdi_thresh is close to zero due to inactivity
1087 * of backing device (see the implementation of bdi_dirty_limit()).
1088 */
1089 if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1090 dirty = bdi_dirty;
1091 if (bdi_dirty < 8)
1092 setpoint = bdi_dirty + 1;
1093 else
1094 setpoint = (bdi_thresh +
1095 bdi_dirty_limit(bdi, bg_thresh)) / 2;
1096 }
1097
1098 if (dirty < setpoint) {
1099 x = min(bdi->balanced_dirty_ratelimit,
1100 min(balanced_dirty_ratelimit, task_ratelimit));
1101 if (dirty_ratelimit < x)
1102 step = x - dirty_ratelimit;
1103 } else {
1104 x = max(bdi->balanced_dirty_ratelimit,
1105 max(balanced_dirty_ratelimit, task_ratelimit));
1106 if (dirty_ratelimit > x)
1107 step = dirty_ratelimit - x;
1108 }
1109
1110 /*
1111 * Don't pursue 100% rate matching. It's impossible since the balanced
1112 * rate itself is constantly fluctuating. So decrease the track speed
1113 * when it gets close to the target. Helps eliminate pointless tremors.
1114 */
1115 step >>= dirty_ratelimit / (2 * step + 1);
1116 /*
1117 * Limit the tracking speed to avoid overshooting.
1118 */
1119 step = (step + 7) / 8;
1120
1121 if (dirty_ratelimit < balanced_dirty_ratelimit)
1122 dirty_ratelimit += step;
1123 else
1124 dirty_ratelimit -= step;
1125
1126 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1127 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1128
1129 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1130}
1131
1132void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1133 unsigned long thresh,
1134 unsigned long bg_thresh,
1135 unsigned long dirty,
1136 unsigned long bdi_thresh,
1137 unsigned long bdi_dirty,
1138 unsigned long start_time)
1139{
1140 unsigned long now = jiffies;
1141 unsigned long elapsed = now - bdi->bw_time_stamp;
1142 unsigned long dirtied;
1143 unsigned long written;
1144
1145 /*
1146 * rate-limit, only update once every 200ms.
1147 */
1148 if (elapsed < BANDWIDTH_INTERVAL)
1149 return;
1150
1151 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1152 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1153
1154 /*
1155 * Skip quiet periods when disk bandwidth is under-utilized.
1156 * (at least 1s idle time between two flusher runs)
1157 */
1158 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1159 goto snapshot;
1160
1161 if (thresh) {
1162 global_update_bandwidth(thresh, dirty, now);
1163 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1164 bdi_thresh, bdi_dirty,
1165 dirtied, elapsed);
1166 }
1167 bdi_update_write_bandwidth(bdi, elapsed, written);
1168
1169snapshot:
1170 bdi->dirtied_stamp = dirtied;
1171 bdi->written_stamp = written;
1172 bdi->bw_time_stamp = now;
1173}
1174
1175static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1176 unsigned long thresh,
1177 unsigned long bg_thresh,
1178 unsigned long dirty,
1179 unsigned long bdi_thresh,
1180 unsigned long bdi_dirty,
1181 unsigned long start_time)
1182{
1183 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1184 return;
1185 spin_lock(&bdi->wb.list_lock);
1186 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1187 bdi_thresh, bdi_dirty, start_time);
1188 spin_unlock(&bdi->wb.list_lock);
1189}
1190
1191/*
1192 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1193 * will look to see if it needs to start dirty throttling.
1194 *
1195 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1196 * global_page_state() too often. So scale it near-sqrt to the safety margin
1197 * (the number of pages we may dirty without exceeding the dirty limits).
1198 */
1199static unsigned long dirty_poll_interval(unsigned long dirty,
1200 unsigned long thresh)
1201{
1202 if (thresh > dirty)
1203 return 1UL << (ilog2(thresh - dirty) >> 1);
1204
1205 return 1;
1206}
1207
1208static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1209 unsigned long bdi_dirty)
1210{
1211 unsigned long bw = bdi->avg_write_bandwidth;
1212 unsigned long t;
1213
1214 /*
1215 * Limit pause time for small memory systems. If sleeping for too long
1216 * time, a small pool of dirty/writeback pages may go empty and disk go
1217 * idle.
1218 *
1219 * 8 serves as the safety ratio.
1220 */
1221 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1222 t++;
1223
1224 return min_t(unsigned long, t, MAX_PAUSE);
1225}
1226
1227static long bdi_min_pause(struct backing_dev_info *bdi,
1228 long max_pause,
1229 unsigned long task_ratelimit,
1230 unsigned long dirty_ratelimit,
1231 int *nr_dirtied_pause)
1232{
1233 long hi = ilog2(bdi->avg_write_bandwidth);
1234 long lo = ilog2(bdi->dirty_ratelimit);
1235 long t; /* target pause */
1236 long pause; /* estimated next pause */
1237 int pages; /* target nr_dirtied_pause */
1238
1239 /* target for 10ms pause on 1-dd case */
1240 t = max(1, HZ / 100);
1241
1242 /*
1243 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1244 * overheads.
1245 *
1246 * (N * 10ms) on 2^N concurrent tasks.
1247 */
1248 if (hi > lo)
1249 t += (hi - lo) * (10 * HZ) / 1024;
1250
1251 /*
1252 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1253 * on the much more stable dirty_ratelimit. However the next pause time
1254 * will be computed based on task_ratelimit and the two rate limits may
1255 * depart considerably at some time. Especially if task_ratelimit goes
1256 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1257 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1258 * result task_ratelimit won't be executed faithfully, which could
1259 * eventually bring down dirty_ratelimit.
1260 *
1261 * We apply two rules to fix it up:
1262 * 1) try to estimate the next pause time and if necessary, use a lower
1263 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1264 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1265 * 2) limit the target pause time to max_pause/2, so that the normal
1266 * small fluctuations of task_ratelimit won't trigger rule (1) and
1267 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1268 */
1269 t = min(t, 1 + max_pause / 2);
1270 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1271
1272 /*
1273 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1274 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1275 * When the 16 consecutive reads are often interrupted by some dirty
1276 * throttling pause during the async writes, cfq will go into idles
1277 * (deadline is fine). So push nr_dirtied_pause as high as possible
1278 * until reaches DIRTY_POLL_THRESH=32 pages.
1279 */
1280 if (pages < DIRTY_POLL_THRESH) {
1281 t = max_pause;
1282 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1283 if (pages > DIRTY_POLL_THRESH) {
1284 pages = DIRTY_POLL_THRESH;
1285 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1286 }
1287 }
1288
1289 pause = HZ * pages / (task_ratelimit + 1);
1290 if (pause > max_pause) {
1291 t = max_pause;
1292 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1293 }
1294
1295 *nr_dirtied_pause = pages;
1296 /*
1297 * The minimal pause time will normally be half the target pause time.
1298 */
1299 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1300}
1301
1302static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
1303 unsigned long dirty_thresh,
1304 unsigned long background_thresh,
1305 unsigned long *bdi_dirty,
1306 unsigned long *bdi_thresh,
1307 unsigned long *bdi_bg_thresh)
1308{
1309 unsigned long bdi_reclaimable;
1310
1311 /*
1312 * bdi_thresh is not treated as some limiting factor as
1313 * dirty_thresh, due to reasons
1314 * - in JBOD setup, bdi_thresh can fluctuate a lot
1315 * - in a system with HDD and USB key, the USB key may somehow
1316 * go into state (bdi_dirty >> bdi_thresh) either because
1317 * bdi_dirty starts high, or because bdi_thresh drops low.
1318 * In this case we don't want to hard throttle the USB key
1319 * dirtiers for 100 seconds until bdi_dirty drops under
1320 * bdi_thresh. Instead the auxiliary bdi control line in
1321 * bdi_position_ratio() will let the dirtier task progress
1322 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1323 */
1324 *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1325
1326 if (bdi_bg_thresh)
1327 *bdi_bg_thresh = div_u64((u64)*bdi_thresh *
1328 background_thresh,
1329 dirty_thresh);
1330
1331 /*
1332 * In order to avoid the stacked BDI deadlock we need
1333 * to ensure we accurately count the 'dirty' pages when
1334 * the threshold is low.
1335 *
1336 * Otherwise it would be possible to get thresh+n pages
1337 * reported dirty, even though there are thresh-m pages
1338 * actually dirty; with m+n sitting in the percpu
1339 * deltas.
1340 */
1341 if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
1342 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1343 *bdi_dirty = bdi_reclaimable +
1344 bdi_stat_sum(bdi, BDI_WRITEBACK);
1345 } else {
1346 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1347 *bdi_dirty = bdi_reclaimable +
1348 bdi_stat(bdi, BDI_WRITEBACK);
1349 }
1350}
1351
1352/*
1353 * balance_dirty_pages() must be called by processes which are generating dirty
1354 * data. It looks at the number of dirty pages in the machine and will force
1355 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1356 * If we're over `background_thresh' then the writeback threads are woken to
1357 * perform some writeout.
1358 */
1359static void balance_dirty_pages(struct address_space *mapping,
1360 unsigned long pages_dirtied)
1361{
1362 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1363 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1364 unsigned long background_thresh;
1365 unsigned long dirty_thresh;
1366 long period;
1367 long pause;
1368 long max_pause;
1369 long min_pause;
1370 int nr_dirtied_pause;
1371 bool dirty_exceeded = false;
1372 unsigned long task_ratelimit;
1373 unsigned long dirty_ratelimit;
1374 unsigned long pos_ratio;
1375 struct backing_dev_info *bdi = mapping->backing_dev_info;
1376 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1377 unsigned long start_time = jiffies;
1378
1379 for (;;) {
1380 unsigned long now = jiffies;
1381 unsigned long uninitialized_var(bdi_thresh);
1382 unsigned long thresh;
1383 unsigned long uninitialized_var(bdi_dirty);
1384 unsigned long dirty;
1385 unsigned long bg_thresh;
1386
1387 /*
1388 * Unstable writes are a feature of certain networked
1389 * filesystems (i.e. NFS) in which data may have been
1390 * written to the server's write cache, but has not yet
1391 * been flushed to permanent storage.
1392 */
1393 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1394 global_page_state(NR_UNSTABLE_NFS);
1395 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1396
1397 global_dirty_limits(&background_thresh, &dirty_thresh);
1398
1399 if (unlikely(strictlimit)) {
1400 bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1401 &bdi_dirty, &bdi_thresh, &bg_thresh);
1402
1403 dirty = bdi_dirty;
1404 thresh = bdi_thresh;
1405 } else {
1406 dirty = nr_dirty;
1407 thresh = dirty_thresh;
1408 bg_thresh = background_thresh;
1409 }
1410
1411 /*
1412 * Throttle it only when the background writeback cannot
1413 * catch-up. This avoids (excessively) small writeouts
1414 * when the bdi limits are ramping up in case of !strictlimit.
1415 *
1416 * In strictlimit case make decision based on the bdi counters
1417 * and limits. Small writeouts when the bdi limits are ramping
1418 * up are the price we consciously pay for strictlimit-ing.
1419 */
1420 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1421 current->dirty_paused_when = now;
1422 current->nr_dirtied = 0;
1423 current->nr_dirtied_pause =
1424 dirty_poll_interval(dirty, thresh);
1425 break;
1426 }
1427
1428 if (unlikely(!writeback_in_progress(bdi)))
1429 bdi_start_background_writeback(bdi);
1430
1431 if (!strictlimit)
1432 bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1433 &bdi_dirty, &bdi_thresh, NULL);
1434
1435 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1436 ((nr_dirty > dirty_thresh) || strictlimit);
1437 if (dirty_exceeded && !bdi->dirty_exceeded)
1438 bdi->dirty_exceeded = 1;
1439
1440 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1441 nr_dirty, bdi_thresh, bdi_dirty,
1442 start_time);
1443
1444 dirty_ratelimit = bdi->dirty_ratelimit;
1445 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1446 background_thresh, nr_dirty,
1447 bdi_thresh, bdi_dirty);
1448 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1449 RATELIMIT_CALC_SHIFT;
1450 max_pause = bdi_max_pause(bdi, bdi_dirty);
1451 min_pause = bdi_min_pause(bdi, max_pause,
1452 task_ratelimit, dirty_ratelimit,
1453 &nr_dirtied_pause);
1454
1455 if (unlikely(task_ratelimit == 0)) {
1456 period = max_pause;
1457 pause = max_pause;
1458 goto pause;
1459 }
1460 period = HZ * pages_dirtied / task_ratelimit;
1461 pause = period;
1462 if (current->dirty_paused_when)
1463 pause -= now - current->dirty_paused_when;
1464 /*
1465 * For less than 1s think time (ext3/4 may block the dirtier
1466 * for up to 800ms from time to time on 1-HDD; so does xfs,
1467 * however at much less frequency), try to compensate it in
1468 * future periods by updating the virtual time; otherwise just
1469 * do a reset, as it may be a light dirtier.
1470 */
1471 if (pause < min_pause) {
1472 trace_balance_dirty_pages(bdi,
1473 dirty_thresh,
1474 background_thresh,
1475 nr_dirty,
1476 bdi_thresh,
1477 bdi_dirty,
1478 dirty_ratelimit,
1479 task_ratelimit,
1480 pages_dirtied,
1481 period,
1482 min(pause, 0L),
1483 start_time);
1484 if (pause < -HZ) {
1485 current->dirty_paused_when = now;
1486 current->nr_dirtied = 0;
1487 } else if (period) {
1488 current->dirty_paused_when += period;
1489 current->nr_dirtied = 0;
1490 } else if (current->nr_dirtied_pause <= pages_dirtied)
1491 current->nr_dirtied_pause += pages_dirtied;
1492 break;
1493 }
1494 if (unlikely(pause > max_pause)) {
1495 /* for occasional dropped task_ratelimit */
1496 now += min(pause - max_pause, max_pause);
1497 pause = max_pause;
1498 }
1499
1500pause:
1501 trace_balance_dirty_pages(bdi,
1502 dirty_thresh,
1503 background_thresh,
1504 nr_dirty,
1505 bdi_thresh,
1506 bdi_dirty,
1507 dirty_ratelimit,
1508 task_ratelimit,
1509 pages_dirtied,
1510 period,
1511 pause,
1512 start_time);
1513 __set_current_state(TASK_KILLABLE);
1514 io_schedule_timeout(pause);
1515
1516 current->dirty_paused_when = now + pause;
1517 current->nr_dirtied = 0;
1518 current->nr_dirtied_pause = nr_dirtied_pause;
1519
1520 /*
1521 * This is typically equal to (nr_dirty < dirty_thresh) and can
1522 * also keep "1000+ dd on a slow USB stick" under control.
1523 */
1524 if (task_ratelimit)
1525 break;
1526
1527 /*
1528 * In the case of an unresponding NFS server and the NFS dirty
1529 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1530 * to go through, so that tasks on them still remain responsive.
1531 *
1532 * In theory 1 page is enough to keep the comsumer-producer
1533 * pipe going: the flusher cleans 1 page => the task dirties 1
1534 * more page. However bdi_dirty has accounting errors. So use
1535 * the larger and more IO friendly bdi_stat_error.
1536 */
1537 if (bdi_dirty <= bdi_stat_error(bdi))
1538 break;
1539
1540 if (fatal_signal_pending(current))
1541 break;
1542 }
1543
1544 if (!dirty_exceeded && bdi->dirty_exceeded)
1545 bdi->dirty_exceeded = 0;
1546
1547 if (writeback_in_progress(bdi))
1548 return;
1549
1550 /*
1551 * In laptop mode, we wait until hitting the higher threshold before
1552 * starting background writeout, and then write out all the way down
1553 * to the lower threshold. So slow writers cause minimal disk activity.
1554 *
1555 * In normal mode, we start background writeout at the lower
1556 * background_thresh, to keep the amount of dirty memory low.
1557 */
1558 if (laptop_mode)
1559 return;
1560
1561 if (nr_reclaimable > background_thresh)
1562 bdi_start_background_writeback(bdi);
1563}
1564
1565void set_page_dirty_balance(struct page *page)
1566{
1567 if (set_page_dirty(page)) {
1568 struct address_space *mapping = page_mapping(page);
1569
1570 if (mapping)
1571 balance_dirty_pages_ratelimited(mapping);
1572 }
1573}
1574
1575static DEFINE_PER_CPU(int, bdp_ratelimits);
1576
1577/*
1578 * Normal tasks are throttled by
1579 * loop {
1580 * dirty tsk->nr_dirtied_pause pages;
1581 * take a snap in balance_dirty_pages();
1582 * }
1583 * However there is a worst case. If every task exit immediately when dirtied
1584 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1585 * called to throttle the page dirties. The solution is to save the not yet
1586 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1587 * randomly into the running tasks. This works well for the above worst case,
1588 * as the new task will pick up and accumulate the old task's leaked dirty
1589 * count and eventually get throttled.
1590 */
1591DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1592
1593/**
1594 * balance_dirty_pages_ratelimited - balance dirty memory state
1595 * @mapping: address_space which was dirtied
1596 *
1597 * Processes which are dirtying memory should call in here once for each page
1598 * which was newly dirtied. The function will periodically check the system's
1599 * dirty state and will initiate writeback if needed.
1600 *
1601 * On really big machines, get_writeback_state is expensive, so try to avoid
1602 * calling it too often (ratelimiting). But once we're over the dirty memory
1603 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1604 * from overshooting the limit by (ratelimit_pages) each.
1605 */
1606void balance_dirty_pages_ratelimited(struct address_space *mapping)
1607{
1608 struct backing_dev_info *bdi = mapping->backing_dev_info;
1609 int ratelimit;
1610 int *p;
1611
1612 if (!bdi_cap_account_dirty(bdi))
1613 return;
1614
1615 ratelimit = current->nr_dirtied_pause;
1616 if (bdi->dirty_exceeded)
1617 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1618
1619 preempt_disable();
1620 /*
1621 * This prevents one CPU to accumulate too many dirtied pages without
1622 * calling into balance_dirty_pages(), which can happen when there are
1623 * 1000+ tasks, all of them start dirtying pages at exactly the same
1624 * time, hence all honoured too large initial task->nr_dirtied_pause.
1625 */
1626 p = &__get_cpu_var(bdp_ratelimits);
1627 if (unlikely(current->nr_dirtied >= ratelimit))
1628 *p = 0;
1629 else if (unlikely(*p >= ratelimit_pages)) {
1630 *p = 0;
1631 ratelimit = 0;
1632 }
1633 /*
1634 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1635 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1636 * the dirty throttling and livelock other long-run dirtiers.
1637 */
1638 p = &__get_cpu_var(dirty_throttle_leaks);
1639 if (*p > 0 && current->nr_dirtied < ratelimit) {
1640 unsigned long nr_pages_dirtied;
1641 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1642 *p -= nr_pages_dirtied;
1643 current->nr_dirtied += nr_pages_dirtied;
1644 }
1645 preempt_enable();
1646
1647 if (unlikely(current->nr_dirtied >= ratelimit))
1648 balance_dirty_pages(mapping, current->nr_dirtied);
1649}
1650EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1651
1652void throttle_vm_writeout(gfp_t gfp_mask)
1653{
1654 unsigned long background_thresh;
1655 unsigned long dirty_thresh;
1656
1657 for ( ; ; ) {
1658 global_dirty_limits(&background_thresh, &dirty_thresh);
1659 dirty_thresh = hard_dirty_limit(dirty_thresh);
1660
1661 /*
1662 * Boost the allowable dirty threshold a bit for page
1663 * allocators so they don't get DoS'ed by heavy writers
1664 */
1665 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1666
1667 if (global_page_state(NR_UNSTABLE_NFS) +
1668 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1669 break;
1670 congestion_wait(BLK_RW_ASYNC, HZ/10);
1671
1672 /*
1673 * The caller might hold locks which can prevent IO completion
1674 * or progress in the filesystem. So we cannot just sit here
1675 * waiting for IO to complete.
1676 */
1677 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1678 break;
1679 }
1680}
1681
1682/*
1683 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1684 */
1685int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1686 void __user *buffer, size_t *length, loff_t *ppos)
1687{
1688 proc_dointvec(table, write, buffer, length, ppos);
1689 return 0;
1690}
1691
1692#ifdef CONFIG_BLOCK
1693void laptop_mode_timer_fn(unsigned long data)
1694{
1695 struct request_queue *q = (struct request_queue *)data;
1696 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1697 global_page_state(NR_UNSTABLE_NFS);
1698
1699 /*
1700 * We want to write everything out, not just down to the dirty
1701 * threshold
1702 */
1703 if (bdi_has_dirty_io(&q->backing_dev_info))
1704 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1705 WB_REASON_LAPTOP_TIMER);
1706}
1707
1708/*
1709 * We've spun up the disk and we're in laptop mode: schedule writeback
1710 * of all dirty data a few seconds from now. If the flush is already scheduled
1711 * then push it back - the user is still using the disk.
1712 */
1713void laptop_io_completion(struct backing_dev_info *info)
1714{
1715 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1716}
1717
1718/*
1719 * We're in laptop mode and we've just synced. The sync's writes will have
1720 * caused another writeback to be scheduled by laptop_io_completion.
1721 * Nothing needs to be written back anymore, so we unschedule the writeback.
1722 */
1723void laptop_sync_completion(void)
1724{
1725 struct backing_dev_info *bdi;
1726
1727 rcu_read_lock();
1728
1729 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1730 del_timer(&bdi->laptop_mode_wb_timer);
1731
1732 rcu_read_unlock();
1733}
1734#endif
1735
1736/*
1737 * If ratelimit_pages is too high then we can get into dirty-data overload
1738 * if a large number of processes all perform writes at the same time.
1739 * If it is too low then SMP machines will call the (expensive)
1740 * get_writeback_state too often.
1741 *
1742 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1743 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1744 * thresholds.
1745 */
1746
1747void writeback_set_ratelimit(void)
1748{
1749 unsigned long background_thresh;
1750 unsigned long dirty_thresh;
1751 global_dirty_limits(&background_thresh, &dirty_thresh);
1752 global_dirty_limit = dirty_thresh;
1753 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1754 if (ratelimit_pages < 16)
1755 ratelimit_pages = 16;
1756}
1757
1758static int
1759ratelimit_handler(struct notifier_block *self, unsigned long action,
1760 void *hcpu)
1761{
1762
1763 switch (action & ~CPU_TASKS_FROZEN) {
1764 case CPU_ONLINE:
1765 case CPU_DEAD:
1766 writeback_set_ratelimit();
1767 return NOTIFY_OK;
1768 default:
1769 return NOTIFY_DONE;
1770 }
1771}
1772
1773static struct notifier_block ratelimit_nb = {
1774 .notifier_call = ratelimit_handler,
1775 .next = NULL,
1776};
1777
1778/*
1779 * Called early on to tune the page writeback dirty limits.
1780 *
1781 * We used to scale dirty pages according to how total memory
1782 * related to pages that could be allocated for buffers (by
1783 * comparing nr_free_buffer_pages() to vm_total_pages.
1784 *
1785 * However, that was when we used "dirty_ratio" to scale with
1786 * all memory, and we don't do that any more. "dirty_ratio"
1787 * is now applied to total non-HIGHPAGE memory (by subtracting
1788 * totalhigh_pages from vm_total_pages), and as such we can't
1789 * get into the old insane situation any more where we had
1790 * large amounts of dirty pages compared to a small amount of
1791 * non-HIGHMEM memory.
1792 *
1793 * But we might still want to scale the dirty_ratio by how
1794 * much memory the box has..
1795 */
1796void __init page_writeback_init(void)
1797{
1798 writeback_set_ratelimit();
1799 register_cpu_notifier(&ratelimit_nb);
1800
1801 fprop_global_init(&writeout_completions);
1802}
1803
1804/**
1805 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1806 * @mapping: address space structure to write
1807 * @start: starting page index
1808 * @end: ending page index (inclusive)
1809 *
1810 * This function scans the page range from @start to @end (inclusive) and tags
1811 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1812 * that write_cache_pages (or whoever calls this function) will then use
1813 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1814 * used to avoid livelocking of writeback by a process steadily creating new
1815 * dirty pages in the file (thus it is important for this function to be quick
1816 * so that it can tag pages faster than a dirtying process can create them).
1817 */
1818/*
1819 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1820 */
1821void tag_pages_for_writeback(struct address_space *mapping,
1822 pgoff_t start, pgoff_t end)
1823{
1824#define WRITEBACK_TAG_BATCH 4096
1825 unsigned long tagged;
1826
1827 do {
1828 spin_lock_irq(&mapping->tree_lock);
1829 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1830 &start, end, WRITEBACK_TAG_BATCH,
1831 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1832 spin_unlock_irq(&mapping->tree_lock);
1833 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1834 cond_resched();
1835 /* We check 'start' to handle wrapping when end == ~0UL */
1836 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1837}
1838EXPORT_SYMBOL(tag_pages_for_writeback);
1839
1840/**
1841 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1842 * @mapping: address space structure to write
1843 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1844 * @writepage: function called for each page
1845 * @data: data passed to writepage function
1846 *
1847 * If a page is already under I/O, write_cache_pages() skips it, even
1848 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1849 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1850 * and msync() need to guarantee that all the data which was dirty at the time
1851 * the call was made get new I/O started against them. If wbc->sync_mode is
1852 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1853 * existing IO to complete.
1854 *
1855 * To avoid livelocks (when other process dirties new pages), we first tag
1856 * pages which should be written back with TOWRITE tag and only then start
1857 * writing them. For data-integrity sync we have to be careful so that we do
1858 * not miss some pages (e.g., because some other process has cleared TOWRITE
1859 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1860 * by the process clearing the DIRTY tag (and submitting the page for IO).
1861 */
1862int write_cache_pages(struct address_space *mapping,
1863 struct writeback_control *wbc, writepage_t writepage,
1864 void *data)
1865{
1866 int ret = 0;
1867 int done = 0;
1868 struct pagevec pvec;
1869 int nr_pages;
1870 pgoff_t uninitialized_var(writeback_index);
1871 pgoff_t index;
1872 pgoff_t end; /* Inclusive */
1873 pgoff_t done_index;
1874 int cycled;
1875 int range_whole = 0;
1876 int tag;
1877
1878 pagevec_init(&pvec, 0);
1879 if (wbc->range_cyclic) {
1880 writeback_index = mapping->writeback_index; /* prev offset */
1881 index = writeback_index;
1882 if (index == 0)
1883 cycled = 1;
1884 else
1885 cycled = 0;
1886 end = -1;
1887 } else {
1888 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1889 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1890 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1891 range_whole = 1;
1892 cycled = 1; /* ignore range_cyclic tests */
1893 }
1894 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1895 tag = PAGECACHE_TAG_TOWRITE;
1896 else
1897 tag = PAGECACHE_TAG_DIRTY;
1898retry:
1899 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1900 tag_pages_for_writeback(mapping, index, end);
1901 done_index = index;
1902 while (!done && (index <= end)) {
1903 int i;
1904
1905 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1906 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1907 if (nr_pages == 0)
1908 break;
1909
1910 for (i = 0; i < nr_pages; i++) {
1911 struct page *page = pvec.pages[i];
1912
1913 /*
1914 * At this point, the page may be truncated or
1915 * invalidated (changing page->mapping to NULL), or
1916 * even swizzled back from swapper_space to tmpfs file
1917 * mapping. However, page->index will not change
1918 * because we have a reference on the page.
1919 */
1920 if (page->index > end) {
1921 /*
1922 * can't be range_cyclic (1st pass) because
1923 * end == -1 in that case.
1924 */
1925 done = 1;
1926 break;
1927 }
1928
1929 done_index = page->index;
1930
1931 lock_page(page);
1932
1933 /*
1934 * Page truncated or invalidated. We can freely skip it
1935 * then, even for data integrity operations: the page
1936 * has disappeared concurrently, so there could be no
1937 * real expectation of this data interity operation
1938 * even if there is now a new, dirty page at the same
1939 * pagecache address.
1940 */
1941 if (unlikely(page->mapping != mapping)) {
1942continue_unlock:
1943 unlock_page(page);
1944 continue;
1945 }
1946
1947 if (!PageDirty(page)) {
1948 /* someone wrote it for us */
1949 goto continue_unlock;
1950 }
1951
1952 if (PageWriteback(page)) {
1953 if (wbc->sync_mode != WB_SYNC_NONE)
1954 wait_on_page_writeback(page);
1955 else
1956 goto continue_unlock;
1957 }
1958
1959 BUG_ON(PageWriteback(page));
1960 if (!clear_page_dirty_for_io(page))
1961 goto continue_unlock;
1962
1963 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1964 ret = (*writepage)(page, wbc, data);
1965 if (unlikely(ret)) {
1966 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1967 unlock_page(page);
1968 ret = 0;
1969 } else {
1970 /*
1971 * done_index is set past this page,
1972 * so media errors will not choke
1973 * background writeout for the entire
1974 * file. This has consequences for
1975 * range_cyclic semantics (ie. it may
1976 * not be suitable for data integrity
1977 * writeout).
1978 */
1979 done_index = page->index + 1;
1980 done = 1;
1981 break;
1982 }
1983 }
1984
1985 /*
1986 * We stop writing back only if we are not doing
1987 * integrity sync. In case of integrity sync we have to
1988 * keep going until we have written all the pages
1989 * we tagged for writeback prior to entering this loop.
1990 */
1991 if (--wbc->nr_to_write <= 0 &&
1992 wbc->sync_mode == WB_SYNC_NONE) {
1993 done = 1;
1994 break;
1995 }
1996 }
1997 pagevec_release(&pvec);
1998 cond_resched();
1999 }
2000 if (!cycled && !done) {
2001 /*
2002 * range_cyclic:
2003 * We hit the last page and there is more work to be done: wrap
2004 * back to the start of the file
2005 */
2006 cycled = 1;
2007 index = 0;
2008 end = writeback_index - 1;
2009 goto retry;
2010 }
2011 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2012 mapping->writeback_index = done_index;
2013
2014 return ret;
2015}
2016EXPORT_SYMBOL(write_cache_pages);
2017
2018/*
2019 * Function used by generic_writepages to call the real writepage
2020 * function and set the mapping flags on error
2021 */
2022static int __writepage(struct page *page, struct writeback_control *wbc,
2023 void *data)
2024{
2025 struct address_space *mapping = data;
2026 int ret = mapping->a_ops->writepage(page, wbc);
2027 mapping_set_error(mapping, ret);
2028 return ret;
2029}
2030
2031/**
2032 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2033 * @mapping: address space structure to write
2034 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2035 *
2036 * This is a library function, which implements the writepages()
2037 * address_space_operation.
2038 */
2039int generic_writepages(struct address_space *mapping,
2040 struct writeback_control *wbc)
2041{
2042 struct blk_plug plug;
2043 int ret;
2044
2045 /* deal with chardevs and other special file */
2046 if (!mapping->a_ops->writepage)
2047 return 0;
2048
2049 blk_start_plug(&plug);
2050 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2051 blk_finish_plug(&plug);
2052 return ret;
2053}
2054
2055EXPORT_SYMBOL(generic_writepages);
2056
2057int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2058{
2059 int ret;
2060
2061 if (wbc->nr_to_write <= 0)
2062 return 0;
2063 if (mapping->a_ops->writepages)
2064 ret = mapping->a_ops->writepages(mapping, wbc);
2065 else
2066 ret = generic_writepages(mapping, wbc);
2067 return ret;
2068}
2069
2070/**
2071 * write_one_page - write out a single page and optionally wait on I/O
2072 * @page: the page to write
2073 * @wait: if true, wait on writeout
2074 *
2075 * The page must be locked by the caller and will be unlocked upon return.
2076 *
2077 * write_one_page() returns a negative error code if I/O failed.
2078 */
2079int write_one_page(struct page *page, int wait)
2080{
2081 struct address_space *mapping = page->mapping;
2082 int ret = 0;
2083 struct writeback_control wbc = {
2084 .sync_mode = WB_SYNC_ALL,
2085 .nr_to_write = 1,
2086 };
2087
2088 BUG_ON(!PageLocked(page));
2089
2090 if (wait)
2091 wait_on_page_writeback(page);
2092
2093 if (clear_page_dirty_for_io(page)) {
2094 page_cache_get(page);
2095 ret = mapping->a_ops->writepage(page, &wbc);
2096 if (ret == 0 && wait) {
2097 wait_on_page_writeback(page);
2098 if (PageError(page))
2099 ret = -EIO;
2100 }
2101 page_cache_release(page);
2102 } else {
2103 unlock_page(page);
2104 }
2105 return ret;
2106}
2107EXPORT_SYMBOL(write_one_page);
2108
2109/*
2110 * For address_spaces which do not use buffers nor write back.
2111 */
2112int __set_page_dirty_no_writeback(struct page *page)
2113{
2114 if (!PageDirty(page))
2115 return !TestSetPageDirty(page);
2116 return 0;
2117}
2118
2119/*
2120 * Helper function for set_page_dirty family.
2121 * NOTE: This relies on being atomic wrt interrupts.
2122 */
2123void account_page_dirtied(struct page *page, struct address_space *mapping)
2124{
2125 trace_writeback_dirty_page(page, mapping);
2126
2127 if (mapping_cap_account_dirty(mapping)) {
2128 __inc_zone_page_state(page, NR_FILE_DIRTY);
2129 __inc_zone_page_state(page, NR_DIRTIED);
2130 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
2131 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2132 task_io_account_write(PAGE_CACHE_SIZE);
2133 current->nr_dirtied++;
2134 this_cpu_inc(bdp_ratelimits);
2135 }
2136}
2137EXPORT_SYMBOL(account_page_dirtied);
2138
2139/*
2140 * Helper function for set_page_writeback family.
2141 *
2142 * The caller must hold mem_cgroup_begin/end_update_page_stat() lock
2143 * while calling this function.
2144 * See test_set_page_writeback for example.
2145 *
2146 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2147 * wrt interrupts.
2148 */
2149void account_page_writeback(struct page *page)
2150{
2151 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2152 inc_zone_page_state(page, NR_WRITEBACK);
2153}
2154EXPORT_SYMBOL(account_page_writeback);
2155
2156/*
2157 * For address_spaces which do not use buffers. Just tag the page as dirty in
2158 * its radix tree.
2159 *
2160 * This is also used when a single buffer is being dirtied: we want to set the
2161 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2162 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2163 *
2164 * Most callers have locked the page, which pins the address_space in memory.
2165 * But zap_pte_range() does not lock the page, however in that case the
2166 * mapping is pinned by the vma's ->vm_file reference.
2167 *
2168 * We take care to handle the case where the page was truncated from the
2169 * mapping by re-checking page_mapping() inside tree_lock.
2170 */
2171int __set_page_dirty_nobuffers(struct page *page)
2172{
2173 if (!TestSetPageDirty(page)) {
2174 struct address_space *mapping = page_mapping(page);
2175 struct address_space *mapping2;
2176 unsigned long flags;
2177
2178 if (!mapping)
2179 return 1;
2180
2181 spin_lock_irqsave(&mapping->tree_lock, flags);
2182 mapping2 = page_mapping(page);
2183 if (mapping2) { /* Race with truncate? */
2184 BUG_ON(mapping2 != mapping);
2185 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2186 account_page_dirtied(page, mapping);
2187 radix_tree_tag_set(&mapping->page_tree,
2188 page_index(page), PAGECACHE_TAG_DIRTY);
2189 }
2190 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2191 if (mapping->host) {
2192 /* !PageAnon && !swapper_space */
2193 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2194 }
2195 return 1;
2196 }
2197 return 0;
2198}
2199EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2200
2201/*
2202 * Call this whenever redirtying a page, to de-account the dirty counters
2203 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2204 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2205 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2206 * control.
2207 */
2208void account_page_redirty(struct page *page)
2209{
2210 struct address_space *mapping = page->mapping;
2211 if (mapping && mapping_cap_account_dirty(mapping)) {
2212 current->nr_dirtied--;
2213 dec_zone_page_state(page, NR_DIRTIED);
2214 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2215 }
2216}
2217EXPORT_SYMBOL(account_page_redirty);
2218
2219/*
2220 * When a writepage implementation decides that it doesn't want to write this
2221 * page for some reason, it should redirty the locked page via
2222 * redirty_page_for_writepage() and it should then unlock the page and return 0
2223 */
2224int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2225{
2226 wbc->pages_skipped++;
2227 account_page_redirty(page);
2228 return __set_page_dirty_nobuffers(page);
2229}
2230EXPORT_SYMBOL(redirty_page_for_writepage);
2231
2232/*
2233 * Dirty a page.
2234 *
2235 * For pages with a mapping this should be done under the page lock
2236 * for the benefit of asynchronous memory errors who prefer a consistent
2237 * dirty state. This rule can be broken in some special cases,
2238 * but should be better not to.
2239 *
2240 * If the mapping doesn't provide a set_page_dirty a_op, then
2241 * just fall through and assume that it wants buffer_heads.
2242 */
2243int set_page_dirty(struct page *page)
2244{
2245 struct address_space *mapping = page_mapping(page);
2246
2247 if (likely(mapping)) {
2248 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2249 /*
2250 * readahead/lru_deactivate_page could remain
2251 * PG_readahead/PG_reclaim due to race with end_page_writeback
2252 * About readahead, if the page is written, the flags would be
2253 * reset. So no problem.
2254 * About lru_deactivate_page, if the page is redirty, the flag
2255 * will be reset. So no problem. but if the page is used by readahead
2256 * it will confuse readahead and make it restart the size rampup
2257 * process. But it's a trivial problem.
2258 */
2259 ClearPageReclaim(page);
2260#ifdef CONFIG_BLOCK
2261 if (!spd)
2262 spd = __set_page_dirty_buffers;
2263#endif
2264 return (*spd)(page);
2265 }
2266 if (!PageDirty(page)) {
2267 if (!TestSetPageDirty(page))
2268 return 1;
2269 }
2270 return 0;
2271}
2272EXPORT_SYMBOL(set_page_dirty);
2273
2274/*
2275 * set_page_dirty() is racy if the caller has no reference against
2276 * page->mapping->host, and if the page is unlocked. This is because another
2277 * CPU could truncate the page off the mapping and then free the mapping.
2278 *
2279 * Usually, the page _is_ locked, or the caller is a user-space process which
2280 * holds a reference on the inode by having an open file.
2281 *
2282 * In other cases, the page should be locked before running set_page_dirty().
2283 */
2284int set_page_dirty_lock(struct page *page)
2285{
2286 int ret;
2287
2288 lock_page(page);
2289 ret = set_page_dirty(page);
2290 unlock_page(page);
2291 return ret;
2292}
2293EXPORT_SYMBOL(set_page_dirty_lock);
2294
2295/*
2296 * Clear a page's dirty flag, while caring for dirty memory accounting.
2297 * Returns true if the page was previously dirty.
2298 *
2299 * This is for preparing to put the page under writeout. We leave the page
2300 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2301 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2302 * implementation will run either set_page_writeback() or set_page_dirty(),
2303 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2304 * back into sync.
2305 *
2306 * This incoherency between the page's dirty flag and radix-tree tag is
2307 * unfortunate, but it only exists while the page is locked.
2308 */
2309int clear_page_dirty_for_io(struct page *page)
2310{
2311 struct address_space *mapping = page_mapping(page);
2312
2313 BUG_ON(!PageLocked(page));
2314
2315 if (mapping && mapping_cap_account_dirty(mapping)) {
2316 /*
2317 * Yes, Virginia, this is indeed insane.
2318 *
2319 * We use this sequence to make sure that
2320 * (a) we account for dirty stats properly
2321 * (b) we tell the low-level filesystem to
2322 * mark the whole page dirty if it was
2323 * dirty in a pagetable. Only to then
2324 * (c) clean the page again and return 1 to
2325 * cause the writeback.
2326 *
2327 * This way we avoid all nasty races with the
2328 * dirty bit in multiple places and clearing
2329 * them concurrently from different threads.
2330 *
2331 * Note! Normally the "set_page_dirty(page)"
2332 * has no effect on the actual dirty bit - since
2333 * that will already usually be set. But we
2334 * need the side effects, and it can help us
2335 * avoid races.
2336 *
2337 * We basically use the page "master dirty bit"
2338 * as a serialization point for all the different
2339 * threads doing their things.
2340 */
2341 if (page_mkclean(page))
2342 set_page_dirty(page);
2343 /*
2344 * We carefully synchronise fault handlers against
2345 * installing a dirty pte and marking the page dirty
2346 * at this point. We do this by having them hold the
2347 * page lock at some point after installing their
2348 * pte, but before marking the page dirty.
2349 * Pages are always locked coming in here, so we get
2350 * the desired exclusion. See mm/memory.c:do_wp_page()
2351 * for more comments.
2352 */
2353 if (TestClearPageDirty(page)) {
2354 dec_zone_page_state(page, NR_FILE_DIRTY);
2355 dec_bdi_stat(mapping->backing_dev_info,
2356 BDI_RECLAIMABLE);
2357 return 1;
2358 }
2359 return 0;
2360 }
2361 return TestClearPageDirty(page);
2362}
2363EXPORT_SYMBOL(clear_page_dirty_for_io);
2364
2365int test_clear_page_writeback(struct page *page)
2366{
2367 struct address_space *mapping = page_mapping(page);
2368 int ret;
2369 bool locked;
2370 unsigned long memcg_flags;
2371
2372 mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2373 if (mapping) {
2374 struct backing_dev_info *bdi = mapping->backing_dev_info;
2375 unsigned long flags;
2376
2377 spin_lock_irqsave(&mapping->tree_lock, flags);
2378 ret = TestClearPageWriteback(page);
2379 if (ret) {
2380 radix_tree_tag_clear(&mapping->page_tree,
2381 page_index(page),
2382 PAGECACHE_TAG_WRITEBACK);
2383 if (bdi_cap_account_writeback(bdi)) {
2384 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2385 __bdi_writeout_inc(bdi);
2386 }
2387 }
2388 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2389 } else {
2390 ret = TestClearPageWriteback(page);
2391 }
2392 if (ret) {
2393 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2394 dec_zone_page_state(page, NR_WRITEBACK);
2395 inc_zone_page_state(page, NR_WRITTEN);
2396 }
2397 mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2398 return ret;
2399}
2400
2401int test_set_page_writeback(struct page *page)
2402{
2403 struct address_space *mapping = page_mapping(page);
2404 int ret;
2405 bool locked;
2406 unsigned long memcg_flags;
2407
2408 mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2409 if (mapping) {
2410 struct backing_dev_info *bdi = mapping->backing_dev_info;
2411 unsigned long flags;
2412
2413 spin_lock_irqsave(&mapping->tree_lock, flags);
2414 ret = TestSetPageWriteback(page);
2415 if (!ret) {
2416 radix_tree_tag_set(&mapping->page_tree,
2417 page_index(page),
2418 PAGECACHE_TAG_WRITEBACK);
2419 if (bdi_cap_account_writeback(bdi))
2420 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2421 }
2422 if (!PageDirty(page))
2423 radix_tree_tag_clear(&mapping->page_tree,
2424 page_index(page),
2425 PAGECACHE_TAG_DIRTY);
2426 radix_tree_tag_clear(&mapping->page_tree,
2427 page_index(page),
2428 PAGECACHE_TAG_TOWRITE);
2429 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2430 } else {
2431 ret = TestSetPageWriteback(page);
2432 }
2433 if (!ret)
2434 account_page_writeback(page);
2435 mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2436 return ret;
2437
2438}
2439EXPORT_SYMBOL(test_set_page_writeback);
2440
2441/*
2442 * Return true if any of the pages in the mapping are marked with the
2443 * passed tag.
2444 */
2445int mapping_tagged(struct address_space *mapping, int tag)
2446{
2447 return radix_tree_tagged(&mapping->page_tree, tag);
2448}
2449EXPORT_SYMBOL(mapping_tagged);
2450
2451/**
2452 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2453 * @page: The page to wait on.
2454 *
2455 * This function determines if the given page is related to a backing device
2456 * that requires page contents to be held stable during writeback. If so, then
2457 * it will wait for any pending writeback to complete.
2458 */
2459void wait_for_stable_page(struct page *page)
2460{
2461 struct address_space *mapping = page_mapping(page);
2462 struct backing_dev_info *bdi = mapping->backing_dev_info;
2463
2464 if (!bdi_cap_stable_pages_required(bdi))
2465 return;
2466
2467 wait_on_page_writeback(page);
2468}
2469EXPORT_SYMBOL_GPL(wait_for_stable_page);