1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * mm/page-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
11 * 10Apr2002 Andrew Morton
12 * Initial version
13 */
14
15 #include <linux/kernel.h>
16 #include <linux/export.h>
17 #include <linux/spinlock.h>
18 #include <linux/fs.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/writeback.h>
24 #include <linux/init.h>
25 #include <linux/backing-dev.h>
26 #include <linux/task_io_accounting_ops.h>
27 #include <linux/blkdev.h>
28 #include <linux/mpage.h>
29 #include <linux/rmap.h>
30 #include <linux/percpu.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/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 */
67 static 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 */
74 int 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 */
80 unsigned 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 */
86 int vm_highmem_is_dirtyable;
87
88 /*
89 * The generator of dirty data starts writeback at this percentage
90 */
91 int 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 */
97 unsigned long vm_dirty_bytes;
98
99 /*
100 * The interval between `kupdate'-style writebacks
101 */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106 /*
107 * The longest time for which data is allowed to remain dirty
108 */
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111 /*
112 * Flag that makes the machine dump writes/reads and block dirtyings.
113 */
114 int 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 */
120 int laptop_mode;
121
122 EXPORT_SYMBOL(laptop_mode);
123
124 /* End of sysctl-exported parameters */
125
126 struct wb_domain global_wb_domain;
127
128 /* consolidated parameters for balance_dirty_pages() and its subroutines */
129 struct 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
169 static bool mdtc_valid(struct dirty_throttle_control *dtc)
170 {
171 return dtc->dom;
172 }
173
174 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175 {
176 return dtc->dom;
177 }
178
179 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180 {
181 return mdtc->gdtc;
182 }
183
184 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185 {
186 return &wb->memcg_completions;
187 }
188
189 static 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 min = div64_ul(min, tot_bw);
205 }
206 if (max < 100) {
207 max *= this_bw;
208 max = div64_ul(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
223 static bool mdtc_valid(struct dirty_throttle_control *dtc)
224 {
225 return false;
226 }
227
228 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229 {
230 return &global_wb_domain;
231 }
232
233 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234 {
235 return NULL;
236 }
237
238 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239 {
240 return NULL;
241 }
242
243 static 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 * Return: 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 */
277 static 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
304 static 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 * Return: the global number of pages potentially available for dirty
360 * page cache. This is the base value for the global dirty limits.
361 */
362 static 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 */
393 static 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 */
459 void 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 * Return: the maximum number of dirty pages allowed in a node, based
475 * on the node's dirtyable memory.
476 */
477 static 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 * Return: %true when the dirty pages in @pgdat are within the node's
500 * dirty limit, %false if the limit is exceeded.
501 */
502 bool 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
514 int 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
526 int 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
538 int 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
553 int 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
568 static 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
577 static 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 */
600 static 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
614 void 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 }
622 EXPORT_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 */
628 static 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
647 int 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
661 void 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 */
673 static unsigned int bdi_min_ratio;
674
675 int 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
696 int 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 }
714 EXPORT_SYMBOL(bdi_set_max_ratio);
715
716 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
717 unsigned long bg_thresh)
718 {
719 return (thresh + bg_thresh) / 2;
720 }
721
722 static 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 */
732 static 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 * Note that balance_dirty_pages() will only seriously take it as a hard limit
748 * when sleeping max_pause per page is not enough to keep the dirty pages under
749 * control. For example, when the device is completely stalled due to some error
750 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
751 * In the other normal situations, it acts more gently by throttling the tasks
752 * more (rather than completely block them) when the wb dirty pages go high.
753 *
754 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
755 * - starving fast devices
756 * - piling up dirty pages (that will take long time to sync) on slow devices
757 *
758 * The wb's share of dirty limit will be adapting to its throughput and
759 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
760 *
761 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
762 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
763 */
764 static 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
791 unsigned 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 */
812 static 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 */
904 static 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
1083 static 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
1121 out:
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
1133 static 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;
1158 update:
1159 dom->dirty_limit = limit;
1160 }
1161
1162 static 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 */
1187 static 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
1345 static 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
1389 snapshot:
1390 wb->dirtied_stamp = dirtied;
1391 wb->written_stamp = written;
1392 wb->bw_time_stamp = now;
1393 }
1394
1395 void 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 */
1410 static 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
1419 static 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
1438 static 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
1513 static 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 */
1561 static 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 mem_cgroup_flush_foreign(wb);
1671
1672 /*
1673 * Calculate global domain's pos_ratio and select the
1674 * global dtc by default.
1675 */
1676 if (!strictlimit)
1677 wb_dirty_limits(gdtc);
1678
1679 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1680 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1681
1682 wb_position_ratio(gdtc);
1683 sdtc = gdtc;
1684
1685 if (mdtc) {
1686 /*
1687 * If memcg domain is in effect, calculate its
1688 * pos_ratio. @wb should satisfy constraints from
1689 * both global and memcg domains. Choose the one
1690 * w/ lower pos_ratio.
1691 */
1692 if (!strictlimit)
1693 wb_dirty_limits(mdtc);
1694
1695 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1696 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1697
1698 wb_position_ratio(mdtc);
1699 if (mdtc->pos_ratio < gdtc->pos_ratio)
1700 sdtc = mdtc;
1701 }
1702
1703 if (dirty_exceeded && !wb->dirty_exceeded)
1704 wb->dirty_exceeded = 1;
1705
1706 if (time_is_before_jiffies(wb->bw_time_stamp +
1707 BANDWIDTH_INTERVAL)) {
1708 spin_lock(&wb->list_lock);
1709 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1710 spin_unlock(&wb->list_lock);
1711 }
1712
1713 /* throttle according to the chosen dtc */
1714 dirty_ratelimit = wb->dirty_ratelimit;
1715 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1716 RATELIMIT_CALC_SHIFT;
1717 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1718 min_pause = wb_min_pause(wb, max_pause,
1719 task_ratelimit, dirty_ratelimit,
1720 &nr_dirtied_pause);
1721
1722 if (unlikely(task_ratelimit == 0)) {
1723 period = max_pause;
1724 pause = max_pause;
1725 goto pause;
1726 }
1727 period = HZ * pages_dirtied / task_ratelimit;
1728 pause = period;
1729 if (current->dirty_paused_when)
1730 pause -= now - current->dirty_paused_when;
1731 /*
1732 * For less than 1s think time (ext3/4 may block the dirtier
1733 * for up to 800ms from time to time on 1-HDD; so does xfs,
1734 * however at much less frequency), try to compensate it in
1735 * future periods by updating the virtual time; otherwise just
1736 * do a reset, as it may be a light dirtier.
1737 */
1738 if (pause < min_pause) {
1739 trace_balance_dirty_pages(wb,
1740 sdtc->thresh,
1741 sdtc->bg_thresh,
1742 sdtc->dirty,
1743 sdtc->wb_thresh,
1744 sdtc->wb_dirty,
1745 dirty_ratelimit,
1746 task_ratelimit,
1747 pages_dirtied,
1748 period,
1749 min(pause, 0L),
1750 start_time);
1751 if (pause < -HZ) {
1752 current->dirty_paused_when = now;
1753 current->nr_dirtied = 0;
1754 } else if (period) {
1755 current->dirty_paused_when += period;
1756 current->nr_dirtied = 0;
1757 } else if (current->nr_dirtied_pause <= pages_dirtied)
1758 current->nr_dirtied_pause += pages_dirtied;
1759 break;
1760 }
1761 if (unlikely(pause > max_pause)) {
1762 /* for occasional dropped task_ratelimit */
1763 now += min(pause - max_pause, max_pause);
1764 pause = max_pause;
1765 }
1766
1767 pause:
1768 trace_balance_dirty_pages(wb,
1769 sdtc->thresh,
1770 sdtc->bg_thresh,
1771 sdtc->dirty,
1772 sdtc->wb_thresh,
1773 sdtc->wb_dirty,
1774 dirty_ratelimit,
1775 task_ratelimit,
1776 pages_dirtied,
1777 period,
1778 pause,
1779 start_time);
1780 __set_current_state(TASK_KILLABLE);
1781 wb->dirty_sleep = now;
1782 io_schedule_timeout(pause);
1783
1784 current->dirty_paused_when = now + pause;
1785 current->nr_dirtied = 0;
1786 current->nr_dirtied_pause = nr_dirtied_pause;
1787
1788 /*
1789 * This is typically equal to (dirty < thresh) and can also
1790 * keep "1000+ dd on a slow USB stick" under control.
1791 */
1792 if (task_ratelimit)
1793 break;
1794
1795 /*
1796 * In the case of an unresponding NFS server and the NFS dirty
1797 * pages exceeds dirty_thresh, give the other good wb's a pipe
1798 * to go through, so that tasks on them still remain responsive.
1799 *
1800 * In theory 1 page is enough to keep the consumer-producer
1801 * pipe going: the flusher cleans 1 page => the task dirties 1
1802 * more page. However wb_dirty has accounting errors. So use
1803 * the larger and more IO friendly wb_stat_error.
1804 */
1805 if (sdtc->wb_dirty <= wb_stat_error())
1806 break;
1807
1808 if (fatal_signal_pending(current))
1809 break;
1810 }
1811
1812 if (!dirty_exceeded && wb->dirty_exceeded)
1813 wb->dirty_exceeded = 0;
1814
1815 if (writeback_in_progress(wb))
1816 return;
1817
1818 /*
1819 * In laptop mode, we wait until hitting the higher threshold before
1820 * starting background writeout, and then write out all the way down
1821 * to the lower threshold. So slow writers cause minimal disk activity.
1822 *
1823 * In normal mode, we start background writeout at the lower
1824 * background_thresh, to keep the amount of dirty memory low.
1825 */
1826 if (laptop_mode)
1827 return;
1828
1829 if (nr_reclaimable > gdtc->bg_thresh)
1830 wb_start_background_writeback(wb);
1831 }
1832
1833 static DEFINE_PER_CPU(int, bdp_ratelimits);
1834
1835 /*
1836 * Normal tasks are throttled by
1837 * loop {
1838 * dirty tsk->nr_dirtied_pause pages;
1839 * take a snap in balance_dirty_pages();
1840 * }
1841 * However there is a worst case. If every task exit immediately when dirtied
1842 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1843 * called to throttle the page dirties. The solution is to save the not yet
1844 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1845 * randomly into the running tasks. This works well for the above worst case,
1846 * as the new task will pick up and accumulate the old task's leaked dirty
1847 * count and eventually get throttled.
1848 */
1849 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1850
1851 /**
1852 * balance_dirty_pages_ratelimited - balance dirty memory state
1853 * @mapping: address_space which was dirtied
1854 *
1855 * Processes which are dirtying memory should call in here once for each page
1856 * which was newly dirtied. The function will periodically check the system's
1857 * dirty state and will initiate writeback if needed.
1858 *
1859 * On really big machines, get_writeback_state is expensive, so try to avoid
1860 * calling it too often (ratelimiting). But once we're over the dirty memory
1861 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1862 * from overshooting the limit by (ratelimit_pages) each.
1863 */
1864 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1865 {
1866 struct inode *inode = mapping->host;
1867 struct backing_dev_info *bdi = inode_to_bdi(inode);
1868 struct bdi_writeback *wb = NULL;
1869 int ratelimit;
1870 int *p;
1871
1872 if (!bdi_cap_account_dirty(bdi))
1873 return;
1874
1875 if (inode_cgwb_enabled(inode))
1876 wb = wb_get_create_current(bdi, GFP_KERNEL);
1877 if (!wb)
1878 wb = &bdi->wb;
1879
1880 ratelimit = current->nr_dirtied_pause;
1881 if (wb->dirty_exceeded)
1882 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1883
1884 preempt_disable();
1885 /*
1886 * This prevents one CPU to accumulate too many dirtied pages without
1887 * calling into balance_dirty_pages(), which can happen when there are
1888 * 1000+ tasks, all of them start dirtying pages at exactly the same
1889 * time, hence all honoured too large initial task->nr_dirtied_pause.
1890 */
1891 p = this_cpu_ptr(&bdp_ratelimits);
1892 if (unlikely(current->nr_dirtied >= ratelimit))
1893 *p = 0;
1894 else if (unlikely(*p >= ratelimit_pages)) {
1895 *p = 0;
1896 ratelimit = 0;
1897 }
1898 /*
1899 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1900 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1901 * the dirty throttling and livelock other long-run dirtiers.
1902 */
1903 p = this_cpu_ptr(&dirty_throttle_leaks);
1904 if (*p > 0 && current->nr_dirtied < ratelimit) {
1905 unsigned long nr_pages_dirtied;
1906 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1907 *p -= nr_pages_dirtied;
1908 current->nr_dirtied += nr_pages_dirtied;
1909 }
1910 preempt_enable();
1911
1912 if (unlikely(current->nr_dirtied >= ratelimit))
1913 balance_dirty_pages(wb, current->nr_dirtied);
1914
1915 wb_put(wb);
1916 }
1917 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1918
1919 /**
1920 * wb_over_bg_thresh - does @wb need to be written back?
1921 * @wb: bdi_writeback of interest
1922 *
1923 * Determines whether background writeback should keep writing @wb or it's
1924 * clean enough.
1925 *
1926 * Return: %true if writeback should continue.
1927 */
1928 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1929 {
1930 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1931 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1932 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1933 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1934 &mdtc_stor : NULL;
1935
1936 /*
1937 * Similar to balance_dirty_pages() but ignores pages being written
1938 * as we're trying to decide whether to put more under writeback.
1939 */
1940 gdtc->avail = global_dirtyable_memory();
1941 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1942 global_node_page_state(NR_UNSTABLE_NFS);
1943 domain_dirty_limits(gdtc);
1944
1945 if (gdtc->dirty > gdtc->bg_thresh)
1946 return true;
1947
1948 if (wb_stat(wb, WB_RECLAIMABLE) >
1949 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1950 return true;
1951
1952 if (mdtc) {
1953 unsigned long filepages, headroom, writeback;
1954
1955 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1956 &writeback);
1957 mdtc_calc_avail(mdtc, filepages, headroom);
1958 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1959
1960 if (mdtc->dirty > mdtc->bg_thresh)
1961 return true;
1962
1963 if (wb_stat(wb, WB_RECLAIMABLE) >
1964 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1965 return true;
1966 }
1967
1968 return false;
1969 }
1970
1971 /*
1972 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1973 */
1974 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1975 void __user *buffer, size_t *length, loff_t *ppos)
1976 {
1977 unsigned int old_interval = dirty_writeback_interval;
1978 int ret;
1979
1980 ret = proc_dointvec(table, write, buffer, length, ppos);
1981
1982 /*
1983 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1984 * and a different non-zero value will wakeup the writeback threads.
1985 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1986 * iterate over all bdis and wbs.
1987 * The reason we do this is to make the change take effect immediately.
1988 */
1989 if (!ret && write && dirty_writeback_interval &&
1990 dirty_writeback_interval != old_interval)
1991 wakeup_flusher_threads(WB_REASON_PERIODIC);
1992
1993 return ret;
1994 }
1995
1996 #ifdef CONFIG_BLOCK
1997 void laptop_mode_timer_fn(struct timer_list *t)
1998 {
1999 struct backing_dev_info *backing_dev_info =
2000 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2001
2002 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2003 }
2004
2005 /*
2006 * We've spun up the disk and we're in laptop mode: schedule writeback
2007 * of all dirty data a few seconds from now. If the flush is already scheduled
2008 * then push it back - the user is still using the disk.
2009 */
2010 void laptop_io_completion(struct backing_dev_info *info)
2011 {
2012 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2013 }
2014
2015 /*
2016 * We're in laptop mode and we've just synced. The sync's writes will have
2017 * caused another writeback to be scheduled by laptop_io_completion.
2018 * Nothing needs to be written back anymore, so we unschedule the writeback.
2019 */
2020 void laptop_sync_completion(void)
2021 {
2022 struct backing_dev_info *bdi;
2023
2024 rcu_read_lock();
2025
2026 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2027 del_timer(&bdi->laptop_mode_wb_timer);
2028
2029 rcu_read_unlock();
2030 }
2031 #endif
2032
2033 /*
2034 * If ratelimit_pages is too high then we can get into dirty-data overload
2035 * if a large number of processes all perform writes at the same time.
2036 * If it is too low then SMP machines will call the (expensive)
2037 * get_writeback_state too often.
2038 *
2039 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2040 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2041 * thresholds.
2042 */
2043
2044 void writeback_set_ratelimit(void)
2045 {
2046 struct wb_domain *dom = &global_wb_domain;
2047 unsigned long background_thresh;
2048 unsigned long dirty_thresh;
2049
2050 global_dirty_limits(&background_thresh, &dirty_thresh);
2051 dom->dirty_limit = dirty_thresh;
2052 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2053 if (ratelimit_pages < 16)
2054 ratelimit_pages = 16;
2055 }
2056
2057 static int page_writeback_cpu_online(unsigned int cpu)
2058 {
2059 writeback_set_ratelimit();
2060 return 0;
2061 }
2062
2063 /*
2064 * Called early on to tune the page writeback dirty limits.
2065 *
2066 * We used to scale dirty pages according to how total memory
2067 * related to pages that could be allocated for buffers (by
2068 * comparing nr_free_buffer_pages() to vm_total_pages.
2069 *
2070 * However, that was when we used "dirty_ratio" to scale with
2071 * all memory, and we don't do that any more. "dirty_ratio"
2072 * is now applied to total non-HIGHPAGE memory (by subtracting
2073 * totalhigh_pages from vm_total_pages), and as such we can't
2074 * get into the old insane situation any more where we had
2075 * large amounts of dirty pages compared to a small amount of
2076 * non-HIGHMEM memory.
2077 *
2078 * But we might still want to scale the dirty_ratio by how
2079 * much memory the box has..
2080 */
2081 void __init page_writeback_init(void)
2082 {
2083 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2084
2085 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2086 page_writeback_cpu_online, NULL);
2087 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2088 page_writeback_cpu_online);
2089 }
2090
2091 /**
2092 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2093 * @mapping: address space structure to write
2094 * @start: starting page index
2095 * @end: ending page index (inclusive)
2096 *
2097 * This function scans the page range from @start to @end (inclusive) and tags
2098 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2099 * that write_cache_pages (or whoever calls this function) will then use
2100 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2101 * used to avoid livelocking of writeback by a process steadily creating new
2102 * dirty pages in the file (thus it is important for this function to be quick
2103 * so that it can tag pages faster than a dirtying process can create them).
2104 */
2105 void tag_pages_for_writeback(struct address_space *mapping,
2106 pgoff_t start, pgoff_t end)
2107 {
2108 XA_STATE(xas, &mapping->i_pages, start);
2109 unsigned int tagged = 0;
2110 void *page;
2111
2112 xas_lock_irq(&xas);
2113 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2114 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2115 if (++tagged % XA_CHECK_SCHED)
2116 continue;
2117
2118 xas_pause(&xas);
2119 xas_unlock_irq(&xas);
2120 cond_resched();
2121 xas_lock_irq(&xas);
2122 }
2123 xas_unlock_irq(&xas);
2124 }
2125 EXPORT_SYMBOL(tag_pages_for_writeback);
2126
2127 /**
2128 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2129 * @mapping: address space structure to write
2130 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2131 * @writepage: function called for each page
2132 * @data: data passed to writepage function
2133 *
2134 * If a page is already under I/O, write_cache_pages() skips it, even
2135 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2136 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2137 * and msync() need to guarantee that all the data which was dirty at the time
2138 * the call was made get new I/O started against them. If wbc->sync_mode is
2139 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2140 * existing IO to complete.
2141 *
2142 * To avoid livelocks (when other process dirties new pages), we first tag
2143 * pages which should be written back with TOWRITE tag and only then start
2144 * writing them. For data-integrity sync we have to be careful so that we do
2145 * not miss some pages (e.g., because some other process has cleared TOWRITE
2146 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2147 * by the process clearing the DIRTY tag (and submitting the page for IO).
2148 *
2149 * To avoid deadlocks between range_cyclic writeback and callers that hold
2150 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2151 * we do not loop back to the start of the file. Doing so causes a page
2152 * lock/page writeback access order inversion - we should only ever lock
2153 * multiple pages in ascending page->index order, and looping back to the start
2154 * of the file violates that rule and causes deadlocks.
2155 *
2156 * Return: %0 on success, negative error code otherwise
2157 */
2158 int write_cache_pages(struct address_space *mapping,
2159 struct writeback_control *wbc, writepage_t writepage,
2160 void *data)
2161 {
2162 int ret = 0;
2163 int done = 0;
2164 int error;
2165 struct pagevec pvec;
2166 int nr_pages;
2167 pgoff_t uninitialized_var(writeback_index);
2168 pgoff_t index;
2169 pgoff_t end; /* Inclusive */
2170 pgoff_t done_index;
2171 int range_whole = 0;
2172 xa_mark_t tag;
2173
2174 pagevec_init(&pvec);
2175 if (wbc->range_cyclic) {
2176 writeback_index = mapping->writeback_index; /* prev offset */
2177 index = writeback_index;
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 }
2185 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2186 tag = PAGECACHE_TAG_TOWRITE;
2187 else
2188 tag = PAGECACHE_TAG_DIRTY;
2189 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2190 tag_pages_for_writeback(mapping, index, end);
2191 done_index = index;
2192 while (!done && (index <= end)) {
2193 int i;
2194
2195 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2196 tag);
2197 if (nr_pages == 0)
2198 break;
2199
2200 for (i = 0; i < nr_pages; i++) {
2201 struct page *page = pvec.pages[i];
2202
2203 done_index = page->index;
2204
2205 lock_page(page);
2206
2207 /*
2208 * Page truncated or invalidated. We can freely skip it
2209 * then, even for data integrity operations: the page
2210 * has disappeared concurrently, so there could be no
2211 * real expectation of this data interity operation
2212 * even if there is now a new, dirty page at the same
2213 * pagecache address.
2214 */
2215 if (unlikely(page->mapping != mapping)) {
2216 continue_unlock:
2217 unlock_page(page);
2218 continue;
2219 }
2220
2221 if (!PageDirty(page)) {
2222 /* someone wrote it for us */
2223 goto continue_unlock;
2224 }
2225
2226 if (PageWriteback(page)) {
2227 if (wbc->sync_mode != WB_SYNC_NONE)
2228 wait_on_page_writeback(page);
2229 else
2230 goto continue_unlock;
2231 }
2232
2233 BUG_ON(PageWriteback(page));
2234 if (!clear_page_dirty_for_io(page))
2235 goto continue_unlock;
2236
2237 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2238 error = (*writepage)(page, wbc, data);
2239 if (unlikely(error)) {
2240 /*
2241 * Handle errors according to the type of
2242 * writeback. There's no need to continue for
2243 * background writeback. Just push done_index
2244 * past this page so media errors won't choke
2245 * writeout for the entire file. For integrity
2246 * writeback, we must process the entire dirty
2247 * set regardless of errors because the fs may
2248 * still have state to clear for each page. In
2249 * that case we continue processing and return
2250 * the first error.
2251 */
2252 if (error == AOP_WRITEPAGE_ACTIVATE) {
2253 unlock_page(page);
2254 error = 0;
2255 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2256 ret = error;
2257 done_index = page->index + 1;
2258 done = 1;
2259 break;
2260 }
2261 if (!ret)
2262 ret = error;
2263 }
2264
2265 /*
2266 * We stop writing back only if we are not doing
2267 * integrity sync. In case of integrity sync we have to
2268 * keep going until we have written all the pages
2269 * we tagged for writeback prior to entering this loop.
2270 */
2271 if (--wbc->nr_to_write <= 0 &&
2272 wbc->sync_mode == WB_SYNC_NONE) {
2273 done = 1;
2274 break;
2275 }
2276 }
2277 pagevec_release(&pvec);
2278 cond_resched();
2279 }
2280
2281 /*
2282 * If we hit the last page and there is more work to be done: wrap
2283 * back the index back to the start of the file for the next
2284 * time we are called.
2285 */
2286 if (wbc->range_cyclic && !done)
2287 done_index = 0;
2288 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2289 mapping->writeback_index = done_index;
2290
2291 return ret;
2292 }
2293 EXPORT_SYMBOL(write_cache_pages);
2294
2295 /*
2296 * Function used by generic_writepages to call the real writepage
2297 * function and set the mapping flags on error
2298 */
2299 static int __writepage(struct page *page, struct writeback_control *wbc,
2300 void *data)
2301 {
2302 struct address_space *mapping = data;
2303 int ret = mapping->a_ops->writepage(page, wbc);
2304 mapping_set_error(mapping, ret);
2305 return ret;
2306 }
2307
2308 /**
2309 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2310 * @mapping: address space structure to write
2311 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2312 *
2313 * This is a library function, which implements the writepages()
2314 * address_space_operation.
2315 *
2316 * Return: %0 on success, negative error code otherwise
2317 */
2318 int generic_writepages(struct address_space *mapping,
2319 struct writeback_control *wbc)
2320 {
2321 struct blk_plug plug;
2322 int ret;
2323
2324 /* deal with chardevs and other special file */
2325 if (!mapping->a_ops->writepage)
2326 return 0;
2327
2328 blk_start_plug(&plug);
2329 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2330 blk_finish_plug(&plug);
2331 return ret;
2332 }
2333
2334 EXPORT_SYMBOL(generic_writepages);
2335
2336 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2337 {
2338 int ret;
2339
2340 if (wbc->nr_to_write <= 0)
2341 return 0;
2342 while (1) {
2343 if (mapping->a_ops->writepages)
2344 ret = mapping->a_ops->writepages(mapping, wbc);
2345 else
2346 ret = generic_writepages(mapping, wbc);
2347 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2348 break;
2349 cond_resched();
2350 congestion_wait(BLK_RW_ASYNC, HZ/50);
2351 }
2352 return ret;
2353 }
2354
2355 /**
2356 * write_one_page - write out a single page and wait on I/O
2357 * @page: the page to write
2358 *
2359 * The page must be locked by the caller and will be unlocked upon return.
2360 *
2361 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2362 * function returns.
2363 *
2364 * Return: %0 on success, negative error code otherwise
2365 */
2366 int write_one_page(struct page *page)
2367 {
2368 struct address_space *mapping = page->mapping;
2369 int ret = 0;
2370 struct writeback_control wbc = {
2371 .sync_mode = WB_SYNC_ALL,
2372 .nr_to_write = 1,
2373 };
2374
2375 BUG_ON(!PageLocked(page));
2376
2377 wait_on_page_writeback(page);
2378
2379 if (clear_page_dirty_for_io(page)) {
2380 get_page(page);
2381 ret = mapping->a_ops->writepage(page, &wbc);
2382 if (ret == 0)
2383 wait_on_page_writeback(page);
2384 put_page(page);
2385 } else {
2386 unlock_page(page);
2387 }
2388
2389 if (!ret)
2390 ret = filemap_check_errors(mapping);
2391 return ret;
2392 }
2393 EXPORT_SYMBOL(write_one_page);
2394
2395 /*
2396 * For address_spaces which do not use buffers nor write back.
2397 */
2398 int __set_page_dirty_no_writeback(struct page *page)
2399 {
2400 if (!PageDirty(page))
2401 return !TestSetPageDirty(page);
2402 return 0;
2403 }
2404
2405 /*
2406 * Helper function for set_page_dirty family.
2407 *
2408 * Caller must hold lock_page_memcg().
2409 *
2410 * NOTE: This relies on being atomic wrt interrupts.
2411 */
2412 void account_page_dirtied(struct page *page, struct address_space *mapping)
2413 {
2414 struct inode *inode = mapping->host;
2415
2416 trace_writeback_dirty_page(page, mapping);
2417
2418 if (mapping_cap_account_dirty(mapping)) {
2419 struct bdi_writeback *wb;
2420
2421 inode_attach_wb(inode, page);
2422 wb = inode_to_wb(inode);
2423
2424 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2425 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2426 __inc_node_page_state(page, NR_DIRTIED);
2427 inc_wb_stat(wb, WB_RECLAIMABLE);
2428 inc_wb_stat(wb, WB_DIRTIED);
2429 task_io_account_write(PAGE_SIZE);
2430 current->nr_dirtied++;
2431 this_cpu_inc(bdp_ratelimits);
2432
2433 mem_cgroup_track_foreign_dirty(page, wb);
2434 }
2435 }
2436
2437 /*
2438 * Helper function for deaccounting dirty page without writeback.
2439 *
2440 * Caller must hold lock_page_memcg().
2441 */
2442 void account_page_cleaned(struct page *page, struct address_space *mapping,
2443 struct bdi_writeback *wb)
2444 {
2445 if (mapping_cap_account_dirty(mapping)) {
2446 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2447 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2448 dec_wb_stat(wb, WB_RECLAIMABLE);
2449 task_io_account_cancelled_write(PAGE_SIZE);
2450 }
2451 }
2452
2453 /*
2454 * For address_spaces which do not use buffers. Just tag the page as dirty in
2455 * the xarray.
2456 *
2457 * This is also used when a single buffer is being dirtied: we want to set the
2458 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2459 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2460 *
2461 * The caller must ensure this doesn't race with truncation. Most will simply
2462 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2463 * the pte lock held, which also locks out truncation.
2464 */
2465 int __set_page_dirty_nobuffers(struct page *page)
2466 {
2467 lock_page_memcg(page);
2468 if (!TestSetPageDirty(page)) {
2469 struct address_space *mapping = page_mapping(page);
2470 unsigned long flags;
2471
2472 if (!mapping) {
2473 unlock_page_memcg(page);
2474 return 1;
2475 }
2476
2477 xa_lock_irqsave(&mapping->i_pages, flags);
2478 BUG_ON(page_mapping(page) != mapping);
2479 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2480 account_page_dirtied(page, mapping);
2481 __xa_set_mark(&mapping->i_pages, page_index(page),
2482 PAGECACHE_TAG_DIRTY);
2483 xa_unlock_irqrestore(&mapping->i_pages, flags);
2484 unlock_page_memcg(page);
2485
2486 if (mapping->host) {
2487 /* !PageAnon && !swapper_space */
2488 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2489 }
2490 return 1;
2491 }
2492 unlock_page_memcg(page);
2493 return 0;
2494 }
2495 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2496
2497 /*
2498 * Call this whenever redirtying a page, to de-account the dirty counters
2499 * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2500 * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2501 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2502 * control.
2503 */
2504 void account_page_redirty(struct page *page)
2505 {
2506 struct address_space *mapping = page->mapping;
2507
2508 if (mapping && mapping_cap_account_dirty(mapping)) {
2509 struct inode *inode = mapping->host;
2510 struct bdi_writeback *wb;
2511 struct wb_lock_cookie cookie = {};
2512
2513 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2514 current->nr_dirtied--;
2515 dec_node_page_state(page, NR_DIRTIED);
2516 dec_wb_stat(wb, WB_DIRTIED);
2517 unlocked_inode_to_wb_end(inode, &cookie);
2518 }
2519 }
2520 EXPORT_SYMBOL(account_page_redirty);
2521
2522 /*
2523 * When a writepage implementation decides that it doesn't want to write this
2524 * page for some reason, it should redirty the locked page via
2525 * redirty_page_for_writepage() and it should then unlock the page and return 0
2526 */
2527 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2528 {
2529 int ret;
2530
2531 wbc->pages_skipped++;
2532 ret = __set_page_dirty_nobuffers(page);
2533 account_page_redirty(page);
2534 return ret;
2535 }
2536 EXPORT_SYMBOL(redirty_page_for_writepage);
2537
2538 /*
2539 * Dirty a page.
2540 *
2541 * For pages with a mapping this should be done under the page lock
2542 * for the benefit of asynchronous memory errors who prefer a consistent
2543 * dirty state. This rule can be broken in some special cases,
2544 * but should be better not to.
2545 *
2546 * If the mapping doesn't provide a set_page_dirty a_op, then
2547 * just fall through and assume that it wants buffer_heads.
2548 */
2549 int set_page_dirty(struct page *page)
2550 {
2551 struct address_space *mapping = page_mapping(page);
2552
2553 page = compound_head(page);
2554 if (likely(mapping)) {
2555 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2556 /*
2557 * readahead/lru_deactivate_page could remain
2558 * PG_readahead/PG_reclaim due to race with end_page_writeback
2559 * About readahead, if the page is written, the flags would be
2560 * reset. So no problem.
2561 * About lru_deactivate_page, if the page is redirty, the flag
2562 * will be reset. So no problem. but if the page is used by readahead
2563 * it will confuse readahead and make it restart the size rampup
2564 * process. But it's a trivial problem.
2565 */
2566 if (PageReclaim(page))
2567 ClearPageReclaim(page);
2568 #ifdef CONFIG_BLOCK
2569 if (!spd)
2570 spd = __set_page_dirty_buffers;
2571 #endif
2572 return (*spd)(page);
2573 }
2574 if (!PageDirty(page)) {
2575 if (!TestSetPageDirty(page))
2576 return 1;
2577 }
2578 return 0;
2579 }
2580 EXPORT_SYMBOL(set_page_dirty);
2581
2582 /*
2583 * set_page_dirty() is racy if the caller has no reference against
2584 * page->mapping->host, and if the page is unlocked. This is because another
2585 * CPU could truncate the page off the mapping and then free the mapping.
2586 *
2587 * Usually, the page _is_ locked, or the caller is a user-space process which
2588 * holds a reference on the inode by having an open file.
2589 *
2590 * In other cases, the page should be locked before running set_page_dirty().
2591 */
2592 int set_page_dirty_lock(struct page *page)
2593 {
2594 int ret;
2595
2596 lock_page(page);
2597 ret = set_page_dirty(page);
2598 unlock_page(page);
2599 return ret;
2600 }
2601 EXPORT_SYMBOL(set_page_dirty_lock);
2602
2603 /*
2604 * This cancels just the dirty bit on the kernel page itself, it does NOT
2605 * actually remove dirty bits on any mmap's that may be around. It also
2606 * leaves the page tagged dirty, so any sync activity will still find it on
2607 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2608 * look at the dirty bits in the VM.
2609 *
2610 * Doing this should *normally* only ever be done when a page is truncated,
2611 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2612 * this when it notices that somebody has cleaned out all the buffers on a
2613 * page without actually doing it through the VM. Can you say "ext3 is
2614 * horribly ugly"? Thought you could.
2615 */
2616 void __cancel_dirty_page(struct page *page)
2617 {
2618 struct address_space *mapping = page_mapping(page);
2619
2620 if (mapping_cap_account_dirty(mapping)) {
2621 struct inode *inode = mapping->host;
2622 struct bdi_writeback *wb;
2623 struct wb_lock_cookie cookie = {};
2624
2625 lock_page_memcg(page);
2626 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2627
2628 if (TestClearPageDirty(page))
2629 account_page_cleaned(page, mapping, wb);
2630
2631 unlocked_inode_to_wb_end(inode, &cookie);
2632 unlock_page_memcg(page);
2633 } else {
2634 ClearPageDirty(page);
2635 }
2636 }
2637 EXPORT_SYMBOL(__cancel_dirty_page);
2638
2639 /*
2640 * Clear a page's dirty flag, while caring for dirty memory accounting.
2641 * Returns true if the page was previously dirty.
2642 *
2643 * This is for preparing to put the page under writeout. We leave the page
2644 * tagged as dirty in the xarray so that a concurrent write-for-sync
2645 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2646 * implementation will run either set_page_writeback() or set_page_dirty(),
2647 * at which stage we bring the page's dirty flag and xarray dirty tag
2648 * back into sync.
2649 *
2650 * This incoherency between the page's dirty flag and xarray tag is
2651 * unfortunate, but it only exists while the page is locked.
2652 */
2653 int clear_page_dirty_for_io(struct page *page)
2654 {
2655 struct address_space *mapping = page_mapping(page);
2656 int ret = 0;
2657
2658 BUG_ON(!PageLocked(page));
2659
2660 if (mapping && mapping_cap_account_dirty(mapping)) {
2661 struct inode *inode = mapping->host;
2662 struct bdi_writeback *wb;
2663 struct wb_lock_cookie cookie = {};
2664
2665 /*
2666 * Yes, Virginia, this is indeed insane.
2667 *
2668 * We use this sequence to make sure that
2669 * (a) we account for dirty stats properly
2670 * (b) we tell the low-level filesystem to
2671 * mark the whole page dirty if it was
2672 * dirty in a pagetable. Only to then
2673 * (c) clean the page again and return 1 to
2674 * cause the writeback.
2675 *
2676 * This way we avoid all nasty races with the
2677 * dirty bit in multiple places and clearing
2678 * them concurrently from different threads.
2679 *
2680 * Note! Normally the "set_page_dirty(page)"
2681 * has no effect on the actual dirty bit - since
2682 * that will already usually be set. But we
2683 * need the side effects, and it can help us
2684 * avoid races.
2685 *
2686 * We basically use the page "master dirty bit"
2687 * as a serialization point for all the different
2688 * threads doing their things.
2689 */
2690 if (page_mkclean(page))
2691 set_page_dirty(page);
2692 /*
2693 * We carefully synchronise fault handlers against
2694 * installing a dirty pte and marking the page dirty
2695 * at this point. We do this by having them hold the
2696 * page lock while dirtying the page, and pages are
2697 * always locked coming in here, so we get the desired
2698 * exclusion.
2699 */
2700 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2701 if (TestClearPageDirty(page)) {
2702 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2703 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2704 dec_wb_stat(wb, WB_RECLAIMABLE);
2705 ret = 1;
2706 }
2707 unlocked_inode_to_wb_end(inode, &cookie);
2708 return ret;
2709 }
2710 return TestClearPageDirty(page);
2711 }
2712 EXPORT_SYMBOL(clear_page_dirty_for_io);
2713
2714 int test_clear_page_writeback(struct page *page)
2715 {
2716 struct address_space *mapping = page_mapping(page);
2717 struct mem_cgroup *memcg;
2718 struct lruvec *lruvec;
2719 int ret;
2720
2721 memcg = lock_page_memcg(page);
2722 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2723 if (mapping && mapping_use_writeback_tags(mapping)) {
2724 struct inode *inode = mapping->host;
2725 struct backing_dev_info *bdi = inode_to_bdi(inode);
2726 unsigned long flags;
2727
2728 xa_lock_irqsave(&mapping->i_pages, flags);
2729 ret = TestClearPageWriteback(page);
2730 if (ret) {
2731 __xa_clear_mark(&mapping->i_pages, page_index(page),
2732 PAGECACHE_TAG_WRITEBACK);
2733 if (bdi_cap_account_writeback(bdi)) {
2734 struct bdi_writeback *wb = inode_to_wb(inode);
2735
2736 dec_wb_stat(wb, WB_WRITEBACK);
2737 __wb_writeout_inc(wb);
2738 }
2739 }
2740
2741 if (mapping->host && !mapping_tagged(mapping,
2742 PAGECACHE_TAG_WRITEBACK))
2743 sb_clear_inode_writeback(mapping->host);
2744
2745 xa_unlock_irqrestore(&mapping->i_pages, flags);
2746 } else {
2747 ret = TestClearPageWriteback(page);
2748 }
2749 /*
2750 * NOTE: Page might be free now! Writeback doesn't hold a page
2751 * reference on its own, it relies on truncation to wait for
2752 * the clearing of PG_writeback. The below can only access
2753 * page state that is static across allocation cycles.
2754 */
2755 if (ret) {
2756 dec_lruvec_state(lruvec, NR_WRITEBACK);
2757 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2758 inc_node_page_state(page, NR_WRITTEN);
2759 }
2760 __unlock_page_memcg(memcg);
2761 return ret;
2762 }
2763
2764 int __test_set_page_writeback(struct page *page, bool keep_write)
2765 {
2766 struct address_space *mapping = page_mapping(page);
2767 int ret;
2768
2769 lock_page_memcg(page);
2770 if (mapping && mapping_use_writeback_tags(mapping)) {
2771 XA_STATE(xas, &mapping->i_pages, page_index(page));
2772 struct inode *inode = mapping->host;
2773 struct backing_dev_info *bdi = inode_to_bdi(inode);
2774 unsigned long flags;
2775
2776 xas_lock_irqsave(&xas, flags);
2777 xas_load(&xas);
2778 ret = TestSetPageWriteback(page);
2779 if (!ret) {
2780 bool on_wblist;
2781
2782 on_wblist = mapping_tagged(mapping,
2783 PAGECACHE_TAG_WRITEBACK);
2784
2785 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2786 if (bdi_cap_account_writeback(bdi))
2787 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2788
2789 /*
2790 * We can come through here when swapping anonymous
2791 * pages, so we don't necessarily have an inode to track
2792 * for sync.
2793 */
2794 if (mapping->host && !on_wblist)
2795 sb_mark_inode_writeback(mapping->host);
2796 }
2797 if (!PageDirty(page))
2798 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2799 if (!keep_write)
2800 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2801 xas_unlock_irqrestore(&xas, flags);
2802 } else {
2803 ret = TestSetPageWriteback(page);
2804 }
2805 if (!ret) {
2806 inc_lruvec_page_state(page, NR_WRITEBACK);
2807 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2808 }
2809 unlock_page_memcg(page);
2810 return ret;
2811
2812 }
2813 EXPORT_SYMBOL(__test_set_page_writeback);
2814
2815 /*
2816 * Wait for a page to complete writeback
2817 */
2818 void wait_on_page_writeback(struct page *page)
2819 {
2820 if (PageWriteback(page)) {
2821 trace_wait_on_page_writeback(page, page_mapping(page));
2822 wait_on_page_bit(page, PG_writeback);
2823 }
2824 }
2825 EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2826
2827 /**
2828 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2829 * @page: The page to wait on.
2830 *
2831 * This function determines if the given page is related to a backing device
2832 * that requires page contents to be held stable during writeback. If so, then
2833 * it will wait for any pending writeback to complete.
2834 */
2835 void wait_for_stable_page(struct page *page)
2836 {
2837 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2838 wait_on_page_writeback(page);
2839 }
2840 EXPORT_SYMBOL_GPL(wait_for_stable_page);