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 static 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 static 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 static int vm_highmem_is_dirtyable;
86 
87 /*
88  * The generator of dirty data starts writeback at this percentage
89  */
90 static 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 static 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 
mdtc_valid(struct dirty_throttle_control * dtc)163 static bool mdtc_valid(struct dirty_throttle_control *dtc)
164 {
165 	return dtc->dom;
166 }
167 
dtc_dom(struct dirty_throttle_control * dtc)168 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
169 {
170 	return dtc->dom;
171 }
172 
mdtc_gdtc(struct dirty_throttle_control * mdtc)173 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
174 {
175 	return mdtc->gdtc;
176 }
177 
wb_memcg_completions(struct bdi_writeback * wb)178 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
179 {
180 	return &wb->memcg_completions;
181 }
182 
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)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 
mdtc_valid(struct dirty_throttle_control * dtc)217 static bool mdtc_valid(struct dirty_throttle_control *dtc)
218 {
219 	return false;
220 }
221 
dtc_dom(struct dirty_throttle_control * dtc)222 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
223 {
224 	return &global_wb_domain;
225 }
226 
mdtc_gdtc(struct dirty_throttle_control * mdtc)227 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
228 {
229 	return NULL;
230 }
231 
wb_memcg_completions(struct bdi_writeback * wb)232 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
233 {
234 	return NULL;
235 }
236 
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)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  */
node_dirtyable_memory(struct pglist_data * pgdat)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 
highmem_dirtyable_memory(unsigned long total)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 	 * Make sure that the number of highmem pages is never larger
328 	 * than the number of the total dirtyable memory. This can only
329 	 * occur in very strange VM situations but we want to make sure
330 	 * that this does not occur.
331 	 */
332 	return min(x, total);
333 #else
334 	return 0;
335 #endif
336 }
337 
338 /**
339  * global_dirtyable_memory - number of globally dirtyable pages
340  *
341  * Return: the global number of pages potentially available for dirty
342  * page cache.  This is the base value for the global dirty limits.
343  */
global_dirtyable_memory(void)344 static unsigned long global_dirtyable_memory(void)
345 {
346 	unsigned long x;
347 
348 	x = global_zone_page_state(NR_FREE_PAGES);
349 	/*
350 	 * Pages reserved for the kernel should not be considered
351 	 * dirtyable, to prevent a situation where reclaim has to
352 	 * clean pages in order to balance the zones.
353 	 */
354 	x -= min(x, totalreserve_pages);
355 
356 	x += global_node_page_state(NR_INACTIVE_FILE);
357 	x += global_node_page_state(NR_ACTIVE_FILE);
358 
359 	if (!vm_highmem_is_dirtyable)
360 		x -= highmem_dirtyable_memory(x);
361 
362 	return x + 1;	/* Ensure that we never return 0 */
363 }
364 
365 /**
366  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
367  * @dtc: dirty_throttle_control of interest
368  *
369  * Calculate @dtc->thresh and ->bg_thresh considering
370  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
371  * must ensure that @dtc->avail is set before calling this function.  The
372  * dirty limits will be lifted by 1/4 for real-time tasks.
373  */
domain_dirty_limits(struct dirty_throttle_control * dtc)374 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
375 {
376 	const unsigned long available_memory = dtc->avail;
377 	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
378 	unsigned long bytes = vm_dirty_bytes;
379 	unsigned long bg_bytes = dirty_background_bytes;
380 	/* convert ratios to per-PAGE_SIZE for higher precision */
381 	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
382 	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
383 	unsigned long thresh;
384 	unsigned long bg_thresh;
385 	struct task_struct *tsk;
386 
387 	/* gdtc is !NULL iff @dtc is for memcg domain */
388 	if (gdtc) {
389 		unsigned long global_avail = gdtc->avail;
390 
391 		/*
392 		 * The byte settings can't be applied directly to memcg
393 		 * domains.  Convert them to ratios by scaling against
394 		 * globally available memory.  As the ratios are in
395 		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
396 		 * number of pages.
397 		 */
398 		if (bytes)
399 			ratio = min(DIV_ROUND_UP(bytes, global_avail),
400 				    PAGE_SIZE);
401 		if (bg_bytes)
402 			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
403 				       PAGE_SIZE);
404 		bytes = bg_bytes = 0;
405 	}
406 
407 	if (bytes)
408 		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
409 	else
410 		thresh = (ratio * available_memory) / PAGE_SIZE;
411 
412 	if (bg_bytes)
413 		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
414 	else
415 		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
416 
417 	if (bg_thresh >= thresh)
418 		bg_thresh = thresh / 2;
419 	tsk = current;
420 	if (rt_task(tsk)) {
421 		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
422 		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
423 	}
424 	dtc->thresh = thresh;
425 	dtc->bg_thresh = bg_thresh;
426 
427 	/* we should eventually report the domain in the TP */
428 	if (!gdtc)
429 		trace_global_dirty_state(bg_thresh, thresh);
430 }
431 
432 /**
433  * global_dirty_limits - background-writeback and dirty-throttling thresholds
434  * @pbackground: out parameter for bg_thresh
435  * @pdirty: out parameter for thresh
436  *
437  * Calculate bg_thresh and thresh for global_wb_domain.  See
438  * domain_dirty_limits() for details.
439  */
global_dirty_limits(unsigned long * pbackground,unsigned long * pdirty)440 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
441 {
442 	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
443 
444 	gdtc.avail = global_dirtyable_memory();
445 	domain_dirty_limits(&gdtc);
446 
447 	*pbackground = gdtc.bg_thresh;
448 	*pdirty = gdtc.thresh;
449 }
450 
451 /**
452  * node_dirty_limit - maximum number of dirty pages allowed in a node
453  * @pgdat: the node
454  *
455  * Return: the maximum number of dirty pages allowed in a node, based
456  * on the node's dirtyable memory.
457  */
node_dirty_limit(struct pglist_data * pgdat)458 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
459 {
460 	unsigned long node_memory = node_dirtyable_memory(pgdat);
461 	struct task_struct *tsk = current;
462 	unsigned long dirty;
463 
464 	if (vm_dirty_bytes)
465 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
466 			node_memory / global_dirtyable_memory();
467 	else
468 		dirty = vm_dirty_ratio * node_memory / 100;
469 
470 	if (rt_task(tsk))
471 		dirty += dirty / 4;
472 
473 	return dirty;
474 }
475 
476 /**
477  * node_dirty_ok - tells whether a node is within its dirty limits
478  * @pgdat: the node to check
479  *
480  * Return: %true when the dirty pages in @pgdat are within the node's
481  * dirty limit, %false if the limit is exceeded.
482  */
node_dirty_ok(struct pglist_data * pgdat)483 bool node_dirty_ok(struct pglist_data *pgdat)
484 {
485 	unsigned long limit = node_dirty_limit(pgdat);
486 	unsigned long nr_pages = 0;
487 
488 	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
489 	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
490 
491 	return nr_pages <= limit;
492 }
493 
494 #ifdef CONFIG_SYSCTL
dirty_background_ratio_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)495 static int dirty_background_ratio_handler(struct ctl_table *table, int write,
496 		void *buffer, size_t *lenp, loff_t *ppos)
497 {
498 	int ret;
499 
500 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
501 	if (ret == 0 && write)
502 		dirty_background_bytes = 0;
503 	return ret;
504 }
505 
dirty_background_bytes_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)506 static int dirty_background_bytes_handler(struct ctl_table *table, int write,
507 		void *buffer, size_t *lenp, loff_t *ppos)
508 {
509 	int ret;
510 
511 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
512 	if (ret == 0 && write)
513 		dirty_background_ratio = 0;
514 	return ret;
515 }
516 
dirty_ratio_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)517 static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
518 		size_t *lenp, loff_t *ppos)
519 {
520 	int old_ratio = vm_dirty_ratio;
521 	int ret;
522 
523 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524 	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
525 		writeback_set_ratelimit();
526 		vm_dirty_bytes = 0;
527 	}
528 	return ret;
529 }
530 
dirty_bytes_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)531 static int dirty_bytes_handler(struct ctl_table *table, int write,
532 		void *buffer, size_t *lenp, loff_t *ppos)
533 {
534 	unsigned long old_bytes = vm_dirty_bytes;
535 	int ret;
536 
537 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
538 	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
539 		writeback_set_ratelimit();
540 		vm_dirty_ratio = 0;
541 	}
542 	return ret;
543 }
544 #endif
545 
wp_next_time(unsigned long cur_time)546 static unsigned long wp_next_time(unsigned long cur_time)
547 {
548 	cur_time += VM_COMPLETIONS_PERIOD_LEN;
549 	/* 0 has a special meaning... */
550 	if (!cur_time)
551 		return 1;
552 	return cur_time;
553 }
554 
wb_domain_writeout_add(struct wb_domain * dom,struct fprop_local_percpu * completions,unsigned int max_prop_frac,long nr)555 static void wb_domain_writeout_add(struct wb_domain *dom,
556 				   struct fprop_local_percpu *completions,
557 				   unsigned int max_prop_frac, long nr)
558 {
559 	__fprop_add_percpu_max(&dom->completions, completions,
560 			       max_prop_frac, nr);
561 	/* First event after period switching was turned off? */
562 	if (unlikely(!dom->period_time)) {
563 		/*
564 		 * We can race with other __bdi_writeout_inc calls here but
565 		 * it does not cause any harm since the resulting time when
566 		 * timer will fire and what is in writeout_period_time will be
567 		 * roughly the same.
568 		 */
569 		dom->period_time = wp_next_time(jiffies);
570 		mod_timer(&dom->period_timer, dom->period_time);
571 	}
572 }
573 
574 /*
575  * Increment @wb's writeout completion count and the global writeout
576  * completion count. Called from __folio_end_writeback().
577  */
__wb_writeout_add(struct bdi_writeback * wb,long nr)578 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
579 {
580 	struct wb_domain *cgdom;
581 
582 	wb_stat_mod(wb, WB_WRITTEN, nr);
583 	wb_domain_writeout_add(&global_wb_domain, &wb->completions,
584 			       wb->bdi->max_prop_frac, nr);
585 
586 	cgdom = mem_cgroup_wb_domain(wb);
587 	if (cgdom)
588 		wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
589 				       wb->bdi->max_prop_frac, nr);
590 }
591 
wb_writeout_inc(struct bdi_writeback * wb)592 void wb_writeout_inc(struct bdi_writeback *wb)
593 {
594 	unsigned long flags;
595 
596 	local_irq_save(flags);
597 	__wb_writeout_add(wb, 1);
598 	local_irq_restore(flags);
599 }
600 EXPORT_SYMBOL_GPL(wb_writeout_inc);
601 
602 /*
603  * On idle system, we can be called long after we scheduled because we use
604  * deferred timers so count with missed periods.
605  */
writeout_period(struct timer_list * t)606 static void writeout_period(struct timer_list *t)
607 {
608 	struct wb_domain *dom = from_timer(dom, t, period_timer);
609 	int miss_periods = (jiffies - dom->period_time) /
610 						 VM_COMPLETIONS_PERIOD_LEN;
611 
612 	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
613 		dom->period_time = wp_next_time(dom->period_time +
614 				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
615 		mod_timer(&dom->period_timer, dom->period_time);
616 	} else {
617 		/*
618 		 * Aging has zeroed all fractions. Stop wasting CPU on period
619 		 * updates.
620 		 */
621 		dom->period_time = 0;
622 	}
623 }
624 
wb_domain_init(struct wb_domain * dom,gfp_t gfp)625 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
626 {
627 	memset(dom, 0, sizeof(*dom));
628 
629 	spin_lock_init(&dom->lock);
630 
631 	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
632 
633 	dom->dirty_limit_tstamp = jiffies;
634 
635 	return fprop_global_init(&dom->completions, gfp);
636 }
637 
638 #ifdef CONFIG_CGROUP_WRITEBACK
wb_domain_exit(struct wb_domain * dom)639 void wb_domain_exit(struct wb_domain *dom)
640 {
641 	del_timer_sync(&dom->period_timer);
642 	fprop_global_destroy(&dom->completions);
643 }
644 #endif
645 
646 /*
647  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
648  * registered backing devices, which, for obvious reasons, can not
649  * exceed 100%.
650  */
651 static unsigned int bdi_min_ratio;
652 
bdi_set_min_ratio(struct backing_dev_info * bdi,unsigned int min_ratio)653 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
654 {
655 	unsigned int delta;
656 	int ret = 0;
657 
658 	spin_lock_bh(&bdi_lock);
659 	if (min_ratio > bdi->max_ratio) {
660 		ret = -EINVAL;
661 	} else {
662 		if (min_ratio < bdi->min_ratio) {
663 			delta = bdi->min_ratio - min_ratio;
664 			bdi_min_ratio -= delta;
665 			bdi->min_ratio = min_ratio;
666 		} else {
667 			delta = min_ratio - bdi->min_ratio;
668 			if (bdi_min_ratio + delta < 100) {
669 				bdi_min_ratio += delta;
670 				bdi->min_ratio = min_ratio;
671 			} else {
672 				ret = -EINVAL;
673 			}
674 		}
675 	}
676 	spin_unlock_bh(&bdi_lock);
677 
678 	return ret;
679 }
680 
bdi_set_max_ratio(struct backing_dev_info * bdi,unsigned max_ratio)681 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
682 {
683 	int ret = 0;
684 
685 	if (max_ratio > 100)
686 		return -EINVAL;
687 
688 	spin_lock_bh(&bdi_lock);
689 	if (bdi->min_ratio > max_ratio) {
690 		ret = -EINVAL;
691 	} else {
692 		bdi->max_ratio = max_ratio;
693 		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
694 	}
695 	spin_unlock_bh(&bdi_lock);
696 
697 	return ret;
698 }
699 EXPORT_SYMBOL(bdi_set_max_ratio);
700 
dirty_freerun_ceiling(unsigned long thresh,unsigned long bg_thresh)701 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
702 					   unsigned long bg_thresh)
703 {
704 	return (thresh + bg_thresh) / 2;
705 }
706 
hard_dirty_limit(struct wb_domain * dom,unsigned long thresh)707 static unsigned long hard_dirty_limit(struct wb_domain *dom,
708 				      unsigned long thresh)
709 {
710 	return max(thresh, dom->dirty_limit);
711 }
712 
713 /*
714  * Memory which can be further allocated to a memcg domain is capped by
715  * system-wide clean memory excluding the amount being used in the domain.
716  */
mdtc_calc_avail(struct dirty_throttle_control * mdtc,unsigned long filepages,unsigned long headroom)717 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
718 			    unsigned long filepages, unsigned long headroom)
719 {
720 	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
721 	unsigned long clean = filepages - min(filepages, mdtc->dirty);
722 	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
723 	unsigned long other_clean = global_clean - min(global_clean, clean);
724 
725 	mdtc->avail = filepages + min(headroom, other_clean);
726 }
727 
728 /**
729  * __wb_calc_thresh - @wb's share of dirty throttling threshold
730  * @dtc: dirty_throttle_context of interest
731  *
732  * Note that balance_dirty_pages() will only seriously take it as a hard limit
733  * when sleeping max_pause per page is not enough to keep the dirty pages under
734  * control. For example, when the device is completely stalled due to some error
735  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
736  * In the other normal situations, it acts more gently by throttling the tasks
737  * more (rather than completely block them) when the wb dirty pages go high.
738  *
739  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
740  * - starving fast devices
741  * - piling up dirty pages (that will take long time to sync) on slow devices
742  *
743  * The wb's share of dirty limit will be adapting to its throughput and
744  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
745  *
746  * Return: @wb's dirty limit in pages. The term "dirty" in the context of
747  * dirty balancing includes all PG_dirty and PG_writeback pages.
748  */
__wb_calc_thresh(struct dirty_throttle_control * dtc)749 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
750 {
751 	struct wb_domain *dom = dtc_dom(dtc);
752 	unsigned long thresh = dtc->thresh;
753 	u64 wb_thresh;
754 	unsigned long numerator, denominator;
755 	unsigned long wb_min_ratio, wb_max_ratio;
756 
757 	/*
758 	 * Calculate this BDI's share of the thresh ratio.
759 	 */
760 	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
761 			      &numerator, &denominator);
762 
763 	wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
764 	wb_thresh *= numerator;
765 	wb_thresh = div64_ul(wb_thresh, denominator);
766 
767 	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
768 
769 	wb_thresh += (thresh * wb_min_ratio) / 100;
770 	if (wb_thresh > (thresh * wb_max_ratio) / 100)
771 		wb_thresh = thresh * wb_max_ratio / 100;
772 
773 	return wb_thresh;
774 }
775 
wb_calc_thresh(struct bdi_writeback * wb,unsigned long thresh)776 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
777 {
778 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
779 					       .thresh = thresh };
780 	return __wb_calc_thresh(&gdtc);
781 }
782 
783 /*
784  *                           setpoint - dirty 3
785  *        f(dirty) := 1.0 + (----------------)
786  *                           limit - setpoint
787  *
788  * it's a 3rd order polynomial that subjects to
789  *
790  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
791  * (2) f(setpoint) = 1.0 => the balance point
792  * (3) f(limit)    = 0   => the hard limit
793  * (4) df/dx      <= 0	 => negative feedback control
794  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
795  *     => fast response on large errors; small oscillation near setpoint
796  */
pos_ratio_polynom(unsigned long setpoint,unsigned long dirty,unsigned long limit)797 static long long pos_ratio_polynom(unsigned long setpoint,
798 					  unsigned long dirty,
799 					  unsigned long limit)
800 {
801 	long long pos_ratio;
802 	long x;
803 
804 	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
805 		      (limit - setpoint) | 1);
806 	pos_ratio = x;
807 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
808 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
809 	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
810 
811 	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
812 }
813 
814 /*
815  * Dirty position control.
816  *
817  * (o) global/bdi setpoints
818  *
819  * We want the dirty pages be balanced around the global/wb setpoints.
820  * When the number of dirty pages is higher/lower than the setpoint, the
821  * dirty position control ratio (and hence task dirty ratelimit) will be
822  * decreased/increased to bring the dirty pages back to the setpoint.
823  *
824  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
825  *
826  *     if (dirty < setpoint) scale up   pos_ratio
827  *     if (dirty > setpoint) scale down pos_ratio
828  *
829  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
830  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
831  *
832  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
833  *
834  * (o) global control line
835  *
836  *     ^ pos_ratio
837  *     |
838  *     |            |<===== global dirty control scope ======>|
839  * 2.0  * * * * * * *
840  *     |            .*
841  *     |            . *
842  *     |            .   *
843  *     |            .     *
844  *     |            .        *
845  *     |            .            *
846  * 1.0 ................................*
847  *     |            .                  .     *
848  *     |            .                  .          *
849  *     |            .                  .              *
850  *     |            .                  .                 *
851  *     |            .                  .                    *
852  *   0 +------------.------------------.----------------------*------------->
853  *           freerun^          setpoint^                 limit^   dirty pages
854  *
855  * (o) wb control line
856  *
857  *     ^ pos_ratio
858  *     |
859  *     |            *
860  *     |              *
861  *     |                *
862  *     |                  *
863  *     |                    * |<=========== span ============>|
864  * 1.0 .......................*
865  *     |                      . *
866  *     |                      .   *
867  *     |                      .     *
868  *     |                      .       *
869  *     |                      .         *
870  *     |                      .           *
871  *     |                      .             *
872  *     |                      .               *
873  *     |                      .                 *
874  *     |                      .                   *
875  *     |                      .                     *
876  * 1/4 ...............................................* * * * * * * * * * * *
877  *     |                      .                         .
878  *     |                      .                           .
879  *     |                      .                             .
880  *   0 +----------------------.-------------------------------.------------->
881  *                wb_setpoint^                    x_intercept^
882  *
883  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
884  * be smoothly throttled down to normal if it starts high in situations like
885  * - start writing to a slow SD card and a fast disk at the same time. The SD
886  *   card's wb_dirty may rush to many times higher than wb_setpoint.
887  * - the wb dirty thresh drops quickly due to change of JBOD workload
888  */
wb_position_ratio(struct dirty_throttle_control * dtc)889 static void wb_position_ratio(struct dirty_throttle_control *dtc)
890 {
891 	struct bdi_writeback *wb = dtc->wb;
892 	unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
893 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
894 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
895 	unsigned long wb_thresh = dtc->wb_thresh;
896 	unsigned long x_intercept;
897 	unsigned long setpoint;		/* dirty pages' target balance point */
898 	unsigned long wb_setpoint;
899 	unsigned long span;
900 	long long pos_ratio;		/* for scaling up/down the rate limit */
901 	long x;
902 
903 	dtc->pos_ratio = 0;
904 
905 	if (unlikely(dtc->dirty >= limit))
906 		return;
907 
908 	/*
909 	 * global setpoint
910 	 *
911 	 * See comment for pos_ratio_polynom().
912 	 */
913 	setpoint = (freerun + limit) / 2;
914 	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
915 
916 	/*
917 	 * The strictlimit feature is a tool preventing mistrusted filesystems
918 	 * from growing a large number of dirty pages before throttling. For
919 	 * such filesystems balance_dirty_pages always checks wb counters
920 	 * against wb limits. Even if global "nr_dirty" is under "freerun".
921 	 * This is especially important for fuse which sets bdi->max_ratio to
922 	 * 1% by default. Without strictlimit feature, fuse writeback may
923 	 * consume arbitrary amount of RAM because it is accounted in
924 	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
925 	 *
926 	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
927 	 * two values: wb_dirty and wb_thresh. Let's consider an example:
928 	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
929 	 * limits are set by default to 10% and 20% (background and throttle).
930 	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
931 	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
932 	 * about ~6K pages (as the average of background and throttle wb
933 	 * limits). The 3rd order polynomial will provide positive feedback if
934 	 * wb_dirty is under wb_setpoint and vice versa.
935 	 *
936 	 * Note, that we cannot use global counters in these calculations
937 	 * because we want to throttle process writing to a strictlimit wb
938 	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
939 	 * in the example above).
940 	 */
941 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
942 		long long wb_pos_ratio;
943 
944 		if (dtc->wb_dirty < 8) {
945 			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
946 					   2 << RATELIMIT_CALC_SHIFT);
947 			return;
948 		}
949 
950 		if (dtc->wb_dirty >= wb_thresh)
951 			return;
952 
953 		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
954 						    dtc->wb_bg_thresh);
955 
956 		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
957 			return;
958 
959 		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
960 						 wb_thresh);
961 
962 		/*
963 		 * Typically, for strictlimit case, wb_setpoint << setpoint
964 		 * and pos_ratio >> wb_pos_ratio. In the other words global
965 		 * state ("dirty") is not limiting factor and we have to
966 		 * make decision based on wb counters. But there is an
967 		 * important case when global pos_ratio should get precedence:
968 		 * global limits are exceeded (e.g. due to activities on other
969 		 * wb's) while given strictlimit wb is below limit.
970 		 *
971 		 * "pos_ratio * wb_pos_ratio" would work for the case above,
972 		 * but it would look too non-natural for the case of all
973 		 * activity in the system coming from a single strictlimit wb
974 		 * with bdi->max_ratio == 100%.
975 		 *
976 		 * Note that min() below somewhat changes the dynamics of the
977 		 * control system. Normally, pos_ratio value can be well over 3
978 		 * (when globally we are at freerun and wb is well below wb
979 		 * setpoint). Now the maximum pos_ratio in the same situation
980 		 * is 2. We might want to tweak this if we observe the control
981 		 * system is too slow to adapt.
982 		 */
983 		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
984 		return;
985 	}
986 
987 	/*
988 	 * We have computed basic pos_ratio above based on global situation. If
989 	 * the wb is over/under its share of dirty pages, we want to scale
990 	 * pos_ratio further down/up. That is done by the following mechanism.
991 	 */
992 
993 	/*
994 	 * wb setpoint
995 	 *
996 	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
997 	 *
998 	 *                        x_intercept - wb_dirty
999 	 *                     := --------------------------
1000 	 *                        x_intercept - wb_setpoint
1001 	 *
1002 	 * The main wb control line is a linear function that subjects to
1003 	 *
1004 	 * (1) f(wb_setpoint) = 1.0
1005 	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1006 	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1007 	 *
1008 	 * For single wb case, the dirty pages are observed to fluctuate
1009 	 * regularly within range
1010 	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1011 	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1012 	 * fluctuation range for pos_ratio.
1013 	 *
1014 	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1015 	 * own size, so move the slope over accordingly and choose a slope that
1016 	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1017 	 */
1018 	if (unlikely(wb_thresh > dtc->thresh))
1019 		wb_thresh = dtc->thresh;
1020 	/*
1021 	 * It's very possible that wb_thresh is close to 0 not because the
1022 	 * device is slow, but that it has remained inactive for long time.
1023 	 * Honour such devices a reasonable good (hopefully IO efficient)
1024 	 * threshold, so that the occasional writes won't be blocked and active
1025 	 * writes can rampup the threshold quickly.
1026 	 */
1027 	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1028 	/*
1029 	 * scale global setpoint to wb's:
1030 	 *	wb_setpoint = setpoint * wb_thresh / thresh
1031 	 */
1032 	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1033 	wb_setpoint = setpoint * (u64)x >> 16;
1034 	/*
1035 	 * Use span=(8*write_bw) in single wb case as indicated by
1036 	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1037 	 *
1038 	 *        wb_thresh                    thresh - wb_thresh
1039 	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1040 	 *         thresh                           thresh
1041 	 */
1042 	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1043 	x_intercept = wb_setpoint + span;
1044 
1045 	if (dtc->wb_dirty < x_intercept - span / 4) {
1046 		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1047 				      (x_intercept - wb_setpoint) | 1);
1048 	} else
1049 		pos_ratio /= 4;
1050 
1051 	/*
1052 	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1053 	 * It may push the desired control point of global dirty pages higher
1054 	 * than setpoint.
1055 	 */
1056 	x_intercept = wb_thresh / 2;
1057 	if (dtc->wb_dirty < x_intercept) {
1058 		if (dtc->wb_dirty > x_intercept / 8)
1059 			pos_ratio = div_u64(pos_ratio * x_intercept,
1060 					    dtc->wb_dirty);
1061 		else
1062 			pos_ratio *= 8;
1063 	}
1064 
1065 	dtc->pos_ratio = pos_ratio;
1066 }
1067 
wb_update_write_bandwidth(struct bdi_writeback * wb,unsigned long elapsed,unsigned long written)1068 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1069 				      unsigned long elapsed,
1070 				      unsigned long written)
1071 {
1072 	const unsigned long period = roundup_pow_of_two(3 * HZ);
1073 	unsigned long avg = wb->avg_write_bandwidth;
1074 	unsigned long old = wb->write_bandwidth;
1075 	u64 bw;
1076 
1077 	/*
1078 	 * bw = written * HZ / elapsed
1079 	 *
1080 	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1081 	 * write_bandwidth = ---------------------------------------------------
1082 	 *                                          period
1083 	 *
1084 	 * @written may have decreased due to folio_account_redirty().
1085 	 * Avoid underflowing @bw calculation.
1086 	 */
1087 	bw = written - min(written, wb->written_stamp);
1088 	bw *= HZ;
1089 	if (unlikely(elapsed > period)) {
1090 		bw = div64_ul(bw, elapsed);
1091 		avg = bw;
1092 		goto out;
1093 	}
1094 	bw += (u64)wb->write_bandwidth * (period - elapsed);
1095 	bw >>= ilog2(period);
1096 
1097 	/*
1098 	 * one more level of smoothing, for filtering out sudden spikes
1099 	 */
1100 	if (avg > old && old >= (unsigned long)bw)
1101 		avg -= (avg - old) >> 3;
1102 
1103 	if (avg < old && old <= (unsigned long)bw)
1104 		avg += (old - avg) >> 3;
1105 
1106 out:
1107 	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1108 	avg = max(avg, 1LU);
1109 	if (wb_has_dirty_io(wb)) {
1110 		long delta = avg - wb->avg_write_bandwidth;
1111 		WARN_ON_ONCE(atomic_long_add_return(delta,
1112 					&wb->bdi->tot_write_bandwidth) <= 0);
1113 	}
1114 	wb->write_bandwidth = bw;
1115 	WRITE_ONCE(wb->avg_write_bandwidth, avg);
1116 }
1117 
update_dirty_limit(struct dirty_throttle_control * dtc)1118 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1119 {
1120 	struct wb_domain *dom = dtc_dom(dtc);
1121 	unsigned long thresh = dtc->thresh;
1122 	unsigned long limit = dom->dirty_limit;
1123 
1124 	/*
1125 	 * Follow up in one step.
1126 	 */
1127 	if (limit < thresh) {
1128 		limit = thresh;
1129 		goto update;
1130 	}
1131 
1132 	/*
1133 	 * Follow down slowly. Use the higher one as the target, because thresh
1134 	 * may drop below dirty. This is exactly the reason to introduce
1135 	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1136 	 */
1137 	thresh = max(thresh, dtc->dirty);
1138 	if (limit > thresh) {
1139 		limit -= (limit - thresh) >> 5;
1140 		goto update;
1141 	}
1142 	return;
1143 update:
1144 	dom->dirty_limit = limit;
1145 }
1146 
domain_update_dirty_limit(struct dirty_throttle_control * dtc,unsigned long now)1147 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1148 				      unsigned long now)
1149 {
1150 	struct wb_domain *dom = dtc_dom(dtc);
1151 
1152 	/*
1153 	 * check locklessly first to optimize away locking for the most time
1154 	 */
1155 	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1156 		return;
1157 
1158 	spin_lock(&dom->lock);
1159 	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1160 		update_dirty_limit(dtc);
1161 		dom->dirty_limit_tstamp = now;
1162 	}
1163 	spin_unlock(&dom->lock);
1164 }
1165 
1166 /*
1167  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1168  *
1169  * Normal wb tasks will be curbed at or below it in long term.
1170  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1171  */
wb_update_dirty_ratelimit(struct dirty_throttle_control * dtc,unsigned long dirtied,unsigned long elapsed)1172 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1173 				      unsigned long dirtied,
1174 				      unsigned long elapsed)
1175 {
1176 	struct bdi_writeback *wb = dtc->wb;
1177 	unsigned long dirty = dtc->dirty;
1178 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1179 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1180 	unsigned long setpoint = (freerun + limit) / 2;
1181 	unsigned long write_bw = wb->avg_write_bandwidth;
1182 	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1183 	unsigned long dirty_rate;
1184 	unsigned long task_ratelimit;
1185 	unsigned long balanced_dirty_ratelimit;
1186 	unsigned long step;
1187 	unsigned long x;
1188 	unsigned long shift;
1189 
1190 	/*
1191 	 * The dirty rate will match the writeout rate in long term, except
1192 	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1193 	 */
1194 	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1195 
1196 	/*
1197 	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1198 	 */
1199 	task_ratelimit = (u64)dirty_ratelimit *
1200 					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1201 	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1202 
1203 	/*
1204 	 * A linear estimation of the "balanced" throttle rate. The theory is,
1205 	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1206 	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1207 	 * formula will yield the balanced rate limit (write_bw / N).
1208 	 *
1209 	 * Note that the expanded form is not a pure rate feedback:
1210 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1211 	 * but also takes pos_ratio into account:
1212 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1213 	 *
1214 	 * (1) is not realistic because pos_ratio also takes part in balancing
1215 	 * the dirty rate.  Consider the state
1216 	 *	pos_ratio = 0.5						     (3)
1217 	 *	rate = 2 * (write_bw / N)				     (4)
1218 	 * If (1) is used, it will stuck in that state! Because each dd will
1219 	 * be throttled at
1220 	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1221 	 * yielding
1222 	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1223 	 * put (6) into (1) we get
1224 	 *	rate_(i+1) = rate_(i)					     (7)
1225 	 *
1226 	 * So we end up using (2) to always keep
1227 	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1228 	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1229 	 * pos_ratio is able to drive itself to 1.0, which is not only where
1230 	 * the dirty count meet the setpoint, but also where the slope of
1231 	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1232 	 */
1233 	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1234 					   dirty_rate | 1);
1235 	/*
1236 	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1237 	 */
1238 	if (unlikely(balanced_dirty_ratelimit > write_bw))
1239 		balanced_dirty_ratelimit = write_bw;
1240 
1241 	/*
1242 	 * We could safely do this and return immediately:
1243 	 *
1244 	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1245 	 *
1246 	 * However to get a more stable dirty_ratelimit, the below elaborated
1247 	 * code makes use of task_ratelimit to filter out singular points and
1248 	 * limit the step size.
1249 	 *
1250 	 * The below code essentially only uses the relative value of
1251 	 *
1252 	 *	task_ratelimit - dirty_ratelimit
1253 	 *	= (pos_ratio - 1) * dirty_ratelimit
1254 	 *
1255 	 * which reflects the direction and size of dirty position error.
1256 	 */
1257 
1258 	/*
1259 	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1260 	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1261 	 * For example, when
1262 	 * - dirty_ratelimit > balanced_dirty_ratelimit
1263 	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1264 	 * lowering dirty_ratelimit will help meet both the position and rate
1265 	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1266 	 * only help meet the rate target. After all, what the users ultimately
1267 	 * feel and care are stable dirty rate and small position error.
1268 	 *
1269 	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1270 	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1271 	 * keeps jumping around randomly and can even leap far away at times
1272 	 * due to the small 200ms estimation period of dirty_rate (we want to
1273 	 * keep that period small to reduce time lags).
1274 	 */
1275 	step = 0;
1276 
1277 	/*
1278 	 * For strictlimit case, calculations above were based on wb counters
1279 	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1280 	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1281 	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1282 	 * "dirty" and wb_setpoint as "setpoint".
1283 	 *
1284 	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1285 	 * it's possible that wb_thresh is close to zero due to inactivity
1286 	 * of backing device.
1287 	 */
1288 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1289 		dirty = dtc->wb_dirty;
1290 		if (dtc->wb_dirty < 8)
1291 			setpoint = dtc->wb_dirty + 1;
1292 		else
1293 			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1294 	}
1295 
1296 	if (dirty < setpoint) {
1297 		x = min3(wb->balanced_dirty_ratelimit,
1298 			 balanced_dirty_ratelimit, task_ratelimit);
1299 		if (dirty_ratelimit < x)
1300 			step = x - dirty_ratelimit;
1301 	} else {
1302 		x = max3(wb->balanced_dirty_ratelimit,
1303 			 balanced_dirty_ratelimit, task_ratelimit);
1304 		if (dirty_ratelimit > x)
1305 			step = dirty_ratelimit - x;
1306 	}
1307 
1308 	/*
1309 	 * Don't pursue 100% rate matching. It's impossible since the balanced
1310 	 * rate itself is constantly fluctuating. So decrease the track speed
1311 	 * when it gets close to the target. Helps eliminate pointless tremors.
1312 	 */
1313 	shift = dirty_ratelimit / (2 * step + 1);
1314 	if (shift < BITS_PER_LONG)
1315 		step = DIV_ROUND_UP(step >> shift, 8);
1316 	else
1317 		step = 0;
1318 
1319 	if (dirty_ratelimit < balanced_dirty_ratelimit)
1320 		dirty_ratelimit += step;
1321 	else
1322 		dirty_ratelimit -= step;
1323 
1324 	WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1325 	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1326 
1327 	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1328 }
1329 
__wb_update_bandwidth(struct dirty_throttle_control * gdtc,struct dirty_throttle_control * mdtc,bool update_ratelimit)1330 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1331 				  struct dirty_throttle_control *mdtc,
1332 				  bool update_ratelimit)
1333 {
1334 	struct bdi_writeback *wb = gdtc->wb;
1335 	unsigned long now = jiffies;
1336 	unsigned long elapsed;
1337 	unsigned long dirtied;
1338 	unsigned long written;
1339 
1340 	spin_lock(&wb->list_lock);
1341 
1342 	/*
1343 	 * Lockless checks for elapsed time are racy and delayed update after
1344 	 * IO completion doesn't do it at all (to make sure written pages are
1345 	 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1346 	 * division errors.
1347 	 */
1348 	elapsed = max(now - wb->bw_time_stamp, 1UL);
1349 	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1350 	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1351 
1352 	if (update_ratelimit) {
1353 		domain_update_dirty_limit(gdtc, now);
1354 		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1355 
1356 		/*
1357 		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1358 		 * compiler has no way to figure that out.  Help it.
1359 		 */
1360 		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1361 			domain_update_dirty_limit(mdtc, now);
1362 			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1363 		}
1364 	}
1365 	wb_update_write_bandwidth(wb, elapsed, written);
1366 
1367 	wb->dirtied_stamp = dirtied;
1368 	wb->written_stamp = written;
1369 	WRITE_ONCE(wb->bw_time_stamp, now);
1370 	spin_unlock(&wb->list_lock);
1371 }
1372 
wb_update_bandwidth(struct bdi_writeback * wb)1373 void wb_update_bandwidth(struct bdi_writeback *wb)
1374 {
1375 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1376 
1377 	__wb_update_bandwidth(&gdtc, NULL, false);
1378 }
1379 
1380 /* Interval after which we consider wb idle and don't estimate bandwidth */
1381 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1382 
wb_bandwidth_estimate_start(struct bdi_writeback * wb)1383 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1384 {
1385 	unsigned long now = jiffies;
1386 	unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1387 
1388 	if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1389 	    !atomic_read(&wb->writeback_inodes)) {
1390 		spin_lock(&wb->list_lock);
1391 		wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1392 		wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1393 		WRITE_ONCE(wb->bw_time_stamp, now);
1394 		spin_unlock(&wb->list_lock);
1395 	}
1396 }
1397 
1398 /*
1399  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1400  * will look to see if it needs to start dirty throttling.
1401  *
1402  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1403  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1404  * (the number of pages we may dirty without exceeding the dirty limits).
1405  */
dirty_poll_interval(unsigned long dirty,unsigned long thresh)1406 static unsigned long dirty_poll_interval(unsigned long dirty,
1407 					 unsigned long thresh)
1408 {
1409 	if (thresh > dirty)
1410 		return 1UL << (ilog2(thresh - dirty) >> 1);
1411 
1412 	return 1;
1413 }
1414 
wb_max_pause(struct bdi_writeback * wb,unsigned long wb_dirty)1415 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1416 				  unsigned long wb_dirty)
1417 {
1418 	unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1419 	unsigned long t;
1420 
1421 	/*
1422 	 * Limit pause time for small memory systems. If sleeping for too long
1423 	 * time, a small pool of dirty/writeback pages may go empty and disk go
1424 	 * idle.
1425 	 *
1426 	 * 8 serves as the safety ratio.
1427 	 */
1428 	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1429 	t++;
1430 
1431 	return min_t(unsigned long, t, MAX_PAUSE);
1432 }
1433 
wb_min_pause(struct bdi_writeback * wb,long max_pause,unsigned long task_ratelimit,unsigned long dirty_ratelimit,int * nr_dirtied_pause)1434 static long wb_min_pause(struct bdi_writeback *wb,
1435 			 long max_pause,
1436 			 unsigned long task_ratelimit,
1437 			 unsigned long dirty_ratelimit,
1438 			 int *nr_dirtied_pause)
1439 {
1440 	long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1441 	long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1442 	long t;		/* target pause */
1443 	long pause;	/* estimated next pause */
1444 	int pages;	/* target nr_dirtied_pause */
1445 
1446 	/* target for 10ms pause on 1-dd case */
1447 	t = max(1, HZ / 100);
1448 
1449 	/*
1450 	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1451 	 * overheads.
1452 	 *
1453 	 * (N * 10ms) on 2^N concurrent tasks.
1454 	 */
1455 	if (hi > lo)
1456 		t += (hi - lo) * (10 * HZ) / 1024;
1457 
1458 	/*
1459 	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1460 	 * on the much more stable dirty_ratelimit. However the next pause time
1461 	 * will be computed based on task_ratelimit and the two rate limits may
1462 	 * depart considerably at some time. Especially if task_ratelimit goes
1463 	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1464 	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1465 	 * result task_ratelimit won't be executed faithfully, which could
1466 	 * eventually bring down dirty_ratelimit.
1467 	 *
1468 	 * We apply two rules to fix it up:
1469 	 * 1) try to estimate the next pause time and if necessary, use a lower
1470 	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1471 	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1472 	 * 2) limit the target pause time to max_pause/2, so that the normal
1473 	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1474 	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1475 	 */
1476 	t = min(t, 1 + max_pause / 2);
1477 	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1478 
1479 	/*
1480 	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1481 	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1482 	 * When the 16 consecutive reads are often interrupted by some dirty
1483 	 * throttling pause during the async writes, cfq will go into idles
1484 	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1485 	 * until reaches DIRTY_POLL_THRESH=32 pages.
1486 	 */
1487 	if (pages < DIRTY_POLL_THRESH) {
1488 		t = max_pause;
1489 		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1490 		if (pages > DIRTY_POLL_THRESH) {
1491 			pages = DIRTY_POLL_THRESH;
1492 			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1493 		}
1494 	}
1495 
1496 	pause = HZ * pages / (task_ratelimit + 1);
1497 	if (pause > max_pause) {
1498 		t = max_pause;
1499 		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1500 	}
1501 
1502 	*nr_dirtied_pause = pages;
1503 	/*
1504 	 * The minimal pause time will normally be half the target pause time.
1505 	 */
1506 	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1507 }
1508 
wb_dirty_limits(struct dirty_throttle_control * dtc)1509 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1510 {
1511 	struct bdi_writeback *wb = dtc->wb;
1512 	unsigned long wb_reclaimable;
1513 
1514 	/*
1515 	 * wb_thresh is not treated as some limiting factor as
1516 	 * dirty_thresh, due to reasons
1517 	 * - in JBOD setup, wb_thresh can fluctuate a lot
1518 	 * - in a system with HDD and USB key, the USB key may somehow
1519 	 *   go into state (wb_dirty >> wb_thresh) either because
1520 	 *   wb_dirty starts high, or because wb_thresh drops low.
1521 	 *   In this case we don't want to hard throttle the USB key
1522 	 *   dirtiers for 100 seconds until wb_dirty drops under
1523 	 *   wb_thresh. Instead the auxiliary wb control line in
1524 	 *   wb_position_ratio() will let the dirtier task progress
1525 	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1526 	 */
1527 	dtc->wb_thresh = __wb_calc_thresh(dtc);
1528 	dtc->wb_bg_thresh = dtc->thresh ?
1529 		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1530 
1531 	/*
1532 	 * In order to avoid the stacked BDI deadlock we need
1533 	 * to ensure we accurately count the 'dirty' pages when
1534 	 * the threshold is low.
1535 	 *
1536 	 * Otherwise it would be possible to get thresh+n pages
1537 	 * reported dirty, even though there are thresh-m pages
1538 	 * actually dirty; with m+n sitting in the percpu
1539 	 * deltas.
1540 	 */
1541 	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1542 		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1543 		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1544 	} else {
1545 		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1546 		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1547 	}
1548 }
1549 
1550 /*
1551  * balance_dirty_pages() must be called by processes which are generating dirty
1552  * data.  It looks at the number of dirty pages in the machine and will force
1553  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1554  * If we're over `background_thresh' then the writeback threads are woken to
1555  * perform some writeout.
1556  */
balance_dirty_pages(struct bdi_writeback * wb,unsigned long pages_dirtied,unsigned int flags)1557 static int balance_dirty_pages(struct bdi_writeback *wb,
1558 			       unsigned long pages_dirtied, unsigned int flags)
1559 {
1560 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1561 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1562 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1563 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1564 						     &mdtc_stor : NULL;
1565 	struct dirty_throttle_control *sdtc;
1566 	unsigned long nr_reclaimable;	/* = file_dirty */
1567 	long period;
1568 	long pause;
1569 	long max_pause;
1570 	long min_pause;
1571 	int nr_dirtied_pause;
1572 	bool dirty_exceeded = false;
1573 	unsigned long task_ratelimit;
1574 	unsigned long dirty_ratelimit;
1575 	struct backing_dev_info *bdi = wb->bdi;
1576 	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1577 	unsigned long start_time = jiffies;
1578 	int ret = 0;
1579 
1580 	for (;;) {
1581 		unsigned long now = jiffies;
1582 		unsigned long dirty, thresh, bg_thresh;
1583 		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1584 		unsigned long m_thresh = 0;
1585 		unsigned long m_bg_thresh = 0;
1586 
1587 		nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1588 		gdtc->avail = global_dirtyable_memory();
1589 		gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1590 
1591 		domain_dirty_limits(gdtc);
1592 
1593 		if (unlikely(strictlimit)) {
1594 			wb_dirty_limits(gdtc);
1595 
1596 			dirty = gdtc->wb_dirty;
1597 			thresh = gdtc->wb_thresh;
1598 			bg_thresh = gdtc->wb_bg_thresh;
1599 		} else {
1600 			dirty = gdtc->dirty;
1601 			thresh = gdtc->thresh;
1602 			bg_thresh = gdtc->bg_thresh;
1603 		}
1604 
1605 		if (mdtc) {
1606 			unsigned long filepages, headroom, writeback;
1607 
1608 			/*
1609 			 * If @wb belongs to !root memcg, repeat the same
1610 			 * basic calculations for the memcg domain.
1611 			 */
1612 			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1613 					    &mdtc->dirty, &writeback);
1614 			mdtc->dirty += writeback;
1615 			mdtc_calc_avail(mdtc, filepages, headroom);
1616 
1617 			domain_dirty_limits(mdtc);
1618 
1619 			if (unlikely(strictlimit)) {
1620 				wb_dirty_limits(mdtc);
1621 				m_dirty = mdtc->wb_dirty;
1622 				m_thresh = mdtc->wb_thresh;
1623 				m_bg_thresh = mdtc->wb_bg_thresh;
1624 			} else {
1625 				m_dirty = mdtc->dirty;
1626 				m_thresh = mdtc->thresh;
1627 				m_bg_thresh = mdtc->bg_thresh;
1628 			}
1629 		}
1630 
1631 		/*
1632 		 * In laptop mode, we wait until hitting the higher threshold
1633 		 * before starting background writeout, and then write out all
1634 		 * the way down to the lower threshold.  So slow writers cause
1635 		 * minimal disk activity.
1636 		 *
1637 		 * In normal mode, we start background writeout at the lower
1638 		 * background_thresh, to keep the amount of dirty memory low.
1639 		 */
1640 		if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh &&
1641 		    !writeback_in_progress(wb))
1642 			wb_start_background_writeback(wb);
1643 
1644 		/*
1645 		 * Throttle it only when the background writeback cannot
1646 		 * catch-up. This avoids (excessively) small writeouts
1647 		 * when the wb limits are ramping up in case of !strictlimit.
1648 		 *
1649 		 * In strictlimit case make decision based on the wb counters
1650 		 * and limits. Small writeouts when the wb limits are ramping
1651 		 * up are the price we consciously pay for strictlimit-ing.
1652 		 *
1653 		 * If memcg domain is in effect, @dirty should be under
1654 		 * both global and memcg freerun ceilings.
1655 		 */
1656 		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1657 		    (!mdtc ||
1658 		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1659 			unsigned long intv;
1660 			unsigned long m_intv;
1661 
1662 free_running:
1663 			intv = dirty_poll_interval(dirty, thresh);
1664 			m_intv = ULONG_MAX;
1665 
1666 			current->dirty_paused_when = now;
1667 			current->nr_dirtied = 0;
1668 			if (mdtc)
1669 				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1670 			current->nr_dirtied_pause = min(intv, m_intv);
1671 			break;
1672 		}
1673 
1674 		/* Start writeback even when in laptop mode */
1675 		if (unlikely(!writeback_in_progress(wb)))
1676 			wb_start_background_writeback(wb);
1677 
1678 		mem_cgroup_flush_foreign(wb);
1679 
1680 		/*
1681 		 * Calculate global domain's pos_ratio and select the
1682 		 * global dtc by default.
1683 		 */
1684 		if (!strictlimit) {
1685 			wb_dirty_limits(gdtc);
1686 
1687 			if ((current->flags & PF_LOCAL_THROTTLE) &&
1688 			    gdtc->wb_dirty <
1689 			    dirty_freerun_ceiling(gdtc->wb_thresh,
1690 						  gdtc->wb_bg_thresh))
1691 				/*
1692 				 * LOCAL_THROTTLE tasks must not be throttled
1693 				 * when below the per-wb freerun ceiling.
1694 				 */
1695 				goto free_running;
1696 		}
1697 
1698 		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1699 			((gdtc->dirty > gdtc->thresh) || strictlimit);
1700 
1701 		wb_position_ratio(gdtc);
1702 		sdtc = gdtc;
1703 
1704 		if (mdtc) {
1705 			/*
1706 			 * If memcg domain is in effect, calculate its
1707 			 * pos_ratio.  @wb should satisfy constraints from
1708 			 * both global and memcg domains.  Choose the one
1709 			 * w/ lower pos_ratio.
1710 			 */
1711 			if (!strictlimit) {
1712 				wb_dirty_limits(mdtc);
1713 
1714 				if ((current->flags & PF_LOCAL_THROTTLE) &&
1715 				    mdtc->wb_dirty <
1716 				    dirty_freerun_ceiling(mdtc->wb_thresh,
1717 							  mdtc->wb_bg_thresh))
1718 					/*
1719 					 * LOCAL_THROTTLE tasks must not be
1720 					 * throttled when below the per-wb
1721 					 * freerun ceiling.
1722 					 */
1723 					goto free_running;
1724 			}
1725 			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1726 				((mdtc->dirty > mdtc->thresh) || strictlimit);
1727 
1728 			wb_position_ratio(mdtc);
1729 			if (mdtc->pos_ratio < gdtc->pos_ratio)
1730 				sdtc = mdtc;
1731 		}
1732 
1733 		if (dirty_exceeded != wb->dirty_exceeded)
1734 			wb->dirty_exceeded = dirty_exceeded;
1735 
1736 		if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1737 					   BANDWIDTH_INTERVAL))
1738 			__wb_update_bandwidth(gdtc, mdtc, true);
1739 
1740 		/* throttle according to the chosen dtc */
1741 		dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1742 		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1743 							RATELIMIT_CALC_SHIFT;
1744 		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1745 		min_pause = wb_min_pause(wb, max_pause,
1746 					 task_ratelimit, dirty_ratelimit,
1747 					 &nr_dirtied_pause);
1748 
1749 		if (unlikely(task_ratelimit == 0)) {
1750 			period = max_pause;
1751 			pause = max_pause;
1752 			goto pause;
1753 		}
1754 		period = HZ * pages_dirtied / task_ratelimit;
1755 		pause = period;
1756 		if (current->dirty_paused_when)
1757 			pause -= now - current->dirty_paused_when;
1758 		/*
1759 		 * For less than 1s think time (ext3/4 may block the dirtier
1760 		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1761 		 * however at much less frequency), try to compensate it in
1762 		 * future periods by updating the virtual time; otherwise just
1763 		 * do a reset, as it may be a light dirtier.
1764 		 */
1765 		if (pause < min_pause) {
1766 			trace_balance_dirty_pages(wb,
1767 						  sdtc->thresh,
1768 						  sdtc->bg_thresh,
1769 						  sdtc->dirty,
1770 						  sdtc->wb_thresh,
1771 						  sdtc->wb_dirty,
1772 						  dirty_ratelimit,
1773 						  task_ratelimit,
1774 						  pages_dirtied,
1775 						  period,
1776 						  min(pause, 0L),
1777 						  start_time);
1778 			if (pause < -HZ) {
1779 				current->dirty_paused_when = now;
1780 				current->nr_dirtied = 0;
1781 			} else if (period) {
1782 				current->dirty_paused_when += period;
1783 				current->nr_dirtied = 0;
1784 			} else if (current->nr_dirtied_pause <= pages_dirtied)
1785 				current->nr_dirtied_pause += pages_dirtied;
1786 			break;
1787 		}
1788 		if (unlikely(pause > max_pause)) {
1789 			/* for occasional dropped task_ratelimit */
1790 			now += min(pause - max_pause, max_pause);
1791 			pause = max_pause;
1792 		}
1793 
1794 pause:
1795 		trace_balance_dirty_pages(wb,
1796 					  sdtc->thresh,
1797 					  sdtc->bg_thresh,
1798 					  sdtc->dirty,
1799 					  sdtc->wb_thresh,
1800 					  sdtc->wb_dirty,
1801 					  dirty_ratelimit,
1802 					  task_ratelimit,
1803 					  pages_dirtied,
1804 					  period,
1805 					  pause,
1806 					  start_time);
1807 		if (flags & BDP_ASYNC) {
1808 			ret = -EAGAIN;
1809 			break;
1810 		}
1811 		__set_current_state(TASK_KILLABLE);
1812 		wb->dirty_sleep = now;
1813 		io_schedule_timeout(pause);
1814 
1815 		current->dirty_paused_when = now + pause;
1816 		current->nr_dirtied = 0;
1817 		current->nr_dirtied_pause = nr_dirtied_pause;
1818 
1819 		/*
1820 		 * This is typically equal to (dirty < thresh) and can also
1821 		 * keep "1000+ dd on a slow USB stick" under control.
1822 		 */
1823 		if (task_ratelimit)
1824 			break;
1825 
1826 		/*
1827 		 * In the case of an unresponsive NFS server and the NFS dirty
1828 		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1829 		 * to go through, so that tasks on them still remain responsive.
1830 		 *
1831 		 * In theory 1 page is enough to keep the consumer-producer
1832 		 * pipe going: the flusher cleans 1 page => the task dirties 1
1833 		 * more page. However wb_dirty has accounting errors.  So use
1834 		 * the larger and more IO friendly wb_stat_error.
1835 		 */
1836 		if (sdtc->wb_dirty <= wb_stat_error())
1837 			break;
1838 
1839 		if (fatal_signal_pending(current))
1840 			break;
1841 	}
1842 	return ret;
1843 }
1844 
1845 static DEFINE_PER_CPU(int, bdp_ratelimits);
1846 
1847 /*
1848  * Normal tasks are throttled by
1849  *	loop {
1850  *		dirty tsk->nr_dirtied_pause pages;
1851  *		take a snap in balance_dirty_pages();
1852  *	}
1853  * However there is a worst case. If every task exit immediately when dirtied
1854  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1855  * called to throttle the page dirties. The solution is to save the not yet
1856  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1857  * randomly into the running tasks. This works well for the above worst case,
1858  * as the new task will pick up and accumulate the old task's leaked dirty
1859  * count and eventually get throttled.
1860  */
1861 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1862 
1863 /**
1864  * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1865  * @mapping: address_space which was dirtied.
1866  * @flags: BDP flags.
1867  *
1868  * Processes which are dirtying memory should call in here once for each page
1869  * which was newly dirtied.  The function will periodically check the system's
1870  * dirty state and will initiate writeback if needed.
1871  *
1872  * See balance_dirty_pages_ratelimited() for details.
1873  *
1874  * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
1875  * indicate that memory is out of balance and the caller must wait
1876  * for I/O to complete.  Otherwise, it will return 0 to indicate
1877  * that either memory was already in balance, or it was able to sleep
1878  * until the amount of dirty memory returned to balance.
1879  */
balance_dirty_pages_ratelimited_flags(struct address_space * mapping,unsigned int flags)1880 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
1881 					unsigned int flags)
1882 {
1883 	struct inode *inode = mapping->host;
1884 	struct backing_dev_info *bdi = inode_to_bdi(inode);
1885 	struct bdi_writeback *wb = NULL;
1886 	int ratelimit;
1887 	int ret = 0;
1888 	int *p;
1889 
1890 	if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1891 		return ret;
1892 
1893 	if (inode_cgwb_enabled(inode))
1894 		wb = wb_get_create_current(bdi, GFP_KERNEL);
1895 	if (!wb)
1896 		wb = &bdi->wb;
1897 
1898 	ratelimit = current->nr_dirtied_pause;
1899 	if (wb->dirty_exceeded)
1900 		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1901 
1902 	preempt_disable();
1903 	/*
1904 	 * This prevents one CPU to accumulate too many dirtied pages without
1905 	 * calling into balance_dirty_pages(), which can happen when there are
1906 	 * 1000+ tasks, all of them start dirtying pages at exactly the same
1907 	 * time, hence all honoured too large initial task->nr_dirtied_pause.
1908 	 */
1909 	p =  this_cpu_ptr(&bdp_ratelimits);
1910 	if (unlikely(current->nr_dirtied >= ratelimit))
1911 		*p = 0;
1912 	else if (unlikely(*p >= ratelimit_pages)) {
1913 		*p = 0;
1914 		ratelimit = 0;
1915 	}
1916 	/*
1917 	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1918 	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1919 	 * the dirty throttling and livelock other long-run dirtiers.
1920 	 */
1921 	p = this_cpu_ptr(&dirty_throttle_leaks);
1922 	if (*p > 0 && current->nr_dirtied < ratelimit) {
1923 		unsigned long nr_pages_dirtied;
1924 		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1925 		*p -= nr_pages_dirtied;
1926 		current->nr_dirtied += nr_pages_dirtied;
1927 	}
1928 	preempt_enable();
1929 
1930 	if (unlikely(current->nr_dirtied >= ratelimit))
1931 		ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
1932 
1933 	wb_put(wb);
1934 	return ret;
1935 }
1936 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
1937 
1938 /**
1939  * balance_dirty_pages_ratelimited - balance dirty memory state.
1940  * @mapping: address_space which was dirtied.
1941  *
1942  * Processes which are dirtying memory should call in here once for each page
1943  * which was newly dirtied.  The function will periodically check the system's
1944  * dirty state and will initiate writeback if needed.
1945  *
1946  * Once we're over the dirty memory limit we decrease the ratelimiting
1947  * by a lot, to prevent individual processes from overshooting the limit
1948  * by (ratelimit_pages) each.
1949  */
balance_dirty_pages_ratelimited(struct address_space * mapping)1950 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1951 {
1952 	balance_dirty_pages_ratelimited_flags(mapping, 0);
1953 }
1954 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1955 
1956 /**
1957  * wb_over_bg_thresh - does @wb need to be written back?
1958  * @wb: bdi_writeback of interest
1959  *
1960  * Determines whether background writeback should keep writing @wb or it's
1961  * clean enough.
1962  *
1963  * Return: %true if writeback should continue.
1964  */
wb_over_bg_thresh(struct bdi_writeback * wb)1965 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1966 {
1967 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1968 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1969 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1970 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1971 						     &mdtc_stor : NULL;
1972 	unsigned long reclaimable;
1973 	unsigned long thresh;
1974 
1975 	/*
1976 	 * Similar to balance_dirty_pages() but ignores pages being written
1977 	 * as we're trying to decide whether to put more under writeback.
1978 	 */
1979 	gdtc->avail = global_dirtyable_memory();
1980 	gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1981 	domain_dirty_limits(gdtc);
1982 
1983 	if (gdtc->dirty > gdtc->bg_thresh)
1984 		return true;
1985 
1986 	thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1987 	if (thresh < 2 * wb_stat_error())
1988 		reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1989 	else
1990 		reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1991 
1992 	if (reclaimable > thresh)
1993 		return true;
1994 
1995 	if (mdtc) {
1996 		unsigned long filepages, headroom, writeback;
1997 
1998 		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1999 				    &writeback);
2000 		mdtc_calc_avail(mdtc, filepages, headroom);
2001 		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
2002 
2003 		if (mdtc->dirty > mdtc->bg_thresh)
2004 			return true;
2005 
2006 		thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
2007 		if (thresh < 2 * wb_stat_error())
2008 			reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2009 		else
2010 			reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2011 
2012 		if (reclaimable > thresh)
2013 			return true;
2014 	}
2015 
2016 	return false;
2017 }
2018 
2019 #ifdef CONFIG_SYSCTL
2020 /*
2021  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2022  */
dirty_writeback_centisecs_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)2023 static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2024 		void *buffer, size_t *length, loff_t *ppos)
2025 {
2026 	unsigned int old_interval = dirty_writeback_interval;
2027 	int ret;
2028 
2029 	ret = proc_dointvec(table, write, buffer, length, ppos);
2030 
2031 	/*
2032 	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2033 	 * and a different non-zero value will wakeup the writeback threads.
2034 	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2035 	 * iterate over all bdis and wbs.
2036 	 * The reason we do this is to make the change take effect immediately.
2037 	 */
2038 	if (!ret && write && dirty_writeback_interval &&
2039 		dirty_writeback_interval != old_interval)
2040 		wakeup_flusher_threads(WB_REASON_PERIODIC);
2041 
2042 	return ret;
2043 }
2044 #endif
2045 
laptop_mode_timer_fn(struct timer_list * t)2046 void laptop_mode_timer_fn(struct timer_list *t)
2047 {
2048 	struct backing_dev_info *backing_dev_info =
2049 		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2050 
2051 	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2052 }
2053 
2054 /*
2055  * We've spun up the disk and we're in laptop mode: schedule writeback
2056  * of all dirty data a few seconds from now.  If the flush is already scheduled
2057  * then push it back - the user is still using the disk.
2058  */
laptop_io_completion(struct backing_dev_info * info)2059 void laptop_io_completion(struct backing_dev_info *info)
2060 {
2061 	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2062 }
2063 
2064 /*
2065  * We're in laptop mode and we've just synced. The sync's writes will have
2066  * caused another writeback to be scheduled by laptop_io_completion.
2067  * Nothing needs to be written back anymore, so we unschedule the writeback.
2068  */
laptop_sync_completion(void)2069 void laptop_sync_completion(void)
2070 {
2071 	struct backing_dev_info *bdi;
2072 
2073 	rcu_read_lock();
2074 
2075 	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2076 		del_timer(&bdi->laptop_mode_wb_timer);
2077 
2078 	rcu_read_unlock();
2079 }
2080 
2081 /*
2082  * If ratelimit_pages is too high then we can get into dirty-data overload
2083  * if a large number of processes all perform writes at the same time.
2084  *
2085  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2086  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2087  * thresholds.
2088  */
2089 
writeback_set_ratelimit(void)2090 void writeback_set_ratelimit(void)
2091 {
2092 	struct wb_domain *dom = &global_wb_domain;
2093 	unsigned long background_thresh;
2094 	unsigned long dirty_thresh;
2095 
2096 	global_dirty_limits(&background_thresh, &dirty_thresh);
2097 	dom->dirty_limit = dirty_thresh;
2098 	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2099 	if (ratelimit_pages < 16)
2100 		ratelimit_pages = 16;
2101 }
2102 
page_writeback_cpu_online(unsigned int cpu)2103 static int page_writeback_cpu_online(unsigned int cpu)
2104 {
2105 	writeback_set_ratelimit();
2106 	return 0;
2107 }
2108 
2109 #ifdef CONFIG_SYSCTL
2110 
2111 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2112 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2113 
2114 static struct ctl_table vm_page_writeback_sysctls[] = {
2115 	{
2116 		.procname   = "dirty_background_ratio",
2117 		.data       = &dirty_background_ratio,
2118 		.maxlen     = sizeof(dirty_background_ratio),
2119 		.mode       = 0644,
2120 		.proc_handler   = dirty_background_ratio_handler,
2121 		.extra1     = SYSCTL_ZERO,
2122 		.extra2     = SYSCTL_ONE_HUNDRED,
2123 	},
2124 	{
2125 		.procname   = "dirty_background_bytes",
2126 		.data       = &dirty_background_bytes,
2127 		.maxlen     = sizeof(dirty_background_bytes),
2128 		.mode       = 0644,
2129 		.proc_handler   = dirty_background_bytes_handler,
2130 		.extra1     = SYSCTL_LONG_ONE,
2131 	},
2132 	{
2133 		.procname   = "dirty_ratio",
2134 		.data       = &vm_dirty_ratio,
2135 		.maxlen     = sizeof(vm_dirty_ratio),
2136 		.mode       = 0644,
2137 		.proc_handler   = dirty_ratio_handler,
2138 		.extra1     = SYSCTL_ZERO,
2139 		.extra2     = SYSCTL_ONE_HUNDRED,
2140 	},
2141 	{
2142 		.procname   = "dirty_bytes",
2143 		.data       = &vm_dirty_bytes,
2144 		.maxlen     = sizeof(vm_dirty_bytes),
2145 		.mode       = 0644,
2146 		.proc_handler   = dirty_bytes_handler,
2147 		.extra1     = (void *)&dirty_bytes_min,
2148 	},
2149 	{
2150 		.procname   = "dirty_writeback_centisecs",
2151 		.data       = &dirty_writeback_interval,
2152 		.maxlen     = sizeof(dirty_writeback_interval),
2153 		.mode       = 0644,
2154 		.proc_handler   = dirty_writeback_centisecs_handler,
2155 	},
2156 	{
2157 		.procname   = "dirty_expire_centisecs",
2158 		.data       = &dirty_expire_interval,
2159 		.maxlen     = sizeof(dirty_expire_interval),
2160 		.mode       = 0644,
2161 		.proc_handler   = proc_dointvec_minmax,
2162 		.extra1     = SYSCTL_ZERO,
2163 	},
2164 #ifdef CONFIG_HIGHMEM
2165 	{
2166 		.procname	= "highmem_is_dirtyable",
2167 		.data		= &vm_highmem_is_dirtyable,
2168 		.maxlen		= sizeof(vm_highmem_is_dirtyable),
2169 		.mode		= 0644,
2170 		.proc_handler	= proc_dointvec_minmax,
2171 		.extra1		= SYSCTL_ZERO,
2172 		.extra2		= SYSCTL_ONE,
2173 	},
2174 #endif
2175 	{
2176 		.procname	= "laptop_mode",
2177 		.data		= &laptop_mode,
2178 		.maxlen		= sizeof(laptop_mode),
2179 		.mode		= 0644,
2180 		.proc_handler	= proc_dointvec_jiffies,
2181 	},
2182 	{}
2183 };
2184 #endif
2185 
2186 /*
2187  * Called early on to tune the page writeback dirty limits.
2188  *
2189  * We used to scale dirty pages according to how total memory
2190  * related to pages that could be allocated for buffers.
2191  *
2192  * However, that was when we used "dirty_ratio" to scale with
2193  * all memory, and we don't do that any more. "dirty_ratio"
2194  * is now applied to total non-HIGHPAGE memory, and as such we can't
2195  * get into the old insane situation any more where we had
2196  * large amounts of dirty pages compared to a small amount of
2197  * non-HIGHMEM memory.
2198  *
2199  * But we might still want to scale the dirty_ratio by how
2200  * much memory the box has..
2201  */
page_writeback_init(void)2202 void __init page_writeback_init(void)
2203 {
2204 	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2205 
2206 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2207 			  page_writeback_cpu_online, NULL);
2208 	cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2209 			  page_writeback_cpu_online);
2210 #ifdef CONFIG_SYSCTL
2211 	register_sysctl_init("vm", vm_page_writeback_sysctls);
2212 #endif
2213 }
2214 
2215 /**
2216  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2217  * @mapping: address space structure to write
2218  * @start: starting page index
2219  * @end: ending page index (inclusive)
2220  *
2221  * This function scans the page range from @start to @end (inclusive) and tags
2222  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2223  * that write_cache_pages (or whoever calls this function) will then use
2224  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2225  * used to avoid livelocking of writeback by a process steadily creating new
2226  * dirty pages in the file (thus it is important for this function to be quick
2227  * so that it can tag pages faster than a dirtying process can create them).
2228  */
tag_pages_for_writeback(struct address_space * mapping,pgoff_t start,pgoff_t end)2229 void tag_pages_for_writeback(struct address_space *mapping,
2230 			     pgoff_t start, pgoff_t end)
2231 {
2232 	XA_STATE(xas, &mapping->i_pages, start);
2233 	unsigned int tagged = 0;
2234 	void *page;
2235 
2236 	xas_lock_irq(&xas);
2237 	xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2238 		xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2239 		if (++tagged % XA_CHECK_SCHED)
2240 			continue;
2241 
2242 		xas_pause(&xas);
2243 		xas_unlock_irq(&xas);
2244 		cond_resched();
2245 		xas_lock_irq(&xas);
2246 	}
2247 	xas_unlock_irq(&xas);
2248 }
2249 EXPORT_SYMBOL(tag_pages_for_writeback);
2250 
2251 /**
2252  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2253  * @mapping: address space structure to write
2254  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2255  * @writepage: function called for each page
2256  * @data: data passed to writepage function
2257  *
2258  * If a page is already under I/O, write_cache_pages() skips it, even
2259  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2260  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2261  * and msync() need to guarantee that all the data which was dirty at the time
2262  * the call was made get new I/O started against them.  If wbc->sync_mode is
2263  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2264  * existing IO to complete.
2265  *
2266  * To avoid livelocks (when other process dirties new pages), we first tag
2267  * pages which should be written back with TOWRITE tag and only then start
2268  * writing them. For data-integrity sync we have to be careful so that we do
2269  * not miss some pages (e.g., because some other process has cleared TOWRITE
2270  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2271  * by the process clearing the DIRTY tag (and submitting the page for IO).
2272  *
2273  * To avoid deadlocks between range_cyclic writeback and callers that hold
2274  * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2275  * we do not loop back to the start of the file. Doing so causes a page
2276  * lock/page writeback access order inversion - we should only ever lock
2277  * multiple pages in ascending page->index order, and looping back to the start
2278  * of the file violates that rule and causes deadlocks.
2279  *
2280  * Return: %0 on success, negative error code otherwise
2281  */
write_cache_pages(struct address_space * mapping,struct writeback_control * wbc,writepage_t writepage,void * data)2282 int write_cache_pages(struct address_space *mapping,
2283 		      struct writeback_control *wbc, writepage_t writepage,
2284 		      void *data)
2285 {
2286 	int ret = 0;
2287 	int done = 0;
2288 	int error;
2289 	struct pagevec pvec;
2290 	int nr_pages;
2291 	pgoff_t index;
2292 	pgoff_t end;		/* Inclusive */
2293 	pgoff_t done_index;
2294 	int range_whole = 0;
2295 	xa_mark_t tag;
2296 
2297 	pagevec_init(&pvec);
2298 	if (wbc->range_cyclic) {
2299 		index = mapping->writeback_index; /* prev offset */
2300 		end = -1;
2301 	} else {
2302 		index = wbc->range_start >> PAGE_SHIFT;
2303 		end = wbc->range_end >> PAGE_SHIFT;
2304 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2305 			range_whole = 1;
2306 	}
2307 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2308 		tag_pages_for_writeback(mapping, index, end);
2309 		tag = PAGECACHE_TAG_TOWRITE;
2310 	} else {
2311 		tag = PAGECACHE_TAG_DIRTY;
2312 	}
2313 	done_index = index;
2314 	while (!done && (index <= end)) {
2315 		int i;
2316 
2317 		nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2318 				tag);
2319 		if (nr_pages == 0)
2320 			break;
2321 
2322 		for (i = 0; i < nr_pages; i++) {
2323 			struct page *page = pvec.pages[i];
2324 
2325 			done_index = page->index;
2326 
2327 			lock_page(page);
2328 
2329 			/*
2330 			 * Page truncated or invalidated. We can freely skip it
2331 			 * then, even for data integrity operations: the page
2332 			 * has disappeared concurrently, so there could be no
2333 			 * real expectation of this data integrity operation
2334 			 * even if there is now a new, dirty page at the same
2335 			 * pagecache address.
2336 			 */
2337 			if (unlikely(page->mapping != mapping)) {
2338 continue_unlock:
2339 				unlock_page(page);
2340 				continue;
2341 			}
2342 
2343 			if (!PageDirty(page)) {
2344 				/* someone wrote it for us */
2345 				goto continue_unlock;
2346 			}
2347 
2348 			if (PageWriteback(page)) {
2349 				if (wbc->sync_mode != WB_SYNC_NONE)
2350 					wait_on_page_writeback(page);
2351 				else
2352 					goto continue_unlock;
2353 			}
2354 
2355 			BUG_ON(PageWriteback(page));
2356 			if (!clear_page_dirty_for_io(page))
2357 				goto continue_unlock;
2358 
2359 			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2360 			error = (*writepage)(page, wbc, data);
2361 			if (unlikely(error)) {
2362 				/*
2363 				 * Handle errors according to the type of
2364 				 * writeback. There's no need to continue for
2365 				 * background writeback. Just push done_index
2366 				 * past this page so media errors won't choke
2367 				 * writeout for the entire file. For integrity
2368 				 * writeback, we must process the entire dirty
2369 				 * set regardless of errors because the fs may
2370 				 * still have state to clear for each page. In
2371 				 * that case we continue processing and return
2372 				 * the first error.
2373 				 */
2374 				if (error == AOP_WRITEPAGE_ACTIVATE) {
2375 					unlock_page(page);
2376 					error = 0;
2377 				} else if (wbc->sync_mode != WB_SYNC_ALL) {
2378 					ret = error;
2379 					done_index = page->index + 1;
2380 					done = 1;
2381 					break;
2382 				}
2383 				if (!ret)
2384 					ret = error;
2385 			}
2386 
2387 			/*
2388 			 * We stop writing back only if we are not doing
2389 			 * integrity sync. In case of integrity sync we have to
2390 			 * keep going until we have written all the pages
2391 			 * we tagged for writeback prior to entering this loop.
2392 			 */
2393 			if (--wbc->nr_to_write <= 0 &&
2394 			    wbc->sync_mode == WB_SYNC_NONE) {
2395 				done = 1;
2396 				break;
2397 			}
2398 		}
2399 		pagevec_release(&pvec);
2400 		cond_resched();
2401 	}
2402 
2403 	/*
2404 	 * If we hit the last page and there is more work to be done: wrap
2405 	 * back the index back to the start of the file for the next
2406 	 * time we are called.
2407 	 */
2408 	if (wbc->range_cyclic && !done)
2409 		done_index = 0;
2410 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2411 		mapping->writeback_index = done_index;
2412 
2413 	return ret;
2414 }
2415 EXPORT_SYMBOL(write_cache_pages);
2416 
2417 /*
2418  * Function used by generic_writepages to call the real writepage
2419  * function and set the mapping flags on error
2420  */
__writepage(struct page * page,struct writeback_control * wbc,void * data)2421 static int __writepage(struct page *page, struct writeback_control *wbc,
2422 		       void *data)
2423 {
2424 	struct address_space *mapping = data;
2425 	int ret = mapping->a_ops->writepage(page, wbc);
2426 	mapping_set_error(mapping, ret);
2427 	return ret;
2428 }
2429 
2430 /**
2431  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2432  * @mapping: address space structure to write
2433  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2434  *
2435  * This is a library function, which implements the writepages()
2436  * address_space_operation.
2437  *
2438  * Return: %0 on success, negative error code otherwise
2439  */
generic_writepages(struct address_space * mapping,struct writeback_control * wbc)2440 int generic_writepages(struct address_space *mapping,
2441 		       struct writeback_control *wbc)
2442 {
2443 	struct blk_plug plug;
2444 	int ret;
2445 
2446 	/* deal with chardevs and other special file */
2447 	if (!mapping->a_ops->writepage)
2448 		return 0;
2449 
2450 	blk_start_plug(&plug);
2451 	ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2452 	blk_finish_plug(&plug);
2453 	return ret;
2454 }
2455 
2456 EXPORT_SYMBOL(generic_writepages);
2457 
do_writepages(struct address_space * mapping,struct writeback_control * wbc)2458 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2459 {
2460 	int ret;
2461 	struct bdi_writeback *wb;
2462 
2463 	if (wbc->nr_to_write <= 0)
2464 		return 0;
2465 	wb = inode_to_wb_wbc(mapping->host, wbc);
2466 	wb_bandwidth_estimate_start(wb);
2467 	while (1) {
2468 		if (mapping->a_ops->writepages)
2469 			ret = mapping->a_ops->writepages(mapping, wbc);
2470 		else
2471 			ret = generic_writepages(mapping, wbc);
2472 		if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2473 			break;
2474 
2475 		/*
2476 		 * Lacking an allocation context or the locality or writeback
2477 		 * state of any of the inode's pages, throttle based on
2478 		 * writeback activity on the local node. It's as good a
2479 		 * guess as any.
2480 		 */
2481 		reclaim_throttle(NODE_DATA(numa_node_id()),
2482 			VMSCAN_THROTTLE_WRITEBACK);
2483 	}
2484 	/*
2485 	 * Usually few pages are written by now from those we've just submitted
2486 	 * but if there's constant writeback being submitted, this makes sure
2487 	 * writeback bandwidth is updated once in a while.
2488 	 */
2489 	if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2490 				   BANDWIDTH_INTERVAL))
2491 		wb_update_bandwidth(wb);
2492 	return ret;
2493 }
2494 
2495 /**
2496  * folio_write_one - write out a single folio and wait on I/O.
2497  * @folio: The folio to write.
2498  *
2499  * The folio must be locked by the caller and will be unlocked upon return.
2500  *
2501  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2502  * function returns.
2503  *
2504  * Return: %0 on success, negative error code otherwise
2505  */
folio_write_one(struct folio * folio)2506 int folio_write_one(struct folio *folio)
2507 {
2508 	struct address_space *mapping = folio->mapping;
2509 	int ret = 0;
2510 	struct writeback_control wbc = {
2511 		.sync_mode = WB_SYNC_ALL,
2512 		.nr_to_write = folio_nr_pages(folio),
2513 	};
2514 
2515 	BUG_ON(!folio_test_locked(folio));
2516 
2517 	folio_wait_writeback(folio);
2518 
2519 	if (folio_clear_dirty_for_io(folio)) {
2520 		folio_get(folio);
2521 		ret = mapping->a_ops->writepage(&folio->page, &wbc);
2522 		if (ret == 0)
2523 			folio_wait_writeback(folio);
2524 		folio_put(folio);
2525 	} else {
2526 		folio_unlock(folio);
2527 	}
2528 
2529 	if (!ret)
2530 		ret = filemap_check_errors(mapping);
2531 	return ret;
2532 }
2533 EXPORT_SYMBOL(folio_write_one);
2534 
2535 /*
2536  * For address_spaces which do not use buffers nor write back.
2537  */
noop_dirty_folio(struct address_space * mapping,struct folio * folio)2538 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2539 {
2540 	if (!folio_test_dirty(folio))
2541 		return !folio_test_set_dirty(folio);
2542 	return false;
2543 }
2544 EXPORT_SYMBOL(noop_dirty_folio);
2545 
2546 /*
2547  * Helper function for set_page_dirty family.
2548  *
2549  * Caller must hold lock_page_memcg().
2550  *
2551  * NOTE: This relies on being atomic wrt interrupts.
2552  */
folio_account_dirtied(struct folio * folio,struct address_space * mapping)2553 static void folio_account_dirtied(struct folio *folio,
2554 		struct address_space *mapping)
2555 {
2556 	struct inode *inode = mapping->host;
2557 
2558 	trace_writeback_dirty_folio(folio, mapping);
2559 
2560 	if (mapping_can_writeback(mapping)) {
2561 		struct bdi_writeback *wb;
2562 		long nr = folio_nr_pages(folio);
2563 
2564 		inode_attach_wb(inode, &folio->page);
2565 		wb = inode_to_wb(inode);
2566 
2567 		__lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2568 		__zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2569 		__node_stat_mod_folio(folio, NR_DIRTIED, nr);
2570 		wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2571 		wb_stat_mod(wb, WB_DIRTIED, nr);
2572 		task_io_account_write(nr * PAGE_SIZE);
2573 		current->nr_dirtied += nr;
2574 		__this_cpu_add(bdp_ratelimits, nr);
2575 
2576 		mem_cgroup_track_foreign_dirty(folio, wb);
2577 	}
2578 }
2579 
2580 /*
2581  * Helper function for deaccounting dirty page without writeback.
2582  *
2583  * Caller must hold lock_page_memcg().
2584  */
folio_account_cleaned(struct folio * folio,struct bdi_writeback * wb)2585 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2586 {
2587 	long nr = folio_nr_pages(folio);
2588 
2589 	lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2590 	zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2591 	wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2592 	task_io_account_cancelled_write(nr * PAGE_SIZE);
2593 }
2594 
2595 /*
2596  * Mark the folio dirty, and set it dirty in the page cache, and mark
2597  * the inode dirty.
2598  *
2599  * If warn is true, then emit a warning if the folio is not uptodate and has
2600  * not been truncated.
2601  *
2602  * The caller must hold lock_page_memcg().  Most callers have the folio
2603  * locked.  A few have the folio blocked from truncation through other
2604  * means (eg zap_page_range() has it mapped and is holding the page table
2605  * lock).  This can also be called from mark_buffer_dirty(), which I
2606  * cannot prove is always protected against truncate.
2607  */
__folio_mark_dirty(struct folio * folio,struct address_space * mapping,int warn)2608 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2609 			     int warn)
2610 {
2611 	unsigned long flags;
2612 
2613 	xa_lock_irqsave(&mapping->i_pages, flags);
2614 	if (folio->mapping) {	/* Race with truncate? */
2615 		WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2616 		folio_account_dirtied(folio, mapping);
2617 		__xa_set_mark(&mapping->i_pages, folio_index(folio),
2618 				PAGECACHE_TAG_DIRTY);
2619 	}
2620 	xa_unlock_irqrestore(&mapping->i_pages, flags);
2621 }
2622 
2623 /**
2624  * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2625  * @mapping: Address space this folio belongs to.
2626  * @folio: Folio to be marked as dirty.
2627  *
2628  * Filesystems which do not use buffer heads should call this function
2629  * from their set_page_dirty address space operation.  It ignores the
2630  * contents of folio_get_private(), so if the filesystem marks individual
2631  * blocks as dirty, the filesystem should handle that itself.
2632  *
2633  * This is also sometimes used by filesystems which use buffer_heads when
2634  * a single buffer is being dirtied: we want to set the folio dirty in
2635  * that case, but not all the buffers.  This is a "bottom-up" dirtying,
2636  * whereas block_dirty_folio() is a "top-down" dirtying.
2637  *
2638  * The caller must ensure this doesn't race with truncation.  Most will
2639  * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2640  * folio mapped and the pte lock held, which also locks out truncation.
2641  */
filemap_dirty_folio(struct address_space * mapping,struct folio * folio)2642 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2643 {
2644 	folio_memcg_lock(folio);
2645 	if (folio_test_set_dirty(folio)) {
2646 		folio_memcg_unlock(folio);
2647 		return false;
2648 	}
2649 
2650 	__folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2651 	folio_memcg_unlock(folio);
2652 
2653 	if (mapping->host) {
2654 		/* !PageAnon && !swapper_space */
2655 		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2656 	}
2657 	return true;
2658 }
2659 EXPORT_SYMBOL(filemap_dirty_folio);
2660 
2661 /**
2662  * folio_account_redirty - Manually account for redirtying a page.
2663  * @folio: The folio which is being redirtied.
2664  *
2665  * Most filesystems should call folio_redirty_for_writepage() instead
2666  * of this fuction.  If your filesystem is doing writeback outside the
2667  * context of a writeback_control(), it can call this when redirtying
2668  * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2669  * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2670  * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2671  * in balanced_dirty_ratelimit and the dirty pages position control.
2672  */
folio_account_redirty(struct folio * folio)2673 void folio_account_redirty(struct folio *folio)
2674 {
2675 	struct address_space *mapping = folio->mapping;
2676 
2677 	if (mapping && mapping_can_writeback(mapping)) {
2678 		struct inode *inode = mapping->host;
2679 		struct bdi_writeback *wb;
2680 		struct wb_lock_cookie cookie = {};
2681 		long nr = folio_nr_pages(folio);
2682 
2683 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2684 		current->nr_dirtied -= nr;
2685 		node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2686 		wb_stat_mod(wb, WB_DIRTIED, -nr);
2687 		unlocked_inode_to_wb_end(inode, &cookie);
2688 	}
2689 }
2690 EXPORT_SYMBOL(folio_account_redirty);
2691 
2692 /**
2693  * folio_redirty_for_writepage - Decline to write a dirty folio.
2694  * @wbc: The writeback control.
2695  * @folio: The folio.
2696  *
2697  * When a writepage implementation decides that it doesn't want to write
2698  * @folio for some reason, it should call this function, unlock @folio and
2699  * return 0.
2700  *
2701  * Return: True if we redirtied the folio.  False if someone else dirtied
2702  * it first.
2703  */
folio_redirty_for_writepage(struct writeback_control * wbc,struct folio * folio)2704 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2705 		struct folio *folio)
2706 {
2707 	bool ret;
2708 	long nr = folio_nr_pages(folio);
2709 
2710 	wbc->pages_skipped += nr;
2711 	ret = filemap_dirty_folio(folio->mapping, folio);
2712 	folio_account_redirty(folio);
2713 
2714 	return ret;
2715 }
2716 EXPORT_SYMBOL(folio_redirty_for_writepage);
2717 
2718 /**
2719  * folio_mark_dirty - Mark a folio as being modified.
2720  * @folio: The folio.
2721  *
2722  * The folio may not be truncated while this function is running.
2723  * Holding the folio lock is sufficient to prevent truncation, but some
2724  * callers cannot acquire a sleeping lock.  These callers instead hold
2725  * the page table lock for a page table which contains at least one page
2726  * in this folio.  Truncation will block on the page table lock as it
2727  * unmaps pages before removing the folio from its mapping.
2728  *
2729  * Return: True if the folio was newly dirtied, false if it was already dirty.
2730  */
folio_mark_dirty(struct folio * folio)2731 bool folio_mark_dirty(struct folio *folio)
2732 {
2733 	struct address_space *mapping = folio_mapping(folio);
2734 
2735 	if (likely(mapping)) {
2736 		/*
2737 		 * readahead/lru_deactivate_page could remain
2738 		 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2739 		 * About readahead, if the folio is written, the flags would be
2740 		 * reset. So no problem.
2741 		 * About lru_deactivate_page, if the folio is redirtied,
2742 		 * the flag will be reset. So no problem. but if the
2743 		 * folio is used by readahead it will confuse readahead
2744 		 * and make it restart the size rampup process. But it's
2745 		 * a trivial problem.
2746 		 */
2747 		if (folio_test_reclaim(folio))
2748 			folio_clear_reclaim(folio);
2749 		return mapping->a_ops->dirty_folio(mapping, folio);
2750 	}
2751 
2752 	return noop_dirty_folio(mapping, folio);
2753 }
2754 EXPORT_SYMBOL(folio_mark_dirty);
2755 
2756 /*
2757  * set_page_dirty() is racy if the caller has no reference against
2758  * page->mapping->host, and if the page is unlocked.  This is because another
2759  * CPU could truncate the page off the mapping and then free the mapping.
2760  *
2761  * Usually, the page _is_ locked, or the caller is a user-space process which
2762  * holds a reference on the inode by having an open file.
2763  *
2764  * In other cases, the page should be locked before running set_page_dirty().
2765  */
set_page_dirty_lock(struct page * page)2766 int set_page_dirty_lock(struct page *page)
2767 {
2768 	int ret;
2769 
2770 	lock_page(page);
2771 	ret = set_page_dirty(page);
2772 	unlock_page(page);
2773 	return ret;
2774 }
2775 EXPORT_SYMBOL(set_page_dirty_lock);
2776 
2777 /*
2778  * This cancels just the dirty bit on the kernel page itself, it does NOT
2779  * actually remove dirty bits on any mmap's that may be around. It also
2780  * leaves the page tagged dirty, so any sync activity will still find it on
2781  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2782  * look at the dirty bits in the VM.
2783  *
2784  * Doing this should *normally* only ever be done when a page is truncated,
2785  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2786  * this when it notices that somebody has cleaned out all the buffers on a
2787  * page without actually doing it through the VM. Can you say "ext3 is
2788  * horribly ugly"? Thought you could.
2789  */
__folio_cancel_dirty(struct folio * folio)2790 void __folio_cancel_dirty(struct folio *folio)
2791 {
2792 	struct address_space *mapping = folio_mapping(folio);
2793 
2794 	if (mapping_can_writeback(mapping)) {
2795 		struct inode *inode = mapping->host;
2796 		struct bdi_writeback *wb;
2797 		struct wb_lock_cookie cookie = {};
2798 
2799 		folio_memcg_lock(folio);
2800 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2801 
2802 		if (folio_test_clear_dirty(folio))
2803 			folio_account_cleaned(folio, wb);
2804 
2805 		unlocked_inode_to_wb_end(inode, &cookie);
2806 		folio_memcg_unlock(folio);
2807 	} else {
2808 		folio_clear_dirty(folio);
2809 	}
2810 }
2811 EXPORT_SYMBOL(__folio_cancel_dirty);
2812 
2813 /*
2814  * Clear a folio's dirty flag, while caring for dirty memory accounting.
2815  * Returns true if the folio was previously dirty.
2816  *
2817  * This is for preparing to put the folio under writeout.  We leave
2818  * the folio tagged as dirty in the xarray so that a concurrent
2819  * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2820  * The ->writepage implementation will run either folio_start_writeback()
2821  * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2822  * and xarray dirty tag back into sync.
2823  *
2824  * This incoherency between the folio's dirty flag and xarray tag is
2825  * unfortunate, but it only exists while the folio is locked.
2826  */
folio_clear_dirty_for_io(struct folio * folio)2827 bool folio_clear_dirty_for_io(struct folio *folio)
2828 {
2829 	struct address_space *mapping = folio_mapping(folio);
2830 	bool ret = false;
2831 
2832 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2833 
2834 	if (mapping && mapping_can_writeback(mapping)) {
2835 		struct inode *inode = mapping->host;
2836 		struct bdi_writeback *wb;
2837 		struct wb_lock_cookie cookie = {};
2838 
2839 		/*
2840 		 * Yes, Virginia, this is indeed insane.
2841 		 *
2842 		 * We use this sequence to make sure that
2843 		 *  (a) we account for dirty stats properly
2844 		 *  (b) we tell the low-level filesystem to
2845 		 *      mark the whole folio dirty if it was
2846 		 *      dirty in a pagetable. Only to then
2847 		 *  (c) clean the folio again and return 1 to
2848 		 *      cause the writeback.
2849 		 *
2850 		 * This way we avoid all nasty races with the
2851 		 * dirty bit in multiple places and clearing
2852 		 * them concurrently from different threads.
2853 		 *
2854 		 * Note! Normally the "folio_mark_dirty(folio)"
2855 		 * has no effect on the actual dirty bit - since
2856 		 * that will already usually be set. But we
2857 		 * need the side effects, and it can help us
2858 		 * avoid races.
2859 		 *
2860 		 * We basically use the folio "master dirty bit"
2861 		 * as a serialization point for all the different
2862 		 * threads doing their things.
2863 		 */
2864 		if (folio_mkclean(folio))
2865 			folio_mark_dirty(folio);
2866 		/*
2867 		 * We carefully synchronise fault handlers against
2868 		 * installing a dirty pte and marking the folio dirty
2869 		 * at this point.  We do this by having them hold the
2870 		 * page lock while dirtying the folio, and folios are
2871 		 * always locked coming in here, so we get the desired
2872 		 * exclusion.
2873 		 */
2874 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2875 		if (folio_test_clear_dirty(folio)) {
2876 			long nr = folio_nr_pages(folio);
2877 			lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2878 			zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2879 			wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2880 			ret = true;
2881 		}
2882 		unlocked_inode_to_wb_end(inode, &cookie);
2883 		return ret;
2884 	}
2885 	return folio_test_clear_dirty(folio);
2886 }
2887 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2888 
wb_inode_writeback_start(struct bdi_writeback * wb)2889 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2890 {
2891 	atomic_inc(&wb->writeback_inodes);
2892 }
2893 
wb_inode_writeback_end(struct bdi_writeback * wb)2894 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2895 {
2896 	unsigned long flags;
2897 	atomic_dec(&wb->writeback_inodes);
2898 	/*
2899 	 * Make sure estimate of writeback throughput gets updated after
2900 	 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2901 	 * (which is the interval other bandwidth updates use for batching) so
2902 	 * that if multiple inodes end writeback at a similar time, they get
2903 	 * batched into one bandwidth update.
2904 	 */
2905 	spin_lock_irqsave(&wb->work_lock, flags);
2906 	if (test_bit(WB_registered, &wb->state))
2907 		queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2908 	spin_unlock_irqrestore(&wb->work_lock, flags);
2909 }
2910 
__folio_end_writeback(struct folio * folio)2911 bool __folio_end_writeback(struct folio *folio)
2912 {
2913 	long nr = folio_nr_pages(folio);
2914 	struct address_space *mapping = folio_mapping(folio);
2915 	bool ret;
2916 
2917 	folio_memcg_lock(folio);
2918 	if (mapping && mapping_use_writeback_tags(mapping)) {
2919 		struct inode *inode = mapping->host;
2920 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2921 		unsigned long flags;
2922 
2923 		xa_lock_irqsave(&mapping->i_pages, flags);
2924 		ret = folio_test_clear_writeback(folio);
2925 		if (ret) {
2926 			__xa_clear_mark(&mapping->i_pages, folio_index(folio),
2927 						PAGECACHE_TAG_WRITEBACK);
2928 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2929 				struct bdi_writeback *wb = inode_to_wb(inode);
2930 
2931 				wb_stat_mod(wb, WB_WRITEBACK, -nr);
2932 				__wb_writeout_add(wb, nr);
2933 				if (!mapping_tagged(mapping,
2934 						    PAGECACHE_TAG_WRITEBACK))
2935 					wb_inode_writeback_end(wb);
2936 			}
2937 		}
2938 
2939 		if (mapping->host && !mapping_tagged(mapping,
2940 						     PAGECACHE_TAG_WRITEBACK))
2941 			sb_clear_inode_writeback(mapping->host);
2942 
2943 		xa_unlock_irqrestore(&mapping->i_pages, flags);
2944 	} else {
2945 		ret = folio_test_clear_writeback(folio);
2946 	}
2947 	if (ret) {
2948 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2949 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2950 		node_stat_mod_folio(folio, NR_WRITTEN, nr);
2951 	}
2952 	folio_memcg_unlock(folio);
2953 	return ret;
2954 }
2955 
__folio_start_writeback(struct folio * folio,bool keep_write)2956 bool __folio_start_writeback(struct folio *folio, bool keep_write)
2957 {
2958 	long nr = folio_nr_pages(folio);
2959 	struct address_space *mapping = folio_mapping(folio);
2960 	bool ret;
2961 	int access_ret;
2962 
2963 	folio_memcg_lock(folio);
2964 	if (mapping && mapping_use_writeback_tags(mapping)) {
2965 		XA_STATE(xas, &mapping->i_pages, folio_index(folio));
2966 		struct inode *inode = mapping->host;
2967 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2968 		unsigned long flags;
2969 
2970 		xas_lock_irqsave(&xas, flags);
2971 		xas_load(&xas);
2972 		ret = folio_test_set_writeback(folio);
2973 		if (!ret) {
2974 			bool on_wblist;
2975 
2976 			on_wblist = mapping_tagged(mapping,
2977 						   PAGECACHE_TAG_WRITEBACK);
2978 
2979 			xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2980 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2981 				struct bdi_writeback *wb = inode_to_wb(inode);
2982 
2983 				wb_stat_mod(wb, WB_WRITEBACK, nr);
2984 				if (!on_wblist)
2985 					wb_inode_writeback_start(wb);
2986 			}
2987 
2988 			/*
2989 			 * We can come through here when swapping
2990 			 * anonymous folios, so we don't necessarily
2991 			 * have an inode to track for sync.
2992 			 */
2993 			if (mapping->host && !on_wblist)
2994 				sb_mark_inode_writeback(mapping->host);
2995 		}
2996 		if (!folio_test_dirty(folio))
2997 			xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2998 		if (!keep_write)
2999 			xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3000 		xas_unlock_irqrestore(&xas, flags);
3001 	} else {
3002 		ret = folio_test_set_writeback(folio);
3003 	}
3004 	if (!ret) {
3005 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3006 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3007 	}
3008 	folio_memcg_unlock(folio);
3009 	access_ret = arch_make_folio_accessible(folio);
3010 	/*
3011 	 * If writeback has been triggered on a page that cannot be made
3012 	 * accessible, it is too late to recover here.
3013 	 */
3014 	VM_BUG_ON_FOLIO(access_ret != 0, folio);
3015 
3016 	return ret;
3017 }
3018 EXPORT_SYMBOL(__folio_start_writeback);
3019 
3020 /**
3021  * folio_wait_writeback - Wait for a folio to finish writeback.
3022  * @folio: The folio to wait for.
3023  *
3024  * If the folio is currently being written back to storage, wait for the
3025  * I/O to complete.
3026  *
3027  * Context: Sleeps.  Must be called in process context and with
3028  * no spinlocks held.  Caller should hold a reference on the folio.
3029  * If the folio is not locked, writeback may start again after writeback
3030  * has finished.
3031  */
folio_wait_writeback(struct folio * folio)3032 void folio_wait_writeback(struct folio *folio)
3033 {
3034 	while (folio_test_writeback(folio)) {
3035 		trace_folio_wait_writeback(folio, folio_mapping(folio));
3036 		folio_wait_bit(folio, PG_writeback);
3037 	}
3038 }
3039 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3040 
3041 /**
3042  * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3043  * @folio: The folio to wait for.
3044  *
3045  * If the folio is currently being written back to storage, wait for the
3046  * I/O to complete or a fatal signal to arrive.
3047  *
3048  * Context: Sleeps.  Must be called in process context and with
3049  * no spinlocks held.  Caller should hold a reference on the folio.
3050  * If the folio is not locked, writeback may start again after writeback
3051  * has finished.
3052  * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3053  */
folio_wait_writeback_killable(struct folio * folio)3054 int folio_wait_writeback_killable(struct folio *folio)
3055 {
3056 	while (folio_test_writeback(folio)) {
3057 		trace_folio_wait_writeback(folio, folio_mapping(folio));
3058 		if (folio_wait_bit_killable(folio, PG_writeback))
3059 			return -EINTR;
3060 	}
3061 
3062 	return 0;
3063 }
3064 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3065 
3066 /**
3067  * folio_wait_stable() - wait for writeback to finish, if necessary.
3068  * @folio: The folio to wait on.
3069  *
3070  * This function determines if the given folio is related to a backing
3071  * device that requires folio contents to be held stable during writeback.
3072  * If so, then it will wait for any pending writeback to complete.
3073  *
3074  * Context: Sleeps.  Must be called in process context and with
3075  * no spinlocks held.  Caller should hold a reference on the folio.
3076  * If the folio is not locked, writeback may start again after writeback
3077  * has finished.
3078  */
folio_wait_stable(struct folio * folio)3079 void folio_wait_stable(struct folio *folio)
3080 {
3081 	if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
3082 		folio_wait_writeback(folio);
3083 }
3084 EXPORT_SYMBOL_GPL(folio_wait_stable);
3085