xref: /openbmc/linux/mm/page-writeback.c (revision 827beb77)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * mm/page-writeback.c
4  *
5  * Copyright (C) 2002, Linus Torvalds.
6  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7  *
8  * Contains functions related to writing back dirty pages at the
9  * address_space level.
10  *
11  * 10Apr2002	Andrew Morton
12  *		Initial version
13  */
14 
15 #include <linux/kernel.h>
16 #include <linux/export.h>
17 #include <linux/spinlock.h>
18 #include <linux/fs.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/writeback.h>
24 #include <linux/init.h>
25 #include <linux/backing-dev.h>
26 #include <linux/task_io_accounting_ops.h>
27 #include <linux/blkdev.h>
28 #include <linux/mpage.h>
29 #include <linux/rmap.h>
30 #include <linux/percpu.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/pagevec.h>
36 #include <linux/timer.h>
37 #include <linux/sched/rt.h>
38 #include <linux/sched/signal.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
41 
42 #include "internal.h"
43 
44 /*
45  * Sleep at most 200ms at a time in balance_dirty_pages().
46  */
47 #define MAX_PAUSE		max(HZ/5, 1)
48 
49 /*
50  * Try to keep balance_dirty_pages() call intervals higher than this many pages
51  * by raising pause time to max_pause when falls below it.
52  */
53 #define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
54 
55 /*
56  * Estimate write bandwidth at 200ms intervals.
57  */
58 #define BANDWIDTH_INTERVAL	max(HZ/5, 1)
59 
60 #define RATELIMIT_CALC_SHIFT	10
61 
62 /*
63  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64  * will look to see if it needs to force writeback or throttling.
65  */
66 static long ratelimit_pages = 32;
67 
68 /* The following parameters are exported via /proc/sys/vm */
69 
70 /*
71  * Start background writeback (via writeback threads) at this percentage
72  */
73 int dirty_background_ratio = 10;
74 
75 /*
76  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77  * dirty_background_ratio * the amount of dirtyable memory
78  */
79 unsigned long dirty_background_bytes;
80 
81 /*
82  * free highmem will not be subtracted from the total free memory
83  * for calculating free ratios if vm_highmem_is_dirtyable is true
84  */
85 int vm_highmem_is_dirtyable;
86 
87 /*
88  * The generator of dirty data starts writeback at this percentage
89  */
90 int vm_dirty_ratio = 20;
91 
92 /*
93  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94  * vm_dirty_ratio * the amount of dirtyable memory
95  */
96 unsigned long vm_dirty_bytes;
97 
98 /*
99  * The interval between `kupdate'-style writebacks
100  */
101 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102 
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104 
105 /*
106  * The longest time for which data is allowed to remain dirty
107  */
108 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109 
110 /*
111  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112  * a full sync is triggered after this time elapses without any disk activity.
113  */
114 int laptop_mode;
115 
116 EXPORT_SYMBOL(laptop_mode);
117 
118 /* End of sysctl-exported parameters */
119 
120 struct wb_domain global_wb_domain;
121 
122 /* consolidated parameters for balance_dirty_pages() and its subroutines */
123 struct dirty_throttle_control {
124 #ifdef CONFIG_CGROUP_WRITEBACK
125 	struct wb_domain	*dom;
126 	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
127 #endif
128 	struct bdi_writeback	*wb;
129 	struct fprop_local_percpu *wb_completions;
130 
131 	unsigned long		avail;		/* dirtyable */
132 	unsigned long		dirty;		/* file_dirty + write + nfs */
133 	unsigned long		thresh;		/* dirty threshold */
134 	unsigned long		bg_thresh;	/* dirty background threshold */
135 
136 	unsigned long		wb_dirty;	/* per-wb counterparts */
137 	unsigned long		wb_thresh;
138 	unsigned long		wb_bg_thresh;
139 
140 	unsigned long		pos_ratio;
141 };
142 
143 /*
144  * Length of period for aging writeout fractions of bdis. This is an
145  * arbitrarily chosen number. The longer the period, the slower fractions will
146  * reflect changes in current writeout rate.
147  */
148 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
149 
150 #ifdef CONFIG_CGROUP_WRITEBACK
151 
152 #define GDTC_INIT(__wb)		.wb = (__wb),				\
153 				.dom = &global_wb_domain,		\
154 				.wb_completions = &(__wb)->completions
155 
156 #define GDTC_INIT_NO_WB		.dom = &global_wb_domain
157 
158 #define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
159 				.dom = mem_cgroup_wb_domain(__wb),	\
160 				.wb_completions = &(__wb)->memcg_completions, \
161 				.gdtc = __gdtc
162 
163 static bool mdtc_valid(struct dirty_throttle_control *dtc)
164 {
165 	return dtc->dom;
166 }
167 
168 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
169 {
170 	return dtc->dom;
171 }
172 
173 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
174 {
175 	return mdtc->gdtc;
176 }
177 
178 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
179 {
180 	return &wb->memcg_completions;
181 }
182 
183 static void wb_min_max_ratio(struct bdi_writeback *wb,
184 			     unsigned long *minp, unsigned long *maxp)
185 {
186 	unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
187 	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
188 	unsigned long long min = wb->bdi->min_ratio;
189 	unsigned long long max = wb->bdi->max_ratio;
190 
191 	/*
192 	 * @wb may already be clean by the time control reaches here and
193 	 * the total may not include its bw.
194 	 */
195 	if (this_bw < tot_bw) {
196 		if (min) {
197 			min *= this_bw;
198 			min = div64_ul(min, tot_bw);
199 		}
200 		if (max < 100) {
201 			max *= this_bw;
202 			max = div64_ul(max, tot_bw);
203 		}
204 	}
205 
206 	*minp = min;
207 	*maxp = max;
208 }
209 
210 #else	/* CONFIG_CGROUP_WRITEBACK */
211 
212 #define GDTC_INIT(__wb)		.wb = (__wb),                           \
213 				.wb_completions = &(__wb)->completions
214 #define GDTC_INIT_NO_WB
215 #define MDTC_INIT(__wb, __gdtc)
216 
217 static bool mdtc_valid(struct dirty_throttle_control *dtc)
218 {
219 	return false;
220 }
221 
222 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
223 {
224 	return &global_wb_domain;
225 }
226 
227 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
228 {
229 	return NULL;
230 }
231 
232 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
233 {
234 	return NULL;
235 }
236 
237 static void wb_min_max_ratio(struct bdi_writeback *wb,
238 			     unsigned long *minp, unsigned long *maxp)
239 {
240 	*minp = wb->bdi->min_ratio;
241 	*maxp = wb->bdi->max_ratio;
242 }
243 
244 #endif	/* CONFIG_CGROUP_WRITEBACK */
245 
246 /*
247  * In a memory zone, there is a certain amount of pages we consider
248  * available for the page cache, which is essentially the number of
249  * free and reclaimable pages, minus some zone reserves to protect
250  * lowmem and the ability to uphold the zone's watermarks without
251  * requiring writeback.
252  *
253  * This number of dirtyable pages is the base value of which the
254  * user-configurable dirty ratio is the effective number of pages that
255  * are allowed to be actually dirtied.  Per individual zone, or
256  * globally by using the sum of dirtyable pages over all zones.
257  *
258  * Because the user is allowed to specify the dirty limit globally as
259  * absolute number of bytes, calculating the per-zone dirty limit can
260  * require translating the configured limit into a percentage of
261  * global dirtyable memory first.
262  */
263 
264 /**
265  * node_dirtyable_memory - number of dirtyable pages in a node
266  * @pgdat: the node
267  *
268  * Return: the node's number of pages potentially available for dirty
269  * page cache.  This is the base value for the per-node dirty limits.
270  */
271 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
272 {
273 	unsigned long nr_pages = 0;
274 	int z;
275 
276 	for (z = 0; z < MAX_NR_ZONES; z++) {
277 		struct zone *zone = pgdat->node_zones + z;
278 
279 		if (!populated_zone(zone))
280 			continue;
281 
282 		nr_pages += zone_page_state(zone, NR_FREE_PAGES);
283 	}
284 
285 	/*
286 	 * Pages reserved for the kernel should not be considered
287 	 * dirtyable, to prevent a situation where reclaim has to
288 	 * clean pages in order to balance the zones.
289 	 */
290 	nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
291 
292 	nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
293 	nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
294 
295 	return nr_pages;
296 }
297 
298 static unsigned long highmem_dirtyable_memory(unsigned long total)
299 {
300 #ifdef CONFIG_HIGHMEM
301 	int node;
302 	unsigned long x = 0;
303 	int i;
304 
305 	for_each_node_state(node, N_HIGH_MEMORY) {
306 		for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
307 			struct zone *z;
308 			unsigned long nr_pages;
309 
310 			if (!is_highmem_idx(i))
311 				continue;
312 
313 			z = &NODE_DATA(node)->node_zones[i];
314 			if (!populated_zone(z))
315 				continue;
316 
317 			nr_pages = zone_page_state(z, NR_FREE_PAGES);
318 			/* watch for underflows */
319 			nr_pages -= min(nr_pages, high_wmark_pages(z));
320 			nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
321 			nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
322 			x += nr_pages;
323 		}
324 	}
325 
326 	/*
327 	 * Unreclaimable memory (kernel memory or anonymous memory
328 	 * without swap) can bring down the dirtyable pages below
329 	 * the zone's dirty balance reserve and the above calculation
330 	 * will underflow.  However we still want to add in nodes
331 	 * which are below threshold (negative values) to get a more
332 	 * accurate calculation but make sure that the total never
333 	 * underflows.
334 	 */
335 	if ((long)x < 0)
336 		x = 0;
337 
338 	/*
339 	 * Make sure that the number of highmem pages is never larger
340 	 * than the number of the total dirtyable memory. This can only
341 	 * occur in very strange VM situations but we want to make sure
342 	 * that this does not occur.
343 	 */
344 	return min(x, total);
345 #else
346 	return 0;
347 #endif
348 }
349 
350 /**
351  * global_dirtyable_memory - number of globally dirtyable pages
352  *
353  * Return: the global number of pages potentially available for dirty
354  * page cache.  This is the base value for the global dirty limits.
355  */
356 static unsigned long global_dirtyable_memory(void)
357 {
358 	unsigned long x;
359 
360 	x = global_zone_page_state(NR_FREE_PAGES);
361 	/*
362 	 * Pages reserved for the kernel should not be considered
363 	 * dirtyable, to prevent a situation where reclaim has to
364 	 * clean pages in order to balance the zones.
365 	 */
366 	x -= min(x, totalreserve_pages);
367 
368 	x += global_node_page_state(NR_INACTIVE_FILE);
369 	x += global_node_page_state(NR_ACTIVE_FILE);
370 
371 	if (!vm_highmem_is_dirtyable)
372 		x -= highmem_dirtyable_memory(x);
373 
374 	return x + 1;	/* Ensure that we never return 0 */
375 }
376 
377 /**
378  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
379  * @dtc: dirty_throttle_control of interest
380  *
381  * Calculate @dtc->thresh and ->bg_thresh considering
382  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
383  * must ensure that @dtc->avail is set before calling this function.  The
384  * dirty limits will be lifted by 1/4 for real-time tasks.
385  */
386 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
387 {
388 	const unsigned long available_memory = dtc->avail;
389 	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
390 	unsigned long bytes = vm_dirty_bytes;
391 	unsigned long bg_bytes = dirty_background_bytes;
392 	/* convert ratios to per-PAGE_SIZE for higher precision */
393 	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
394 	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
395 	unsigned long thresh;
396 	unsigned long bg_thresh;
397 	struct task_struct *tsk;
398 
399 	/* gdtc is !NULL iff @dtc is for memcg domain */
400 	if (gdtc) {
401 		unsigned long global_avail = gdtc->avail;
402 
403 		/*
404 		 * The byte settings can't be applied directly to memcg
405 		 * domains.  Convert them to ratios by scaling against
406 		 * globally available memory.  As the ratios are in
407 		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
408 		 * number of pages.
409 		 */
410 		if (bytes)
411 			ratio = min(DIV_ROUND_UP(bytes, global_avail),
412 				    PAGE_SIZE);
413 		if (bg_bytes)
414 			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
415 				       PAGE_SIZE);
416 		bytes = bg_bytes = 0;
417 	}
418 
419 	if (bytes)
420 		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
421 	else
422 		thresh = (ratio * available_memory) / PAGE_SIZE;
423 
424 	if (bg_bytes)
425 		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
426 	else
427 		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
428 
429 	if (bg_thresh >= thresh)
430 		bg_thresh = thresh / 2;
431 	tsk = current;
432 	if (rt_task(tsk)) {
433 		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
434 		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
435 	}
436 	dtc->thresh = thresh;
437 	dtc->bg_thresh = bg_thresh;
438 
439 	/* we should eventually report the domain in the TP */
440 	if (!gdtc)
441 		trace_global_dirty_state(bg_thresh, thresh);
442 }
443 
444 /**
445  * global_dirty_limits - background-writeback and dirty-throttling thresholds
446  * @pbackground: out parameter for bg_thresh
447  * @pdirty: out parameter for thresh
448  *
449  * Calculate bg_thresh and thresh for global_wb_domain.  See
450  * domain_dirty_limits() for details.
451  */
452 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
453 {
454 	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
455 
456 	gdtc.avail = global_dirtyable_memory();
457 	domain_dirty_limits(&gdtc);
458 
459 	*pbackground = gdtc.bg_thresh;
460 	*pdirty = gdtc.thresh;
461 }
462 
463 /**
464  * node_dirty_limit - maximum number of dirty pages allowed in a node
465  * @pgdat: the node
466  *
467  * Return: the maximum number of dirty pages allowed in a node, based
468  * on the node's dirtyable memory.
469  */
470 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
471 {
472 	unsigned long node_memory = node_dirtyable_memory(pgdat);
473 	struct task_struct *tsk = current;
474 	unsigned long dirty;
475 
476 	if (vm_dirty_bytes)
477 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
478 			node_memory / global_dirtyable_memory();
479 	else
480 		dirty = vm_dirty_ratio * node_memory / 100;
481 
482 	if (rt_task(tsk))
483 		dirty += dirty / 4;
484 
485 	return dirty;
486 }
487 
488 /**
489  * node_dirty_ok - tells whether a node is within its dirty limits
490  * @pgdat: the node to check
491  *
492  * Return: %true when the dirty pages in @pgdat are within the node's
493  * dirty limit, %false if the limit is exceeded.
494  */
495 bool node_dirty_ok(struct pglist_data *pgdat)
496 {
497 	unsigned long limit = node_dirty_limit(pgdat);
498 	unsigned long nr_pages = 0;
499 
500 	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
501 	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
502 
503 	return nr_pages <= limit;
504 }
505 
506 int dirty_background_ratio_handler(struct ctl_table *table, int write,
507 		void *buffer, size_t *lenp, loff_t *ppos)
508 {
509 	int ret;
510 
511 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
512 	if (ret == 0 && write)
513 		dirty_background_bytes = 0;
514 	return ret;
515 }
516 
517 int dirty_background_bytes_handler(struct ctl_table *table, int write,
518 		void *buffer, size_t *lenp, loff_t *ppos)
519 {
520 	int ret;
521 
522 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
523 	if (ret == 0 && write)
524 		dirty_background_ratio = 0;
525 	return ret;
526 }
527 
528 int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
529 		size_t *lenp, loff_t *ppos)
530 {
531 	int old_ratio = vm_dirty_ratio;
532 	int ret;
533 
534 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
535 	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
536 		writeback_set_ratelimit();
537 		vm_dirty_bytes = 0;
538 	}
539 	return ret;
540 }
541 
542 int dirty_bytes_handler(struct ctl_table *table, int write,
543 		void *buffer, size_t *lenp, loff_t *ppos)
544 {
545 	unsigned long old_bytes = vm_dirty_bytes;
546 	int ret;
547 
548 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
549 	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
550 		writeback_set_ratelimit();
551 		vm_dirty_ratio = 0;
552 	}
553 	return ret;
554 }
555 
556 static unsigned long wp_next_time(unsigned long cur_time)
557 {
558 	cur_time += VM_COMPLETIONS_PERIOD_LEN;
559 	/* 0 has a special meaning... */
560 	if (!cur_time)
561 		return 1;
562 	return cur_time;
563 }
564 
565 static void wb_domain_writeout_add(struct wb_domain *dom,
566 				   struct fprop_local_percpu *completions,
567 				   unsigned int max_prop_frac, long nr)
568 {
569 	__fprop_add_percpu_max(&dom->completions, completions,
570 			       max_prop_frac, nr);
571 	/* First event after period switching was turned off? */
572 	if (unlikely(!dom->period_time)) {
573 		/*
574 		 * We can race with other __bdi_writeout_inc calls here but
575 		 * it does not cause any harm since the resulting time when
576 		 * timer will fire and what is in writeout_period_time will be
577 		 * roughly the same.
578 		 */
579 		dom->period_time = wp_next_time(jiffies);
580 		mod_timer(&dom->period_timer, dom->period_time);
581 	}
582 }
583 
584 /*
585  * Increment @wb's writeout completion count and the global writeout
586  * completion count. Called from __folio_end_writeback().
587  */
588 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
589 {
590 	struct wb_domain *cgdom;
591 
592 	wb_stat_mod(wb, WB_WRITTEN, nr);
593 	wb_domain_writeout_add(&global_wb_domain, &wb->completions,
594 			       wb->bdi->max_prop_frac, nr);
595 
596 	cgdom = mem_cgroup_wb_domain(wb);
597 	if (cgdom)
598 		wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
599 				       wb->bdi->max_prop_frac, nr);
600 }
601 
602 void wb_writeout_inc(struct bdi_writeback *wb)
603 {
604 	unsigned long flags;
605 
606 	local_irq_save(flags);
607 	__wb_writeout_add(wb, 1);
608 	local_irq_restore(flags);
609 }
610 EXPORT_SYMBOL_GPL(wb_writeout_inc);
611 
612 /*
613  * On idle system, we can be called long after we scheduled because we use
614  * deferred timers so count with missed periods.
615  */
616 static void writeout_period(struct timer_list *t)
617 {
618 	struct wb_domain *dom = from_timer(dom, t, period_timer);
619 	int miss_periods = (jiffies - dom->period_time) /
620 						 VM_COMPLETIONS_PERIOD_LEN;
621 
622 	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
623 		dom->period_time = wp_next_time(dom->period_time +
624 				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
625 		mod_timer(&dom->period_timer, dom->period_time);
626 	} else {
627 		/*
628 		 * Aging has zeroed all fractions. Stop wasting CPU on period
629 		 * updates.
630 		 */
631 		dom->period_time = 0;
632 	}
633 }
634 
635 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
636 {
637 	memset(dom, 0, sizeof(*dom));
638 
639 	spin_lock_init(&dom->lock);
640 
641 	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
642 
643 	dom->dirty_limit_tstamp = jiffies;
644 
645 	return fprop_global_init(&dom->completions, gfp);
646 }
647 
648 #ifdef CONFIG_CGROUP_WRITEBACK
649 void wb_domain_exit(struct wb_domain *dom)
650 {
651 	del_timer_sync(&dom->period_timer);
652 	fprop_global_destroy(&dom->completions);
653 }
654 #endif
655 
656 /*
657  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
658  * registered backing devices, which, for obvious reasons, can not
659  * exceed 100%.
660  */
661 static unsigned int bdi_min_ratio;
662 
663 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
664 {
665 	int ret = 0;
666 
667 	spin_lock_bh(&bdi_lock);
668 	if (min_ratio > bdi->max_ratio) {
669 		ret = -EINVAL;
670 	} else {
671 		min_ratio -= bdi->min_ratio;
672 		if (bdi_min_ratio + min_ratio < 100) {
673 			bdi_min_ratio += min_ratio;
674 			bdi->min_ratio += min_ratio;
675 		} else {
676 			ret = -EINVAL;
677 		}
678 	}
679 	spin_unlock_bh(&bdi_lock);
680 
681 	return ret;
682 }
683 
684 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
685 {
686 	int ret = 0;
687 
688 	if (max_ratio > 100)
689 		return -EINVAL;
690 
691 	spin_lock_bh(&bdi_lock);
692 	if (bdi->min_ratio > max_ratio) {
693 		ret = -EINVAL;
694 	} else {
695 		bdi->max_ratio = max_ratio;
696 		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
697 	}
698 	spin_unlock_bh(&bdi_lock);
699 
700 	return ret;
701 }
702 EXPORT_SYMBOL(bdi_set_max_ratio);
703 
704 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
705 					   unsigned long bg_thresh)
706 {
707 	return (thresh + bg_thresh) / 2;
708 }
709 
710 static unsigned long hard_dirty_limit(struct wb_domain *dom,
711 				      unsigned long thresh)
712 {
713 	return max(thresh, dom->dirty_limit);
714 }
715 
716 /*
717  * Memory which can be further allocated to a memcg domain is capped by
718  * system-wide clean memory excluding the amount being used in the domain.
719  */
720 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
721 			    unsigned long filepages, unsigned long headroom)
722 {
723 	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
724 	unsigned long clean = filepages - min(filepages, mdtc->dirty);
725 	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
726 	unsigned long other_clean = global_clean - min(global_clean, clean);
727 
728 	mdtc->avail = filepages + min(headroom, other_clean);
729 }
730 
731 /**
732  * __wb_calc_thresh - @wb's share of dirty throttling threshold
733  * @dtc: dirty_throttle_context of interest
734  *
735  * Note that balance_dirty_pages() will only seriously take it as a hard limit
736  * when sleeping max_pause per page is not enough to keep the dirty pages under
737  * control. For example, when the device is completely stalled due to some error
738  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
739  * In the other normal situations, it acts more gently by throttling the tasks
740  * more (rather than completely block them) when the wb dirty pages go high.
741  *
742  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
743  * - starving fast devices
744  * - piling up dirty pages (that will take long time to sync) on slow devices
745  *
746  * The wb's share of dirty limit will be adapting to its throughput and
747  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
748  *
749  * Return: @wb's dirty limit in pages. The term "dirty" in the context of
750  * dirty balancing includes all PG_dirty and PG_writeback pages.
751  */
752 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
753 {
754 	struct wb_domain *dom = dtc_dom(dtc);
755 	unsigned long thresh = dtc->thresh;
756 	u64 wb_thresh;
757 	unsigned long numerator, denominator;
758 	unsigned long wb_min_ratio, wb_max_ratio;
759 
760 	/*
761 	 * Calculate this BDI's share of the thresh ratio.
762 	 */
763 	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
764 			      &numerator, &denominator);
765 
766 	wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
767 	wb_thresh *= numerator;
768 	wb_thresh = div64_ul(wb_thresh, denominator);
769 
770 	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
771 
772 	wb_thresh += (thresh * wb_min_ratio) / 100;
773 	if (wb_thresh > (thresh * wb_max_ratio) / 100)
774 		wb_thresh = thresh * wb_max_ratio / 100;
775 
776 	return wb_thresh;
777 }
778 
779 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
780 {
781 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
782 					       .thresh = thresh };
783 	return __wb_calc_thresh(&gdtc);
784 }
785 
786 /*
787  *                           setpoint - dirty 3
788  *        f(dirty) := 1.0 + (----------------)
789  *                           limit - setpoint
790  *
791  * it's a 3rd order polynomial that subjects to
792  *
793  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
794  * (2) f(setpoint) = 1.0 => the balance point
795  * (3) f(limit)    = 0   => the hard limit
796  * (4) df/dx      <= 0	 => negative feedback control
797  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
798  *     => fast response on large errors; small oscillation near setpoint
799  */
800 static long long pos_ratio_polynom(unsigned long setpoint,
801 					  unsigned long dirty,
802 					  unsigned long limit)
803 {
804 	long long pos_ratio;
805 	long x;
806 
807 	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
808 		      (limit - setpoint) | 1);
809 	pos_ratio = x;
810 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
811 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
812 	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
813 
814 	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
815 }
816 
817 /*
818  * Dirty position control.
819  *
820  * (o) global/bdi setpoints
821  *
822  * We want the dirty pages be balanced around the global/wb setpoints.
823  * When the number of dirty pages is higher/lower than the setpoint, the
824  * dirty position control ratio (and hence task dirty ratelimit) will be
825  * decreased/increased to bring the dirty pages back to the setpoint.
826  *
827  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
828  *
829  *     if (dirty < setpoint) scale up   pos_ratio
830  *     if (dirty > setpoint) scale down pos_ratio
831  *
832  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
833  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
834  *
835  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
836  *
837  * (o) global control line
838  *
839  *     ^ pos_ratio
840  *     |
841  *     |            |<===== global dirty control scope ======>|
842  * 2.0  * * * * * * *
843  *     |            .*
844  *     |            . *
845  *     |            .   *
846  *     |            .     *
847  *     |            .        *
848  *     |            .            *
849  * 1.0 ................................*
850  *     |            .                  .     *
851  *     |            .                  .          *
852  *     |            .                  .              *
853  *     |            .                  .                 *
854  *     |            .                  .                    *
855  *   0 +------------.------------------.----------------------*------------->
856  *           freerun^          setpoint^                 limit^   dirty pages
857  *
858  * (o) wb control line
859  *
860  *     ^ pos_ratio
861  *     |
862  *     |            *
863  *     |              *
864  *     |                *
865  *     |                  *
866  *     |                    * |<=========== span ============>|
867  * 1.0 .......................*
868  *     |                      . *
869  *     |                      .   *
870  *     |                      .     *
871  *     |                      .       *
872  *     |                      .         *
873  *     |                      .           *
874  *     |                      .             *
875  *     |                      .               *
876  *     |                      .                 *
877  *     |                      .                   *
878  *     |                      .                     *
879  * 1/4 ...............................................* * * * * * * * * * * *
880  *     |                      .                         .
881  *     |                      .                           .
882  *     |                      .                             .
883  *   0 +----------------------.-------------------------------.------------->
884  *                wb_setpoint^                    x_intercept^
885  *
886  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
887  * be smoothly throttled down to normal if it starts high in situations like
888  * - start writing to a slow SD card and a fast disk at the same time. The SD
889  *   card's wb_dirty may rush to many times higher than wb_setpoint.
890  * - the wb dirty thresh drops quickly due to change of JBOD workload
891  */
892 static void wb_position_ratio(struct dirty_throttle_control *dtc)
893 {
894 	struct bdi_writeback *wb = dtc->wb;
895 	unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
896 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
897 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
898 	unsigned long wb_thresh = dtc->wb_thresh;
899 	unsigned long x_intercept;
900 	unsigned long setpoint;		/* dirty pages' target balance point */
901 	unsigned long wb_setpoint;
902 	unsigned long span;
903 	long long pos_ratio;		/* for scaling up/down the rate limit */
904 	long x;
905 
906 	dtc->pos_ratio = 0;
907 
908 	if (unlikely(dtc->dirty >= limit))
909 		return;
910 
911 	/*
912 	 * global setpoint
913 	 *
914 	 * See comment for pos_ratio_polynom().
915 	 */
916 	setpoint = (freerun + limit) / 2;
917 	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
918 
919 	/*
920 	 * The strictlimit feature is a tool preventing mistrusted filesystems
921 	 * from growing a large number of dirty pages before throttling. For
922 	 * such filesystems balance_dirty_pages always checks wb counters
923 	 * against wb limits. Even if global "nr_dirty" is under "freerun".
924 	 * This is especially important for fuse which sets bdi->max_ratio to
925 	 * 1% by default. Without strictlimit feature, fuse writeback may
926 	 * consume arbitrary amount of RAM because it is accounted in
927 	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
928 	 *
929 	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
930 	 * two values: wb_dirty and wb_thresh. Let's consider an example:
931 	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
932 	 * limits are set by default to 10% and 20% (background and throttle).
933 	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
934 	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
935 	 * about ~6K pages (as the average of background and throttle wb
936 	 * limits). The 3rd order polynomial will provide positive feedback if
937 	 * wb_dirty is under wb_setpoint and vice versa.
938 	 *
939 	 * Note, that we cannot use global counters in these calculations
940 	 * because we want to throttle process writing to a strictlimit wb
941 	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
942 	 * in the example above).
943 	 */
944 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
945 		long long wb_pos_ratio;
946 
947 		if (dtc->wb_dirty < 8) {
948 			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
949 					   2 << RATELIMIT_CALC_SHIFT);
950 			return;
951 		}
952 
953 		if (dtc->wb_dirty >= wb_thresh)
954 			return;
955 
956 		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
957 						    dtc->wb_bg_thresh);
958 
959 		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
960 			return;
961 
962 		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
963 						 wb_thresh);
964 
965 		/*
966 		 * Typically, for strictlimit case, wb_setpoint << setpoint
967 		 * and pos_ratio >> wb_pos_ratio. In the other words global
968 		 * state ("dirty") is not limiting factor and we have to
969 		 * make decision based on wb counters. But there is an
970 		 * important case when global pos_ratio should get precedence:
971 		 * global limits are exceeded (e.g. due to activities on other
972 		 * wb's) while given strictlimit wb is below limit.
973 		 *
974 		 * "pos_ratio * wb_pos_ratio" would work for the case above,
975 		 * but it would look too non-natural for the case of all
976 		 * activity in the system coming from a single strictlimit wb
977 		 * with bdi->max_ratio == 100%.
978 		 *
979 		 * Note that min() below somewhat changes the dynamics of the
980 		 * control system. Normally, pos_ratio value can be well over 3
981 		 * (when globally we are at freerun and wb is well below wb
982 		 * setpoint). Now the maximum pos_ratio in the same situation
983 		 * is 2. We might want to tweak this if we observe the control
984 		 * system is too slow to adapt.
985 		 */
986 		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
987 		return;
988 	}
989 
990 	/*
991 	 * We have computed basic pos_ratio above based on global situation. If
992 	 * the wb is over/under its share of dirty pages, we want to scale
993 	 * pos_ratio further down/up. That is done by the following mechanism.
994 	 */
995 
996 	/*
997 	 * wb setpoint
998 	 *
999 	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1000 	 *
1001 	 *                        x_intercept - wb_dirty
1002 	 *                     := --------------------------
1003 	 *                        x_intercept - wb_setpoint
1004 	 *
1005 	 * The main wb control line is a linear function that subjects to
1006 	 *
1007 	 * (1) f(wb_setpoint) = 1.0
1008 	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1009 	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1010 	 *
1011 	 * For single wb case, the dirty pages are observed to fluctuate
1012 	 * regularly within range
1013 	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1014 	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1015 	 * fluctuation range for pos_ratio.
1016 	 *
1017 	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1018 	 * own size, so move the slope over accordingly and choose a slope that
1019 	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1020 	 */
1021 	if (unlikely(wb_thresh > dtc->thresh))
1022 		wb_thresh = dtc->thresh;
1023 	/*
1024 	 * It's very possible that wb_thresh is close to 0 not because the
1025 	 * device is slow, but that it has remained inactive for long time.
1026 	 * Honour such devices a reasonable good (hopefully IO efficient)
1027 	 * threshold, so that the occasional writes won't be blocked and active
1028 	 * writes can rampup the threshold quickly.
1029 	 */
1030 	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1031 	/*
1032 	 * scale global setpoint to wb's:
1033 	 *	wb_setpoint = setpoint * wb_thresh / thresh
1034 	 */
1035 	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1036 	wb_setpoint = setpoint * (u64)x >> 16;
1037 	/*
1038 	 * Use span=(8*write_bw) in single wb case as indicated by
1039 	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1040 	 *
1041 	 *        wb_thresh                    thresh - wb_thresh
1042 	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1043 	 *         thresh                           thresh
1044 	 */
1045 	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1046 	x_intercept = wb_setpoint + span;
1047 
1048 	if (dtc->wb_dirty < x_intercept - span / 4) {
1049 		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1050 				      (x_intercept - wb_setpoint) | 1);
1051 	} else
1052 		pos_ratio /= 4;
1053 
1054 	/*
1055 	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1056 	 * It may push the desired control point of global dirty pages higher
1057 	 * than setpoint.
1058 	 */
1059 	x_intercept = wb_thresh / 2;
1060 	if (dtc->wb_dirty < x_intercept) {
1061 		if (dtc->wb_dirty > x_intercept / 8)
1062 			pos_ratio = div_u64(pos_ratio * x_intercept,
1063 					    dtc->wb_dirty);
1064 		else
1065 			pos_ratio *= 8;
1066 	}
1067 
1068 	dtc->pos_ratio = pos_ratio;
1069 }
1070 
1071 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1072 				      unsigned long elapsed,
1073 				      unsigned long written)
1074 {
1075 	const unsigned long period = roundup_pow_of_two(3 * HZ);
1076 	unsigned long avg = wb->avg_write_bandwidth;
1077 	unsigned long old = wb->write_bandwidth;
1078 	u64 bw;
1079 
1080 	/*
1081 	 * bw = written * HZ / elapsed
1082 	 *
1083 	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1084 	 * write_bandwidth = ---------------------------------------------------
1085 	 *                                          period
1086 	 *
1087 	 * @written may have decreased due to folio_account_redirty().
1088 	 * Avoid underflowing @bw calculation.
1089 	 */
1090 	bw = written - min(written, wb->written_stamp);
1091 	bw *= HZ;
1092 	if (unlikely(elapsed > period)) {
1093 		bw = div64_ul(bw, elapsed);
1094 		avg = bw;
1095 		goto out;
1096 	}
1097 	bw += (u64)wb->write_bandwidth * (period - elapsed);
1098 	bw >>= ilog2(period);
1099 
1100 	/*
1101 	 * one more level of smoothing, for filtering out sudden spikes
1102 	 */
1103 	if (avg > old && old >= (unsigned long)bw)
1104 		avg -= (avg - old) >> 3;
1105 
1106 	if (avg < old && old <= (unsigned long)bw)
1107 		avg += (old - avg) >> 3;
1108 
1109 out:
1110 	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1111 	avg = max(avg, 1LU);
1112 	if (wb_has_dirty_io(wb)) {
1113 		long delta = avg - wb->avg_write_bandwidth;
1114 		WARN_ON_ONCE(atomic_long_add_return(delta,
1115 					&wb->bdi->tot_write_bandwidth) <= 0);
1116 	}
1117 	wb->write_bandwidth = bw;
1118 	WRITE_ONCE(wb->avg_write_bandwidth, avg);
1119 }
1120 
1121 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1122 {
1123 	struct wb_domain *dom = dtc_dom(dtc);
1124 	unsigned long thresh = dtc->thresh;
1125 	unsigned long limit = dom->dirty_limit;
1126 
1127 	/*
1128 	 * Follow up in one step.
1129 	 */
1130 	if (limit < thresh) {
1131 		limit = thresh;
1132 		goto update;
1133 	}
1134 
1135 	/*
1136 	 * Follow down slowly. Use the higher one as the target, because thresh
1137 	 * may drop below dirty. This is exactly the reason to introduce
1138 	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1139 	 */
1140 	thresh = max(thresh, dtc->dirty);
1141 	if (limit > thresh) {
1142 		limit -= (limit - thresh) >> 5;
1143 		goto update;
1144 	}
1145 	return;
1146 update:
1147 	dom->dirty_limit = limit;
1148 }
1149 
1150 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1151 				      unsigned long now)
1152 {
1153 	struct wb_domain *dom = dtc_dom(dtc);
1154 
1155 	/*
1156 	 * check locklessly first to optimize away locking for the most time
1157 	 */
1158 	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1159 		return;
1160 
1161 	spin_lock(&dom->lock);
1162 	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1163 		update_dirty_limit(dtc);
1164 		dom->dirty_limit_tstamp = now;
1165 	}
1166 	spin_unlock(&dom->lock);
1167 }
1168 
1169 /*
1170  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1171  *
1172  * Normal wb tasks will be curbed at or below it in long term.
1173  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1174  */
1175 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1176 				      unsigned long dirtied,
1177 				      unsigned long elapsed)
1178 {
1179 	struct bdi_writeback *wb = dtc->wb;
1180 	unsigned long dirty = dtc->dirty;
1181 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1182 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1183 	unsigned long setpoint = (freerun + limit) / 2;
1184 	unsigned long write_bw = wb->avg_write_bandwidth;
1185 	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1186 	unsigned long dirty_rate;
1187 	unsigned long task_ratelimit;
1188 	unsigned long balanced_dirty_ratelimit;
1189 	unsigned long step;
1190 	unsigned long x;
1191 	unsigned long shift;
1192 
1193 	/*
1194 	 * The dirty rate will match the writeout rate in long term, except
1195 	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1196 	 */
1197 	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1198 
1199 	/*
1200 	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1201 	 */
1202 	task_ratelimit = (u64)dirty_ratelimit *
1203 					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1204 	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1205 
1206 	/*
1207 	 * A linear estimation of the "balanced" throttle rate. The theory is,
1208 	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1209 	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1210 	 * formula will yield the balanced rate limit (write_bw / N).
1211 	 *
1212 	 * Note that the expanded form is not a pure rate feedback:
1213 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1214 	 * but also takes pos_ratio into account:
1215 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1216 	 *
1217 	 * (1) is not realistic because pos_ratio also takes part in balancing
1218 	 * the dirty rate.  Consider the state
1219 	 *	pos_ratio = 0.5						     (3)
1220 	 *	rate = 2 * (write_bw / N)				     (4)
1221 	 * If (1) is used, it will stuck in that state! Because each dd will
1222 	 * be throttled at
1223 	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1224 	 * yielding
1225 	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1226 	 * put (6) into (1) we get
1227 	 *	rate_(i+1) = rate_(i)					     (7)
1228 	 *
1229 	 * So we end up using (2) to always keep
1230 	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1231 	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1232 	 * pos_ratio is able to drive itself to 1.0, which is not only where
1233 	 * the dirty count meet the setpoint, but also where the slope of
1234 	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1235 	 */
1236 	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1237 					   dirty_rate | 1);
1238 	/*
1239 	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1240 	 */
1241 	if (unlikely(balanced_dirty_ratelimit > write_bw))
1242 		balanced_dirty_ratelimit = write_bw;
1243 
1244 	/*
1245 	 * We could safely do this and return immediately:
1246 	 *
1247 	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1248 	 *
1249 	 * However to get a more stable dirty_ratelimit, the below elaborated
1250 	 * code makes use of task_ratelimit to filter out singular points and
1251 	 * limit the step size.
1252 	 *
1253 	 * The below code essentially only uses the relative value of
1254 	 *
1255 	 *	task_ratelimit - dirty_ratelimit
1256 	 *	= (pos_ratio - 1) * dirty_ratelimit
1257 	 *
1258 	 * which reflects the direction and size of dirty position error.
1259 	 */
1260 
1261 	/*
1262 	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1263 	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1264 	 * For example, when
1265 	 * - dirty_ratelimit > balanced_dirty_ratelimit
1266 	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1267 	 * lowering dirty_ratelimit will help meet both the position and rate
1268 	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1269 	 * only help meet the rate target. After all, what the users ultimately
1270 	 * feel and care are stable dirty rate and small position error.
1271 	 *
1272 	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1273 	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1274 	 * keeps jumping around randomly and can even leap far away at times
1275 	 * due to the small 200ms estimation period of dirty_rate (we want to
1276 	 * keep that period small to reduce time lags).
1277 	 */
1278 	step = 0;
1279 
1280 	/*
1281 	 * For strictlimit case, calculations above were based on wb counters
1282 	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1283 	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1284 	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1285 	 * "dirty" and wb_setpoint as "setpoint".
1286 	 *
1287 	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1288 	 * it's possible that wb_thresh is close to zero due to inactivity
1289 	 * of backing device.
1290 	 */
1291 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1292 		dirty = dtc->wb_dirty;
1293 		if (dtc->wb_dirty < 8)
1294 			setpoint = dtc->wb_dirty + 1;
1295 		else
1296 			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1297 	}
1298 
1299 	if (dirty < setpoint) {
1300 		x = min3(wb->balanced_dirty_ratelimit,
1301 			 balanced_dirty_ratelimit, task_ratelimit);
1302 		if (dirty_ratelimit < x)
1303 			step = x - dirty_ratelimit;
1304 	} else {
1305 		x = max3(wb->balanced_dirty_ratelimit,
1306 			 balanced_dirty_ratelimit, task_ratelimit);
1307 		if (dirty_ratelimit > x)
1308 			step = dirty_ratelimit - x;
1309 	}
1310 
1311 	/*
1312 	 * Don't pursue 100% rate matching. It's impossible since the balanced
1313 	 * rate itself is constantly fluctuating. So decrease the track speed
1314 	 * when it gets close to the target. Helps eliminate pointless tremors.
1315 	 */
1316 	shift = dirty_ratelimit / (2 * step + 1);
1317 	if (shift < BITS_PER_LONG)
1318 		step = DIV_ROUND_UP(step >> shift, 8);
1319 	else
1320 		step = 0;
1321 
1322 	if (dirty_ratelimit < balanced_dirty_ratelimit)
1323 		dirty_ratelimit += step;
1324 	else
1325 		dirty_ratelimit -= step;
1326 
1327 	WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1328 	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1329 
1330 	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1331 }
1332 
1333 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1334 				  struct dirty_throttle_control *mdtc,
1335 				  bool update_ratelimit)
1336 {
1337 	struct bdi_writeback *wb = gdtc->wb;
1338 	unsigned long now = jiffies;
1339 	unsigned long elapsed;
1340 	unsigned long dirtied;
1341 	unsigned long written;
1342 
1343 	spin_lock(&wb->list_lock);
1344 
1345 	/*
1346 	 * Lockless checks for elapsed time are racy and delayed update after
1347 	 * IO completion doesn't do it at all (to make sure written pages are
1348 	 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1349 	 * division errors.
1350 	 */
1351 	elapsed = max(now - wb->bw_time_stamp, 1UL);
1352 	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1353 	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1354 
1355 	if (update_ratelimit) {
1356 		domain_update_dirty_limit(gdtc, now);
1357 		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1358 
1359 		/*
1360 		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1361 		 * compiler has no way to figure that out.  Help it.
1362 		 */
1363 		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1364 			domain_update_dirty_limit(mdtc, now);
1365 			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1366 		}
1367 	}
1368 	wb_update_write_bandwidth(wb, elapsed, written);
1369 
1370 	wb->dirtied_stamp = dirtied;
1371 	wb->written_stamp = written;
1372 	WRITE_ONCE(wb->bw_time_stamp, now);
1373 	spin_unlock(&wb->list_lock);
1374 }
1375 
1376 void wb_update_bandwidth(struct bdi_writeback *wb)
1377 {
1378 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1379 
1380 	__wb_update_bandwidth(&gdtc, NULL, false);
1381 }
1382 
1383 /* Interval after which we consider wb idle and don't estimate bandwidth */
1384 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1385 
1386 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1387 {
1388 	unsigned long now = jiffies;
1389 	unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1390 
1391 	if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1392 	    !atomic_read(&wb->writeback_inodes)) {
1393 		spin_lock(&wb->list_lock);
1394 		wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1395 		wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1396 		WRITE_ONCE(wb->bw_time_stamp, now);
1397 		spin_unlock(&wb->list_lock);
1398 	}
1399 }
1400 
1401 /*
1402  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1403  * will look to see if it needs to start dirty throttling.
1404  *
1405  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1406  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1407  * (the number of pages we may dirty without exceeding the dirty limits).
1408  */
1409 static unsigned long dirty_poll_interval(unsigned long dirty,
1410 					 unsigned long thresh)
1411 {
1412 	if (thresh > dirty)
1413 		return 1UL << (ilog2(thresh - dirty) >> 1);
1414 
1415 	return 1;
1416 }
1417 
1418 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1419 				  unsigned long wb_dirty)
1420 {
1421 	unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1422 	unsigned long t;
1423 
1424 	/*
1425 	 * Limit pause time for small memory systems. If sleeping for too long
1426 	 * time, a small pool of dirty/writeback pages may go empty and disk go
1427 	 * idle.
1428 	 *
1429 	 * 8 serves as the safety ratio.
1430 	 */
1431 	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1432 	t++;
1433 
1434 	return min_t(unsigned long, t, MAX_PAUSE);
1435 }
1436 
1437 static long wb_min_pause(struct bdi_writeback *wb,
1438 			 long max_pause,
1439 			 unsigned long task_ratelimit,
1440 			 unsigned long dirty_ratelimit,
1441 			 int *nr_dirtied_pause)
1442 {
1443 	long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1444 	long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1445 	long t;		/* target pause */
1446 	long pause;	/* estimated next pause */
1447 	int pages;	/* target nr_dirtied_pause */
1448 
1449 	/* target for 10ms pause on 1-dd case */
1450 	t = max(1, HZ / 100);
1451 
1452 	/*
1453 	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1454 	 * overheads.
1455 	 *
1456 	 * (N * 10ms) on 2^N concurrent tasks.
1457 	 */
1458 	if (hi > lo)
1459 		t += (hi - lo) * (10 * HZ) / 1024;
1460 
1461 	/*
1462 	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1463 	 * on the much more stable dirty_ratelimit. However the next pause time
1464 	 * will be computed based on task_ratelimit and the two rate limits may
1465 	 * depart considerably at some time. Especially if task_ratelimit goes
1466 	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1467 	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1468 	 * result task_ratelimit won't be executed faithfully, which could
1469 	 * eventually bring down dirty_ratelimit.
1470 	 *
1471 	 * We apply two rules to fix it up:
1472 	 * 1) try to estimate the next pause time and if necessary, use a lower
1473 	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1474 	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1475 	 * 2) limit the target pause time to max_pause/2, so that the normal
1476 	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1477 	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1478 	 */
1479 	t = min(t, 1 + max_pause / 2);
1480 	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1481 
1482 	/*
1483 	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1484 	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1485 	 * When the 16 consecutive reads are often interrupted by some dirty
1486 	 * throttling pause during the async writes, cfq will go into idles
1487 	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1488 	 * until reaches DIRTY_POLL_THRESH=32 pages.
1489 	 */
1490 	if (pages < DIRTY_POLL_THRESH) {
1491 		t = max_pause;
1492 		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1493 		if (pages > DIRTY_POLL_THRESH) {
1494 			pages = DIRTY_POLL_THRESH;
1495 			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1496 		}
1497 	}
1498 
1499 	pause = HZ * pages / (task_ratelimit + 1);
1500 	if (pause > max_pause) {
1501 		t = max_pause;
1502 		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1503 	}
1504 
1505 	*nr_dirtied_pause = pages;
1506 	/*
1507 	 * The minimal pause time will normally be half the target pause time.
1508 	 */
1509 	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1510 }
1511 
1512 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1513 {
1514 	struct bdi_writeback *wb = dtc->wb;
1515 	unsigned long wb_reclaimable;
1516 
1517 	/*
1518 	 * wb_thresh is not treated as some limiting factor as
1519 	 * dirty_thresh, due to reasons
1520 	 * - in JBOD setup, wb_thresh can fluctuate a lot
1521 	 * - in a system with HDD and USB key, the USB key may somehow
1522 	 *   go into state (wb_dirty >> wb_thresh) either because
1523 	 *   wb_dirty starts high, or because wb_thresh drops low.
1524 	 *   In this case we don't want to hard throttle the USB key
1525 	 *   dirtiers for 100 seconds until wb_dirty drops under
1526 	 *   wb_thresh. Instead the auxiliary wb control line in
1527 	 *   wb_position_ratio() will let the dirtier task progress
1528 	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1529 	 */
1530 	dtc->wb_thresh = __wb_calc_thresh(dtc);
1531 	dtc->wb_bg_thresh = dtc->thresh ?
1532 		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1533 
1534 	/*
1535 	 * In order to avoid the stacked BDI deadlock we need
1536 	 * to ensure we accurately count the 'dirty' pages when
1537 	 * the threshold is low.
1538 	 *
1539 	 * Otherwise it would be possible to get thresh+n pages
1540 	 * reported dirty, even though there are thresh-m pages
1541 	 * actually dirty; with m+n sitting in the percpu
1542 	 * deltas.
1543 	 */
1544 	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1545 		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1546 		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1547 	} else {
1548 		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1549 		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1550 	}
1551 }
1552 
1553 /*
1554  * balance_dirty_pages() must be called by processes which are generating dirty
1555  * data.  It looks at the number of dirty pages in the machine and will force
1556  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1557  * If we're over `background_thresh' then the writeback threads are woken to
1558  * perform some writeout.
1559  */
1560 static void balance_dirty_pages(struct bdi_writeback *wb,
1561 				unsigned long pages_dirtied)
1562 {
1563 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1564 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1565 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1566 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1567 						     &mdtc_stor : NULL;
1568 	struct dirty_throttle_control *sdtc;
1569 	unsigned long nr_reclaimable;	/* = file_dirty */
1570 	long period;
1571 	long pause;
1572 	long max_pause;
1573 	long min_pause;
1574 	int nr_dirtied_pause;
1575 	bool dirty_exceeded = false;
1576 	unsigned long task_ratelimit;
1577 	unsigned long dirty_ratelimit;
1578 	struct backing_dev_info *bdi = wb->bdi;
1579 	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1580 	unsigned long start_time = jiffies;
1581 
1582 	for (;;) {
1583 		unsigned long now = jiffies;
1584 		unsigned long dirty, thresh, bg_thresh;
1585 		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1586 		unsigned long m_thresh = 0;
1587 		unsigned long m_bg_thresh = 0;
1588 
1589 		nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1590 		gdtc->avail = global_dirtyable_memory();
1591 		gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1592 
1593 		domain_dirty_limits(gdtc);
1594 
1595 		if (unlikely(strictlimit)) {
1596 			wb_dirty_limits(gdtc);
1597 
1598 			dirty = gdtc->wb_dirty;
1599 			thresh = gdtc->wb_thresh;
1600 			bg_thresh = gdtc->wb_bg_thresh;
1601 		} else {
1602 			dirty = gdtc->dirty;
1603 			thresh = gdtc->thresh;
1604 			bg_thresh = gdtc->bg_thresh;
1605 		}
1606 
1607 		if (mdtc) {
1608 			unsigned long filepages, headroom, writeback;
1609 
1610 			/*
1611 			 * If @wb belongs to !root memcg, repeat the same
1612 			 * basic calculations for the memcg domain.
1613 			 */
1614 			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1615 					    &mdtc->dirty, &writeback);
1616 			mdtc->dirty += writeback;
1617 			mdtc_calc_avail(mdtc, filepages, headroom);
1618 
1619 			domain_dirty_limits(mdtc);
1620 
1621 			if (unlikely(strictlimit)) {
1622 				wb_dirty_limits(mdtc);
1623 				m_dirty = mdtc->wb_dirty;
1624 				m_thresh = mdtc->wb_thresh;
1625 				m_bg_thresh = mdtc->wb_bg_thresh;
1626 			} else {
1627 				m_dirty = mdtc->dirty;
1628 				m_thresh = mdtc->thresh;
1629 				m_bg_thresh = mdtc->bg_thresh;
1630 			}
1631 		}
1632 
1633 		/*
1634 		 * Throttle it only when the background writeback cannot
1635 		 * catch-up. This avoids (excessively) small writeouts
1636 		 * when the wb limits are ramping up in case of !strictlimit.
1637 		 *
1638 		 * In strictlimit case make decision based on the wb counters
1639 		 * and limits. Small writeouts when the wb limits are ramping
1640 		 * up are the price we consciously pay for strictlimit-ing.
1641 		 *
1642 		 * If memcg domain is in effect, @dirty should be under
1643 		 * both global and memcg freerun ceilings.
1644 		 */
1645 		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1646 		    (!mdtc ||
1647 		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1648 			unsigned long intv;
1649 			unsigned long m_intv;
1650 
1651 free_running:
1652 			intv = dirty_poll_interval(dirty, thresh);
1653 			m_intv = ULONG_MAX;
1654 
1655 			current->dirty_paused_when = now;
1656 			current->nr_dirtied = 0;
1657 			if (mdtc)
1658 				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1659 			current->nr_dirtied_pause = min(intv, m_intv);
1660 			break;
1661 		}
1662 
1663 		if (unlikely(!writeback_in_progress(wb)))
1664 			wb_start_background_writeback(wb);
1665 
1666 		mem_cgroup_flush_foreign(wb);
1667 
1668 		/*
1669 		 * Calculate global domain's pos_ratio and select the
1670 		 * global dtc by default.
1671 		 */
1672 		if (!strictlimit) {
1673 			wb_dirty_limits(gdtc);
1674 
1675 			if ((current->flags & PF_LOCAL_THROTTLE) &&
1676 			    gdtc->wb_dirty <
1677 			    dirty_freerun_ceiling(gdtc->wb_thresh,
1678 						  gdtc->wb_bg_thresh))
1679 				/*
1680 				 * LOCAL_THROTTLE tasks must not be throttled
1681 				 * when below the per-wb freerun ceiling.
1682 				 */
1683 				goto free_running;
1684 		}
1685 
1686 		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1687 			((gdtc->dirty > gdtc->thresh) || strictlimit);
1688 
1689 		wb_position_ratio(gdtc);
1690 		sdtc = gdtc;
1691 
1692 		if (mdtc) {
1693 			/*
1694 			 * If memcg domain is in effect, calculate its
1695 			 * pos_ratio.  @wb should satisfy constraints from
1696 			 * both global and memcg domains.  Choose the one
1697 			 * w/ lower pos_ratio.
1698 			 */
1699 			if (!strictlimit) {
1700 				wb_dirty_limits(mdtc);
1701 
1702 				if ((current->flags & PF_LOCAL_THROTTLE) &&
1703 				    mdtc->wb_dirty <
1704 				    dirty_freerun_ceiling(mdtc->wb_thresh,
1705 							  mdtc->wb_bg_thresh))
1706 					/*
1707 					 * LOCAL_THROTTLE tasks must not be
1708 					 * throttled when below the per-wb
1709 					 * freerun ceiling.
1710 					 */
1711 					goto free_running;
1712 			}
1713 			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1714 				((mdtc->dirty > mdtc->thresh) || strictlimit);
1715 
1716 			wb_position_ratio(mdtc);
1717 			if (mdtc->pos_ratio < gdtc->pos_ratio)
1718 				sdtc = mdtc;
1719 		}
1720 
1721 		if (dirty_exceeded && !wb->dirty_exceeded)
1722 			wb->dirty_exceeded = 1;
1723 
1724 		if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1725 					   BANDWIDTH_INTERVAL))
1726 			__wb_update_bandwidth(gdtc, mdtc, true);
1727 
1728 		/* throttle according to the chosen dtc */
1729 		dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1730 		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1731 							RATELIMIT_CALC_SHIFT;
1732 		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1733 		min_pause = wb_min_pause(wb, max_pause,
1734 					 task_ratelimit, dirty_ratelimit,
1735 					 &nr_dirtied_pause);
1736 
1737 		if (unlikely(task_ratelimit == 0)) {
1738 			period = max_pause;
1739 			pause = max_pause;
1740 			goto pause;
1741 		}
1742 		period = HZ * pages_dirtied / task_ratelimit;
1743 		pause = period;
1744 		if (current->dirty_paused_when)
1745 			pause -= now - current->dirty_paused_when;
1746 		/*
1747 		 * For less than 1s think time (ext3/4 may block the dirtier
1748 		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1749 		 * however at much less frequency), try to compensate it in
1750 		 * future periods by updating the virtual time; otherwise just
1751 		 * do a reset, as it may be a light dirtier.
1752 		 */
1753 		if (pause < min_pause) {
1754 			trace_balance_dirty_pages(wb,
1755 						  sdtc->thresh,
1756 						  sdtc->bg_thresh,
1757 						  sdtc->dirty,
1758 						  sdtc->wb_thresh,
1759 						  sdtc->wb_dirty,
1760 						  dirty_ratelimit,
1761 						  task_ratelimit,
1762 						  pages_dirtied,
1763 						  period,
1764 						  min(pause, 0L),
1765 						  start_time);
1766 			if (pause < -HZ) {
1767 				current->dirty_paused_when = now;
1768 				current->nr_dirtied = 0;
1769 			} else if (period) {
1770 				current->dirty_paused_when += period;
1771 				current->nr_dirtied = 0;
1772 			} else if (current->nr_dirtied_pause <= pages_dirtied)
1773 				current->nr_dirtied_pause += pages_dirtied;
1774 			break;
1775 		}
1776 		if (unlikely(pause > max_pause)) {
1777 			/* for occasional dropped task_ratelimit */
1778 			now += min(pause - max_pause, max_pause);
1779 			pause = max_pause;
1780 		}
1781 
1782 pause:
1783 		trace_balance_dirty_pages(wb,
1784 					  sdtc->thresh,
1785 					  sdtc->bg_thresh,
1786 					  sdtc->dirty,
1787 					  sdtc->wb_thresh,
1788 					  sdtc->wb_dirty,
1789 					  dirty_ratelimit,
1790 					  task_ratelimit,
1791 					  pages_dirtied,
1792 					  period,
1793 					  pause,
1794 					  start_time);
1795 		__set_current_state(TASK_KILLABLE);
1796 		wb->dirty_sleep = now;
1797 		io_schedule_timeout(pause);
1798 
1799 		current->dirty_paused_when = now + pause;
1800 		current->nr_dirtied = 0;
1801 		current->nr_dirtied_pause = nr_dirtied_pause;
1802 
1803 		/*
1804 		 * This is typically equal to (dirty < thresh) and can also
1805 		 * keep "1000+ dd on a slow USB stick" under control.
1806 		 */
1807 		if (task_ratelimit)
1808 			break;
1809 
1810 		/*
1811 		 * In the case of an unresponsive NFS server and the NFS dirty
1812 		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1813 		 * to go through, so that tasks on them still remain responsive.
1814 		 *
1815 		 * In theory 1 page is enough to keep the consumer-producer
1816 		 * pipe going: the flusher cleans 1 page => the task dirties 1
1817 		 * more page. However wb_dirty has accounting errors.  So use
1818 		 * the larger and more IO friendly wb_stat_error.
1819 		 */
1820 		if (sdtc->wb_dirty <= wb_stat_error())
1821 			break;
1822 
1823 		if (fatal_signal_pending(current))
1824 			break;
1825 	}
1826 
1827 	if (!dirty_exceeded && wb->dirty_exceeded)
1828 		wb->dirty_exceeded = 0;
1829 
1830 	if (writeback_in_progress(wb))
1831 		return;
1832 
1833 	/*
1834 	 * In laptop mode, we wait until hitting the higher threshold before
1835 	 * starting background writeout, and then write out all the way down
1836 	 * to the lower threshold.  So slow writers cause minimal disk activity.
1837 	 *
1838 	 * In normal mode, we start background writeout at the lower
1839 	 * background_thresh, to keep the amount of dirty memory low.
1840 	 */
1841 	if (laptop_mode)
1842 		return;
1843 
1844 	if (nr_reclaimable > gdtc->bg_thresh)
1845 		wb_start_background_writeback(wb);
1846 }
1847 
1848 static DEFINE_PER_CPU(int, bdp_ratelimits);
1849 
1850 /*
1851  * Normal tasks are throttled by
1852  *	loop {
1853  *		dirty tsk->nr_dirtied_pause pages;
1854  *		take a snap in balance_dirty_pages();
1855  *	}
1856  * However there is a worst case. If every task exit immediately when dirtied
1857  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1858  * called to throttle the page dirties. The solution is to save the not yet
1859  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1860  * randomly into the running tasks. This works well for the above worst case,
1861  * as the new task will pick up and accumulate the old task's leaked dirty
1862  * count and eventually get throttled.
1863  */
1864 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1865 
1866 /**
1867  * balance_dirty_pages_ratelimited - balance dirty memory state
1868  * @mapping: address_space which was dirtied
1869  *
1870  * Processes which are dirtying memory should call in here once for each page
1871  * which was newly dirtied.  The function will periodically check the system's
1872  * dirty state and will initiate writeback if needed.
1873  *
1874  * Once we're over the dirty memory limit we decrease the ratelimiting
1875  * by a lot, to prevent individual processes from overshooting the limit
1876  * by (ratelimit_pages) each.
1877  */
1878 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1879 {
1880 	struct inode *inode = mapping->host;
1881 	struct backing_dev_info *bdi = inode_to_bdi(inode);
1882 	struct bdi_writeback *wb = NULL;
1883 	int ratelimit;
1884 	int *p;
1885 
1886 	if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1887 		return;
1888 
1889 	if (inode_cgwb_enabled(inode))
1890 		wb = wb_get_create_current(bdi, GFP_KERNEL);
1891 	if (!wb)
1892 		wb = &bdi->wb;
1893 
1894 	ratelimit = current->nr_dirtied_pause;
1895 	if (wb->dirty_exceeded)
1896 		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1897 
1898 	preempt_disable();
1899 	/*
1900 	 * This prevents one CPU to accumulate too many dirtied pages without
1901 	 * calling into balance_dirty_pages(), which can happen when there are
1902 	 * 1000+ tasks, all of them start dirtying pages at exactly the same
1903 	 * time, hence all honoured too large initial task->nr_dirtied_pause.
1904 	 */
1905 	p =  this_cpu_ptr(&bdp_ratelimits);
1906 	if (unlikely(current->nr_dirtied >= ratelimit))
1907 		*p = 0;
1908 	else if (unlikely(*p >= ratelimit_pages)) {
1909 		*p = 0;
1910 		ratelimit = 0;
1911 	}
1912 	/*
1913 	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1914 	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1915 	 * the dirty throttling and livelock other long-run dirtiers.
1916 	 */
1917 	p = this_cpu_ptr(&dirty_throttle_leaks);
1918 	if (*p > 0 && current->nr_dirtied < ratelimit) {
1919 		unsigned long nr_pages_dirtied;
1920 		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1921 		*p -= nr_pages_dirtied;
1922 		current->nr_dirtied += nr_pages_dirtied;
1923 	}
1924 	preempt_enable();
1925 
1926 	if (unlikely(current->nr_dirtied >= ratelimit))
1927 		balance_dirty_pages(wb, current->nr_dirtied);
1928 
1929 	wb_put(wb);
1930 }
1931 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1932 
1933 /**
1934  * wb_over_bg_thresh - does @wb need to be written back?
1935  * @wb: bdi_writeback of interest
1936  *
1937  * Determines whether background writeback should keep writing @wb or it's
1938  * clean enough.
1939  *
1940  * Return: %true if writeback should continue.
1941  */
1942 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1943 {
1944 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1945 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1946 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1947 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1948 						     &mdtc_stor : NULL;
1949 	unsigned long reclaimable;
1950 	unsigned long thresh;
1951 
1952 	/*
1953 	 * Similar to balance_dirty_pages() but ignores pages being written
1954 	 * as we're trying to decide whether to put more under writeback.
1955 	 */
1956 	gdtc->avail = global_dirtyable_memory();
1957 	gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1958 	domain_dirty_limits(gdtc);
1959 
1960 	if (gdtc->dirty > gdtc->bg_thresh)
1961 		return true;
1962 
1963 	thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1964 	if (thresh < 2 * wb_stat_error())
1965 		reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1966 	else
1967 		reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1968 
1969 	if (reclaimable > thresh)
1970 		return true;
1971 
1972 	if (mdtc) {
1973 		unsigned long filepages, headroom, writeback;
1974 
1975 		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1976 				    &writeback);
1977 		mdtc_calc_avail(mdtc, filepages, headroom);
1978 		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
1979 
1980 		if (mdtc->dirty > mdtc->bg_thresh)
1981 			return true;
1982 
1983 		thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1984 		if (thresh < 2 * wb_stat_error())
1985 			reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1986 		else
1987 			reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1988 
1989 		if (reclaimable > thresh)
1990 			return true;
1991 	}
1992 
1993 	return false;
1994 }
1995 
1996 /*
1997  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1998  */
1999 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2000 		void *buffer, size_t *length, loff_t *ppos)
2001 {
2002 	unsigned int old_interval = dirty_writeback_interval;
2003 	int ret;
2004 
2005 	ret = proc_dointvec(table, write, buffer, length, ppos);
2006 
2007 	/*
2008 	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2009 	 * and a different non-zero value will wakeup the writeback threads.
2010 	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2011 	 * iterate over all bdis and wbs.
2012 	 * The reason we do this is to make the change take effect immediately.
2013 	 */
2014 	if (!ret && write && dirty_writeback_interval &&
2015 		dirty_writeback_interval != old_interval)
2016 		wakeup_flusher_threads(WB_REASON_PERIODIC);
2017 
2018 	return ret;
2019 }
2020 
2021 void laptop_mode_timer_fn(struct timer_list *t)
2022 {
2023 	struct backing_dev_info *backing_dev_info =
2024 		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2025 
2026 	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2027 }
2028 
2029 /*
2030  * We've spun up the disk and we're in laptop mode: schedule writeback
2031  * of all dirty data a few seconds from now.  If the flush is already scheduled
2032  * then push it back - the user is still using the disk.
2033  */
2034 void laptop_io_completion(struct backing_dev_info *info)
2035 {
2036 	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2037 }
2038 
2039 /*
2040  * We're in laptop mode and we've just synced. The sync's writes will have
2041  * caused another writeback to be scheduled by laptop_io_completion.
2042  * Nothing needs to be written back anymore, so we unschedule the writeback.
2043  */
2044 void laptop_sync_completion(void)
2045 {
2046 	struct backing_dev_info *bdi;
2047 
2048 	rcu_read_lock();
2049 
2050 	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2051 		del_timer(&bdi->laptop_mode_wb_timer);
2052 
2053 	rcu_read_unlock();
2054 }
2055 
2056 /*
2057  * If ratelimit_pages is too high then we can get into dirty-data overload
2058  * if a large number of processes all perform writes at the same time.
2059  *
2060  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2061  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2062  * thresholds.
2063  */
2064 
2065 void writeback_set_ratelimit(void)
2066 {
2067 	struct wb_domain *dom = &global_wb_domain;
2068 	unsigned long background_thresh;
2069 	unsigned long dirty_thresh;
2070 
2071 	global_dirty_limits(&background_thresh, &dirty_thresh);
2072 	dom->dirty_limit = dirty_thresh;
2073 	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2074 	if (ratelimit_pages < 16)
2075 		ratelimit_pages = 16;
2076 }
2077 
2078 static int page_writeback_cpu_online(unsigned int cpu)
2079 {
2080 	writeback_set_ratelimit();
2081 	return 0;
2082 }
2083 
2084 /*
2085  * Called early on to tune the page writeback dirty limits.
2086  *
2087  * We used to scale dirty pages according to how total memory
2088  * related to pages that could be allocated for buffers.
2089  *
2090  * However, that was when we used "dirty_ratio" to scale with
2091  * all memory, and we don't do that any more. "dirty_ratio"
2092  * is now applied to total non-HIGHPAGE memory, and as such we can't
2093  * get into the old insane situation any more where we had
2094  * large amounts of dirty pages compared to a small amount of
2095  * non-HIGHMEM memory.
2096  *
2097  * But we might still want to scale the dirty_ratio by how
2098  * much memory the box has..
2099  */
2100 void __init page_writeback_init(void)
2101 {
2102 	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2103 
2104 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2105 			  page_writeback_cpu_online, NULL);
2106 	cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2107 			  page_writeback_cpu_online);
2108 }
2109 
2110 /**
2111  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2112  * @mapping: address space structure to write
2113  * @start: starting page index
2114  * @end: ending page index (inclusive)
2115  *
2116  * This function scans the page range from @start to @end (inclusive) and tags
2117  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2118  * that write_cache_pages (or whoever calls this function) will then use
2119  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2120  * used to avoid livelocking of writeback by a process steadily creating new
2121  * dirty pages in the file (thus it is important for this function to be quick
2122  * so that it can tag pages faster than a dirtying process can create them).
2123  */
2124 void tag_pages_for_writeback(struct address_space *mapping,
2125 			     pgoff_t start, pgoff_t end)
2126 {
2127 	XA_STATE(xas, &mapping->i_pages, start);
2128 	unsigned int tagged = 0;
2129 	void *page;
2130 
2131 	xas_lock_irq(&xas);
2132 	xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2133 		xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2134 		if (++tagged % XA_CHECK_SCHED)
2135 			continue;
2136 
2137 		xas_pause(&xas);
2138 		xas_unlock_irq(&xas);
2139 		cond_resched();
2140 		xas_lock_irq(&xas);
2141 	}
2142 	xas_unlock_irq(&xas);
2143 }
2144 EXPORT_SYMBOL(tag_pages_for_writeback);
2145 
2146 /**
2147  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2148  * @mapping: address space structure to write
2149  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2150  * @writepage: function called for each page
2151  * @data: data passed to writepage function
2152  *
2153  * If a page is already under I/O, write_cache_pages() skips it, even
2154  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2155  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2156  * and msync() need to guarantee that all the data which was dirty at the time
2157  * the call was made get new I/O started against them.  If wbc->sync_mode is
2158  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2159  * existing IO to complete.
2160  *
2161  * To avoid livelocks (when other process dirties new pages), we first tag
2162  * pages which should be written back with TOWRITE tag and only then start
2163  * writing them. For data-integrity sync we have to be careful so that we do
2164  * not miss some pages (e.g., because some other process has cleared TOWRITE
2165  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2166  * by the process clearing the DIRTY tag (and submitting the page for IO).
2167  *
2168  * To avoid deadlocks between range_cyclic writeback and callers that hold
2169  * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2170  * we do not loop back to the start of the file. Doing so causes a page
2171  * lock/page writeback access order inversion - we should only ever lock
2172  * multiple pages in ascending page->index order, and looping back to the start
2173  * of the file violates that rule and causes deadlocks.
2174  *
2175  * Return: %0 on success, negative error code otherwise
2176  */
2177 int write_cache_pages(struct address_space *mapping,
2178 		      struct writeback_control *wbc, writepage_t writepage,
2179 		      void *data)
2180 {
2181 	int ret = 0;
2182 	int done = 0;
2183 	int error;
2184 	struct pagevec pvec;
2185 	int nr_pages;
2186 	pgoff_t index;
2187 	pgoff_t end;		/* Inclusive */
2188 	pgoff_t done_index;
2189 	int range_whole = 0;
2190 	xa_mark_t tag;
2191 
2192 	pagevec_init(&pvec);
2193 	if (wbc->range_cyclic) {
2194 		index = mapping->writeback_index; /* prev offset */
2195 		end = -1;
2196 	} else {
2197 		index = wbc->range_start >> PAGE_SHIFT;
2198 		end = wbc->range_end >> PAGE_SHIFT;
2199 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2200 			range_whole = 1;
2201 	}
2202 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2203 		tag_pages_for_writeback(mapping, index, end);
2204 		tag = PAGECACHE_TAG_TOWRITE;
2205 	} else {
2206 		tag = PAGECACHE_TAG_DIRTY;
2207 	}
2208 	done_index = index;
2209 	while (!done && (index <= end)) {
2210 		int i;
2211 
2212 		nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2213 				tag);
2214 		if (nr_pages == 0)
2215 			break;
2216 
2217 		for (i = 0; i < nr_pages; i++) {
2218 			struct page *page = pvec.pages[i];
2219 
2220 			done_index = page->index;
2221 
2222 			lock_page(page);
2223 
2224 			/*
2225 			 * Page truncated or invalidated. We can freely skip it
2226 			 * then, even for data integrity operations: the page
2227 			 * has disappeared concurrently, so there could be no
2228 			 * real expectation of this data integrity operation
2229 			 * even if there is now a new, dirty page at the same
2230 			 * pagecache address.
2231 			 */
2232 			if (unlikely(page->mapping != mapping)) {
2233 continue_unlock:
2234 				unlock_page(page);
2235 				continue;
2236 			}
2237 
2238 			if (!PageDirty(page)) {
2239 				/* someone wrote it for us */
2240 				goto continue_unlock;
2241 			}
2242 
2243 			if (PageWriteback(page)) {
2244 				if (wbc->sync_mode != WB_SYNC_NONE)
2245 					wait_on_page_writeback(page);
2246 				else
2247 					goto continue_unlock;
2248 			}
2249 
2250 			BUG_ON(PageWriteback(page));
2251 			if (!clear_page_dirty_for_io(page))
2252 				goto continue_unlock;
2253 
2254 			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2255 			error = (*writepage)(page, wbc, data);
2256 			if (unlikely(error)) {
2257 				/*
2258 				 * Handle errors according to the type of
2259 				 * writeback. There's no need to continue for
2260 				 * background writeback. Just push done_index
2261 				 * past this page so media errors won't choke
2262 				 * writeout for the entire file. For integrity
2263 				 * writeback, we must process the entire dirty
2264 				 * set regardless of errors because the fs may
2265 				 * still have state to clear for each page. In
2266 				 * that case we continue processing and return
2267 				 * the first error.
2268 				 */
2269 				if (error == AOP_WRITEPAGE_ACTIVATE) {
2270 					unlock_page(page);
2271 					error = 0;
2272 				} else if (wbc->sync_mode != WB_SYNC_ALL) {
2273 					ret = error;
2274 					done_index = page->index + 1;
2275 					done = 1;
2276 					break;
2277 				}
2278 				if (!ret)
2279 					ret = error;
2280 			}
2281 
2282 			/*
2283 			 * We stop writing back only if we are not doing
2284 			 * integrity sync. In case of integrity sync we have to
2285 			 * keep going until we have written all the pages
2286 			 * we tagged for writeback prior to entering this loop.
2287 			 */
2288 			if (--wbc->nr_to_write <= 0 &&
2289 			    wbc->sync_mode == WB_SYNC_NONE) {
2290 				done = 1;
2291 				break;
2292 			}
2293 		}
2294 		pagevec_release(&pvec);
2295 		cond_resched();
2296 	}
2297 
2298 	/*
2299 	 * If we hit the last page and there is more work to be done: wrap
2300 	 * back the index back to the start of the file for the next
2301 	 * time we are called.
2302 	 */
2303 	if (wbc->range_cyclic && !done)
2304 		done_index = 0;
2305 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2306 		mapping->writeback_index = done_index;
2307 
2308 	return ret;
2309 }
2310 EXPORT_SYMBOL(write_cache_pages);
2311 
2312 /*
2313  * Function used by generic_writepages to call the real writepage
2314  * function and set the mapping flags on error
2315  */
2316 static int __writepage(struct page *page, struct writeback_control *wbc,
2317 		       void *data)
2318 {
2319 	struct address_space *mapping = data;
2320 	int ret = mapping->a_ops->writepage(page, wbc);
2321 	mapping_set_error(mapping, ret);
2322 	return ret;
2323 }
2324 
2325 /**
2326  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2327  * @mapping: address space structure to write
2328  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2329  *
2330  * This is a library function, which implements the writepages()
2331  * address_space_operation.
2332  *
2333  * Return: %0 on success, negative error code otherwise
2334  */
2335 int generic_writepages(struct address_space *mapping,
2336 		       struct writeback_control *wbc)
2337 {
2338 	struct blk_plug plug;
2339 	int ret;
2340 
2341 	/* deal with chardevs and other special file */
2342 	if (!mapping->a_ops->writepage)
2343 		return 0;
2344 
2345 	blk_start_plug(&plug);
2346 	ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2347 	blk_finish_plug(&plug);
2348 	return ret;
2349 }
2350 
2351 EXPORT_SYMBOL(generic_writepages);
2352 
2353 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2354 {
2355 	int ret;
2356 	struct bdi_writeback *wb;
2357 
2358 	if (wbc->nr_to_write <= 0)
2359 		return 0;
2360 	wb = inode_to_wb_wbc(mapping->host, wbc);
2361 	wb_bandwidth_estimate_start(wb);
2362 	while (1) {
2363 		if (mapping->a_ops->writepages)
2364 			ret = mapping->a_ops->writepages(mapping, wbc);
2365 		else
2366 			ret = generic_writepages(mapping, wbc);
2367 		if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2368 			break;
2369 		cond_resched();
2370 		congestion_wait(BLK_RW_ASYNC, HZ/50);
2371 	}
2372 	/*
2373 	 * Usually few pages are written by now from those we've just submitted
2374 	 * but if there's constant writeback being submitted, this makes sure
2375 	 * writeback bandwidth is updated once in a while.
2376 	 */
2377 	if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2378 				   BANDWIDTH_INTERVAL))
2379 		wb_update_bandwidth(wb);
2380 	return ret;
2381 }
2382 
2383 /**
2384  * folio_write_one - write out a single folio and wait on I/O.
2385  * @folio: The folio to write.
2386  *
2387  * The folio must be locked by the caller and will be unlocked upon return.
2388  *
2389  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2390  * function returns.
2391  *
2392  * Return: %0 on success, negative error code otherwise
2393  */
2394 int folio_write_one(struct folio *folio)
2395 {
2396 	struct address_space *mapping = folio->mapping;
2397 	int ret = 0;
2398 	struct writeback_control wbc = {
2399 		.sync_mode = WB_SYNC_ALL,
2400 		.nr_to_write = folio_nr_pages(folio),
2401 	};
2402 
2403 	BUG_ON(!folio_test_locked(folio));
2404 
2405 	folio_wait_writeback(folio);
2406 
2407 	if (folio_clear_dirty_for_io(folio)) {
2408 		folio_get(folio);
2409 		ret = mapping->a_ops->writepage(&folio->page, &wbc);
2410 		if (ret == 0)
2411 			folio_wait_writeback(folio);
2412 		folio_put(folio);
2413 	} else {
2414 		folio_unlock(folio);
2415 	}
2416 
2417 	if (!ret)
2418 		ret = filemap_check_errors(mapping);
2419 	return ret;
2420 }
2421 EXPORT_SYMBOL(folio_write_one);
2422 
2423 /*
2424  * For address_spaces which do not use buffers nor write back.
2425  */
2426 int __set_page_dirty_no_writeback(struct page *page)
2427 {
2428 	if (!PageDirty(page))
2429 		return !TestSetPageDirty(page);
2430 	return 0;
2431 }
2432 EXPORT_SYMBOL(__set_page_dirty_no_writeback);
2433 
2434 /*
2435  * Helper function for set_page_dirty family.
2436  *
2437  * Caller must hold lock_page_memcg().
2438  *
2439  * NOTE: This relies on being atomic wrt interrupts.
2440  */
2441 static void folio_account_dirtied(struct folio *folio,
2442 		struct address_space *mapping)
2443 {
2444 	struct inode *inode = mapping->host;
2445 
2446 	trace_writeback_dirty_folio(folio, mapping);
2447 
2448 	if (mapping_can_writeback(mapping)) {
2449 		struct bdi_writeback *wb;
2450 		long nr = folio_nr_pages(folio);
2451 
2452 		inode_attach_wb(inode, &folio->page);
2453 		wb = inode_to_wb(inode);
2454 
2455 		__lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2456 		__zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2457 		__node_stat_mod_folio(folio, NR_DIRTIED, nr);
2458 		wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2459 		wb_stat_mod(wb, WB_DIRTIED, nr);
2460 		task_io_account_write(nr * PAGE_SIZE);
2461 		current->nr_dirtied += nr;
2462 		__this_cpu_add(bdp_ratelimits, nr);
2463 
2464 		mem_cgroup_track_foreign_dirty(folio, wb);
2465 	}
2466 }
2467 
2468 /*
2469  * Helper function for deaccounting dirty page without writeback.
2470  *
2471  * Caller must hold lock_page_memcg().
2472  */
2473 void folio_account_cleaned(struct folio *folio, struct address_space *mapping,
2474 			  struct bdi_writeback *wb)
2475 {
2476 	if (mapping_can_writeback(mapping)) {
2477 		long nr = folio_nr_pages(folio);
2478 		lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2479 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2480 		wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2481 		task_io_account_cancelled_write(nr * PAGE_SIZE);
2482 	}
2483 }
2484 
2485 /*
2486  * Mark the folio dirty, and set it dirty in the page cache, and mark
2487  * the inode dirty.
2488  *
2489  * If warn is true, then emit a warning if the folio is not uptodate and has
2490  * not been truncated.
2491  *
2492  * The caller must hold lock_page_memcg().
2493  */
2494 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2495 			     int warn)
2496 {
2497 	unsigned long flags;
2498 
2499 	xa_lock_irqsave(&mapping->i_pages, flags);
2500 	if (folio->mapping) {	/* Race with truncate? */
2501 		WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2502 		folio_account_dirtied(folio, mapping);
2503 		__xa_set_mark(&mapping->i_pages, folio_index(folio),
2504 				PAGECACHE_TAG_DIRTY);
2505 	}
2506 	xa_unlock_irqrestore(&mapping->i_pages, flags);
2507 }
2508 
2509 /**
2510  * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2511  * @mapping: Address space this folio belongs to.
2512  * @folio: Folio to be marked as dirty.
2513  *
2514  * Filesystems which do not use buffer heads should call this function
2515  * from their set_page_dirty address space operation.  It ignores the
2516  * contents of folio_get_private(), so if the filesystem marks individual
2517  * blocks as dirty, the filesystem should handle that itself.
2518  *
2519  * This is also sometimes used by filesystems which use buffer_heads when
2520  * a single buffer is being dirtied: we want to set the folio dirty in
2521  * that case, but not all the buffers.  This is a "bottom-up" dirtying,
2522  * whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2523  *
2524  * The caller must ensure this doesn't race with truncation.  Most will
2525  * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2526  * folio mapped and the pte lock held, which also locks out truncation.
2527  */
2528 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2529 {
2530 	folio_memcg_lock(folio);
2531 	if (folio_test_set_dirty(folio)) {
2532 		folio_memcg_unlock(folio);
2533 		return false;
2534 	}
2535 
2536 	__folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2537 	folio_memcg_unlock(folio);
2538 
2539 	if (mapping->host) {
2540 		/* !PageAnon && !swapper_space */
2541 		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2542 	}
2543 	return true;
2544 }
2545 EXPORT_SYMBOL(filemap_dirty_folio);
2546 
2547 /**
2548  * folio_account_redirty - Manually account for redirtying a page.
2549  * @folio: The folio which is being redirtied.
2550  *
2551  * Most filesystems should call folio_redirty_for_writepage() instead
2552  * of this fuction.  If your filesystem is doing writeback outside the
2553  * context of a writeback_control(), it can call this when redirtying
2554  * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2555  * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2556  * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2557  * in balanced_dirty_ratelimit and the dirty pages position control.
2558  */
2559 void folio_account_redirty(struct folio *folio)
2560 {
2561 	struct address_space *mapping = folio->mapping;
2562 
2563 	if (mapping && mapping_can_writeback(mapping)) {
2564 		struct inode *inode = mapping->host;
2565 		struct bdi_writeback *wb;
2566 		struct wb_lock_cookie cookie = {};
2567 		long nr = folio_nr_pages(folio);
2568 
2569 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2570 		current->nr_dirtied -= nr;
2571 		node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2572 		wb_stat_mod(wb, WB_DIRTIED, -nr);
2573 		unlocked_inode_to_wb_end(inode, &cookie);
2574 	}
2575 }
2576 EXPORT_SYMBOL(folio_account_redirty);
2577 
2578 /**
2579  * folio_redirty_for_writepage - Decline to write a dirty folio.
2580  * @wbc: The writeback control.
2581  * @folio: The folio.
2582  *
2583  * When a writepage implementation decides that it doesn't want to write
2584  * @folio for some reason, it should call this function, unlock @folio and
2585  * return 0.
2586  *
2587  * Return: True if we redirtied the folio.  False if someone else dirtied
2588  * it first.
2589  */
2590 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2591 		struct folio *folio)
2592 {
2593 	bool ret;
2594 	long nr = folio_nr_pages(folio);
2595 
2596 	wbc->pages_skipped += nr;
2597 	ret = filemap_dirty_folio(folio->mapping, folio);
2598 	folio_account_redirty(folio);
2599 
2600 	return ret;
2601 }
2602 EXPORT_SYMBOL(folio_redirty_for_writepage);
2603 
2604 /**
2605  * folio_mark_dirty - Mark a folio as being modified.
2606  * @folio: The folio.
2607  *
2608  * For folios with a mapping this should be done under the page lock
2609  * for the benefit of asynchronous memory errors who prefer a consistent
2610  * dirty state. This rule can be broken in some special cases,
2611  * but should be better not to.
2612  *
2613  * Return: True if the folio was newly dirtied, false if it was already dirty.
2614  */
2615 bool folio_mark_dirty(struct folio *folio)
2616 {
2617 	struct address_space *mapping = folio_mapping(folio);
2618 
2619 	if (likely(mapping)) {
2620 		/*
2621 		 * readahead/lru_deactivate_page could remain
2622 		 * PG_readahead/PG_reclaim due to race with end_page_writeback
2623 		 * About readahead, if the page is written, the flags would be
2624 		 * reset. So no problem.
2625 		 * About lru_deactivate_page, if the page is redirty, the flag
2626 		 * will be reset. So no problem. but if the page is used by readahead
2627 		 * it will confuse readahead and make it restart the size rampup
2628 		 * process. But it's a trivial problem.
2629 		 */
2630 		if (folio_test_reclaim(folio))
2631 			folio_clear_reclaim(folio);
2632 		return mapping->a_ops->set_page_dirty(&folio->page);
2633 	}
2634 	if (!folio_test_dirty(folio)) {
2635 		if (!folio_test_set_dirty(folio))
2636 			return true;
2637 	}
2638 	return false;
2639 }
2640 EXPORT_SYMBOL(folio_mark_dirty);
2641 
2642 /*
2643  * set_page_dirty() is racy if the caller has no reference against
2644  * page->mapping->host, and if the page is unlocked.  This is because another
2645  * CPU could truncate the page off the mapping and then free the mapping.
2646  *
2647  * Usually, the page _is_ locked, or the caller is a user-space process which
2648  * holds a reference on the inode by having an open file.
2649  *
2650  * In other cases, the page should be locked before running set_page_dirty().
2651  */
2652 int set_page_dirty_lock(struct page *page)
2653 {
2654 	int ret;
2655 
2656 	lock_page(page);
2657 	ret = set_page_dirty(page);
2658 	unlock_page(page);
2659 	return ret;
2660 }
2661 EXPORT_SYMBOL(set_page_dirty_lock);
2662 
2663 /*
2664  * This cancels just the dirty bit on the kernel page itself, it does NOT
2665  * actually remove dirty bits on any mmap's that may be around. It also
2666  * leaves the page tagged dirty, so any sync activity will still find it on
2667  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2668  * look at the dirty bits in the VM.
2669  *
2670  * Doing this should *normally* only ever be done when a page is truncated,
2671  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2672  * this when it notices that somebody has cleaned out all the buffers on a
2673  * page without actually doing it through the VM. Can you say "ext3 is
2674  * horribly ugly"? Thought you could.
2675  */
2676 void __folio_cancel_dirty(struct folio *folio)
2677 {
2678 	struct address_space *mapping = folio_mapping(folio);
2679 
2680 	if (mapping_can_writeback(mapping)) {
2681 		struct inode *inode = mapping->host;
2682 		struct bdi_writeback *wb;
2683 		struct wb_lock_cookie cookie = {};
2684 
2685 		folio_memcg_lock(folio);
2686 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2687 
2688 		if (folio_test_clear_dirty(folio))
2689 			folio_account_cleaned(folio, mapping, wb);
2690 
2691 		unlocked_inode_to_wb_end(inode, &cookie);
2692 		folio_memcg_unlock(folio);
2693 	} else {
2694 		folio_clear_dirty(folio);
2695 	}
2696 }
2697 EXPORT_SYMBOL(__folio_cancel_dirty);
2698 
2699 /*
2700  * Clear a folio's dirty flag, while caring for dirty memory accounting.
2701  * Returns true if the folio was previously dirty.
2702  *
2703  * This is for preparing to put the folio under writeout.  We leave
2704  * the folio tagged as dirty in the xarray so that a concurrent
2705  * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2706  * The ->writepage implementation will run either folio_start_writeback()
2707  * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2708  * and xarray dirty tag back into sync.
2709  *
2710  * This incoherency between the folio's dirty flag and xarray tag is
2711  * unfortunate, but it only exists while the folio is locked.
2712  */
2713 bool folio_clear_dirty_for_io(struct folio *folio)
2714 {
2715 	struct address_space *mapping = folio_mapping(folio);
2716 	bool ret = false;
2717 
2718 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2719 
2720 	if (mapping && mapping_can_writeback(mapping)) {
2721 		struct inode *inode = mapping->host;
2722 		struct bdi_writeback *wb;
2723 		struct wb_lock_cookie cookie = {};
2724 
2725 		/*
2726 		 * Yes, Virginia, this is indeed insane.
2727 		 *
2728 		 * We use this sequence to make sure that
2729 		 *  (a) we account for dirty stats properly
2730 		 *  (b) we tell the low-level filesystem to
2731 		 *      mark the whole folio dirty if it was
2732 		 *      dirty in a pagetable. Only to then
2733 		 *  (c) clean the folio again and return 1 to
2734 		 *      cause the writeback.
2735 		 *
2736 		 * This way we avoid all nasty races with the
2737 		 * dirty bit in multiple places and clearing
2738 		 * them concurrently from different threads.
2739 		 *
2740 		 * Note! Normally the "folio_mark_dirty(folio)"
2741 		 * has no effect on the actual dirty bit - since
2742 		 * that will already usually be set. But we
2743 		 * need the side effects, and it can help us
2744 		 * avoid races.
2745 		 *
2746 		 * We basically use the folio "master dirty bit"
2747 		 * as a serialization point for all the different
2748 		 * threads doing their things.
2749 		 */
2750 		if (folio_mkclean(folio))
2751 			folio_mark_dirty(folio);
2752 		/*
2753 		 * We carefully synchronise fault handlers against
2754 		 * installing a dirty pte and marking the folio dirty
2755 		 * at this point.  We do this by having them hold the
2756 		 * page lock while dirtying the folio, and folios are
2757 		 * always locked coming in here, so we get the desired
2758 		 * exclusion.
2759 		 */
2760 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2761 		if (folio_test_clear_dirty(folio)) {
2762 			long nr = folio_nr_pages(folio);
2763 			lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2764 			zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2765 			wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2766 			ret = true;
2767 		}
2768 		unlocked_inode_to_wb_end(inode, &cookie);
2769 		return ret;
2770 	}
2771 	return folio_test_clear_dirty(folio);
2772 }
2773 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2774 
2775 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2776 {
2777 	atomic_inc(&wb->writeback_inodes);
2778 }
2779 
2780 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2781 {
2782 	atomic_dec(&wb->writeback_inodes);
2783 	/*
2784 	 * Make sure estimate of writeback throughput gets updated after
2785 	 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2786 	 * (which is the interval other bandwidth updates use for batching) so
2787 	 * that if multiple inodes end writeback at a similar time, they get
2788 	 * batched into one bandwidth update.
2789 	 */
2790 	queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2791 }
2792 
2793 bool __folio_end_writeback(struct folio *folio)
2794 {
2795 	long nr = folio_nr_pages(folio);
2796 	struct address_space *mapping = folio_mapping(folio);
2797 	bool ret;
2798 
2799 	folio_memcg_lock(folio);
2800 	if (mapping && mapping_use_writeback_tags(mapping)) {
2801 		struct inode *inode = mapping->host;
2802 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2803 		unsigned long flags;
2804 
2805 		xa_lock_irqsave(&mapping->i_pages, flags);
2806 		ret = folio_test_clear_writeback(folio);
2807 		if (ret) {
2808 			__xa_clear_mark(&mapping->i_pages, folio_index(folio),
2809 						PAGECACHE_TAG_WRITEBACK);
2810 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2811 				struct bdi_writeback *wb = inode_to_wb(inode);
2812 
2813 				wb_stat_mod(wb, WB_WRITEBACK, -nr);
2814 				__wb_writeout_add(wb, nr);
2815 				if (!mapping_tagged(mapping,
2816 						    PAGECACHE_TAG_WRITEBACK))
2817 					wb_inode_writeback_end(wb);
2818 			}
2819 		}
2820 
2821 		if (mapping->host && !mapping_tagged(mapping,
2822 						     PAGECACHE_TAG_WRITEBACK))
2823 			sb_clear_inode_writeback(mapping->host);
2824 
2825 		xa_unlock_irqrestore(&mapping->i_pages, flags);
2826 	} else {
2827 		ret = folio_test_clear_writeback(folio);
2828 	}
2829 	if (ret) {
2830 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2831 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2832 		node_stat_mod_folio(folio, NR_WRITTEN, nr);
2833 	}
2834 	folio_memcg_unlock(folio);
2835 	return ret;
2836 }
2837 
2838 bool __folio_start_writeback(struct folio *folio, bool keep_write)
2839 {
2840 	long nr = folio_nr_pages(folio);
2841 	struct address_space *mapping = folio_mapping(folio);
2842 	bool ret;
2843 	int access_ret;
2844 
2845 	folio_memcg_lock(folio);
2846 	if (mapping && mapping_use_writeback_tags(mapping)) {
2847 		XA_STATE(xas, &mapping->i_pages, folio_index(folio));
2848 		struct inode *inode = mapping->host;
2849 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2850 		unsigned long flags;
2851 
2852 		xas_lock_irqsave(&xas, flags);
2853 		xas_load(&xas);
2854 		ret = folio_test_set_writeback(folio);
2855 		if (!ret) {
2856 			bool on_wblist;
2857 
2858 			on_wblist = mapping_tagged(mapping,
2859 						   PAGECACHE_TAG_WRITEBACK);
2860 
2861 			xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2862 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2863 				struct bdi_writeback *wb = inode_to_wb(inode);
2864 
2865 				wb_stat_mod(wb, WB_WRITEBACK, nr);
2866 				if (!on_wblist)
2867 					wb_inode_writeback_start(wb);
2868 			}
2869 
2870 			/*
2871 			 * We can come through here when swapping
2872 			 * anonymous folios, so we don't necessarily
2873 			 * have an inode to track for sync.
2874 			 */
2875 			if (mapping->host && !on_wblist)
2876 				sb_mark_inode_writeback(mapping->host);
2877 		}
2878 		if (!folio_test_dirty(folio))
2879 			xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2880 		if (!keep_write)
2881 			xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2882 		xas_unlock_irqrestore(&xas, flags);
2883 	} else {
2884 		ret = folio_test_set_writeback(folio);
2885 	}
2886 	if (!ret) {
2887 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
2888 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2889 	}
2890 	folio_memcg_unlock(folio);
2891 	access_ret = arch_make_folio_accessible(folio);
2892 	/*
2893 	 * If writeback has been triggered on a page that cannot be made
2894 	 * accessible, it is too late to recover here.
2895 	 */
2896 	VM_BUG_ON_FOLIO(access_ret != 0, folio);
2897 
2898 	return ret;
2899 }
2900 EXPORT_SYMBOL(__folio_start_writeback);
2901 
2902 /**
2903  * folio_wait_writeback - Wait for a folio to finish writeback.
2904  * @folio: The folio to wait for.
2905  *
2906  * If the folio is currently being written back to storage, wait for the
2907  * I/O to complete.
2908  *
2909  * Context: Sleeps.  Must be called in process context and with
2910  * no spinlocks held.  Caller should hold a reference on the folio.
2911  * If the folio is not locked, writeback may start again after writeback
2912  * has finished.
2913  */
2914 void folio_wait_writeback(struct folio *folio)
2915 {
2916 	while (folio_test_writeback(folio)) {
2917 		trace_folio_wait_writeback(folio, folio_mapping(folio));
2918 		folio_wait_bit(folio, PG_writeback);
2919 	}
2920 }
2921 EXPORT_SYMBOL_GPL(folio_wait_writeback);
2922 
2923 /**
2924  * folio_wait_writeback_killable - Wait for a folio to finish writeback.
2925  * @folio: The folio to wait for.
2926  *
2927  * If the folio is currently being written back to storage, wait for the
2928  * I/O to complete or a fatal signal to arrive.
2929  *
2930  * Context: Sleeps.  Must be called in process context and with
2931  * no spinlocks held.  Caller should hold a reference on the folio.
2932  * If the folio is not locked, writeback may start again after writeback
2933  * has finished.
2934  * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
2935  */
2936 int folio_wait_writeback_killable(struct folio *folio)
2937 {
2938 	while (folio_test_writeback(folio)) {
2939 		trace_folio_wait_writeback(folio, folio_mapping(folio));
2940 		if (folio_wait_bit_killable(folio, PG_writeback))
2941 			return -EINTR;
2942 	}
2943 
2944 	return 0;
2945 }
2946 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
2947 
2948 /**
2949  * folio_wait_stable() - wait for writeback to finish, if necessary.
2950  * @folio: The folio to wait on.
2951  *
2952  * This function determines if the given folio is related to a backing
2953  * device that requires folio contents to be held stable during writeback.
2954  * If so, then it will wait for any pending writeback to complete.
2955  *
2956  * Context: Sleeps.  Must be called in process context and with
2957  * no spinlocks held.  Caller should hold a reference on the folio.
2958  * If the folio is not locked, writeback may start again after writeback
2959  * has finished.
2960  */
2961 void folio_wait_stable(struct folio *folio)
2962 {
2963 	if (folio->mapping->host->i_sb->s_iflags & SB_I_STABLE_WRITES)
2964 		folio_wait_writeback(folio);
2965 }
2966 EXPORT_SYMBOL_GPL(folio_wait_stable);
2967