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/math64.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/fs.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/slab.h>
23 #include <linux/pagemap.h>
24 #include <linux/writeback.h>
25 #include <linux/init.h>
26 #include <linux/backing-dev.h>
27 #include <linux/task_io_accounting_ops.h>
28 #include <linux/blkdev.h>
29 #include <linux/mpage.h>
30 #include <linux/rmap.h>
31 #include <linux/percpu.h>
32 #include <linux/smp.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/syscalls.h>
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/sched/signal.h>
40 #include <linux/mm_inline.h>
41 #include <trace/events/writeback.h>
42 
43 #include "internal.h"
44 
45 /*
46  * Sleep at most 200ms at a time in balance_dirty_pages().
47  */
48 #define MAX_PAUSE		max(HZ/5, 1)
49 
50 /*
51  * Try to keep balance_dirty_pages() call intervals higher than this many pages
52  * by raising pause time to max_pause when falls below it.
53  */
54 #define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
55 
56 /*
57  * Estimate write bandwidth at 200ms intervals.
58  */
59 #define BANDWIDTH_INTERVAL	max(HZ/5, 1)
60 
61 #define RATELIMIT_CALC_SHIFT	10
62 
63 /*
64  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65  * will look to see if it needs to force writeback or throttling.
66  */
67 static long ratelimit_pages = 32;
68 
69 /* The following parameters are exported via /proc/sys/vm */
70 
71 /*
72  * Start background writeback (via writeback threads) at this percentage
73  */
74 static int dirty_background_ratio = 10;
75 
76 /*
77  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78  * dirty_background_ratio * the amount of dirtyable memory
79  */
80 static unsigned long dirty_background_bytes;
81 
82 /*
83  * free highmem will not be subtracted from the total free memory
84  * for calculating free ratios if vm_highmem_is_dirtyable is true
85  */
86 static int vm_highmem_is_dirtyable;
87 
88 /*
89  * The generator of dirty data starts writeback at this percentage
90  */
91 static int vm_dirty_ratio = 20;
92 
93 /*
94  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95  * vm_dirty_ratio * the amount of dirtyable memory
96  */
97 static unsigned long vm_dirty_bytes;
98 
99 /*
100  * The interval between `kupdate'-style writebacks
101  */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103 
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105 
106 /*
107  * The longest time for which data is allowed to remain dirty
108  */
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110 
111 /*
112  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113  * a full sync is triggered after this time elapses without any disk activity.
114  */
115 int laptop_mode;
116 
117 EXPORT_SYMBOL(laptop_mode);
118 
119 /* End of sysctl-exported parameters */
120 
121 struct wb_domain global_wb_domain;
122 
123 /* consolidated parameters for balance_dirty_pages() and its subroutines */
124 struct dirty_throttle_control {
125 #ifdef CONFIG_CGROUP_WRITEBACK
126 	struct wb_domain	*dom;
127 	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
128 #endif
129 	struct bdi_writeback	*wb;
130 	struct fprop_local_percpu *wb_completions;
131 
132 	unsigned long		avail;		/* dirtyable */
133 	unsigned long		dirty;		/* file_dirty + write + nfs */
134 	unsigned long		thresh;		/* dirty threshold */
135 	unsigned long		bg_thresh;	/* dirty background threshold */
136 
137 	unsigned long		wb_dirty;	/* per-wb counterparts */
138 	unsigned long		wb_thresh;
139 	unsigned long		wb_bg_thresh;
140 
141 	unsigned long		pos_ratio;
142 };
143 
144 /*
145  * Length of period for aging writeout fractions of bdis. This is an
146  * arbitrarily chosen number. The longer the period, the slower fractions will
147  * reflect changes in current writeout rate.
148  */
149 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
150 
151 #ifdef CONFIG_CGROUP_WRITEBACK
152 
153 #define GDTC_INIT(__wb)		.wb = (__wb),				\
154 				.dom = &global_wb_domain,		\
155 				.wb_completions = &(__wb)->completions
156 
157 #define GDTC_INIT_NO_WB		.dom = &global_wb_domain
158 
159 #define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
160 				.dom = mem_cgroup_wb_domain(__wb),	\
161 				.wb_completions = &(__wb)->memcg_completions, \
162 				.gdtc = __gdtc
163 
mdtc_valid(struct dirty_throttle_control * dtc)164 static bool mdtc_valid(struct dirty_throttle_control *dtc)
165 {
166 	return dtc->dom;
167 }
168 
dtc_dom(struct dirty_throttle_control * dtc)169 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
170 {
171 	return dtc->dom;
172 }
173 
mdtc_gdtc(struct dirty_throttle_control * mdtc)174 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
175 {
176 	return mdtc->gdtc;
177 }
178 
wb_memcg_completions(struct bdi_writeback * wb)179 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
180 {
181 	return &wb->memcg_completions;
182 }
183 
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)184 static void wb_min_max_ratio(struct bdi_writeback *wb,
185 			     unsigned long *minp, unsigned long *maxp)
186 {
187 	unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
188 	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
189 	unsigned long long min = wb->bdi->min_ratio;
190 	unsigned long long max = wb->bdi->max_ratio;
191 
192 	/*
193 	 * @wb may already be clean by the time control reaches here and
194 	 * the total may not include its bw.
195 	 */
196 	if (this_bw < tot_bw) {
197 		if (min) {
198 			min *= this_bw;
199 			min = div64_ul(min, tot_bw);
200 		}
201 		if (max < 100 * BDI_RATIO_SCALE) {
202 			max *= this_bw;
203 			max = div64_ul(max, tot_bw);
204 		}
205 	}
206 
207 	*minp = min;
208 	*maxp = max;
209 }
210 
211 #else	/* CONFIG_CGROUP_WRITEBACK */
212 
213 #define GDTC_INIT(__wb)		.wb = (__wb),                           \
214 				.wb_completions = &(__wb)->completions
215 #define GDTC_INIT_NO_WB
216 #define MDTC_INIT(__wb, __gdtc)
217 
mdtc_valid(struct dirty_throttle_control * dtc)218 static bool mdtc_valid(struct dirty_throttle_control *dtc)
219 {
220 	return false;
221 }
222 
dtc_dom(struct dirty_throttle_control * dtc)223 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
224 {
225 	return &global_wb_domain;
226 }
227 
mdtc_gdtc(struct dirty_throttle_control * mdtc)228 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
229 {
230 	return NULL;
231 }
232 
wb_memcg_completions(struct bdi_writeback * wb)233 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
234 {
235 	return NULL;
236 }
237 
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)238 static void wb_min_max_ratio(struct bdi_writeback *wb,
239 			     unsigned long *minp, unsigned long *maxp)
240 {
241 	*minp = wb->bdi->min_ratio;
242 	*maxp = wb->bdi->max_ratio;
243 }
244 
245 #endif	/* CONFIG_CGROUP_WRITEBACK */
246 
247 /*
248  * In a memory zone, there is a certain amount of pages we consider
249  * available for the page cache, which is essentially the number of
250  * free and reclaimable pages, minus some zone reserves to protect
251  * lowmem and the ability to uphold the zone's watermarks without
252  * requiring writeback.
253  *
254  * This number of dirtyable pages is the base value of which the
255  * user-configurable dirty ratio is the effective number of pages that
256  * are allowed to be actually dirtied.  Per individual zone, or
257  * globally by using the sum of dirtyable pages over all zones.
258  *
259  * Because the user is allowed to specify the dirty limit globally as
260  * absolute number of bytes, calculating the per-zone dirty limit can
261  * require translating the configured limit into a percentage of
262  * global dirtyable memory first.
263  */
264 
265 /**
266  * node_dirtyable_memory - number of dirtyable pages in a node
267  * @pgdat: the node
268  *
269  * Return: the node's number of pages potentially available for dirty
270  * page cache.  This is the base value for the per-node dirty limits.
271  */
node_dirtyable_memory(struct pglist_data * pgdat)272 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
273 {
274 	unsigned long nr_pages = 0;
275 	int z;
276 
277 	for (z = 0; z < MAX_NR_ZONES; z++) {
278 		struct zone *zone = pgdat->node_zones + z;
279 
280 		if (!populated_zone(zone))
281 			continue;
282 
283 		nr_pages += zone_page_state(zone, NR_FREE_PAGES);
284 	}
285 
286 	/*
287 	 * Pages reserved for the kernel should not be considered
288 	 * dirtyable, to prevent a situation where reclaim has to
289 	 * clean pages in order to balance the zones.
290 	 */
291 	nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
292 
293 	nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
294 	nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
295 
296 	return nr_pages;
297 }
298 
highmem_dirtyable_memory(unsigned long total)299 static unsigned long highmem_dirtyable_memory(unsigned long total)
300 {
301 #ifdef CONFIG_HIGHMEM
302 	int node;
303 	unsigned long x = 0;
304 	int i;
305 
306 	for_each_node_state(node, N_HIGH_MEMORY) {
307 		for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
308 			struct zone *z;
309 			unsigned long nr_pages;
310 
311 			if (!is_highmem_idx(i))
312 				continue;
313 
314 			z = &NODE_DATA(node)->node_zones[i];
315 			if (!populated_zone(z))
316 				continue;
317 
318 			nr_pages = zone_page_state(z, NR_FREE_PAGES);
319 			/* watch for underflows */
320 			nr_pages -= min(nr_pages, high_wmark_pages(z));
321 			nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
322 			nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
323 			x += nr_pages;
324 		}
325 	}
326 
327 	/*
328 	 * Make sure that the number of highmem pages is never larger
329 	 * than the number of the total dirtyable memory. This can only
330 	 * occur in very strange VM situations but we want to make sure
331 	 * that this does not occur.
332 	 */
333 	return min(x, total);
334 #else
335 	return 0;
336 #endif
337 }
338 
339 /**
340  * global_dirtyable_memory - number of globally dirtyable pages
341  *
342  * Return: the global number of pages potentially available for dirty
343  * page cache.  This is the base value for the global dirty limits.
344  */
global_dirtyable_memory(void)345 static unsigned long global_dirtyable_memory(void)
346 {
347 	unsigned long x;
348 
349 	x = global_zone_page_state(NR_FREE_PAGES);
350 	/*
351 	 * Pages reserved for the kernel should not be considered
352 	 * dirtyable, to prevent a situation where reclaim has to
353 	 * clean pages in order to balance the zones.
354 	 */
355 	x -= min(x, totalreserve_pages);
356 
357 	x += global_node_page_state(NR_INACTIVE_FILE);
358 	x += global_node_page_state(NR_ACTIVE_FILE);
359 
360 	if (!vm_highmem_is_dirtyable)
361 		x -= highmem_dirtyable_memory(x);
362 
363 	return x + 1;	/* Ensure that we never return 0 */
364 }
365 
366 /**
367  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
368  * @dtc: dirty_throttle_control of interest
369  *
370  * Calculate @dtc->thresh and ->bg_thresh considering
371  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
372  * must ensure that @dtc->avail is set before calling this function.  The
373  * dirty limits will be lifted by 1/4 for real-time tasks.
374  */
domain_dirty_limits(struct dirty_throttle_control * dtc)375 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
376 {
377 	const unsigned long available_memory = dtc->avail;
378 	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
379 	unsigned long bytes = vm_dirty_bytes;
380 	unsigned long bg_bytes = dirty_background_bytes;
381 	/* convert ratios to per-PAGE_SIZE for higher precision */
382 	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
383 	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
384 	unsigned long thresh;
385 	unsigned long bg_thresh;
386 	struct task_struct *tsk;
387 
388 	/* gdtc is !NULL iff @dtc is for memcg domain */
389 	if (gdtc) {
390 		unsigned long global_avail = gdtc->avail;
391 
392 		/*
393 		 * The byte settings can't be applied directly to memcg
394 		 * domains.  Convert them to ratios by scaling against
395 		 * globally available memory.  As the ratios are in
396 		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
397 		 * number of pages.
398 		 */
399 		if (bytes)
400 			ratio = min(DIV_ROUND_UP(bytes, global_avail),
401 				    PAGE_SIZE);
402 		if (bg_bytes)
403 			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
404 				       PAGE_SIZE);
405 		bytes = bg_bytes = 0;
406 	}
407 
408 	if (bytes)
409 		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
410 	else
411 		thresh = (ratio * available_memory) / PAGE_SIZE;
412 
413 	if (bg_bytes)
414 		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
415 	else
416 		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
417 
418 	if (bg_thresh >= thresh)
419 		bg_thresh = thresh / 2;
420 	tsk = current;
421 	if (rt_task(tsk)) {
422 		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
423 		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
424 	}
425 	dtc->thresh = thresh;
426 	dtc->bg_thresh = bg_thresh;
427 
428 	/* we should eventually report the domain in the TP */
429 	if (!gdtc)
430 		trace_global_dirty_state(bg_thresh, thresh);
431 }
432 
433 /**
434  * global_dirty_limits - background-writeback and dirty-throttling thresholds
435  * @pbackground: out parameter for bg_thresh
436  * @pdirty: out parameter for thresh
437  *
438  * Calculate bg_thresh and thresh for global_wb_domain.  See
439  * domain_dirty_limits() for details.
440  */
global_dirty_limits(unsigned long * pbackground,unsigned long * pdirty)441 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
442 {
443 	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
444 
445 	gdtc.avail = global_dirtyable_memory();
446 	domain_dirty_limits(&gdtc);
447 
448 	*pbackground = gdtc.bg_thresh;
449 	*pdirty = gdtc.thresh;
450 }
451 
452 /**
453  * node_dirty_limit - maximum number of dirty pages allowed in a node
454  * @pgdat: the node
455  *
456  * Return: the maximum number of dirty pages allowed in a node, based
457  * on the node's dirtyable memory.
458  */
node_dirty_limit(struct pglist_data * pgdat)459 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
460 {
461 	unsigned long node_memory = node_dirtyable_memory(pgdat);
462 	struct task_struct *tsk = current;
463 	unsigned long dirty;
464 
465 	if (vm_dirty_bytes)
466 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
467 			node_memory / global_dirtyable_memory();
468 	else
469 		dirty = vm_dirty_ratio * node_memory / 100;
470 
471 	if (rt_task(tsk))
472 		dirty += dirty / 4;
473 
474 	return dirty;
475 }
476 
477 /**
478  * node_dirty_ok - tells whether a node is within its dirty limits
479  * @pgdat: the node to check
480  *
481  * Return: %true when the dirty pages in @pgdat are within the node's
482  * dirty limit, %false if the limit is exceeded.
483  */
node_dirty_ok(struct pglist_data * pgdat)484 bool node_dirty_ok(struct pglist_data *pgdat)
485 {
486 	unsigned long limit = node_dirty_limit(pgdat);
487 	unsigned long nr_pages = 0;
488 
489 	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
490 	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
491 
492 	return nr_pages <= limit;
493 }
494 
495 #ifdef CONFIG_SYSCTL
dirty_background_ratio_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)496 static int dirty_background_ratio_handler(struct ctl_table *table, int write,
497 		void *buffer, size_t *lenp, loff_t *ppos)
498 {
499 	int ret;
500 
501 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
502 	if (ret == 0 && write)
503 		dirty_background_bytes = 0;
504 	return ret;
505 }
506 
dirty_background_bytes_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)507 static int dirty_background_bytes_handler(struct ctl_table *table, int write,
508 		void *buffer, size_t *lenp, loff_t *ppos)
509 {
510 	int ret;
511 
512 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
513 	if (ret == 0 && write)
514 		dirty_background_ratio = 0;
515 	return ret;
516 }
517 
dirty_ratio_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)518 static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
519 		size_t *lenp, loff_t *ppos)
520 {
521 	int old_ratio = vm_dirty_ratio;
522 	int ret;
523 
524 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
525 	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
526 		writeback_set_ratelimit();
527 		vm_dirty_bytes = 0;
528 	}
529 	return ret;
530 }
531 
dirty_bytes_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)532 static int dirty_bytes_handler(struct ctl_table *table, int write,
533 		void *buffer, size_t *lenp, loff_t *ppos)
534 {
535 	unsigned long old_bytes = vm_dirty_bytes;
536 	int ret;
537 
538 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
539 	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
540 		writeback_set_ratelimit();
541 		vm_dirty_ratio = 0;
542 	}
543 	return ret;
544 }
545 #endif
546 
wp_next_time(unsigned long cur_time)547 static unsigned long wp_next_time(unsigned long cur_time)
548 {
549 	cur_time += VM_COMPLETIONS_PERIOD_LEN;
550 	/* 0 has a special meaning... */
551 	if (!cur_time)
552 		return 1;
553 	return cur_time;
554 }
555 
wb_domain_writeout_add(struct wb_domain * dom,struct fprop_local_percpu * completions,unsigned int max_prop_frac,long nr)556 static void wb_domain_writeout_add(struct wb_domain *dom,
557 				   struct fprop_local_percpu *completions,
558 				   unsigned int max_prop_frac, long nr)
559 {
560 	__fprop_add_percpu_max(&dom->completions, completions,
561 			       max_prop_frac, nr);
562 	/* First event after period switching was turned off? */
563 	if (unlikely(!dom->period_time)) {
564 		/*
565 		 * We can race with other __bdi_writeout_inc calls here but
566 		 * it does not cause any harm since the resulting time when
567 		 * timer will fire and what is in writeout_period_time will be
568 		 * roughly the same.
569 		 */
570 		dom->period_time = wp_next_time(jiffies);
571 		mod_timer(&dom->period_timer, dom->period_time);
572 	}
573 }
574 
575 /*
576  * Increment @wb's writeout completion count and the global writeout
577  * completion count. Called from __folio_end_writeback().
578  */
__wb_writeout_add(struct bdi_writeback * wb,long nr)579 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
580 {
581 	struct wb_domain *cgdom;
582 
583 	wb_stat_mod(wb, WB_WRITTEN, nr);
584 	wb_domain_writeout_add(&global_wb_domain, &wb->completions,
585 			       wb->bdi->max_prop_frac, nr);
586 
587 	cgdom = mem_cgroup_wb_domain(wb);
588 	if (cgdom)
589 		wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
590 				       wb->bdi->max_prop_frac, nr);
591 }
592 
wb_writeout_inc(struct bdi_writeback * wb)593 void wb_writeout_inc(struct bdi_writeback *wb)
594 {
595 	unsigned long flags;
596 
597 	local_irq_save(flags);
598 	__wb_writeout_add(wb, 1);
599 	local_irq_restore(flags);
600 }
601 EXPORT_SYMBOL_GPL(wb_writeout_inc);
602 
603 /*
604  * On idle system, we can be called long after we scheduled because we use
605  * deferred timers so count with missed periods.
606  */
writeout_period(struct timer_list * t)607 static void writeout_period(struct timer_list *t)
608 {
609 	struct wb_domain *dom = from_timer(dom, t, period_timer);
610 	int miss_periods = (jiffies - dom->period_time) /
611 						 VM_COMPLETIONS_PERIOD_LEN;
612 
613 	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
614 		dom->period_time = wp_next_time(dom->period_time +
615 				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
616 		mod_timer(&dom->period_timer, dom->period_time);
617 	} else {
618 		/*
619 		 * Aging has zeroed all fractions. Stop wasting CPU on period
620 		 * updates.
621 		 */
622 		dom->period_time = 0;
623 	}
624 }
625 
wb_domain_init(struct wb_domain * dom,gfp_t gfp)626 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
627 {
628 	memset(dom, 0, sizeof(*dom));
629 
630 	spin_lock_init(&dom->lock);
631 
632 	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
633 
634 	dom->dirty_limit_tstamp = jiffies;
635 
636 	return fprop_global_init(&dom->completions, gfp);
637 }
638 
639 #ifdef CONFIG_CGROUP_WRITEBACK
wb_domain_exit(struct wb_domain * dom)640 void wb_domain_exit(struct wb_domain *dom)
641 {
642 	del_timer_sync(&dom->period_timer);
643 	fprop_global_destroy(&dom->completions);
644 }
645 #endif
646 
647 /*
648  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
649  * registered backing devices, which, for obvious reasons, can not
650  * exceed 100%.
651  */
652 static unsigned int bdi_min_ratio;
653 
bdi_check_pages_limit(unsigned long pages)654 static int bdi_check_pages_limit(unsigned long pages)
655 {
656 	unsigned long max_dirty_pages = global_dirtyable_memory();
657 
658 	if (pages > max_dirty_pages)
659 		return -EINVAL;
660 
661 	return 0;
662 }
663 
bdi_ratio_from_pages(unsigned long pages)664 static unsigned long bdi_ratio_from_pages(unsigned long pages)
665 {
666 	unsigned long background_thresh;
667 	unsigned long dirty_thresh;
668 	unsigned long ratio;
669 
670 	global_dirty_limits(&background_thresh, &dirty_thresh);
671 	ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
672 
673 	return ratio;
674 }
675 
bdi_get_bytes(unsigned int ratio)676 static u64 bdi_get_bytes(unsigned int ratio)
677 {
678 	unsigned long background_thresh;
679 	unsigned long dirty_thresh;
680 	u64 bytes;
681 
682 	global_dirty_limits(&background_thresh, &dirty_thresh);
683 	bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
684 
685 	return bytes;
686 }
687 
__bdi_set_min_ratio(struct backing_dev_info * bdi,unsigned int min_ratio)688 static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
689 {
690 	unsigned int delta;
691 	int ret = 0;
692 
693 	if (min_ratio > 100 * BDI_RATIO_SCALE)
694 		return -EINVAL;
695 	min_ratio *= BDI_RATIO_SCALE;
696 
697 	spin_lock_bh(&bdi_lock);
698 	if (min_ratio > bdi->max_ratio) {
699 		ret = -EINVAL;
700 	} else {
701 		if (min_ratio < bdi->min_ratio) {
702 			delta = bdi->min_ratio - min_ratio;
703 			bdi_min_ratio -= delta;
704 			bdi->min_ratio = min_ratio;
705 		} else {
706 			delta = min_ratio - bdi->min_ratio;
707 			if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
708 				bdi_min_ratio += delta;
709 				bdi->min_ratio = min_ratio;
710 			} else {
711 				ret = -EINVAL;
712 			}
713 		}
714 	}
715 	spin_unlock_bh(&bdi_lock);
716 
717 	return ret;
718 }
719 
__bdi_set_max_ratio(struct backing_dev_info * bdi,unsigned int max_ratio)720 static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
721 {
722 	int ret = 0;
723 
724 	if (max_ratio > 100 * BDI_RATIO_SCALE)
725 		return -EINVAL;
726 
727 	spin_lock_bh(&bdi_lock);
728 	if (bdi->min_ratio > max_ratio) {
729 		ret = -EINVAL;
730 	} else {
731 		bdi->max_ratio = max_ratio;
732 		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
733 	}
734 	spin_unlock_bh(&bdi_lock);
735 
736 	return ret;
737 }
738 
bdi_set_min_ratio_no_scale(struct backing_dev_info * bdi,unsigned int min_ratio)739 int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
740 {
741 	return __bdi_set_min_ratio(bdi, min_ratio);
742 }
743 
bdi_set_max_ratio_no_scale(struct backing_dev_info * bdi,unsigned int max_ratio)744 int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
745 {
746 	return __bdi_set_max_ratio(bdi, max_ratio);
747 }
748 
bdi_set_min_ratio(struct backing_dev_info * bdi,unsigned int min_ratio)749 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
750 {
751 	return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
752 }
753 
bdi_set_max_ratio(struct backing_dev_info * bdi,unsigned int max_ratio)754 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
755 {
756 	return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
757 }
758 EXPORT_SYMBOL(bdi_set_max_ratio);
759 
bdi_get_min_bytes(struct backing_dev_info * bdi)760 u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
761 {
762 	return bdi_get_bytes(bdi->min_ratio);
763 }
764 
bdi_set_min_bytes(struct backing_dev_info * bdi,u64 min_bytes)765 int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
766 {
767 	int ret;
768 	unsigned long pages = min_bytes >> PAGE_SHIFT;
769 	unsigned long min_ratio;
770 
771 	ret = bdi_check_pages_limit(pages);
772 	if (ret)
773 		return ret;
774 
775 	min_ratio = bdi_ratio_from_pages(pages);
776 	return __bdi_set_min_ratio(bdi, min_ratio);
777 }
778 
bdi_get_max_bytes(struct backing_dev_info * bdi)779 u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
780 {
781 	return bdi_get_bytes(bdi->max_ratio);
782 }
783 
bdi_set_max_bytes(struct backing_dev_info * bdi,u64 max_bytes)784 int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
785 {
786 	int ret;
787 	unsigned long pages = max_bytes >> PAGE_SHIFT;
788 	unsigned long max_ratio;
789 
790 	ret = bdi_check_pages_limit(pages);
791 	if (ret)
792 		return ret;
793 
794 	max_ratio = bdi_ratio_from_pages(pages);
795 	return __bdi_set_max_ratio(bdi, max_ratio);
796 }
797 
bdi_set_strict_limit(struct backing_dev_info * bdi,unsigned int strict_limit)798 int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
799 {
800 	if (strict_limit > 1)
801 		return -EINVAL;
802 
803 	spin_lock_bh(&bdi_lock);
804 	if (strict_limit)
805 		bdi->capabilities |= BDI_CAP_STRICTLIMIT;
806 	else
807 		bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
808 	spin_unlock_bh(&bdi_lock);
809 
810 	return 0;
811 }
812 
dirty_freerun_ceiling(unsigned long thresh,unsigned long bg_thresh)813 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
814 					   unsigned long bg_thresh)
815 {
816 	return (thresh + bg_thresh) / 2;
817 }
818 
hard_dirty_limit(struct wb_domain * dom,unsigned long thresh)819 static unsigned long hard_dirty_limit(struct wb_domain *dom,
820 				      unsigned long thresh)
821 {
822 	return max(thresh, dom->dirty_limit);
823 }
824 
825 /*
826  * Memory which can be further allocated to a memcg domain is capped by
827  * system-wide clean memory excluding the amount being used in the domain.
828  */
mdtc_calc_avail(struct dirty_throttle_control * mdtc,unsigned long filepages,unsigned long headroom)829 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
830 			    unsigned long filepages, unsigned long headroom)
831 {
832 	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
833 	unsigned long clean = filepages - min(filepages, mdtc->dirty);
834 	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
835 	unsigned long other_clean = global_clean - min(global_clean, clean);
836 
837 	mdtc->avail = filepages + min(headroom, other_clean);
838 }
839 
840 /**
841  * __wb_calc_thresh - @wb's share of dirty throttling threshold
842  * @dtc: dirty_throttle_context of interest
843  *
844  * Note that balance_dirty_pages() will only seriously take it as a hard limit
845  * when sleeping max_pause per page is not enough to keep the dirty pages under
846  * control. For example, when the device is completely stalled due to some error
847  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
848  * In the other normal situations, it acts more gently by throttling the tasks
849  * more (rather than completely block them) when the wb dirty pages go high.
850  *
851  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
852  * - starving fast devices
853  * - piling up dirty pages (that will take long time to sync) on slow devices
854  *
855  * The wb's share of dirty limit will be adapting to its throughput and
856  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
857  *
858  * Return: @wb's dirty limit in pages. The term "dirty" in the context of
859  * dirty balancing includes all PG_dirty and PG_writeback pages.
860  */
__wb_calc_thresh(struct dirty_throttle_control * dtc)861 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
862 {
863 	struct wb_domain *dom = dtc_dom(dtc);
864 	unsigned long thresh = dtc->thresh;
865 	u64 wb_thresh;
866 	unsigned long numerator, denominator;
867 	unsigned long wb_min_ratio, wb_max_ratio;
868 
869 	/*
870 	 * Calculate this BDI's share of the thresh ratio.
871 	 */
872 	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
873 			      &numerator, &denominator);
874 
875 	wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
876 	wb_thresh *= numerator;
877 	wb_thresh = div64_ul(wb_thresh, denominator);
878 
879 	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
880 
881 	wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
882 	if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE))
883 		wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
884 
885 	return wb_thresh;
886 }
887 
wb_calc_thresh(struct bdi_writeback * wb,unsigned long thresh)888 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
889 {
890 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
891 					       .thresh = thresh };
892 	return __wb_calc_thresh(&gdtc);
893 }
894 
895 /*
896  *                           setpoint - dirty 3
897  *        f(dirty) := 1.0 + (----------------)
898  *                           limit - setpoint
899  *
900  * it's a 3rd order polynomial that subjects to
901  *
902  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
903  * (2) f(setpoint) = 1.0 => the balance point
904  * (3) f(limit)    = 0   => the hard limit
905  * (4) df/dx      <= 0	 => negative feedback control
906  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
907  *     => fast response on large errors; small oscillation near setpoint
908  */
pos_ratio_polynom(unsigned long setpoint,unsigned long dirty,unsigned long limit)909 static long long pos_ratio_polynom(unsigned long setpoint,
910 					  unsigned long dirty,
911 					  unsigned long limit)
912 {
913 	long long pos_ratio;
914 	long x;
915 
916 	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
917 		      (limit - setpoint) | 1);
918 	pos_ratio = x;
919 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
920 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
921 	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
922 
923 	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
924 }
925 
926 /*
927  * Dirty position control.
928  *
929  * (o) global/bdi setpoints
930  *
931  * We want the dirty pages be balanced around the global/wb setpoints.
932  * When the number of dirty pages is higher/lower than the setpoint, the
933  * dirty position control ratio (and hence task dirty ratelimit) will be
934  * decreased/increased to bring the dirty pages back to the setpoint.
935  *
936  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
937  *
938  *     if (dirty < setpoint) scale up   pos_ratio
939  *     if (dirty > setpoint) scale down pos_ratio
940  *
941  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
942  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
943  *
944  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
945  *
946  * (o) global control line
947  *
948  *     ^ pos_ratio
949  *     |
950  *     |            |<===== global dirty control scope ======>|
951  * 2.0  * * * * * * *
952  *     |            .*
953  *     |            . *
954  *     |            .   *
955  *     |            .     *
956  *     |            .        *
957  *     |            .            *
958  * 1.0 ................................*
959  *     |            .                  .     *
960  *     |            .                  .          *
961  *     |            .                  .              *
962  *     |            .                  .                 *
963  *     |            .                  .                    *
964  *   0 +------------.------------------.----------------------*------------->
965  *           freerun^          setpoint^                 limit^   dirty pages
966  *
967  * (o) wb control line
968  *
969  *     ^ pos_ratio
970  *     |
971  *     |            *
972  *     |              *
973  *     |                *
974  *     |                  *
975  *     |                    * |<=========== span ============>|
976  * 1.0 .......................*
977  *     |                      . *
978  *     |                      .   *
979  *     |                      .     *
980  *     |                      .       *
981  *     |                      .         *
982  *     |                      .           *
983  *     |                      .             *
984  *     |                      .               *
985  *     |                      .                 *
986  *     |                      .                   *
987  *     |                      .                     *
988  * 1/4 ...............................................* * * * * * * * * * * *
989  *     |                      .                         .
990  *     |                      .                           .
991  *     |                      .                             .
992  *   0 +----------------------.-------------------------------.------------->
993  *                wb_setpoint^                    x_intercept^
994  *
995  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
996  * be smoothly throttled down to normal if it starts high in situations like
997  * - start writing to a slow SD card and a fast disk at the same time. The SD
998  *   card's wb_dirty may rush to many times higher than wb_setpoint.
999  * - the wb dirty thresh drops quickly due to change of JBOD workload
1000  */
wb_position_ratio(struct dirty_throttle_control * dtc)1001 static void wb_position_ratio(struct dirty_throttle_control *dtc)
1002 {
1003 	struct bdi_writeback *wb = dtc->wb;
1004 	unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
1005 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1006 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1007 	unsigned long wb_thresh = dtc->wb_thresh;
1008 	unsigned long x_intercept;
1009 	unsigned long setpoint;		/* dirty pages' target balance point */
1010 	unsigned long wb_setpoint;
1011 	unsigned long span;
1012 	long long pos_ratio;		/* for scaling up/down the rate limit */
1013 	long x;
1014 
1015 	dtc->pos_ratio = 0;
1016 
1017 	if (unlikely(dtc->dirty >= limit))
1018 		return;
1019 
1020 	/*
1021 	 * global setpoint
1022 	 *
1023 	 * See comment for pos_ratio_polynom().
1024 	 */
1025 	setpoint = (freerun + limit) / 2;
1026 	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
1027 
1028 	/*
1029 	 * The strictlimit feature is a tool preventing mistrusted filesystems
1030 	 * from growing a large number of dirty pages before throttling. For
1031 	 * such filesystems balance_dirty_pages always checks wb counters
1032 	 * against wb limits. Even if global "nr_dirty" is under "freerun".
1033 	 * This is especially important for fuse which sets bdi->max_ratio to
1034 	 * 1% by default. Without strictlimit feature, fuse writeback may
1035 	 * consume arbitrary amount of RAM because it is accounted in
1036 	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1037 	 *
1038 	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1039 	 * two values: wb_dirty and wb_thresh. Let's consider an example:
1040 	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1041 	 * limits are set by default to 10% and 20% (background and throttle).
1042 	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1043 	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1044 	 * about ~6K pages (as the average of background and throttle wb
1045 	 * limits). The 3rd order polynomial will provide positive feedback if
1046 	 * wb_dirty is under wb_setpoint and vice versa.
1047 	 *
1048 	 * Note, that we cannot use global counters in these calculations
1049 	 * because we want to throttle process writing to a strictlimit wb
1050 	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1051 	 * in the example above).
1052 	 */
1053 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1054 		long long wb_pos_ratio;
1055 
1056 		if (dtc->wb_dirty < 8) {
1057 			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1058 					   2 << RATELIMIT_CALC_SHIFT);
1059 			return;
1060 		}
1061 
1062 		if (dtc->wb_dirty >= wb_thresh)
1063 			return;
1064 
1065 		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1066 						    dtc->wb_bg_thresh);
1067 
1068 		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1069 			return;
1070 
1071 		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1072 						 wb_thresh);
1073 
1074 		/*
1075 		 * Typically, for strictlimit case, wb_setpoint << setpoint
1076 		 * and pos_ratio >> wb_pos_ratio. In the other words global
1077 		 * state ("dirty") is not limiting factor and we have to
1078 		 * make decision based on wb counters. But there is an
1079 		 * important case when global pos_ratio should get precedence:
1080 		 * global limits are exceeded (e.g. due to activities on other
1081 		 * wb's) while given strictlimit wb is below limit.
1082 		 *
1083 		 * "pos_ratio * wb_pos_ratio" would work for the case above,
1084 		 * but it would look too non-natural for the case of all
1085 		 * activity in the system coming from a single strictlimit wb
1086 		 * with bdi->max_ratio == 100%.
1087 		 *
1088 		 * Note that min() below somewhat changes the dynamics of the
1089 		 * control system. Normally, pos_ratio value can be well over 3
1090 		 * (when globally we are at freerun and wb is well below wb
1091 		 * setpoint). Now the maximum pos_ratio in the same situation
1092 		 * is 2. We might want to tweak this if we observe the control
1093 		 * system is too slow to adapt.
1094 		 */
1095 		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1096 		return;
1097 	}
1098 
1099 	/*
1100 	 * We have computed basic pos_ratio above based on global situation. If
1101 	 * the wb is over/under its share of dirty pages, we want to scale
1102 	 * pos_ratio further down/up. That is done by the following mechanism.
1103 	 */
1104 
1105 	/*
1106 	 * wb setpoint
1107 	 *
1108 	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1109 	 *
1110 	 *                        x_intercept - wb_dirty
1111 	 *                     := --------------------------
1112 	 *                        x_intercept - wb_setpoint
1113 	 *
1114 	 * The main wb control line is a linear function that subjects to
1115 	 *
1116 	 * (1) f(wb_setpoint) = 1.0
1117 	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1118 	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1119 	 *
1120 	 * For single wb case, the dirty pages are observed to fluctuate
1121 	 * regularly within range
1122 	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1123 	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1124 	 * fluctuation range for pos_ratio.
1125 	 *
1126 	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1127 	 * own size, so move the slope over accordingly and choose a slope that
1128 	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1129 	 */
1130 	if (unlikely(wb_thresh > dtc->thresh))
1131 		wb_thresh = dtc->thresh;
1132 	/*
1133 	 * It's very possible that wb_thresh is close to 0 not because the
1134 	 * device is slow, but that it has remained inactive for long time.
1135 	 * Honour such devices a reasonable good (hopefully IO efficient)
1136 	 * threshold, so that the occasional writes won't be blocked and active
1137 	 * writes can rampup the threshold quickly.
1138 	 */
1139 	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1140 	/*
1141 	 * scale global setpoint to wb's:
1142 	 *	wb_setpoint = setpoint * wb_thresh / thresh
1143 	 */
1144 	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1145 	wb_setpoint = setpoint * (u64)x >> 16;
1146 	/*
1147 	 * Use span=(8*write_bw) in single wb case as indicated by
1148 	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1149 	 *
1150 	 *        wb_thresh                    thresh - wb_thresh
1151 	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1152 	 *         thresh                           thresh
1153 	 */
1154 	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1155 	x_intercept = wb_setpoint + span;
1156 
1157 	if (dtc->wb_dirty < x_intercept - span / 4) {
1158 		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1159 				      (x_intercept - wb_setpoint) | 1);
1160 	} else
1161 		pos_ratio /= 4;
1162 
1163 	/*
1164 	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1165 	 * It may push the desired control point of global dirty pages higher
1166 	 * than setpoint.
1167 	 */
1168 	x_intercept = wb_thresh / 2;
1169 	if (dtc->wb_dirty < x_intercept) {
1170 		if (dtc->wb_dirty > x_intercept / 8)
1171 			pos_ratio = div_u64(pos_ratio * x_intercept,
1172 					    dtc->wb_dirty);
1173 		else
1174 			pos_ratio *= 8;
1175 	}
1176 
1177 	dtc->pos_ratio = pos_ratio;
1178 }
1179 
wb_update_write_bandwidth(struct bdi_writeback * wb,unsigned long elapsed,unsigned long written)1180 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1181 				      unsigned long elapsed,
1182 				      unsigned long written)
1183 {
1184 	const unsigned long period = roundup_pow_of_two(3 * HZ);
1185 	unsigned long avg = wb->avg_write_bandwidth;
1186 	unsigned long old = wb->write_bandwidth;
1187 	u64 bw;
1188 
1189 	/*
1190 	 * bw = written * HZ / elapsed
1191 	 *
1192 	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1193 	 * write_bandwidth = ---------------------------------------------------
1194 	 *                                          period
1195 	 *
1196 	 * @written may have decreased due to folio_redirty_for_writepage().
1197 	 * Avoid underflowing @bw calculation.
1198 	 */
1199 	bw = written - min(written, wb->written_stamp);
1200 	bw *= HZ;
1201 	if (unlikely(elapsed > period)) {
1202 		bw = div64_ul(bw, elapsed);
1203 		avg = bw;
1204 		goto out;
1205 	}
1206 	bw += (u64)wb->write_bandwidth * (period - elapsed);
1207 	bw >>= ilog2(period);
1208 
1209 	/*
1210 	 * one more level of smoothing, for filtering out sudden spikes
1211 	 */
1212 	if (avg > old && old >= (unsigned long)bw)
1213 		avg -= (avg - old) >> 3;
1214 
1215 	if (avg < old && old <= (unsigned long)bw)
1216 		avg += (old - avg) >> 3;
1217 
1218 out:
1219 	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1220 	avg = max(avg, 1LU);
1221 	if (wb_has_dirty_io(wb)) {
1222 		long delta = avg - wb->avg_write_bandwidth;
1223 		WARN_ON_ONCE(atomic_long_add_return(delta,
1224 					&wb->bdi->tot_write_bandwidth) <= 0);
1225 	}
1226 	wb->write_bandwidth = bw;
1227 	WRITE_ONCE(wb->avg_write_bandwidth, avg);
1228 }
1229 
update_dirty_limit(struct dirty_throttle_control * dtc)1230 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1231 {
1232 	struct wb_domain *dom = dtc_dom(dtc);
1233 	unsigned long thresh = dtc->thresh;
1234 	unsigned long limit = dom->dirty_limit;
1235 
1236 	/*
1237 	 * Follow up in one step.
1238 	 */
1239 	if (limit < thresh) {
1240 		limit = thresh;
1241 		goto update;
1242 	}
1243 
1244 	/*
1245 	 * Follow down slowly. Use the higher one as the target, because thresh
1246 	 * may drop below dirty. This is exactly the reason to introduce
1247 	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1248 	 */
1249 	thresh = max(thresh, dtc->dirty);
1250 	if (limit > thresh) {
1251 		limit -= (limit - thresh) >> 5;
1252 		goto update;
1253 	}
1254 	return;
1255 update:
1256 	dom->dirty_limit = limit;
1257 }
1258 
domain_update_dirty_limit(struct dirty_throttle_control * dtc,unsigned long now)1259 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1260 				      unsigned long now)
1261 {
1262 	struct wb_domain *dom = dtc_dom(dtc);
1263 
1264 	/*
1265 	 * check locklessly first to optimize away locking for the most time
1266 	 */
1267 	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1268 		return;
1269 
1270 	spin_lock(&dom->lock);
1271 	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1272 		update_dirty_limit(dtc);
1273 		dom->dirty_limit_tstamp = now;
1274 	}
1275 	spin_unlock(&dom->lock);
1276 }
1277 
1278 /*
1279  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1280  *
1281  * Normal wb tasks will be curbed at or below it in long term.
1282  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1283  */
wb_update_dirty_ratelimit(struct dirty_throttle_control * dtc,unsigned long dirtied,unsigned long elapsed)1284 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1285 				      unsigned long dirtied,
1286 				      unsigned long elapsed)
1287 {
1288 	struct bdi_writeback *wb = dtc->wb;
1289 	unsigned long dirty = dtc->dirty;
1290 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1291 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1292 	unsigned long setpoint = (freerun + limit) / 2;
1293 	unsigned long write_bw = wb->avg_write_bandwidth;
1294 	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1295 	unsigned long dirty_rate;
1296 	unsigned long task_ratelimit;
1297 	unsigned long balanced_dirty_ratelimit;
1298 	unsigned long step;
1299 	unsigned long x;
1300 	unsigned long shift;
1301 
1302 	/*
1303 	 * The dirty rate will match the writeout rate in long term, except
1304 	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1305 	 */
1306 	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1307 
1308 	/*
1309 	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1310 	 */
1311 	task_ratelimit = (u64)dirty_ratelimit *
1312 					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1313 	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1314 
1315 	/*
1316 	 * A linear estimation of the "balanced" throttle rate. The theory is,
1317 	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1318 	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1319 	 * formula will yield the balanced rate limit (write_bw / N).
1320 	 *
1321 	 * Note that the expanded form is not a pure rate feedback:
1322 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1323 	 * but also takes pos_ratio into account:
1324 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1325 	 *
1326 	 * (1) is not realistic because pos_ratio also takes part in balancing
1327 	 * the dirty rate.  Consider the state
1328 	 *	pos_ratio = 0.5						     (3)
1329 	 *	rate = 2 * (write_bw / N)				     (4)
1330 	 * If (1) is used, it will stuck in that state! Because each dd will
1331 	 * be throttled at
1332 	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1333 	 * yielding
1334 	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1335 	 * put (6) into (1) we get
1336 	 *	rate_(i+1) = rate_(i)					     (7)
1337 	 *
1338 	 * So we end up using (2) to always keep
1339 	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1340 	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1341 	 * pos_ratio is able to drive itself to 1.0, which is not only where
1342 	 * the dirty count meet the setpoint, but also where the slope of
1343 	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1344 	 */
1345 	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1346 					   dirty_rate | 1);
1347 	/*
1348 	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1349 	 */
1350 	if (unlikely(balanced_dirty_ratelimit > write_bw))
1351 		balanced_dirty_ratelimit = write_bw;
1352 
1353 	/*
1354 	 * We could safely do this and return immediately:
1355 	 *
1356 	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1357 	 *
1358 	 * However to get a more stable dirty_ratelimit, the below elaborated
1359 	 * code makes use of task_ratelimit to filter out singular points and
1360 	 * limit the step size.
1361 	 *
1362 	 * The below code essentially only uses the relative value of
1363 	 *
1364 	 *	task_ratelimit - dirty_ratelimit
1365 	 *	= (pos_ratio - 1) * dirty_ratelimit
1366 	 *
1367 	 * which reflects the direction and size of dirty position error.
1368 	 */
1369 
1370 	/*
1371 	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1372 	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1373 	 * For example, when
1374 	 * - dirty_ratelimit > balanced_dirty_ratelimit
1375 	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1376 	 * lowering dirty_ratelimit will help meet both the position and rate
1377 	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1378 	 * only help meet the rate target. After all, what the users ultimately
1379 	 * feel and care are stable dirty rate and small position error.
1380 	 *
1381 	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1382 	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1383 	 * keeps jumping around randomly and can even leap far away at times
1384 	 * due to the small 200ms estimation period of dirty_rate (we want to
1385 	 * keep that period small to reduce time lags).
1386 	 */
1387 	step = 0;
1388 
1389 	/*
1390 	 * For strictlimit case, calculations above were based on wb counters
1391 	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1392 	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1393 	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1394 	 * "dirty" and wb_setpoint as "setpoint".
1395 	 *
1396 	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1397 	 * it's possible that wb_thresh is close to zero due to inactivity
1398 	 * of backing device.
1399 	 */
1400 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1401 		dirty = dtc->wb_dirty;
1402 		if (dtc->wb_dirty < 8)
1403 			setpoint = dtc->wb_dirty + 1;
1404 		else
1405 			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1406 	}
1407 
1408 	if (dirty < setpoint) {
1409 		x = min3(wb->balanced_dirty_ratelimit,
1410 			 balanced_dirty_ratelimit, task_ratelimit);
1411 		if (dirty_ratelimit < x)
1412 			step = x - dirty_ratelimit;
1413 	} else {
1414 		x = max3(wb->balanced_dirty_ratelimit,
1415 			 balanced_dirty_ratelimit, task_ratelimit);
1416 		if (dirty_ratelimit > x)
1417 			step = dirty_ratelimit - x;
1418 	}
1419 
1420 	/*
1421 	 * Don't pursue 100% rate matching. It's impossible since the balanced
1422 	 * rate itself is constantly fluctuating. So decrease the track speed
1423 	 * when it gets close to the target. Helps eliminate pointless tremors.
1424 	 */
1425 	shift = dirty_ratelimit / (2 * step + 1);
1426 	if (shift < BITS_PER_LONG)
1427 		step = DIV_ROUND_UP(step >> shift, 8);
1428 	else
1429 		step = 0;
1430 
1431 	if (dirty_ratelimit < balanced_dirty_ratelimit)
1432 		dirty_ratelimit += step;
1433 	else
1434 		dirty_ratelimit -= step;
1435 
1436 	WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1437 	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1438 
1439 	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1440 }
1441 
__wb_update_bandwidth(struct dirty_throttle_control * gdtc,struct dirty_throttle_control * mdtc,bool update_ratelimit)1442 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1443 				  struct dirty_throttle_control *mdtc,
1444 				  bool update_ratelimit)
1445 {
1446 	struct bdi_writeback *wb = gdtc->wb;
1447 	unsigned long now = jiffies;
1448 	unsigned long elapsed;
1449 	unsigned long dirtied;
1450 	unsigned long written;
1451 
1452 	spin_lock(&wb->list_lock);
1453 
1454 	/*
1455 	 * Lockless checks for elapsed time are racy and delayed update after
1456 	 * IO completion doesn't do it at all (to make sure written pages are
1457 	 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1458 	 * division errors.
1459 	 */
1460 	elapsed = max(now - wb->bw_time_stamp, 1UL);
1461 	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1462 	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1463 
1464 	if (update_ratelimit) {
1465 		domain_update_dirty_limit(gdtc, now);
1466 		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1467 
1468 		/*
1469 		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1470 		 * compiler has no way to figure that out.  Help it.
1471 		 */
1472 		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1473 			domain_update_dirty_limit(mdtc, now);
1474 			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1475 		}
1476 	}
1477 	wb_update_write_bandwidth(wb, elapsed, written);
1478 
1479 	wb->dirtied_stamp = dirtied;
1480 	wb->written_stamp = written;
1481 	WRITE_ONCE(wb->bw_time_stamp, now);
1482 	spin_unlock(&wb->list_lock);
1483 }
1484 
wb_update_bandwidth(struct bdi_writeback * wb)1485 void wb_update_bandwidth(struct bdi_writeback *wb)
1486 {
1487 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1488 
1489 	__wb_update_bandwidth(&gdtc, NULL, false);
1490 }
1491 
1492 /* Interval after which we consider wb idle and don't estimate bandwidth */
1493 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1494 
wb_bandwidth_estimate_start(struct bdi_writeback * wb)1495 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1496 {
1497 	unsigned long now = jiffies;
1498 	unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1499 
1500 	if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1501 	    !atomic_read(&wb->writeback_inodes)) {
1502 		spin_lock(&wb->list_lock);
1503 		wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1504 		wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1505 		WRITE_ONCE(wb->bw_time_stamp, now);
1506 		spin_unlock(&wb->list_lock);
1507 	}
1508 }
1509 
1510 /*
1511  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1512  * will look to see if it needs to start dirty throttling.
1513  *
1514  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1515  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1516  * (the number of pages we may dirty without exceeding the dirty limits).
1517  */
dirty_poll_interval(unsigned long dirty,unsigned long thresh)1518 static unsigned long dirty_poll_interval(unsigned long dirty,
1519 					 unsigned long thresh)
1520 {
1521 	if (thresh > dirty)
1522 		return 1UL << (ilog2(thresh - dirty) >> 1);
1523 
1524 	return 1;
1525 }
1526 
wb_max_pause(struct bdi_writeback * wb,unsigned long wb_dirty)1527 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1528 				  unsigned long wb_dirty)
1529 {
1530 	unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1531 	unsigned long t;
1532 
1533 	/*
1534 	 * Limit pause time for small memory systems. If sleeping for too long
1535 	 * time, a small pool of dirty/writeback pages may go empty and disk go
1536 	 * idle.
1537 	 *
1538 	 * 8 serves as the safety ratio.
1539 	 */
1540 	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1541 	t++;
1542 
1543 	return min_t(unsigned long, t, MAX_PAUSE);
1544 }
1545 
wb_min_pause(struct bdi_writeback * wb,long max_pause,unsigned long task_ratelimit,unsigned long dirty_ratelimit,int * nr_dirtied_pause)1546 static long wb_min_pause(struct bdi_writeback *wb,
1547 			 long max_pause,
1548 			 unsigned long task_ratelimit,
1549 			 unsigned long dirty_ratelimit,
1550 			 int *nr_dirtied_pause)
1551 {
1552 	long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1553 	long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1554 	long t;		/* target pause */
1555 	long pause;	/* estimated next pause */
1556 	int pages;	/* target nr_dirtied_pause */
1557 
1558 	/* target for 10ms pause on 1-dd case */
1559 	t = max(1, HZ / 100);
1560 
1561 	/*
1562 	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1563 	 * overheads.
1564 	 *
1565 	 * (N * 10ms) on 2^N concurrent tasks.
1566 	 */
1567 	if (hi > lo)
1568 		t += (hi - lo) * (10 * HZ) / 1024;
1569 
1570 	/*
1571 	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1572 	 * on the much more stable dirty_ratelimit. However the next pause time
1573 	 * will be computed based on task_ratelimit and the two rate limits may
1574 	 * depart considerably at some time. Especially if task_ratelimit goes
1575 	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1576 	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1577 	 * result task_ratelimit won't be executed faithfully, which could
1578 	 * eventually bring down dirty_ratelimit.
1579 	 *
1580 	 * We apply two rules to fix it up:
1581 	 * 1) try to estimate the next pause time and if necessary, use a lower
1582 	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1583 	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1584 	 * 2) limit the target pause time to max_pause/2, so that the normal
1585 	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1586 	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1587 	 */
1588 	t = min(t, 1 + max_pause / 2);
1589 	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1590 
1591 	/*
1592 	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1593 	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1594 	 * When the 16 consecutive reads are often interrupted by some dirty
1595 	 * throttling pause during the async writes, cfq will go into idles
1596 	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1597 	 * until reaches DIRTY_POLL_THRESH=32 pages.
1598 	 */
1599 	if (pages < DIRTY_POLL_THRESH) {
1600 		t = max_pause;
1601 		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1602 		if (pages > DIRTY_POLL_THRESH) {
1603 			pages = DIRTY_POLL_THRESH;
1604 			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1605 		}
1606 	}
1607 
1608 	pause = HZ * pages / (task_ratelimit + 1);
1609 	if (pause > max_pause) {
1610 		t = max_pause;
1611 		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1612 	}
1613 
1614 	*nr_dirtied_pause = pages;
1615 	/*
1616 	 * The minimal pause time will normally be half the target pause time.
1617 	 */
1618 	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1619 }
1620 
wb_dirty_limits(struct dirty_throttle_control * dtc)1621 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1622 {
1623 	struct bdi_writeback *wb = dtc->wb;
1624 	unsigned long wb_reclaimable;
1625 
1626 	/*
1627 	 * wb_thresh is not treated as some limiting factor as
1628 	 * dirty_thresh, due to reasons
1629 	 * - in JBOD setup, wb_thresh can fluctuate a lot
1630 	 * - in a system with HDD and USB key, the USB key may somehow
1631 	 *   go into state (wb_dirty >> wb_thresh) either because
1632 	 *   wb_dirty starts high, or because wb_thresh drops low.
1633 	 *   In this case we don't want to hard throttle the USB key
1634 	 *   dirtiers for 100 seconds until wb_dirty drops under
1635 	 *   wb_thresh. Instead the auxiliary wb control line in
1636 	 *   wb_position_ratio() will let the dirtier task progress
1637 	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1638 	 */
1639 	dtc->wb_thresh = __wb_calc_thresh(dtc);
1640 	dtc->wb_bg_thresh = dtc->thresh ?
1641 		div64_u64(dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1642 
1643 	/*
1644 	 * In order to avoid the stacked BDI deadlock we need
1645 	 * to ensure we accurately count the 'dirty' pages when
1646 	 * the threshold is low.
1647 	 *
1648 	 * Otherwise it would be possible to get thresh+n pages
1649 	 * reported dirty, even though there are thresh-m pages
1650 	 * actually dirty; with m+n sitting in the percpu
1651 	 * deltas.
1652 	 */
1653 	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1654 		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1655 		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1656 	} else {
1657 		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1658 		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1659 	}
1660 }
1661 
1662 /*
1663  * balance_dirty_pages() must be called by processes which are generating dirty
1664  * data.  It looks at the number of dirty pages in the machine and will force
1665  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1666  * If we're over `background_thresh' then the writeback threads are woken to
1667  * perform some writeout.
1668  */
balance_dirty_pages(struct bdi_writeback * wb,unsigned long pages_dirtied,unsigned int flags)1669 static int balance_dirty_pages(struct bdi_writeback *wb,
1670 			       unsigned long pages_dirtied, unsigned int flags)
1671 {
1672 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1673 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1674 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1675 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1676 						     &mdtc_stor : NULL;
1677 	struct dirty_throttle_control *sdtc;
1678 	unsigned long nr_reclaimable;	/* = file_dirty */
1679 	long period;
1680 	long pause;
1681 	long max_pause;
1682 	long min_pause;
1683 	int nr_dirtied_pause;
1684 	bool dirty_exceeded = false;
1685 	unsigned long task_ratelimit;
1686 	unsigned long dirty_ratelimit;
1687 	struct backing_dev_info *bdi = wb->bdi;
1688 	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1689 	unsigned long start_time = jiffies;
1690 	int ret = 0;
1691 
1692 	for (;;) {
1693 		unsigned long now = jiffies;
1694 		unsigned long dirty, thresh, bg_thresh;
1695 		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1696 		unsigned long m_thresh = 0;
1697 		unsigned long m_bg_thresh = 0;
1698 
1699 		nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1700 		gdtc->avail = global_dirtyable_memory();
1701 		gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1702 
1703 		domain_dirty_limits(gdtc);
1704 
1705 		if (unlikely(strictlimit)) {
1706 			wb_dirty_limits(gdtc);
1707 
1708 			dirty = gdtc->wb_dirty;
1709 			thresh = gdtc->wb_thresh;
1710 			bg_thresh = gdtc->wb_bg_thresh;
1711 		} else {
1712 			dirty = gdtc->dirty;
1713 			thresh = gdtc->thresh;
1714 			bg_thresh = gdtc->bg_thresh;
1715 		}
1716 
1717 		if (mdtc) {
1718 			unsigned long filepages, headroom, writeback;
1719 
1720 			/*
1721 			 * If @wb belongs to !root memcg, repeat the same
1722 			 * basic calculations for the memcg domain.
1723 			 */
1724 			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1725 					    &mdtc->dirty, &writeback);
1726 			mdtc->dirty += writeback;
1727 			mdtc_calc_avail(mdtc, filepages, headroom);
1728 
1729 			domain_dirty_limits(mdtc);
1730 
1731 			if (unlikely(strictlimit)) {
1732 				wb_dirty_limits(mdtc);
1733 				m_dirty = mdtc->wb_dirty;
1734 				m_thresh = mdtc->wb_thresh;
1735 				m_bg_thresh = mdtc->wb_bg_thresh;
1736 			} else {
1737 				m_dirty = mdtc->dirty;
1738 				m_thresh = mdtc->thresh;
1739 				m_bg_thresh = mdtc->bg_thresh;
1740 			}
1741 		}
1742 
1743 		/*
1744 		 * In laptop mode, we wait until hitting the higher threshold
1745 		 * before starting background writeout, and then write out all
1746 		 * the way down to the lower threshold.  So slow writers cause
1747 		 * minimal disk activity.
1748 		 *
1749 		 * In normal mode, we start background writeout at the lower
1750 		 * background_thresh, to keep the amount of dirty memory low.
1751 		 */
1752 		if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh &&
1753 		    !writeback_in_progress(wb))
1754 			wb_start_background_writeback(wb);
1755 
1756 		/*
1757 		 * Throttle it only when the background writeback cannot
1758 		 * catch-up. This avoids (excessively) small writeouts
1759 		 * when the wb limits are ramping up in case of !strictlimit.
1760 		 *
1761 		 * In strictlimit case make decision based on the wb counters
1762 		 * and limits. Small writeouts when the wb limits are ramping
1763 		 * up are the price we consciously pay for strictlimit-ing.
1764 		 *
1765 		 * If memcg domain is in effect, @dirty should be under
1766 		 * both global and memcg freerun ceilings.
1767 		 */
1768 		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1769 		    (!mdtc ||
1770 		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1771 			unsigned long intv;
1772 			unsigned long m_intv;
1773 
1774 free_running:
1775 			intv = dirty_poll_interval(dirty, thresh);
1776 			m_intv = ULONG_MAX;
1777 
1778 			current->dirty_paused_when = now;
1779 			current->nr_dirtied = 0;
1780 			if (mdtc)
1781 				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1782 			current->nr_dirtied_pause = min(intv, m_intv);
1783 			break;
1784 		}
1785 
1786 		/* Start writeback even when in laptop mode */
1787 		if (unlikely(!writeback_in_progress(wb)))
1788 			wb_start_background_writeback(wb);
1789 
1790 		mem_cgroup_flush_foreign(wb);
1791 
1792 		/*
1793 		 * Calculate global domain's pos_ratio and select the
1794 		 * global dtc by default.
1795 		 */
1796 		if (!strictlimit) {
1797 			wb_dirty_limits(gdtc);
1798 
1799 			if ((current->flags & PF_LOCAL_THROTTLE) &&
1800 			    gdtc->wb_dirty <
1801 			    dirty_freerun_ceiling(gdtc->wb_thresh,
1802 						  gdtc->wb_bg_thresh))
1803 				/*
1804 				 * LOCAL_THROTTLE tasks must not be throttled
1805 				 * when below the per-wb freerun ceiling.
1806 				 */
1807 				goto free_running;
1808 		}
1809 
1810 		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1811 			((gdtc->dirty > gdtc->thresh) || strictlimit);
1812 
1813 		wb_position_ratio(gdtc);
1814 		sdtc = gdtc;
1815 
1816 		if (mdtc) {
1817 			/*
1818 			 * If memcg domain is in effect, calculate its
1819 			 * pos_ratio.  @wb should satisfy constraints from
1820 			 * both global and memcg domains.  Choose the one
1821 			 * w/ lower pos_ratio.
1822 			 */
1823 			if (!strictlimit) {
1824 				wb_dirty_limits(mdtc);
1825 
1826 				if ((current->flags & PF_LOCAL_THROTTLE) &&
1827 				    mdtc->wb_dirty <
1828 				    dirty_freerun_ceiling(mdtc->wb_thresh,
1829 							  mdtc->wb_bg_thresh))
1830 					/*
1831 					 * LOCAL_THROTTLE tasks must not be
1832 					 * throttled when below the per-wb
1833 					 * freerun ceiling.
1834 					 */
1835 					goto free_running;
1836 			}
1837 			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1838 				((mdtc->dirty > mdtc->thresh) || strictlimit);
1839 
1840 			wb_position_ratio(mdtc);
1841 			if (mdtc->pos_ratio < gdtc->pos_ratio)
1842 				sdtc = mdtc;
1843 		}
1844 
1845 		if (dirty_exceeded != wb->dirty_exceeded)
1846 			wb->dirty_exceeded = dirty_exceeded;
1847 
1848 		if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1849 					   BANDWIDTH_INTERVAL))
1850 			__wb_update_bandwidth(gdtc, mdtc, true);
1851 
1852 		/* throttle according to the chosen dtc */
1853 		dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1854 		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1855 							RATELIMIT_CALC_SHIFT;
1856 		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1857 		min_pause = wb_min_pause(wb, max_pause,
1858 					 task_ratelimit, dirty_ratelimit,
1859 					 &nr_dirtied_pause);
1860 
1861 		if (unlikely(task_ratelimit == 0)) {
1862 			period = max_pause;
1863 			pause = max_pause;
1864 			goto pause;
1865 		}
1866 		period = HZ * pages_dirtied / task_ratelimit;
1867 		pause = period;
1868 		if (current->dirty_paused_when)
1869 			pause -= now - current->dirty_paused_when;
1870 		/*
1871 		 * For less than 1s think time (ext3/4 may block the dirtier
1872 		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1873 		 * however at much less frequency), try to compensate it in
1874 		 * future periods by updating the virtual time; otherwise just
1875 		 * do a reset, as it may be a light dirtier.
1876 		 */
1877 		if (pause < min_pause) {
1878 			trace_balance_dirty_pages(wb,
1879 						  sdtc->thresh,
1880 						  sdtc->bg_thresh,
1881 						  sdtc->dirty,
1882 						  sdtc->wb_thresh,
1883 						  sdtc->wb_dirty,
1884 						  dirty_ratelimit,
1885 						  task_ratelimit,
1886 						  pages_dirtied,
1887 						  period,
1888 						  min(pause, 0L),
1889 						  start_time);
1890 			if (pause < -HZ) {
1891 				current->dirty_paused_when = now;
1892 				current->nr_dirtied = 0;
1893 			} else if (period) {
1894 				current->dirty_paused_when += period;
1895 				current->nr_dirtied = 0;
1896 			} else if (current->nr_dirtied_pause <= pages_dirtied)
1897 				current->nr_dirtied_pause += pages_dirtied;
1898 			break;
1899 		}
1900 		if (unlikely(pause > max_pause)) {
1901 			/* for occasional dropped task_ratelimit */
1902 			now += min(pause - max_pause, max_pause);
1903 			pause = max_pause;
1904 		}
1905 
1906 pause:
1907 		trace_balance_dirty_pages(wb,
1908 					  sdtc->thresh,
1909 					  sdtc->bg_thresh,
1910 					  sdtc->dirty,
1911 					  sdtc->wb_thresh,
1912 					  sdtc->wb_dirty,
1913 					  dirty_ratelimit,
1914 					  task_ratelimit,
1915 					  pages_dirtied,
1916 					  period,
1917 					  pause,
1918 					  start_time);
1919 		if (flags & BDP_ASYNC) {
1920 			ret = -EAGAIN;
1921 			break;
1922 		}
1923 		__set_current_state(TASK_KILLABLE);
1924 		bdi->last_bdp_sleep = jiffies;
1925 		io_schedule_timeout(pause);
1926 
1927 		current->dirty_paused_when = now + pause;
1928 		current->nr_dirtied = 0;
1929 		current->nr_dirtied_pause = nr_dirtied_pause;
1930 
1931 		/*
1932 		 * This is typically equal to (dirty < thresh) and can also
1933 		 * keep "1000+ dd on a slow USB stick" under control.
1934 		 */
1935 		if (task_ratelimit)
1936 			break;
1937 
1938 		/*
1939 		 * In the case of an unresponsive NFS server and the NFS dirty
1940 		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1941 		 * to go through, so that tasks on them still remain responsive.
1942 		 *
1943 		 * In theory 1 page is enough to keep the consumer-producer
1944 		 * pipe going: the flusher cleans 1 page => the task dirties 1
1945 		 * more page. However wb_dirty has accounting errors.  So use
1946 		 * the larger and more IO friendly wb_stat_error.
1947 		 */
1948 		if (sdtc->wb_dirty <= wb_stat_error())
1949 			break;
1950 
1951 		if (fatal_signal_pending(current))
1952 			break;
1953 	}
1954 	return ret;
1955 }
1956 
1957 static DEFINE_PER_CPU(int, bdp_ratelimits);
1958 
1959 /*
1960  * Normal tasks are throttled by
1961  *	loop {
1962  *		dirty tsk->nr_dirtied_pause pages;
1963  *		take a snap in balance_dirty_pages();
1964  *	}
1965  * However there is a worst case. If every task exit immediately when dirtied
1966  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1967  * called to throttle the page dirties. The solution is to save the not yet
1968  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1969  * randomly into the running tasks. This works well for the above worst case,
1970  * as the new task will pick up and accumulate the old task's leaked dirty
1971  * count and eventually get throttled.
1972  */
1973 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1974 
1975 /**
1976  * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1977  * @mapping: address_space which was dirtied.
1978  * @flags: BDP flags.
1979  *
1980  * Processes which are dirtying memory should call in here once for each page
1981  * which was newly dirtied.  The function will periodically check the system's
1982  * dirty state and will initiate writeback if needed.
1983  *
1984  * See balance_dirty_pages_ratelimited() for details.
1985  *
1986  * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
1987  * indicate that memory is out of balance and the caller must wait
1988  * for I/O to complete.  Otherwise, it will return 0 to indicate
1989  * that either memory was already in balance, or it was able to sleep
1990  * until the amount of dirty memory returned to balance.
1991  */
balance_dirty_pages_ratelimited_flags(struct address_space * mapping,unsigned int flags)1992 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
1993 					unsigned int flags)
1994 {
1995 	struct inode *inode = mapping->host;
1996 	struct backing_dev_info *bdi = inode_to_bdi(inode);
1997 	struct bdi_writeback *wb = NULL;
1998 	int ratelimit;
1999 	int ret = 0;
2000 	int *p;
2001 
2002 	if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
2003 		return ret;
2004 
2005 	if (inode_cgwb_enabled(inode))
2006 		wb = wb_get_create_current(bdi, GFP_KERNEL);
2007 	if (!wb)
2008 		wb = &bdi->wb;
2009 
2010 	ratelimit = current->nr_dirtied_pause;
2011 	if (wb->dirty_exceeded)
2012 		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
2013 
2014 	preempt_disable();
2015 	/*
2016 	 * This prevents one CPU to accumulate too many dirtied pages without
2017 	 * calling into balance_dirty_pages(), which can happen when there are
2018 	 * 1000+ tasks, all of them start dirtying pages at exactly the same
2019 	 * time, hence all honoured too large initial task->nr_dirtied_pause.
2020 	 */
2021 	p =  this_cpu_ptr(&bdp_ratelimits);
2022 	if (unlikely(current->nr_dirtied >= ratelimit))
2023 		*p = 0;
2024 	else if (unlikely(*p >= ratelimit_pages)) {
2025 		*p = 0;
2026 		ratelimit = 0;
2027 	}
2028 	/*
2029 	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2030 	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2031 	 * the dirty throttling and livelock other long-run dirtiers.
2032 	 */
2033 	p = this_cpu_ptr(&dirty_throttle_leaks);
2034 	if (*p > 0 && current->nr_dirtied < ratelimit) {
2035 		unsigned long nr_pages_dirtied;
2036 		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2037 		*p -= nr_pages_dirtied;
2038 		current->nr_dirtied += nr_pages_dirtied;
2039 	}
2040 	preempt_enable();
2041 
2042 	if (unlikely(current->nr_dirtied >= ratelimit))
2043 		ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2044 
2045 	wb_put(wb);
2046 	return ret;
2047 }
2048 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2049 
2050 /**
2051  * balance_dirty_pages_ratelimited - balance dirty memory state.
2052  * @mapping: address_space which was dirtied.
2053  *
2054  * Processes which are dirtying memory should call in here once for each page
2055  * which was newly dirtied.  The function will periodically check the system's
2056  * dirty state and will initiate writeback if needed.
2057  *
2058  * Once we're over the dirty memory limit we decrease the ratelimiting
2059  * by a lot, to prevent individual processes from overshooting the limit
2060  * by (ratelimit_pages) each.
2061  */
balance_dirty_pages_ratelimited(struct address_space * mapping)2062 void balance_dirty_pages_ratelimited(struct address_space *mapping)
2063 {
2064 	balance_dirty_pages_ratelimited_flags(mapping, 0);
2065 }
2066 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
2067 
2068 /**
2069  * wb_over_bg_thresh - does @wb need to be written back?
2070  * @wb: bdi_writeback of interest
2071  *
2072  * Determines whether background writeback should keep writing @wb or it's
2073  * clean enough.
2074  *
2075  * Return: %true if writeback should continue.
2076  */
wb_over_bg_thresh(struct bdi_writeback * wb)2077 bool wb_over_bg_thresh(struct bdi_writeback *wb)
2078 {
2079 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
2080 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
2081 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
2082 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
2083 						     &mdtc_stor : NULL;
2084 	unsigned long reclaimable;
2085 	unsigned long thresh;
2086 
2087 	/*
2088 	 * Similar to balance_dirty_pages() but ignores pages being written
2089 	 * as we're trying to decide whether to put more under writeback.
2090 	 */
2091 	gdtc->avail = global_dirtyable_memory();
2092 	gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
2093 	domain_dirty_limits(gdtc);
2094 
2095 	if (gdtc->dirty > gdtc->bg_thresh)
2096 		return true;
2097 
2098 	thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
2099 	if (thresh < 2 * wb_stat_error())
2100 		reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2101 	else
2102 		reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2103 
2104 	if (reclaimable > thresh)
2105 		return true;
2106 
2107 	if (mdtc) {
2108 		unsigned long filepages, headroom, writeback;
2109 
2110 		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
2111 				    &writeback);
2112 		mdtc_calc_avail(mdtc, filepages, headroom);
2113 		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
2114 
2115 		if (mdtc->dirty > mdtc->bg_thresh)
2116 			return true;
2117 
2118 		thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
2119 		if (thresh < 2 * wb_stat_error())
2120 			reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2121 		else
2122 			reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2123 
2124 		if (reclaimable > thresh)
2125 			return true;
2126 	}
2127 
2128 	return false;
2129 }
2130 
2131 #ifdef CONFIG_SYSCTL
2132 /*
2133  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2134  */
dirty_writeback_centisecs_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)2135 static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2136 		void *buffer, size_t *length, loff_t *ppos)
2137 {
2138 	unsigned int old_interval = dirty_writeback_interval;
2139 	int ret;
2140 
2141 	ret = proc_dointvec(table, write, buffer, length, ppos);
2142 
2143 	/*
2144 	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2145 	 * and a different non-zero value will wakeup the writeback threads.
2146 	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2147 	 * iterate over all bdis and wbs.
2148 	 * The reason we do this is to make the change take effect immediately.
2149 	 */
2150 	if (!ret && write && dirty_writeback_interval &&
2151 		dirty_writeback_interval != old_interval)
2152 		wakeup_flusher_threads(WB_REASON_PERIODIC);
2153 
2154 	return ret;
2155 }
2156 #endif
2157 
laptop_mode_timer_fn(struct timer_list * t)2158 void laptop_mode_timer_fn(struct timer_list *t)
2159 {
2160 	struct backing_dev_info *backing_dev_info =
2161 		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2162 
2163 	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2164 }
2165 
2166 /*
2167  * We've spun up the disk and we're in laptop mode: schedule writeback
2168  * of all dirty data a few seconds from now.  If the flush is already scheduled
2169  * then push it back - the user is still using the disk.
2170  */
laptop_io_completion(struct backing_dev_info * info)2171 void laptop_io_completion(struct backing_dev_info *info)
2172 {
2173 	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2174 }
2175 
2176 /*
2177  * We're in laptop mode and we've just synced. The sync's writes will have
2178  * caused another writeback to be scheduled by laptop_io_completion.
2179  * Nothing needs to be written back anymore, so we unschedule the writeback.
2180  */
laptop_sync_completion(void)2181 void laptop_sync_completion(void)
2182 {
2183 	struct backing_dev_info *bdi;
2184 
2185 	rcu_read_lock();
2186 
2187 	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2188 		del_timer(&bdi->laptop_mode_wb_timer);
2189 
2190 	rcu_read_unlock();
2191 }
2192 
2193 /*
2194  * If ratelimit_pages is too high then we can get into dirty-data overload
2195  * if a large number of processes all perform writes at the same time.
2196  *
2197  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2198  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2199  * thresholds.
2200  */
2201 
writeback_set_ratelimit(void)2202 void writeback_set_ratelimit(void)
2203 {
2204 	struct wb_domain *dom = &global_wb_domain;
2205 	unsigned long background_thresh;
2206 	unsigned long dirty_thresh;
2207 
2208 	global_dirty_limits(&background_thresh, &dirty_thresh);
2209 	dom->dirty_limit = dirty_thresh;
2210 	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2211 	if (ratelimit_pages < 16)
2212 		ratelimit_pages = 16;
2213 }
2214 
page_writeback_cpu_online(unsigned int cpu)2215 static int page_writeback_cpu_online(unsigned int cpu)
2216 {
2217 	writeback_set_ratelimit();
2218 	return 0;
2219 }
2220 
2221 #ifdef CONFIG_SYSCTL
2222 
2223 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2224 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2225 
2226 static struct ctl_table vm_page_writeback_sysctls[] = {
2227 	{
2228 		.procname   = "dirty_background_ratio",
2229 		.data       = &dirty_background_ratio,
2230 		.maxlen     = sizeof(dirty_background_ratio),
2231 		.mode       = 0644,
2232 		.proc_handler   = dirty_background_ratio_handler,
2233 		.extra1     = SYSCTL_ZERO,
2234 		.extra2     = SYSCTL_ONE_HUNDRED,
2235 	},
2236 	{
2237 		.procname   = "dirty_background_bytes",
2238 		.data       = &dirty_background_bytes,
2239 		.maxlen     = sizeof(dirty_background_bytes),
2240 		.mode       = 0644,
2241 		.proc_handler   = dirty_background_bytes_handler,
2242 		.extra1     = SYSCTL_LONG_ONE,
2243 	},
2244 	{
2245 		.procname   = "dirty_ratio",
2246 		.data       = &vm_dirty_ratio,
2247 		.maxlen     = sizeof(vm_dirty_ratio),
2248 		.mode       = 0644,
2249 		.proc_handler   = dirty_ratio_handler,
2250 		.extra1     = SYSCTL_ZERO,
2251 		.extra2     = SYSCTL_ONE_HUNDRED,
2252 	},
2253 	{
2254 		.procname   = "dirty_bytes",
2255 		.data       = &vm_dirty_bytes,
2256 		.maxlen     = sizeof(vm_dirty_bytes),
2257 		.mode       = 0644,
2258 		.proc_handler   = dirty_bytes_handler,
2259 		.extra1     = (void *)&dirty_bytes_min,
2260 	},
2261 	{
2262 		.procname   = "dirty_writeback_centisecs",
2263 		.data       = &dirty_writeback_interval,
2264 		.maxlen     = sizeof(dirty_writeback_interval),
2265 		.mode       = 0644,
2266 		.proc_handler   = dirty_writeback_centisecs_handler,
2267 	},
2268 	{
2269 		.procname   = "dirty_expire_centisecs",
2270 		.data       = &dirty_expire_interval,
2271 		.maxlen     = sizeof(dirty_expire_interval),
2272 		.mode       = 0644,
2273 		.proc_handler   = proc_dointvec_minmax,
2274 		.extra1     = SYSCTL_ZERO,
2275 	},
2276 #ifdef CONFIG_HIGHMEM
2277 	{
2278 		.procname	= "highmem_is_dirtyable",
2279 		.data		= &vm_highmem_is_dirtyable,
2280 		.maxlen		= sizeof(vm_highmem_is_dirtyable),
2281 		.mode		= 0644,
2282 		.proc_handler	= proc_dointvec_minmax,
2283 		.extra1		= SYSCTL_ZERO,
2284 		.extra2		= SYSCTL_ONE,
2285 	},
2286 #endif
2287 	{
2288 		.procname	= "laptop_mode",
2289 		.data		= &laptop_mode,
2290 		.maxlen		= sizeof(laptop_mode),
2291 		.mode		= 0644,
2292 		.proc_handler	= proc_dointvec_jiffies,
2293 	},
2294 	{}
2295 };
2296 #endif
2297 
2298 /*
2299  * Called early on to tune the page writeback dirty limits.
2300  *
2301  * We used to scale dirty pages according to how total memory
2302  * related to pages that could be allocated for buffers.
2303  *
2304  * However, that was when we used "dirty_ratio" to scale with
2305  * all memory, and we don't do that any more. "dirty_ratio"
2306  * is now applied to total non-HIGHPAGE memory, and as such we can't
2307  * get into the old insane situation any more where we had
2308  * large amounts of dirty pages compared to a small amount of
2309  * non-HIGHMEM memory.
2310  *
2311  * But we might still want to scale the dirty_ratio by how
2312  * much memory the box has..
2313  */
page_writeback_init(void)2314 void __init page_writeback_init(void)
2315 {
2316 	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2317 
2318 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2319 			  page_writeback_cpu_online, NULL);
2320 	cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2321 			  page_writeback_cpu_online);
2322 #ifdef CONFIG_SYSCTL
2323 	register_sysctl_init("vm", vm_page_writeback_sysctls);
2324 #endif
2325 }
2326 
2327 /**
2328  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2329  * @mapping: address space structure to write
2330  * @start: starting page index
2331  * @end: ending page index (inclusive)
2332  *
2333  * This function scans the page range from @start to @end (inclusive) and tags
2334  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2335  * that write_cache_pages (or whoever calls this function) will then use
2336  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2337  * used to avoid livelocking of writeback by a process steadily creating new
2338  * dirty pages in the file (thus it is important for this function to be quick
2339  * so that it can tag pages faster than a dirtying process can create them).
2340  */
tag_pages_for_writeback(struct address_space * mapping,pgoff_t start,pgoff_t end)2341 void tag_pages_for_writeback(struct address_space *mapping,
2342 			     pgoff_t start, pgoff_t end)
2343 {
2344 	XA_STATE(xas, &mapping->i_pages, start);
2345 	unsigned int tagged = 0;
2346 	void *page;
2347 
2348 	xas_lock_irq(&xas);
2349 	xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2350 		xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2351 		if (++tagged % XA_CHECK_SCHED)
2352 			continue;
2353 
2354 		xas_pause(&xas);
2355 		xas_unlock_irq(&xas);
2356 		cond_resched();
2357 		xas_lock_irq(&xas);
2358 	}
2359 	xas_unlock_irq(&xas);
2360 }
2361 EXPORT_SYMBOL(tag_pages_for_writeback);
2362 
2363 /**
2364  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2365  * @mapping: address space structure to write
2366  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2367  * @writepage: function called for each page
2368  * @data: data passed to writepage function
2369  *
2370  * If a page is already under I/O, write_cache_pages() skips it, even
2371  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2372  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2373  * and msync() need to guarantee that all the data which was dirty at the time
2374  * the call was made get new I/O started against them.  If wbc->sync_mode is
2375  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2376  * existing IO to complete.
2377  *
2378  * To avoid livelocks (when other process dirties new pages), we first tag
2379  * pages which should be written back with TOWRITE tag and only then start
2380  * writing them. For data-integrity sync we have to be careful so that we do
2381  * not miss some pages (e.g., because some other process has cleared TOWRITE
2382  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2383  * by the process clearing the DIRTY tag (and submitting the page for IO).
2384  *
2385  * To avoid deadlocks between range_cyclic writeback and callers that hold
2386  * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2387  * we do not loop back to the start of the file. Doing so causes a page
2388  * lock/page writeback access order inversion - we should only ever lock
2389  * multiple pages in ascending page->index order, and looping back to the start
2390  * of the file violates that rule and causes deadlocks.
2391  *
2392  * Return: %0 on success, negative error code otherwise
2393  */
write_cache_pages(struct address_space * mapping,struct writeback_control * wbc,writepage_t writepage,void * data)2394 int write_cache_pages(struct address_space *mapping,
2395 		      struct writeback_control *wbc, writepage_t writepage,
2396 		      void *data)
2397 {
2398 	int ret = 0;
2399 	int done = 0;
2400 	int error;
2401 	struct folio_batch fbatch;
2402 	int nr_folios;
2403 	pgoff_t index;
2404 	pgoff_t end;		/* Inclusive */
2405 	pgoff_t done_index;
2406 	int range_whole = 0;
2407 	xa_mark_t tag;
2408 
2409 	folio_batch_init(&fbatch);
2410 	if (wbc->range_cyclic) {
2411 		index = mapping->writeback_index; /* prev offset */
2412 		end = -1;
2413 	} else {
2414 		index = wbc->range_start >> PAGE_SHIFT;
2415 		end = wbc->range_end >> PAGE_SHIFT;
2416 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2417 			range_whole = 1;
2418 	}
2419 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2420 		tag_pages_for_writeback(mapping, index, end);
2421 		tag = PAGECACHE_TAG_TOWRITE;
2422 	} else {
2423 		tag = PAGECACHE_TAG_DIRTY;
2424 	}
2425 	done_index = index;
2426 	while (!done && (index <= end)) {
2427 		int i;
2428 
2429 		nr_folios = filemap_get_folios_tag(mapping, &index, end,
2430 				tag, &fbatch);
2431 
2432 		if (nr_folios == 0)
2433 			break;
2434 
2435 		for (i = 0; i < nr_folios; i++) {
2436 			struct folio *folio = fbatch.folios[i];
2437 			unsigned long nr;
2438 
2439 			done_index = folio->index;
2440 
2441 			folio_lock(folio);
2442 
2443 			/*
2444 			 * Page truncated or invalidated. We can freely skip it
2445 			 * then, even for data integrity operations: the page
2446 			 * has disappeared concurrently, so there could be no
2447 			 * real expectation of this data integrity operation
2448 			 * even if there is now a new, dirty page at the same
2449 			 * pagecache address.
2450 			 */
2451 			if (unlikely(folio->mapping != mapping)) {
2452 continue_unlock:
2453 				folio_unlock(folio);
2454 				continue;
2455 			}
2456 
2457 			if (!folio_test_dirty(folio)) {
2458 				/* someone wrote it for us */
2459 				goto continue_unlock;
2460 			}
2461 
2462 			if (folio_test_writeback(folio)) {
2463 				if (wbc->sync_mode != WB_SYNC_NONE)
2464 					folio_wait_writeback(folio);
2465 				else
2466 					goto continue_unlock;
2467 			}
2468 
2469 			BUG_ON(folio_test_writeback(folio));
2470 			if (!folio_clear_dirty_for_io(folio))
2471 				goto continue_unlock;
2472 
2473 			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2474 			error = writepage(folio, wbc, data);
2475 			nr = folio_nr_pages(folio);
2476 			if (unlikely(error)) {
2477 				/*
2478 				 * Handle errors according to the type of
2479 				 * writeback. There's no need to continue for
2480 				 * background writeback. Just push done_index
2481 				 * past this page so media errors won't choke
2482 				 * writeout for the entire file. For integrity
2483 				 * writeback, we must process the entire dirty
2484 				 * set regardless of errors because the fs may
2485 				 * still have state to clear for each page. In
2486 				 * that case we continue processing and return
2487 				 * the first error.
2488 				 */
2489 				if (error == AOP_WRITEPAGE_ACTIVATE) {
2490 					folio_unlock(folio);
2491 					error = 0;
2492 				} else if (wbc->sync_mode != WB_SYNC_ALL) {
2493 					ret = error;
2494 					done_index = folio->index + nr;
2495 					done = 1;
2496 					break;
2497 				}
2498 				if (!ret)
2499 					ret = error;
2500 			}
2501 
2502 			/*
2503 			 * We stop writing back only if we are not doing
2504 			 * integrity sync. In case of integrity sync we have to
2505 			 * keep going until we have written all the pages
2506 			 * we tagged for writeback prior to entering this loop.
2507 			 */
2508 			wbc->nr_to_write -= nr;
2509 			if (wbc->nr_to_write <= 0 &&
2510 			    wbc->sync_mode == WB_SYNC_NONE) {
2511 				done = 1;
2512 				break;
2513 			}
2514 		}
2515 		folio_batch_release(&fbatch);
2516 		cond_resched();
2517 	}
2518 
2519 	/*
2520 	 * If we hit the last page and there is more work to be done: wrap
2521 	 * back the index back to the start of the file for the next
2522 	 * time we are called.
2523 	 */
2524 	if (wbc->range_cyclic && !done)
2525 		done_index = 0;
2526 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2527 		mapping->writeback_index = done_index;
2528 
2529 	return ret;
2530 }
2531 EXPORT_SYMBOL(write_cache_pages);
2532 
writepage_cb(struct folio * folio,struct writeback_control * wbc,void * data)2533 static int writepage_cb(struct folio *folio, struct writeback_control *wbc,
2534 		void *data)
2535 {
2536 	struct address_space *mapping = data;
2537 	int ret = mapping->a_ops->writepage(&folio->page, wbc);
2538 	mapping_set_error(mapping, ret);
2539 	return ret;
2540 }
2541 
do_writepages(struct address_space * mapping,struct writeback_control * wbc)2542 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2543 {
2544 	int ret;
2545 	struct bdi_writeback *wb;
2546 
2547 	if (wbc->nr_to_write <= 0)
2548 		return 0;
2549 	wb = inode_to_wb_wbc(mapping->host, wbc);
2550 	wb_bandwidth_estimate_start(wb);
2551 	while (1) {
2552 		if (mapping->a_ops->writepages) {
2553 			ret = mapping->a_ops->writepages(mapping, wbc);
2554 		} else if (mapping->a_ops->writepage) {
2555 			struct blk_plug plug;
2556 
2557 			blk_start_plug(&plug);
2558 			ret = write_cache_pages(mapping, wbc, writepage_cb,
2559 						mapping);
2560 			blk_finish_plug(&plug);
2561 		} else {
2562 			/* deal with chardevs and other special files */
2563 			ret = 0;
2564 		}
2565 		if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
2566 			break;
2567 
2568 		/*
2569 		 * Lacking an allocation context or the locality or writeback
2570 		 * state of any of the inode's pages, throttle based on
2571 		 * writeback activity on the local node. It's as good a
2572 		 * guess as any.
2573 		 */
2574 		reclaim_throttle(NODE_DATA(numa_node_id()),
2575 			VMSCAN_THROTTLE_WRITEBACK);
2576 	}
2577 	/*
2578 	 * Usually few pages are written by now from those we've just submitted
2579 	 * but if there's constant writeback being submitted, this makes sure
2580 	 * writeback bandwidth is updated once in a while.
2581 	 */
2582 	if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2583 				   BANDWIDTH_INTERVAL))
2584 		wb_update_bandwidth(wb);
2585 	return ret;
2586 }
2587 
2588 /*
2589  * For address_spaces which do not use buffers nor write back.
2590  */
noop_dirty_folio(struct address_space * mapping,struct folio * folio)2591 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2592 {
2593 	if (!folio_test_dirty(folio))
2594 		return !folio_test_set_dirty(folio);
2595 	return false;
2596 }
2597 EXPORT_SYMBOL(noop_dirty_folio);
2598 
2599 /*
2600  * Helper function for set_page_dirty family.
2601  *
2602  * Caller must hold folio_memcg_lock().
2603  *
2604  * NOTE: This relies on being atomic wrt interrupts.
2605  */
folio_account_dirtied(struct folio * folio,struct address_space * mapping)2606 static void folio_account_dirtied(struct folio *folio,
2607 		struct address_space *mapping)
2608 {
2609 	struct inode *inode = mapping->host;
2610 
2611 	trace_writeback_dirty_folio(folio, mapping);
2612 
2613 	if (mapping_can_writeback(mapping)) {
2614 		struct bdi_writeback *wb;
2615 		long nr = folio_nr_pages(folio);
2616 
2617 		inode_attach_wb(inode, folio);
2618 		wb = inode_to_wb(inode);
2619 
2620 		__lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2621 		__zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2622 		__node_stat_mod_folio(folio, NR_DIRTIED, nr);
2623 		wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2624 		wb_stat_mod(wb, WB_DIRTIED, nr);
2625 		task_io_account_write(nr * PAGE_SIZE);
2626 		current->nr_dirtied += nr;
2627 		__this_cpu_add(bdp_ratelimits, nr);
2628 
2629 		mem_cgroup_track_foreign_dirty(folio, wb);
2630 	}
2631 }
2632 
2633 /*
2634  * Helper function for deaccounting dirty page without writeback.
2635  *
2636  * Caller must hold folio_memcg_lock().
2637  */
folio_account_cleaned(struct folio * folio,struct bdi_writeback * wb)2638 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2639 {
2640 	long nr = folio_nr_pages(folio);
2641 
2642 	lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2643 	zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2644 	wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2645 	task_io_account_cancelled_write(nr * PAGE_SIZE);
2646 }
2647 
2648 /*
2649  * Mark the folio dirty, and set it dirty in the page cache, and mark
2650  * the inode dirty.
2651  *
2652  * If warn is true, then emit a warning if the folio is not uptodate and has
2653  * not been truncated.
2654  *
2655  * The caller must hold folio_memcg_lock().  Most callers have the folio
2656  * locked.  A few have the folio blocked from truncation through other
2657  * means (eg zap_vma_pages() has it mapped and is holding the page table
2658  * lock).  This can also be called from mark_buffer_dirty(), which I
2659  * cannot prove is always protected against truncate.
2660  */
__folio_mark_dirty(struct folio * folio,struct address_space * mapping,int warn)2661 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2662 			     int warn)
2663 {
2664 	unsigned long flags;
2665 
2666 	xa_lock_irqsave(&mapping->i_pages, flags);
2667 	if (folio->mapping) {	/* Race with truncate? */
2668 		WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2669 		folio_account_dirtied(folio, mapping);
2670 		__xa_set_mark(&mapping->i_pages, folio_index(folio),
2671 				PAGECACHE_TAG_DIRTY);
2672 	}
2673 	xa_unlock_irqrestore(&mapping->i_pages, flags);
2674 }
2675 
2676 /**
2677  * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2678  * @mapping: Address space this folio belongs to.
2679  * @folio: Folio to be marked as dirty.
2680  *
2681  * Filesystems which do not use buffer heads should call this function
2682  * from their set_page_dirty address space operation.  It ignores the
2683  * contents of folio_get_private(), so if the filesystem marks individual
2684  * blocks as dirty, the filesystem should handle that itself.
2685  *
2686  * This is also sometimes used by filesystems which use buffer_heads when
2687  * a single buffer is being dirtied: we want to set the folio dirty in
2688  * that case, but not all the buffers.  This is a "bottom-up" dirtying,
2689  * whereas block_dirty_folio() is a "top-down" dirtying.
2690  *
2691  * The caller must ensure this doesn't race with truncation.  Most will
2692  * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2693  * folio mapped and the pte lock held, which also locks out truncation.
2694  */
filemap_dirty_folio(struct address_space * mapping,struct folio * folio)2695 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2696 {
2697 	folio_memcg_lock(folio);
2698 	if (folio_test_set_dirty(folio)) {
2699 		folio_memcg_unlock(folio);
2700 		return false;
2701 	}
2702 
2703 	__folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2704 	folio_memcg_unlock(folio);
2705 
2706 	if (mapping->host) {
2707 		/* !PageAnon && !swapper_space */
2708 		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2709 	}
2710 	return true;
2711 }
2712 EXPORT_SYMBOL(filemap_dirty_folio);
2713 
2714 /**
2715  * folio_redirty_for_writepage - Decline to write a dirty folio.
2716  * @wbc: The writeback control.
2717  * @folio: The folio.
2718  *
2719  * When a writepage implementation decides that it doesn't want to write
2720  * @folio for some reason, it should call this function, unlock @folio and
2721  * return 0.
2722  *
2723  * Return: True if we redirtied the folio.  False if someone else dirtied
2724  * it first.
2725  */
folio_redirty_for_writepage(struct writeback_control * wbc,struct folio * folio)2726 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2727 		struct folio *folio)
2728 {
2729 	struct address_space *mapping = folio->mapping;
2730 	long nr = folio_nr_pages(folio);
2731 	bool ret;
2732 
2733 	wbc->pages_skipped += nr;
2734 	ret = filemap_dirty_folio(mapping, folio);
2735 	if (mapping && mapping_can_writeback(mapping)) {
2736 		struct inode *inode = mapping->host;
2737 		struct bdi_writeback *wb;
2738 		struct wb_lock_cookie cookie = {};
2739 
2740 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2741 		current->nr_dirtied -= nr;
2742 		node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2743 		wb_stat_mod(wb, WB_DIRTIED, -nr);
2744 		unlocked_inode_to_wb_end(inode, &cookie);
2745 	}
2746 	return ret;
2747 }
2748 EXPORT_SYMBOL(folio_redirty_for_writepage);
2749 
2750 /**
2751  * folio_mark_dirty - Mark a folio as being modified.
2752  * @folio: The folio.
2753  *
2754  * The folio may not be truncated while this function is running.
2755  * Holding the folio lock is sufficient to prevent truncation, but some
2756  * callers cannot acquire a sleeping lock.  These callers instead hold
2757  * the page table lock for a page table which contains at least one page
2758  * in this folio.  Truncation will block on the page table lock as it
2759  * unmaps pages before removing the folio from its mapping.
2760  *
2761  * Return: True if the folio was newly dirtied, false if it was already dirty.
2762  */
folio_mark_dirty(struct folio * folio)2763 bool folio_mark_dirty(struct folio *folio)
2764 {
2765 	struct address_space *mapping = folio_mapping(folio);
2766 
2767 	if (likely(mapping)) {
2768 		/*
2769 		 * readahead/folio_deactivate could remain
2770 		 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2771 		 * About readahead, if the folio is written, the flags would be
2772 		 * reset. So no problem.
2773 		 * About folio_deactivate, if the folio is redirtied,
2774 		 * the flag will be reset. So no problem. but if the
2775 		 * folio is used by readahead it will confuse readahead
2776 		 * and make it restart the size rampup process. But it's
2777 		 * a trivial problem.
2778 		 */
2779 		if (folio_test_reclaim(folio))
2780 			folio_clear_reclaim(folio);
2781 		return mapping->a_ops->dirty_folio(mapping, folio);
2782 	}
2783 
2784 	return noop_dirty_folio(mapping, folio);
2785 }
2786 EXPORT_SYMBOL(folio_mark_dirty);
2787 
2788 /*
2789  * set_page_dirty() is racy if the caller has no reference against
2790  * page->mapping->host, and if the page is unlocked.  This is because another
2791  * CPU could truncate the page off the mapping and then free the mapping.
2792  *
2793  * Usually, the page _is_ locked, or the caller is a user-space process which
2794  * holds a reference on the inode by having an open file.
2795  *
2796  * In other cases, the page should be locked before running set_page_dirty().
2797  */
set_page_dirty_lock(struct page * page)2798 int set_page_dirty_lock(struct page *page)
2799 {
2800 	int ret;
2801 
2802 	lock_page(page);
2803 	ret = set_page_dirty(page);
2804 	unlock_page(page);
2805 	return ret;
2806 }
2807 EXPORT_SYMBOL(set_page_dirty_lock);
2808 
2809 /*
2810  * This cancels just the dirty bit on the kernel page itself, it does NOT
2811  * actually remove dirty bits on any mmap's that may be around. It also
2812  * leaves the page tagged dirty, so any sync activity will still find it on
2813  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2814  * look at the dirty bits in the VM.
2815  *
2816  * Doing this should *normally* only ever be done when a page is truncated,
2817  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2818  * this when it notices that somebody has cleaned out all the buffers on a
2819  * page without actually doing it through the VM. Can you say "ext3 is
2820  * horribly ugly"? Thought you could.
2821  */
__folio_cancel_dirty(struct folio * folio)2822 void __folio_cancel_dirty(struct folio *folio)
2823 {
2824 	struct address_space *mapping = folio_mapping(folio);
2825 
2826 	if (mapping_can_writeback(mapping)) {
2827 		struct inode *inode = mapping->host;
2828 		struct bdi_writeback *wb;
2829 		struct wb_lock_cookie cookie = {};
2830 
2831 		folio_memcg_lock(folio);
2832 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2833 
2834 		if (folio_test_clear_dirty(folio))
2835 			folio_account_cleaned(folio, wb);
2836 
2837 		unlocked_inode_to_wb_end(inode, &cookie);
2838 		folio_memcg_unlock(folio);
2839 	} else {
2840 		folio_clear_dirty(folio);
2841 	}
2842 }
2843 EXPORT_SYMBOL(__folio_cancel_dirty);
2844 
2845 /*
2846  * Clear a folio's dirty flag, while caring for dirty memory accounting.
2847  * Returns true if the folio was previously dirty.
2848  *
2849  * This is for preparing to put the folio under writeout.  We leave
2850  * the folio tagged as dirty in the xarray so that a concurrent
2851  * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2852  * The ->writepage implementation will run either folio_start_writeback()
2853  * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2854  * and xarray dirty tag back into sync.
2855  *
2856  * This incoherency between the folio's dirty flag and xarray tag is
2857  * unfortunate, but it only exists while the folio is locked.
2858  */
folio_clear_dirty_for_io(struct folio * folio)2859 bool folio_clear_dirty_for_io(struct folio *folio)
2860 {
2861 	struct address_space *mapping = folio_mapping(folio);
2862 	bool ret = false;
2863 
2864 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2865 
2866 	if (mapping && mapping_can_writeback(mapping)) {
2867 		struct inode *inode = mapping->host;
2868 		struct bdi_writeback *wb;
2869 		struct wb_lock_cookie cookie = {};
2870 
2871 		/*
2872 		 * Yes, Virginia, this is indeed insane.
2873 		 *
2874 		 * We use this sequence to make sure that
2875 		 *  (a) we account for dirty stats properly
2876 		 *  (b) we tell the low-level filesystem to
2877 		 *      mark the whole folio dirty if it was
2878 		 *      dirty in a pagetable. Only to then
2879 		 *  (c) clean the folio again and return 1 to
2880 		 *      cause the writeback.
2881 		 *
2882 		 * This way we avoid all nasty races with the
2883 		 * dirty bit in multiple places and clearing
2884 		 * them concurrently from different threads.
2885 		 *
2886 		 * Note! Normally the "folio_mark_dirty(folio)"
2887 		 * has no effect on the actual dirty bit - since
2888 		 * that will already usually be set. But we
2889 		 * need the side effects, and it can help us
2890 		 * avoid races.
2891 		 *
2892 		 * We basically use the folio "master dirty bit"
2893 		 * as a serialization point for all the different
2894 		 * threads doing their things.
2895 		 */
2896 		if (folio_mkclean(folio))
2897 			folio_mark_dirty(folio);
2898 		/*
2899 		 * We carefully synchronise fault handlers against
2900 		 * installing a dirty pte and marking the folio dirty
2901 		 * at this point.  We do this by having them hold the
2902 		 * page lock while dirtying the folio, and folios are
2903 		 * always locked coming in here, so we get the desired
2904 		 * exclusion.
2905 		 */
2906 		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2907 		if (folio_test_clear_dirty(folio)) {
2908 			long nr = folio_nr_pages(folio);
2909 			lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2910 			zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2911 			wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2912 			ret = true;
2913 		}
2914 		unlocked_inode_to_wb_end(inode, &cookie);
2915 		return ret;
2916 	}
2917 	return folio_test_clear_dirty(folio);
2918 }
2919 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2920 
wb_inode_writeback_start(struct bdi_writeback * wb)2921 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2922 {
2923 	atomic_inc(&wb->writeback_inodes);
2924 }
2925 
wb_inode_writeback_end(struct bdi_writeback * wb)2926 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2927 {
2928 	unsigned long flags;
2929 	atomic_dec(&wb->writeback_inodes);
2930 	/*
2931 	 * Make sure estimate of writeback throughput gets updated after
2932 	 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2933 	 * (which is the interval other bandwidth updates use for batching) so
2934 	 * that if multiple inodes end writeback at a similar time, they get
2935 	 * batched into one bandwidth update.
2936 	 */
2937 	spin_lock_irqsave(&wb->work_lock, flags);
2938 	if (test_bit(WB_registered, &wb->state))
2939 		queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2940 	spin_unlock_irqrestore(&wb->work_lock, flags);
2941 }
2942 
__folio_end_writeback(struct folio * folio)2943 bool __folio_end_writeback(struct folio *folio)
2944 {
2945 	long nr = folio_nr_pages(folio);
2946 	struct address_space *mapping = folio_mapping(folio);
2947 	bool ret;
2948 
2949 	folio_memcg_lock(folio);
2950 	if (mapping && mapping_use_writeback_tags(mapping)) {
2951 		struct inode *inode = mapping->host;
2952 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2953 		unsigned long flags;
2954 
2955 		xa_lock_irqsave(&mapping->i_pages, flags);
2956 		ret = folio_test_clear_writeback(folio);
2957 		if (ret) {
2958 			__xa_clear_mark(&mapping->i_pages, folio_index(folio),
2959 						PAGECACHE_TAG_WRITEBACK);
2960 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2961 				struct bdi_writeback *wb = inode_to_wb(inode);
2962 
2963 				wb_stat_mod(wb, WB_WRITEBACK, -nr);
2964 				__wb_writeout_add(wb, nr);
2965 				if (!mapping_tagged(mapping,
2966 						    PAGECACHE_TAG_WRITEBACK))
2967 					wb_inode_writeback_end(wb);
2968 			}
2969 		}
2970 
2971 		if (mapping->host && !mapping_tagged(mapping,
2972 						     PAGECACHE_TAG_WRITEBACK))
2973 			sb_clear_inode_writeback(mapping->host);
2974 
2975 		xa_unlock_irqrestore(&mapping->i_pages, flags);
2976 	} else {
2977 		ret = folio_test_clear_writeback(folio);
2978 	}
2979 	if (ret) {
2980 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2981 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2982 		node_stat_mod_folio(folio, NR_WRITTEN, nr);
2983 	}
2984 	folio_memcg_unlock(folio);
2985 	return ret;
2986 }
2987 
__folio_start_writeback(struct folio * folio,bool keep_write)2988 bool __folio_start_writeback(struct folio *folio, bool keep_write)
2989 {
2990 	long nr = folio_nr_pages(folio);
2991 	struct address_space *mapping = folio_mapping(folio);
2992 	bool ret;
2993 	int access_ret;
2994 
2995 	folio_memcg_lock(folio);
2996 	if (mapping && mapping_use_writeback_tags(mapping)) {
2997 		XA_STATE(xas, &mapping->i_pages, folio_index(folio));
2998 		struct inode *inode = mapping->host;
2999 		struct backing_dev_info *bdi = inode_to_bdi(inode);
3000 		unsigned long flags;
3001 
3002 		xas_lock_irqsave(&xas, flags);
3003 		xas_load(&xas);
3004 		ret = folio_test_set_writeback(folio);
3005 		if (!ret) {
3006 			bool on_wblist;
3007 
3008 			on_wblist = mapping_tagged(mapping,
3009 						   PAGECACHE_TAG_WRITEBACK);
3010 
3011 			xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3012 			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3013 				struct bdi_writeback *wb = inode_to_wb(inode);
3014 
3015 				wb_stat_mod(wb, WB_WRITEBACK, nr);
3016 				if (!on_wblist)
3017 					wb_inode_writeback_start(wb);
3018 			}
3019 
3020 			/*
3021 			 * We can come through here when swapping
3022 			 * anonymous folios, so we don't necessarily
3023 			 * have an inode to track for sync.
3024 			 */
3025 			if (mapping->host && !on_wblist)
3026 				sb_mark_inode_writeback(mapping->host);
3027 		}
3028 		if (!folio_test_dirty(folio))
3029 			xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3030 		if (!keep_write)
3031 			xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3032 		xas_unlock_irqrestore(&xas, flags);
3033 	} else {
3034 		ret = folio_test_set_writeback(folio);
3035 	}
3036 	if (!ret) {
3037 		lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3038 		zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3039 	}
3040 	folio_memcg_unlock(folio);
3041 	access_ret = arch_make_folio_accessible(folio);
3042 	/*
3043 	 * If writeback has been triggered on a page that cannot be made
3044 	 * accessible, it is too late to recover here.
3045 	 */
3046 	VM_BUG_ON_FOLIO(access_ret != 0, folio);
3047 
3048 	return ret;
3049 }
3050 EXPORT_SYMBOL(__folio_start_writeback);
3051 
3052 /**
3053  * folio_wait_writeback - Wait for a folio to finish writeback.
3054  * @folio: The folio to wait for.
3055  *
3056  * If the folio is currently being written back to storage, wait for the
3057  * I/O to complete.
3058  *
3059  * Context: Sleeps.  Must be called in process context and with
3060  * no spinlocks held.  Caller should hold a reference on the folio.
3061  * If the folio is not locked, writeback may start again after writeback
3062  * has finished.
3063  */
folio_wait_writeback(struct folio * folio)3064 void folio_wait_writeback(struct folio *folio)
3065 {
3066 	while (folio_test_writeback(folio)) {
3067 		trace_folio_wait_writeback(folio, folio_mapping(folio));
3068 		folio_wait_bit(folio, PG_writeback);
3069 	}
3070 }
3071 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3072 
3073 /**
3074  * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3075  * @folio: The folio to wait for.
3076  *
3077  * If the folio is currently being written back to storage, wait for the
3078  * I/O to complete or a fatal signal to arrive.
3079  *
3080  * Context: Sleeps.  Must be called in process context and with
3081  * no spinlocks held.  Caller should hold a reference on the folio.
3082  * If the folio is not locked, writeback may start again after writeback
3083  * has finished.
3084  * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3085  */
folio_wait_writeback_killable(struct folio * folio)3086 int folio_wait_writeback_killable(struct folio *folio)
3087 {
3088 	while (folio_test_writeback(folio)) {
3089 		trace_folio_wait_writeback(folio, folio_mapping(folio));
3090 		if (folio_wait_bit_killable(folio, PG_writeback))
3091 			return -EINTR;
3092 	}
3093 
3094 	return 0;
3095 }
3096 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3097 
3098 /**
3099  * folio_wait_stable() - wait for writeback to finish, if necessary.
3100  * @folio: The folio to wait on.
3101  *
3102  * This function determines if the given folio is related to a backing
3103  * device that requires folio contents to be held stable during writeback.
3104  * If so, then it will wait for any pending writeback to complete.
3105  *
3106  * Context: Sleeps.  Must be called in process context and with
3107  * no spinlocks held.  Caller should hold a reference on the folio.
3108  * If the folio is not locked, writeback may start again after writeback
3109  * has finished.
3110  */
folio_wait_stable(struct folio * folio)3111 void folio_wait_stable(struct folio *folio)
3112 {
3113 	if (mapping_stable_writes(folio_mapping(folio)))
3114 		folio_wait_writeback(folio);
3115 }
3116 EXPORT_SYMBOL_GPL(folio_wait_stable);
3117