xref: /openbmc/linux/mm/page-writeback.c (revision 9b9c2cd4)
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002	Andrew Morton
11  *		Initial version
12  */
13 
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
41 
42 #include "internal.h"
43 
44 /*
45  * Sleep at most 200ms at a time in balance_dirty_pages().
46  */
47 #define MAX_PAUSE		max(HZ/5, 1)
48 
49 /*
50  * Try to keep balance_dirty_pages() call intervals higher than this many pages
51  * by raising pause time to max_pause when falls below it.
52  */
53 #define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
54 
55 /*
56  * Estimate write bandwidth at 200ms intervals.
57  */
58 #define BANDWIDTH_INTERVAL	max(HZ/5, 1)
59 
60 #define RATELIMIT_CALC_SHIFT	10
61 
62 /*
63  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64  * will look to see if it needs to force writeback or throttling.
65  */
66 static long ratelimit_pages = 32;
67 
68 /* The following parameters are exported via /proc/sys/vm */
69 
70 /*
71  * Start background writeback (via writeback threads) at this percentage
72  */
73 int dirty_background_ratio = 10;
74 
75 /*
76  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77  * dirty_background_ratio * the amount of dirtyable memory
78  */
79 unsigned long dirty_background_bytes;
80 
81 /*
82  * free highmem will not be subtracted from the total free memory
83  * for calculating free ratios if vm_highmem_is_dirtyable is true
84  */
85 int vm_highmem_is_dirtyable;
86 
87 /*
88  * The generator of dirty data starts writeback at this percentage
89  */
90 int vm_dirty_ratio = 20;
91 
92 /*
93  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94  * vm_dirty_ratio * the amount of dirtyable memory
95  */
96 unsigned long vm_dirty_bytes;
97 
98 /*
99  * The interval between `kupdate'-style writebacks
100  */
101 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102 
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104 
105 /*
106  * The longest time for which data is allowed to remain dirty
107  */
108 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109 
110 /*
111  * Flag that makes the machine dump writes/reads and block dirtyings.
112  */
113 int block_dump;
114 
115 /*
116  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117  * a full sync is triggered after this time elapses without any disk activity.
118  */
119 int laptop_mode;
120 
121 EXPORT_SYMBOL(laptop_mode);
122 
123 /* End of sysctl-exported parameters */
124 
125 struct wb_domain global_wb_domain;
126 
127 /* consolidated parameters for balance_dirty_pages() and its subroutines */
128 struct dirty_throttle_control {
129 #ifdef CONFIG_CGROUP_WRITEBACK
130 	struct wb_domain	*dom;
131 	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
132 #endif
133 	struct bdi_writeback	*wb;
134 	struct fprop_local_percpu *wb_completions;
135 
136 	unsigned long		avail;		/* dirtyable */
137 	unsigned long		dirty;		/* file_dirty + write + nfs */
138 	unsigned long		thresh;		/* dirty threshold */
139 	unsigned long		bg_thresh;	/* dirty background threshold */
140 
141 	unsigned long		wb_dirty;	/* per-wb counterparts */
142 	unsigned long		wb_thresh;
143 	unsigned long		wb_bg_thresh;
144 
145 	unsigned long		pos_ratio;
146 };
147 
148 /*
149  * Length of period for aging writeout fractions of bdis. This is an
150  * arbitrarily chosen number. The longer the period, the slower fractions will
151  * reflect changes in current writeout rate.
152  */
153 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
154 
155 #ifdef CONFIG_CGROUP_WRITEBACK
156 
157 #define GDTC_INIT(__wb)		.wb = (__wb),				\
158 				.dom = &global_wb_domain,		\
159 				.wb_completions = &(__wb)->completions
160 
161 #define GDTC_INIT_NO_WB		.dom = &global_wb_domain
162 
163 #define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
164 				.dom = mem_cgroup_wb_domain(__wb),	\
165 				.wb_completions = &(__wb)->memcg_completions, \
166 				.gdtc = __gdtc
167 
168 static bool mdtc_valid(struct dirty_throttle_control *dtc)
169 {
170 	return dtc->dom;
171 }
172 
173 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174 {
175 	return dtc->dom;
176 }
177 
178 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179 {
180 	return mdtc->gdtc;
181 }
182 
183 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184 {
185 	return &wb->memcg_completions;
186 }
187 
188 static void wb_min_max_ratio(struct bdi_writeback *wb,
189 			     unsigned long *minp, unsigned long *maxp)
190 {
191 	unsigned long this_bw = wb->avg_write_bandwidth;
192 	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193 	unsigned long long min = wb->bdi->min_ratio;
194 	unsigned long long max = wb->bdi->max_ratio;
195 
196 	/*
197 	 * @wb may already be clean by the time control reaches here and
198 	 * the total may not include its bw.
199 	 */
200 	if (this_bw < tot_bw) {
201 		if (min) {
202 			min *= this_bw;
203 			do_div(min, tot_bw);
204 		}
205 		if (max < 100) {
206 			max *= this_bw;
207 			do_div(max, tot_bw);
208 		}
209 	}
210 
211 	*minp = min;
212 	*maxp = max;
213 }
214 
215 #else	/* CONFIG_CGROUP_WRITEBACK */
216 
217 #define GDTC_INIT(__wb)		.wb = (__wb),                           \
218 				.wb_completions = &(__wb)->completions
219 #define GDTC_INIT_NO_WB
220 #define MDTC_INIT(__wb, __gdtc)
221 
222 static bool mdtc_valid(struct dirty_throttle_control *dtc)
223 {
224 	return false;
225 }
226 
227 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228 {
229 	return &global_wb_domain;
230 }
231 
232 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233 {
234 	return NULL;
235 }
236 
237 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238 {
239 	return NULL;
240 }
241 
242 static void wb_min_max_ratio(struct bdi_writeback *wb,
243 			     unsigned long *minp, unsigned long *maxp)
244 {
245 	*minp = wb->bdi->min_ratio;
246 	*maxp = wb->bdi->max_ratio;
247 }
248 
249 #endif	/* CONFIG_CGROUP_WRITEBACK */
250 
251 /*
252  * In a memory zone, there is a certain amount of pages we consider
253  * available for the page cache, which is essentially the number of
254  * free and reclaimable pages, minus some zone reserves to protect
255  * lowmem and the ability to uphold the zone's watermarks without
256  * requiring writeback.
257  *
258  * This number of dirtyable pages is the base value of which the
259  * user-configurable dirty ratio is the effictive number of pages that
260  * are allowed to be actually dirtied.  Per individual zone, or
261  * globally by using the sum of dirtyable pages over all zones.
262  *
263  * Because the user is allowed to specify the dirty limit globally as
264  * absolute number of bytes, calculating the per-zone dirty limit can
265  * require translating the configured limit into a percentage of
266  * global dirtyable memory first.
267  */
268 
269 /**
270  * zone_dirtyable_memory - number of dirtyable pages in a zone
271  * @zone: the zone
272  *
273  * Returns the zone's number of pages potentially available for dirty
274  * page cache.  This is the base value for the per-zone dirty limits.
275  */
276 static unsigned long zone_dirtyable_memory(struct zone *zone)
277 {
278 	unsigned long nr_pages;
279 
280 	nr_pages = zone_page_state(zone, NR_FREE_PAGES);
281 	nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
282 
283 	nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
284 	nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
285 
286 	return nr_pages;
287 }
288 
289 static unsigned long highmem_dirtyable_memory(unsigned long total)
290 {
291 #ifdef CONFIG_HIGHMEM
292 	int node;
293 	unsigned long x = 0;
294 
295 	for_each_node_state(node, N_HIGH_MEMORY) {
296 		struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
297 
298 		x += zone_dirtyable_memory(z);
299 	}
300 	/*
301 	 * Unreclaimable memory (kernel memory or anonymous memory
302 	 * without swap) can bring down the dirtyable pages below
303 	 * the zone's dirty balance reserve and the above calculation
304 	 * will underflow.  However we still want to add in nodes
305 	 * which are below threshold (negative values) to get a more
306 	 * accurate calculation but make sure that the total never
307 	 * underflows.
308 	 */
309 	if ((long)x < 0)
310 		x = 0;
311 
312 	/*
313 	 * Make sure that the number of highmem pages is never larger
314 	 * than the number of the total dirtyable memory. This can only
315 	 * occur in very strange VM situations but we want to make sure
316 	 * that this does not occur.
317 	 */
318 	return min(x, total);
319 #else
320 	return 0;
321 #endif
322 }
323 
324 /**
325  * global_dirtyable_memory - number of globally dirtyable pages
326  *
327  * Returns the global number of pages potentially available for dirty
328  * page cache.  This is the base value for the global dirty limits.
329  */
330 static unsigned long global_dirtyable_memory(void)
331 {
332 	unsigned long x;
333 
334 	x = global_page_state(NR_FREE_PAGES);
335 	x -= min(x, dirty_balance_reserve);
336 
337 	x += global_page_state(NR_INACTIVE_FILE);
338 	x += global_page_state(NR_ACTIVE_FILE);
339 
340 	if (!vm_highmem_is_dirtyable)
341 		x -= highmem_dirtyable_memory(x);
342 
343 	return x + 1;	/* Ensure that we never return 0 */
344 }
345 
346 /**
347  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
348  * @dtc: dirty_throttle_control of interest
349  *
350  * Calculate @dtc->thresh and ->bg_thresh considering
351  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
352  * must ensure that @dtc->avail is set before calling this function.  The
353  * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
354  * real-time tasks.
355  */
356 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
357 {
358 	const unsigned long available_memory = dtc->avail;
359 	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
360 	unsigned long bytes = vm_dirty_bytes;
361 	unsigned long bg_bytes = dirty_background_bytes;
362 	unsigned long ratio = vm_dirty_ratio;
363 	unsigned long bg_ratio = dirty_background_ratio;
364 	unsigned long thresh;
365 	unsigned long bg_thresh;
366 	struct task_struct *tsk;
367 
368 	/* gdtc is !NULL iff @dtc is for memcg domain */
369 	if (gdtc) {
370 		unsigned long global_avail = gdtc->avail;
371 
372 		/*
373 		 * The byte settings can't be applied directly to memcg
374 		 * domains.  Convert them to ratios by scaling against
375 		 * globally available memory.
376 		 */
377 		if (bytes)
378 			ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
379 				    global_avail, 100UL);
380 		if (bg_bytes)
381 			bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
382 				       global_avail, 100UL);
383 		bytes = bg_bytes = 0;
384 	}
385 
386 	if (bytes)
387 		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
388 	else
389 		thresh = (ratio * available_memory) / 100;
390 
391 	if (bg_bytes)
392 		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
393 	else
394 		bg_thresh = (bg_ratio * available_memory) / 100;
395 
396 	if (bg_thresh >= thresh)
397 		bg_thresh = thresh / 2;
398 	tsk = current;
399 	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
400 		bg_thresh += bg_thresh / 4;
401 		thresh += thresh / 4;
402 	}
403 	dtc->thresh = thresh;
404 	dtc->bg_thresh = bg_thresh;
405 
406 	/* we should eventually report the domain in the TP */
407 	if (!gdtc)
408 		trace_global_dirty_state(bg_thresh, thresh);
409 }
410 
411 /**
412  * global_dirty_limits - background-writeback and dirty-throttling thresholds
413  * @pbackground: out parameter for bg_thresh
414  * @pdirty: out parameter for thresh
415  *
416  * Calculate bg_thresh and thresh for global_wb_domain.  See
417  * domain_dirty_limits() for details.
418  */
419 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
420 {
421 	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
422 
423 	gdtc.avail = global_dirtyable_memory();
424 	domain_dirty_limits(&gdtc);
425 
426 	*pbackground = gdtc.bg_thresh;
427 	*pdirty = gdtc.thresh;
428 }
429 
430 /**
431  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
432  * @zone: the zone
433  *
434  * Returns the maximum number of dirty pages allowed in a zone, based
435  * on the zone's dirtyable memory.
436  */
437 static unsigned long zone_dirty_limit(struct zone *zone)
438 {
439 	unsigned long zone_memory = zone_dirtyable_memory(zone);
440 	struct task_struct *tsk = current;
441 	unsigned long dirty;
442 
443 	if (vm_dirty_bytes)
444 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
445 			zone_memory / global_dirtyable_memory();
446 	else
447 		dirty = vm_dirty_ratio * zone_memory / 100;
448 
449 	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
450 		dirty += dirty / 4;
451 
452 	return dirty;
453 }
454 
455 /**
456  * zone_dirty_ok - tells whether a zone is within its dirty limits
457  * @zone: the zone to check
458  *
459  * Returns %true when the dirty pages in @zone are within the zone's
460  * dirty limit, %false if the limit is exceeded.
461  */
462 bool zone_dirty_ok(struct zone *zone)
463 {
464 	unsigned long limit = zone_dirty_limit(zone);
465 
466 	return zone_page_state(zone, NR_FILE_DIRTY) +
467 	       zone_page_state(zone, NR_UNSTABLE_NFS) +
468 	       zone_page_state(zone, NR_WRITEBACK) <= limit;
469 }
470 
471 int dirty_background_ratio_handler(struct ctl_table *table, int write,
472 		void __user *buffer, size_t *lenp,
473 		loff_t *ppos)
474 {
475 	int ret;
476 
477 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
478 	if (ret == 0 && write)
479 		dirty_background_bytes = 0;
480 	return ret;
481 }
482 
483 int dirty_background_bytes_handler(struct ctl_table *table, int write,
484 		void __user *buffer, size_t *lenp,
485 		loff_t *ppos)
486 {
487 	int ret;
488 
489 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
490 	if (ret == 0 && write)
491 		dirty_background_ratio = 0;
492 	return ret;
493 }
494 
495 int dirty_ratio_handler(struct ctl_table *table, int write,
496 		void __user *buffer, size_t *lenp,
497 		loff_t *ppos)
498 {
499 	int old_ratio = vm_dirty_ratio;
500 	int ret;
501 
502 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
503 	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
504 		writeback_set_ratelimit();
505 		vm_dirty_bytes = 0;
506 	}
507 	return ret;
508 }
509 
510 int dirty_bytes_handler(struct ctl_table *table, int write,
511 		void __user *buffer, size_t *lenp,
512 		loff_t *ppos)
513 {
514 	unsigned long old_bytes = vm_dirty_bytes;
515 	int ret;
516 
517 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
518 	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
519 		writeback_set_ratelimit();
520 		vm_dirty_ratio = 0;
521 	}
522 	return ret;
523 }
524 
525 static unsigned long wp_next_time(unsigned long cur_time)
526 {
527 	cur_time += VM_COMPLETIONS_PERIOD_LEN;
528 	/* 0 has a special meaning... */
529 	if (!cur_time)
530 		return 1;
531 	return cur_time;
532 }
533 
534 static void wb_domain_writeout_inc(struct wb_domain *dom,
535 				   struct fprop_local_percpu *completions,
536 				   unsigned int max_prop_frac)
537 {
538 	__fprop_inc_percpu_max(&dom->completions, completions,
539 			       max_prop_frac);
540 	/* First event after period switching was turned off? */
541 	if (!unlikely(dom->period_time)) {
542 		/*
543 		 * We can race with other __bdi_writeout_inc calls here but
544 		 * it does not cause any harm since the resulting time when
545 		 * timer will fire and what is in writeout_period_time will be
546 		 * roughly the same.
547 		 */
548 		dom->period_time = wp_next_time(jiffies);
549 		mod_timer(&dom->period_timer, dom->period_time);
550 	}
551 }
552 
553 /*
554  * Increment @wb's writeout completion count and the global writeout
555  * completion count. Called from test_clear_page_writeback().
556  */
557 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
558 {
559 	struct wb_domain *cgdom;
560 
561 	__inc_wb_stat(wb, WB_WRITTEN);
562 	wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
563 			       wb->bdi->max_prop_frac);
564 
565 	cgdom = mem_cgroup_wb_domain(wb);
566 	if (cgdom)
567 		wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
568 				       wb->bdi->max_prop_frac);
569 }
570 
571 void wb_writeout_inc(struct bdi_writeback *wb)
572 {
573 	unsigned long flags;
574 
575 	local_irq_save(flags);
576 	__wb_writeout_inc(wb);
577 	local_irq_restore(flags);
578 }
579 EXPORT_SYMBOL_GPL(wb_writeout_inc);
580 
581 /*
582  * On idle system, we can be called long after we scheduled because we use
583  * deferred timers so count with missed periods.
584  */
585 static void writeout_period(unsigned long t)
586 {
587 	struct wb_domain *dom = (void *)t;
588 	int miss_periods = (jiffies - dom->period_time) /
589 						 VM_COMPLETIONS_PERIOD_LEN;
590 
591 	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
592 		dom->period_time = wp_next_time(dom->period_time +
593 				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
594 		mod_timer(&dom->period_timer, dom->period_time);
595 	} else {
596 		/*
597 		 * Aging has zeroed all fractions. Stop wasting CPU on period
598 		 * updates.
599 		 */
600 		dom->period_time = 0;
601 	}
602 }
603 
604 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
605 {
606 	memset(dom, 0, sizeof(*dom));
607 
608 	spin_lock_init(&dom->lock);
609 
610 	init_timer_deferrable(&dom->period_timer);
611 	dom->period_timer.function = writeout_period;
612 	dom->period_timer.data = (unsigned long)dom;
613 
614 	dom->dirty_limit_tstamp = jiffies;
615 
616 	return fprop_global_init(&dom->completions, gfp);
617 }
618 
619 #ifdef CONFIG_CGROUP_WRITEBACK
620 void wb_domain_exit(struct wb_domain *dom)
621 {
622 	del_timer_sync(&dom->period_timer);
623 	fprop_global_destroy(&dom->completions);
624 }
625 #endif
626 
627 /*
628  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
629  * registered backing devices, which, for obvious reasons, can not
630  * exceed 100%.
631  */
632 static unsigned int bdi_min_ratio;
633 
634 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
635 {
636 	int ret = 0;
637 
638 	spin_lock_bh(&bdi_lock);
639 	if (min_ratio > bdi->max_ratio) {
640 		ret = -EINVAL;
641 	} else {
642 		min_ratio -= bdi->min_ratio;
643 		if (bdi_min_ratio + min_ratio < 100) {
644 			bdi_min_ratio += min_ratio;
645 			bdi->min_ratio += min_ratio;
646 		} else {
647 			ret = -EINVAL;
648 		}
649 	}
650 	spin_unlock_bh(&bdi_lock);
651 
652 	return ret;
653 }
654 
655 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
656 {
657 	int ret = 0;
658 
659 	if (max_ratio > 100)
660 		return -EINVAL;
661 
662 	spin_lock_bh(&bdi_lock);
663 	if (bdi->min_ratio > max_ratio) {
664 		ret = -EINVAL;
665 	} else {
666 		bdi->max_ratio = max_ratio;
667 		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
668 	}
669 	spin_unlock_bh(&bdi_lock);
670 
671 	return ret;
672 }
673 EXPORT_SYMBOL(bdi_set_max_ratio);
674 
675 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
676 					   unsigned long bg_thresh)
677 {
678 	return (thresh + bg_thresh) / 2;
679 }
680 
681 static unsigned long hard_dirty_limit(struct wb_domain *dom,
682 				      unsigned long thresh)
683 {
684 	return max(thresh, dom->dirty_limit);
685 }
686 
687 /*
688  * Memory which can be further allocated to a memcg domain is capped by
689  * system-wide clean memory excluding the amount being used in the domain.
690  */
691 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
692 			    unsigned long filepages, unsigned long headroom)
693 {
694 	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
695 	unsigned long clean = filepages - min(filepages, mdtc->dirty);
696 	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
697 	unsigned long other_clean = global_clean - min(global_clean, clean);
698 
699 	mdtc->avail = filepages + min(headroom, other_clean);
700 }
701 
702 /**
703  * __wb_calc_thresh - @wb's share of dirty throttling threshold
704  * @dtc: dirty_throttle_context of interest
705  *
706  * Returns @wb's dirty limit in pages. The term "dirty" in the context of
707  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
708  *
709  * Note that balance_dirty_pages() will only seriously take it as a hard limit
710  * when sleeping max_pause per page is not enough to keep the dirty pages under
711  * control. For example, when the device is completely stalled due to some error
712  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
713  * In the other normal situations, it acts more gently by throttling the tasks
714  * more (rather than completely block them) when the wb dirty pages go high.
715  *
716  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
717  * - starving fast devices
718  * - piling up dirty pages (that will take long time to sync) on slow devices
719  *
720  * The wb's share of dirty limit will be adapting to its throughput and
721  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
722  */
723 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
724 {
725 	struct wb_domain *dom = dtc_dom(dtc);
726 	unsigned long thresh = dtc->thresh;
727 	u64 wb_thresh;
728 	long numerator, denominator;
729 	unsigned long wb_min_ratio, wb_max_ratio;
730 
731 	/*
732 	 * Calculate this BDI's share of the thresh ratio.
733 	 */
734 	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
735 			      &numerator, &denominator);
736 
737 	wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
738 	wb_thresh *= numerator;
739 	do_div(wb_thresh, denominator);
740 
741 	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
742 
743 	wb_thresh += (thresh * wb_min_ratio) / 100;
744 	if (wb_thresh > (thresh * wb_max_ratio) / 100)
745 		wb_thresh = thresh * wb_max_ratio / 100;
746 
747 	return wb_thresh;
748 }
749 
750 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
751 {
752 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
753 					       .thresh = thresh };
754 	return __wb_calc_thresh(&gdtc);
755 }
756 
757 /*
758  *                           setpoint - dirty 3
759  *        f(dirty) := 1.0 + (----------------)
760  *                           limit - setpoint
761  *
762  * it's a 3rd order polynomial that subjects to
763  *
764  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
765  * (2) f(setpoint) = 1.0 => the balance point
766  * (3) f(limit)    = 0   => the hard limit
767  * (4) df/dx      <= 0	 => negative feedback control
768  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
769  *     => fast response on large errors; small oscillation near setpoint
770  */
771 static long long pos_ratio_polynom(unsigned long setpoint,
772 					  unsigned long dirty,
773 					  unsigned long limit)
774 {
775 	long long pos_ratio;
776 	long x;
777 
778 	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
779 		      (limit - setpoint) | 1);
780 	pos_ratio = x;
781 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
782 	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
783 	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
784 
785 	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
786 }
787 
788 /*
789  * Dirty position control.
790  *
791  * (o) global/bdi setpoints
792  *
793  * We want the dirty pages be balanced around the global/wb setpoints.
794  * When the number of dirty pages is higher/lower than the setpoint, the
795  * dirty position control ratio (and hence task dirty ratelimit) will be
796  * decreased/increased to bring the dirty pages back to the setpoint.
797  *
798  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
799  *
800  *     if (dirty < setpoint) scale up   pos_ratio
801  *     if (dirty > setpoint) scale down pos_ratio
802  *
803  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
804  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
805  *
806  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
807  *
808  * (o) global control line
809  *
810  *     ^ pos_ratio
811  *     |
812  *     |            |<===== global dirty control scope ======>|
813  * 2.0 .............*
814  *     |            .*
815  *     |            . *
816  *     |            .   *
817  *     |            .     *
818  *     |            .        *
819  *     |            .            *
820  * 1.0 ................................*
821  *     |            .                  .     *
822  *     |            .                  .          *
823  *     |            .                  .              *
824  *     |            .                  .                 *
825  *     |            .                  .                    *
826  *   0 +------------.------------------.----------------------*------------->
827  *           freerun^          setpoint^                 limit^   dirty pages
828  *
829  * (o) wb control line
830  *
831  *     ^ pos_ratio
832  *     |
833  *     |            *
834  *     |              *
835  *     |                *
836  *     |                  *
837  *     |                    * |<=========== span ============>|
838  * 1.0 .......................*
839  *     |                      . *
840  *     |                      .   *
841  *     |                      .     *
842  *     |                      .       *
843  *     |                      .         *
844  *     |                      .           *
845  *     |                      .             *
846  *     |                      .               *
847  *     |                      .                 *
848  *     |                      .                   *
849  *     |                      .                     *
850  * 1/4 ...............................................* * * * * * * * * * * *
851  *     |                      .                         .
852  *     |                      .                           .
853  *     |                      .                             .
854  *   0 +----------------------.-------------------------------.------------->
855  *                wb_setpoint^                    x_intercept^
856  *
857  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
858  * be smoothly throttled down to normal if it starts high in situations like
859  * - start writing to a slow SD card and a fast disk at the same time. The SD
860  *   card's wb_dirty may rush to many times higher than wb_setpoint.
861  * - the wb dirty thresh drops quickly due to change of JBOD workload
862  */
863 static void wb_position_ratio(struct dirty_throttle_control *dtc)
864 {
865 	struct bdi_writeback *wb = dtc->wb;
866 	unsigned long write_bw = wb->avg_write_bandwidth;
867 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
868 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
869 	unsigned long wb_thresh = dtc->wb_thresh;
870 	unsigned long x_intercept;
871 	unsigned long setpoint;		/* dirty pages' target balance point */
872 	unsigned long wb_setpoint;
873 	unsigned long span;
874 	long long pos_ratio;		/* for scaling up/down the rate limit */
875 	long x;
876 
877 	dtc->pos_ratio = 0;
878 
879 	if (unlikely(dtc->dirty >= limit))
880 		return;
881 
882 	/*
883 	 * global setpoint
884 	 *
885 	 * See comment for pos_ratio_polynom().
886 	 */
887 	setpoint = (freerun + limit) / 2;
888 	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
889 
890 	/*
891 	 * The strictlimit feature is a tool preventing mistrusted filesystems
892 	 * from growing a large number of dirty pages before throttling. For
893 	 * such filesystems balance_dirty_pages always checks wb counters
894 	 * against wb limits. Even if global "nr_dirty" is under "freerun".
895 	 * This is especially important for fuse which sets bdi->max_ratio to
896 	 * 1% by default. Without strictlimit feature, fuse writeback may
897 	 * consume arbitrary amount of RAM because it is accounted in
898 	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
899 	 *
900 	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
901 	 * two values: wb_dirty and wb_thresh. Let's consider an example:
902 	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
903 	 * limits are set by default to 10% and 20% (background and throttle).
904 	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
905 	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
906 	 * about ~6K pages (as the average of background and throttle wb
907 	 * limits). The 3rd order polynomial will provide positive feedback if
908 	 * wb_dirty is under wb_setpoint and vice versa.
909 	 *
910 	 * Note, that we cannot use global counters in these calculations
911 	 * because we want to throttle process writing to a strictlimit wb
912 	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
913 	 * in the example above).
914 	 */
915 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
916 		long long wb_pos_ratio;
917 
918 		if (dtc->wb_dirty < 8) {
919 			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
920 					   2 << RATELIMIT_CALC_SHIFT);
921 			return;
922 		}
923 
924 		if (dtc->wb_dirty >= wb_thresh)
925 			return;
926 
927 		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
928 						    dtc->wb_bg_thresh);
929 
930 		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
931 			return;
932 
933 		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
934 						 wb_thresh);
935 
936 		/*
937 		 * Typically, for strictlimit case, wb_setpoint << setpoint
938 		 * and pos_ratio >> wb_pos_ratio. In the other words global
939 		 * state ("dirty") is not limiting factor and we have to
940 		 * make decision based on wb counters. But there is an
941 		 * important case when global pos_ratio should get precedence:
942 		 * global limits are exceeded (e.g. due to activities on other
943 		 * wb's) while given strictlimit wb is below limit.
944 		 *
945 		 * "pos_ratio * wb_pos_ratio" would work for the case above,
946 		 * but it would look too non-natural for the case of all
947 		 * activity in the system coming from a single strictlimit wb
948 		 * with bdi->max_ratio == 100%.
949 		 *
950 		 * Note that min() below somewhat changes the dynamics of the
951 		 * control system. Normally, pos_ratio value can be well over 3
952 		 * (when globally we are at freerun and wb is well below wb
953 		 * setpoint). Now the maximum pos_ratio in the same situation
954 		 * is 2. We might want to tweak this if we observe the control
955 		 * system is too slow to adapt.
956 		 */
957 		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
958 		return;
959 	}
960 
961 	/*
962 	 * We have computed basic pos_ratio above based on global situation. If
963 	 * the wb is over/under its share of dirty pages, we want to scale
964 	 * pos_ratio further down/up. That is done by the following mechanism.
965 	 */
966 
967 	/*
968 	 * wb setpoint
969 	 *
970 	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
971 	 *
972 	 *                        x_intercept - wb_dirty
973 	 *                     := --------------------------
974 	 *                        x_intercept - wb_setpoint
975 	 *
976 	 * The main wb control line is a linear function that subjects to
977 	 *
978 	 * (1) f(wb_setpoint) = 1.0
979 	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
980 	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
981 	 *
982 	 * For single wb case, the dirty pages are observed to fluctuate
983 	 * regularly within range
984 	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
985 	 * for various filesystems, where (2) can yield in a reasonable 12.5%
986 	 * fluctuation range for pos_ratio.
987 	 *
988 	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
989 	 * own size, so move the slope over accordingly and choose a slope that
990 	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
991 	 */
992 	if (unlikely(wb_thresh > dtc->thresh))
993 		wb_thresh = dtc->thresh;
994 	/*
995 	 * It's very possible that wb_thresh is close to 0 not because the
996 	 * device is slow, but that it has remained inactive for long time.
997 	 * Honour such devices a reasonable good (hopefully IO efficient)
998 	 * threshold, so that the occasional writes won't be blocked and active
999 	 * writes can rampup the threshold quickly.
1000 	 */
1001 	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1002 	/*
1003 	 * scale global setpoint to wb's:
1004 	 *	wb_setpoint = setpoint * wb_thresh / thresh
1005 	 */
1006 	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1007 	wb_setpoint = setpoint * (u64)x >> 16;
1008 	/*
1009 	 * Use span=(8*write_bw) in single wb case as indicated by
1010 	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1011 	 *
1012 	 *        wb_thresh                    thresh - wb_thresh
1013 	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1014 	 *         thresh                           thresh
1015 	 */
1016 	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1017 	x_intercept = wb_setpoint + span;
1018 
1019 	if (dtc->wb_dirty < x_intercept - span / 4) {
1020 		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1021 				      (x_intercept - wb_setpoint) | 1);
1022 	} else
1023 		pos_ratio /= 4;
1024 
1025 	/*
1026 	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1027 	 * It may push the desired control point of global dirty pages higher
1028 	 * than setpoint.
1029 	 */
1030 	x_intercept = wb_thresh / 2;
1031 	if (dtc->wb_dirty < x_intercept) {
1032 		if (dtc->wb_dirty > x_intercept / 8)
1033 			pos_ratio = div_u64(pos_ratio * x_intercept,
1034 					    dtc->wb_dirty);
1035 		else
1036 			pos_ratio *= 8;
1037 	}
1038 
1039 	dtc->pos_ratio = pos_ratio;
1040 }
1041 
1042 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1043 				      unsigned long elapsed,
1044 				      unsigned long written)
1045 {
1046 	const unsigned long period = roundup_pow_of_two(3 * HZ);
1047 	unsigned long avg = wb->avg_write_bandwidth;
1048 	unsigned long old = wb->write_bandwidth;
1049 	u64 bw;
1050 
1051 	/*
1052 	 * bw = written * HZ / elapsed
1053 	 *
1054 	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1055 	 * write_bandwidth = ---------------------------------------------------
1056 	 *                                          period
1057 	 *
1058 	 * @written may have decreased due to account_page_redirty().
1059 	 * Avoid underflowing @bw calculation.
1060 	 */
1061 	bw = written - min(written, wb->written_stamp);
1062 	bw *= HZ;
1063 	if (unlikely(elapsed > period)) {
1064 		do_div(bw, elapsed);
1065 		avg = bw;
1066 		goto out;
1067 	}
1068 	bw += (u64)wb->write_bandwidth * (period - elapsed);
1069 	bw >>= ilog2(period);
1070 
1071 	/*
1072 	 * one more level of smoothing, for filtering out sudden spikes
1073 	 */
1074 	if (avg > old && old >= (unsigned long)bw)
1075 		avg -= (avg - old) >> 3;
1076 
1077 	if (avg < old && old <= (unsigned long)bw)
1078 		avg += (old - avg) >> 3;
1079 
1080 out:
1081 	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1082 	avg = max(avg, 1LU);
1083 	if (wb_has_dirty_io(wb)) {
1084 		long delta = avg - wb->avg_write_bandwidth;
1085 		WARN_ON_ONCE(atomic_long_add_return(delta,
1086 					&wb->bdi->tot_write_bandwidth) <= 0);
1087 	}
1088 	wb->write_bandwidth = bw;
1089 	wb->avg_write_bandwidth = avg;
1090 }
1091 
1092 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1093 {
1094 	struct wb_domain *dom = dtc_dom(dtc);
1095 	unsigned long thresh = dtc->thresh;
1096 	unsigned long limit = dom->dirty_limit;
1097 
1098 	/*
1099 	 * Follow up in one step.
1100 	 */
1101 	if (limit < thresh) {
1102 		limit = thresh;
1103 		goto update;
1104 	}
1105 
1106 	/*
1107 	 * Follow down slowly. Use the higher one as the target, because thresh
1108 	 * may drop below dirty. This is exactly the reason to introduce
1109 	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1110 	 */
1111 	thresh = max(thresh, dtc->dirty);
1112 	if (limit > thresh) {
1113 		limit -= (limit - thresh) >> 5;
1114 		goto update;
1115 	}
1116 	return;
1117 update:
1118 	dom->dirty_limit = limit;
1119 }
1120 
1121 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1122 				    unsigned long now)
1123 {
1124 	struct wb_domain *dom = dtc_dom(dtc);
1125 
1126 	/*
1127 	 * check locklessly first to optimize away locking for the most time
1128 	 */
1129 	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1130 		return;
1131 
1132 	spin_lock(&dom->lock);
1133 	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1134 		update_dirty_limit(dtc);
1135 		dom->dirty_limit_tstamp = now;
1136 	}
1137 	spin_unlock(&dom->lock);
1138 }
1139 
1140 /*
1141  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1142  *
1143  * Normal wb tasks will be curbed at or below it in long term.
1144  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1145  */
1146 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1147 				      unsigned long dirtied,
1148 				      unsigned long elapsed)
1149 {
1150 	struct bdi_writeback *wb = dtc->wb;
1151 	unsigned long dirty = dtc->dirty;
1152 	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1153 	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1154 	unsigned long setpoint = (freerun + limit) / 2;
1155 	unsigned long write_bw = wb->avg_write_bandwidth;
1156 	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1157 	unsigned long dirty_rate;
1158 	unsigned long task_ratelimit;
1159 	unsigned long balanced_dirty_ratelimit;
1160 	unsigned long step;
1161 	unsigned long x;
1162 
1163 	/*
1164 	 * The dirty rate will match the writeout rate in long term, except
1165 	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1166 	 */
1167 	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1168 
1169 	/*
1170 	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1171 	 */
1172 	task_ratelimit = (u64)dirty_ratelimit *
1173 					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1174 	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1175 
1176 	/*
1177 	 * A linear estimation of the "balanced" throttle rate. The theory is,
1178 	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1179 	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1180 	 * formula will yield the balanced rate limit (write_bw / N).
1181 	 *
1182 	 * Note that the expanded form is not a pure rate feedback:
1183 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1184 	 * but also takes pos_ratio into account:
1185 	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1186 	 *
1187 	 * (1) is not realistic because pos_ratio also takes part in balancing
1188 	 * the dirty rate.  Consider the state
1189 	 *	pos_ratio = 0.5						     (3)
1190 	 *	rate = 2 * (write_bw / N)				     (4)
1191 	 * If (1) is used, it will stuck in that state! Because each dd will
1192 	 * be throttled at
1193 	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1194 	 * yielding
1195 	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1196 	 * put (6) into (1) we get
1197 	 *	rate_(i+1) = rate_(i)					     (7)
1198 	 *
1199 	 * So we end up using (2) to always keep
1200 	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1201 	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1202 	 * pos_ratio is able to drive itself to 1.0, which is not only where
1203 	 * the dirty count meet the setpoint, but also where the slope of
1204 	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1205 	 */
1206 	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1207 					   dirty_rate | 1);
1208 	/*
1209 	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1210 	 */
1211 	if (unlikely(balanced_dirty_ratelimit > write_bw))
1212 		balanced_dirty_ratelimit = write_bw;
1213 
1214 	/*
1215 	 * We could safely do this and return immediately:
1216 	 *
1217 	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1218 	 *
1219 	 * However to get a more stable dirty_ratelimit, the below elaborated
1220 	 * code makes use of task_ratelimit to filter out singular points and
1221 	 * limit the step size.
1222 	 *
1223 	 * The below code essentially only uses the relative value of
1224 	 *
1225 	 *	task_ratelimit - dirty_ratelimit
1226 	 *	= (pos_ratio - 1) * dirty_ratelimit
1227 	 *
1228 	 * which reflects the direction and size of dirty position error.
1229 	 */
1230 
1231 	/*
1232 	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1233 	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1234 	 * For example, when
1235 	 * - dirty_ratelimit > balanced_dirty_ratelimit
1236 	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1237 	 * lowering dirty_ratelimit will help meet both the position and rate
1238 	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1239 	 * only help meet the rate target. After all, what the users ultimately
1240 	 * feel and care are stable dirty rate and small position error.
1241 	 *
1242 	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1243 	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1244 	 * keeps jumping around randomly and can even leap far away at times
1245 	 * due to the small 200ms estimation period of dirty_rate (we want to
1246 	 * keep that period small to reduce time lags).
1247 	 */
1248 	step = 0;
1249 
1250 	/*
1251 	 * For strictlimit case, calculations above were based on wb counters
1252 	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1253 	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1254 	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1255 	 * "dirty" and wb_setpoint as "setpoint".
1256 	 *
1257 	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1258 	 * it's possible that wb_thresh is close to zero due to inactivity
1259 	 * of backing device.
1260 	 */
1261 	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1262 		dirty = dtc->wb_dirty;
1263 		if (dtc->wb_dirty < 8)
1264 			setpoint = dtc->wb_dirty + 1;
1265 		else
1266 			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1267 	}
1268 
1269 	if (dirty < setpoint) {
1270 		x = min3(wb->balanced_dirty_ratelimit,
1271 			 balanced_dirty_ratelimit, task_ratelimit);
1272 		if (dirty_ratelimit < x)
1273 			step = x - dirty_ratelimit;
1274 	} else {
1275 		x = max3(wb->balanced_dirty_ratelimit,
1276 			 balanced_dirty_ratelimit, task_ratelimit);
1277 		if (dirty_ratelimit > x)
1278 			step = dirty_ratelimit - x;
1279 	}
1280 
1281 	/*
1282 	 * Don't pursue 100% rate matching. It's impossible since the balanced
1283 	 * rate itself is constantly fluctuating. So decrease the track speed
1284 	 * when it gets close to the target. Helps eliminate pointless tremors.
1285 	 */
1286 	step >>= dirty_ratelimit / (2 * step + 1);
1287 	/*
1288 	 * Limit the tracking speed to avoid overshooting.
1289 	 */
1290 	step = (step + 7) / 8;
1291 
1292 	if (dirty_ratelimit < balanced_dirty_ratelimit)
1293 		dirty_ratelimit += step;
1294 	else
1295 		dirty_ratelimit -= step;
1296 
1297 	wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1298 	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1299 
1300 	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1301 }
1302 
1303 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1304 				  struct dirty_throttle_control *mdtc,
1305 				  unsigned long start_time,
1306 				  bool update_ratelimit)
1307 {
1308 	struct bdi_writeback *wb = gdtc->wb;
1309 	unsigned long now = jiffies;
1310 	unsigned long elapsed = now - wb->bw_time_stamp;
1311 	unsigned long dirtied;
1312 	unsigned long written;
1313 
1314 	lockdep_assert_held(&wb->list_lock);
1315 
1316 	/*
1317 	 * rate-limit, only update once every 200ms.
1318 	 */
1319 	if (elapsed < BANDWIDTH_INTERVAL)
1320 		return;
1321 
1322 	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1323 	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1324 
1325 	/*
1326 	 * Skip quiet periods when disk bandwidth is under-utilized.
1327 	 * (at least 1s idle time between two flusher runs)
1328 	 */
1329 	if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1330 		goto snapshot;
1331 
1332 	if (update_ratelimit) {
1333 		domain_update_bandwidth(gdtc, now);
1334 		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1335 
1336 		/*
1337 		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1338 		 * compiler has no way to figure that out.  Help it.
1339 		 */
1340 		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1341 			domain_update_bandwidth(mdtc, now);
1342 			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1343 		}
1344 	}
1345 	wb_update_write_bandwidth(wb, elapsed, written);
1346 
1347 snapshot:
1348 	wb->dirtied_stamp = dirtied;
1349 	wb->written_stamp = written;
1350 	wb->bw_time_stamp = now;
1351 }
1352 
1353 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1354 {
1355 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1356 
1357 	__wb_update_bandwidth(&gdtc, NULL, start_time, false);
1358 }
1359 
1360 /*
1361  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1362  * will look to see if it needs to start dirty throttling.
1363  *
1364  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1365  * global_page_state() too often. So scale it near-sqrt to the safety margin
1366  * (the number of pages we may dirty without exceeding the dirty limits).
1367  */
1368 static unsigned long dirty_poll_interval(unsigned long dirty,
1369 					 unsigned long thresh)
1370 {
1371 	if (thresh > dirty)
1372 		return 1UL << (ilog2(thresh - dirty) >> 1);
1373 
1374 	return 1;
1375 }
1376 
1377 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1378 				  unsigned long wb_dirty)
1379 {
1380 	unsigned long bw = wb->avg_write_bandwidth;
1381 	unsigned long t;
1382 
1383 	/*
1384 	 * Limit pause time for small memory systems. If sleeping for too long
1385 	 * time, a small pool of dirty/writeback pages may go empty and disk go
1386 	 * idle.
1387 	 *
1388 	 * 8 serves as the safety ratio.
1389 	 */
1390 	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1391 	t++;
1392 
1393 	return min_t(unsigned long, t, MAX_PAUSE);
1394 }
1395 
1396 static long wb_min_pause(struct bdi_writeback *wb,
1397 			 long max_pause,
1398 			 unsigned long task_ratelimit,
1399 			 unsigned long dirty_ratelimit,
1400 			 int *nr_dirtied_pause)
1401 {
1402 	long hi = ilog2(wb->avg_write_bandwidth);
1403 	long lo = ilog2(wb->dirty_ratelimit);
1404 	long t;		/* target pause */
1405 	long pause;	/* estimated next pause */
1406 	int pages;	/* target nr_dirtied_pause */
1407 
1408 	/* target for 10ms pause on 1-dd case */
1409 	t = max(1, HZ / 100);
1410 
1411 	/*
1412 	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1413 	 * overheads.
1414 	 *
1415 	 * (N * 10ms) on 2^N concurrent tasks.
1416 	 */
1417 	if (hi > lo)
1418 		t += (hi - lo) * (10 * HZ) / 1024;
1419 
1420 	/*
1421 	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1422 	 * on the much more stable dirty_ratelimit. However the next pause time
1423 	 * will be computed based on task_ratelimit and the two rate limits may
1424 	 * depart considerably at some time. Especially if task_ratelimit goes
1425 	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1426 	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1427 	 * result task_ratelimit won't be executed faithfully, which could
1428 	 * eventually bring down dirty_ratelimit.
1429 	 *
1430 	 * We apply two rules to fix it up:
1431 	 * 1) try to estimate the next pause time and if necessary, use a lower
1432 	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1433 	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1434 	 * 2) limit the target pause time to max_pause/2, so that the normal
1435 	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1436 	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1437 	 */
1438 	t = min(t, 1 + max_pause / 2);
1439 	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1440 
1441 	/*
1442 	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1443 	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1444 	 * When the 16 consecutive reads are often interrupted by some dirty
1445 	 * throttling pause during the async writes, cfq will go into idles
1446 	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1447 	 * until reaches DIRTY_POLL_THRESH=32 pages.
1448 	 */
1449 	if (pages < DIRTY_POLL_THRESH) {
1450 		t = max_pause;
1451 		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1452 		if (pages > DIRTY_POLL_THRESH) {
1453 			pages = DIRTY_POLL_THRESH;
1454 			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1455 		}
1456 	}
1457 
1458 	pause = HZ * pages / (task_ratelimit + 1);
1459 	if (pause > max_pause) {
1460 		t = max_pause;
1461 		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1462 	}
1463 
1464 	*nr_dirtied_pause = pages;
1465 	/*
1466 	 * The minimal pause time will normally be half the target pause time.
1467 	 */
1468 	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1469 }
1470 
1471 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1472 {
1473 	struct bdi_writeback *wb = dtc->wb;
1474 	unsigned long wb_reclaimable;
1475 
1476 	/*
1477 	 * wb_thresh is not treated as some limiting factor as
1478 	 * dirty_thresh, due to reasons
1479 	 * - in JBOD setup, wb_thresh can fluctuate a lot
1480 	 * - in a system with HDD and USB key, the USB key may somehow
1481 	 *   go into state (wb_dirty >> wb_thresh) either because
1482 	 *   wb_dirty starts high, or because wb_thresh drops low.
1483 	 *   In this case we don't want to hard throttle the USB key
1484 	 *   dirtiers for 100 seconds until wb_dirty drops under
1485 	 *   wb_thresh. Instead the auxiliary wb control line in
1486 	 *   wb_position_ratio() will let the dirtier task progress
1487 	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1488 	 */
1489 	dtc->wb_thresh = __wb_calc_thresh(dtc);
1490 	dtc->wb_bg_thresh = dtc->thresh ?
1491 		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1492 
1493 	/*
1494 	 * In order to avoid the stacked BDI deadlock we need
1495 	 * to ensure we accurately count the 'dirty' pages when
1496 	 * the threshold is low.
1497 	 *
1498 	 * Otherwise it would be possible to get thresh+n pages
1499 	 * reported dirty, even though there are thresh-m pages
1500 	 * actually dirty; with m+n sitting in the percpu
1501 	 * deltas.
1502 	 */
1503 	if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1504 		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1505 		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1506 	} else {
1507 		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1508 		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1509 	}
1510 }
1511 
1512 /*
1513  * balance_dirty_pages() must be called by processes which are generating dirty
1514  * data.  It looks at the number of dirty pages in the machine and will force
1515  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1516  * If we're over `background_thresh' then the writeback threads are woken to
1517  * perform some writeout.
1518  */
1519 static void balance_dirty_pages(struct address_space *mapping,
1520 				struct bdi_writeback *wb,
1521 				unsigned long pages_dirtied)
1522 {
1523 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1524 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1525 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1526 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1527 						     &mdtc_stor : NULL;
1528 	struct dirty_throttle_control *sdtc;
1529 	unsigned long nr_reclaimable;	/* = file_dirty + unstable_nfs */
1530 	long period;
1531 	long pause;
1532 	long max_pause;
1533 	long min_pause;
1534 	int nr_dirtied_pause;
1535 	bool dirty_exceeded = false;
1536 	unsigned long task_ratelimit;
1537 	unsigned long dirty_ratelimit;
1538 	struct backing_dev_info *bdi = wb->bdi;
1539 	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1540 	unsigned long start_time = jiffies;
1541 
1542 	for (;;) {
1543 		unsigned long now = jiffies;
1544 		unsigned long dirty, thresh, bg_thresh;
1545 		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1546 		unsigned long m_thresh = 0;
1547 		unsigned long m_bg_thresh = 0;
1548 
1549 		/*
1550 		 * Unstable writes are a feature of certain networked
1551 		 * filesystems (i.e. NFS) in which data may have been
1552 		 * written to the server's write cache, but has not yet
1553 		 * been flushed to permanent storage.
1554 		 */
1555 		nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1556 					global_page_state(NR_UNSTABLE_NFS);
1557 		gdtc->avail = global_dirtyable_memory();
1558 		gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1559 
1560 		domain_dirty_limits(gdtc);
1561 
1562 		if (unlikely(strictlimit)) {
1563 			wb_dirty_limits(gdtc);
1564 
1565 			dirty = gdtc->wb_dirty;
1566 			thresh = gdtc->wb_thresh;
1567 			bg_thresh = gdtc->wb_bg_thresh;
1568 		} else {
1569 			dirty = gdtc->dirty;
1570 			thresh = gdtc->thresh;
1571 			bg_thresh = gdtc->bg_thresh;
1572 		}
1573 
1574 		if (mdtc) {
1575 			unsigned long filepages, headroom, writeback;
1576 
1577 			/*
1578 			 * If @wb belongs to !root memcg, repeat the same
1579 			 * basic calculations for the memcg domain.
1580 			 */
1581 			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1582 					    &mdtc->dirty, &writeback);
1583 			mdtc->dirty += writeback;
1584 			mdtc_calc_avail(mdtc, filepages, headroom);
1585 
1586 			domain_dirty_limits(mdtc);
1587 
1588 			if (unlikely(strictlimit)) {
1589 				wb_dirty_limits(mdtc);
1590 				m_dirty = mdtc->wb_dirty;
1591 				m_thresh = mdtc->wb_thresh;
1592 				m_bg_thresh = mdtc->wb_bg_thresh;
1593 			} else {
1594 				m_dirty = mdtc->dirty;
1595 				m_thresh = mdtc->thresh;
1596 				m_bg_thresh = mdtc->bg_thresh;
1597 			}
1598 		}
1599 
1600 		/*
1601 		 * Throttle it only when the background writeback cannot
1602 		 * catch-up. This avoids (excessively) small writeouts
1603 		 * when the wb limits are ramping up in case of !strictlimit.
1604 		 *
1605 		 * In strictlimit case make decision based on the wb counters
1606 		 * and limits. Small writeouts when the wb limits are ramping
1607 		 * up are the price we consciously pay for strictlimit-ing.
1608 		 *
1609 		 * If memcg domain is in effect, @dirty should be under
1610 		 * both global and memcg freerun ceilings.
1611 		 */
1612 		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1613 		    (!mdtc ||
1614 		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1615 			unsigned long intv = dirty_poll_interval(dirty, thresh);
1616 			unsigned long m_intv = ULONG_MAX;
1617 
1618 			current->dirty_paused_when = now;
1619 			current->nr_dirtied = 0;
1620 			if (mdtc)
1621 				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1622 			current->nr_dirtied_pause = min(intv, m_intv);
1623 			break;
1624 		}
1625 
1626 		if (unlikely(!writeback_in_progress(wb)))
1627 			wb_start_background_writeback(wb);
1628 
1629 		/*
1630 		 * Calculate global domain's pos_ratio and select the
1631 		 * global dtc by default.
1632 		 */
1633 		if (!strictlimit)
1634 			wb_dirty_limits(gdtc);
1635 
1636 		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1637 			((gdtc->dirty > gdtc->thresh) || strictlimit);
1638 
1639 		wb_position_ratio(gdtc);
1640 		sdtc = gdtc;
1641 
1642 		if (mdtc) {
1643 			/*
1644 			 * If memcg domain is in effect, calculate its
1645 			 * pos_ratio.  @wb should satisfy constraints from
1646 			 * both global and memcg domains.  Choose the one
1647 			 * w/ lower pos_ratio.
1648 			 */
1649 			if (!strictlimit)
1650 				wb_dirty_limits(mdtc);
1651 
1652 			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1653 				((mdtc->dirty > mdtc->thresh) || strictlimit);
1654 
1655 			wb_position_ratio(mdtc);
1656 			if (mdtc->pos_ratio < gdtc->pos_ratio)
1657 				sdtc = mdtc;
1658 		}
1659 
1660 		if (dirty_exceeded && !wb->dirty_exceeded)
1661 			wb->dirty_exceeded = 1;
1662 
1663 		if (time_is_before_jiffies(wb->bw_time_stamp +
1664 					   BANDWIDTH_INTERVAL)) {
1665 			spin_lock(&wb->list_lock);
1666 			__wb_update_bandwidth(gdtc, mdtc, start_time, true);
1667 			spin_unlock(&wb->list_lock);
1668 		}
1669 
1670 		/* throttle according to the chosen dtc */
1671 		dirty_ratelimit = wb->dirty_ratelimit;
1672 		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1673 							RATELIMIT_CALC_SHIFT;
1674 		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1675 		min_pause = wb_min_pause(wb, max_pause,
1676 					 task_ratelimit, dirty_ratelimit,
1677 					 &nr_dirtied_pause);
1678 
1679 		if (unlikely(task_ratelimit == 0)) {
1680 			period = max_pause;
1681 			pause = max_pause;
1682 			goto pause;
1683 		}
1684 		period = HZ * pages_dirtied / task_ratelimit;
1685 		pause = period;
1686 		if (current->dirty_paused_when)
1687 			pause -= now - current->dirty_paused_when;
1688 		/*
1689 		 * For less than 1s think time (ext3/4 may block the dirtier
1690 		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1691 		 * however at much less frequency), try to compensate it in
1692 		 * future periods by updating the virtual time; otherwise just
1693 		 * do a reset, as it may be a light dirtier.
1694 		 */
1695 		if (pause < min_pause) {
1696 			trace_balance_dirty_pages(wb,
1697 						  sdtc->thresh,
1698 						  sdtc->bg_thresh,
1699 						  sdtc->dirty,
1700 						  sdtc->wb_thresh,
1701 						  sdtc->wb_dirty,
1702 						  dirty_ratelimit,
1703 						  task_ratelimit,
1704 						  pages_dirtied,
1705 						  period,
1706 						  min(pause, 0L),
1707 						  start_time);
1708 			if (pause < -HZ) {
1709 				current->dirty_paused_when = now;
1710 				current->nr_dirtied = 0;
1711 			} else if (period) {
1712 				current->dirty_paused_when += period;
1713 				current->nr_dirtied = 0;
1714 			} else if (current->nr_dirtied_pause <= pages_dirtied)
1715 				current->nr_dirtied_pause += pages_dirtied;
1716 			break;
1717 		}
1718 		if (unlikely(pause > max_pause)) {
1719 			/* for occasional dropped task_ratelimit */
1720 			now += min(pause - max_pause, max_pause);
1721 			pause = max_pause;
1722 		}
1723 
1724 pause:
1725 		trace_balance_dirty_pages(wb,
1726 					  sdtc->thresh,
1727 					  sdtc->bg_thresh,
1728 					  sdtc->dirty,
1729 					  sdtc->wb_thresh,
1730 					  sdtc->wb_dirty,
1731 					  dirty_ratelimit,
1732 					  task_ratelimit,
1733 					  pages_dirtied,
1734 					  period,
1735 					  pause,
1736 					  start_time);
1737 		__set_current_state(TASK_KILLABLE);
1738 		io_schedule_timeout(pause);
1739 
1740 		current->dirty_paused_when = now + pause;
1741 		current->nr_dirtied = 0;
1742 		current->nr_dirtied_pause = nr_dirtied_pause;
1743 
1744 		/*
1745 		 * This is typically equal to (dirty < thresh) and can also
1746 		 * keep "1000+ dd on a slow USB stick" under control.
1747 		 */
1748 		if (task_ratelimit)
1749 			break;
1750 
1751 		/*
1752 		 * In the case of an unresponding NFS server and the NFS dirty
1753 		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1754 		 * to go through, so that tasks on them still remain responsive.
1755 		 *
1756 		 * In theory 1 page is enough to keep the comsumer-producer
1757 		 * pipe going: the flusher cleans 1 page => the task dirties 1
1758 		 * more page. However wb_dirty has accounting errors.  So use
1759 		 * the larger and more IO friendly wb_stat_error.
1760 		 */
1761 		if (sdtc->wb_dirty <= wb_stat_error(wb))
1762 			break;
1763 
1764 		if (fatal_signal_pending(current))
1765 			break;
1766 	}
1767 
1768 	if (!dirty_exceeded && wb->dirty_exceeded)
1769 		wb->dirty_exceeded = 0;
1770 
1771 	if (writeback_in_progress(wb))
1772 		return;
1773 
1774 	/*
1775 	 * In laptop mode, we wait until hitting the higher threshold before
1776 	 * starting background writeout, and then write out all the way down
1777 	 * to the lower threshold.  So slow writers cause minimal disk activity.
1778 	 *
1779 	 * In normal mode, we start background writeout at the lower
1780 	 * background_thresh, to keep the amount of dirty memory low.
1781 	 */
1782 	if (laptop_mode)
1783 		return;
1784 
1785 	if (nr_reclaimable > gdtc->bg_thresh)
1786 		wb_start_background_writeback(wb);
1787 }
1788 
1789 static DEFINE_PER_CPU(int, bdp_ratelimits);
1790 
1791 /*
1792  * Normal tasks are throttled by
1793  *	loop {
1794  *		dirty tsk->nr_dirtied_pause pages;
1795  *		take a snap in balance_dirty_pages();
1796  *	}
1797  * However there is a worst case. If every task exit immediately when dirtied
1798  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1799  * called to throttle the page dirties. The solution is to save the not yet
1800  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1801  * randomly into the running tasks. This works well for the above worst case,
1802  * as the new task will pick up and accumulate the old task's leaked dirty
1803  * count and eventually get throttled.
1804  */
1805 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1806 
1807 /**
1808  * balance_dirty_pages_ratelimited - balance dirty memory state
1809  * @mapping: address_space which was dirtied
1810  *
1811  * Processes which are dirtying memory should call in here once for each page
1812  * which was newly dirtied.  The function will periodically check the system's
1813  * dirty state and will initiate writeback if needed.
1814  *
1815  * On really big machines, get_writeback_state is expensive, so try to avoid
1816  * calling it too often (ratelimiting).  But once we're over the dirty memory
1817  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1818  * from overshooting the limit by (ratelimit_pages) each.
1819  */
1820 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1821 {
1822 	struct inode *inode = mapping->host;
1823 	struct backing_dev_info *bdi = inode_to_bdi(inode);
1824 	struct bdi_writeback *wb = NULL;
1825 	int ratelimit;
1826 	int *p;
1827 
1828 	if (!bdi_cap_account_dirty(bdi))
1829 		return;
1830 
1831 	if (inode_cgwb_enabled(inode))
1832 		wb = wb_get_create_current(bdi, GFP_KERNEL);
1833 	if (!wb)
1834 		wb = &bdi->wb;
1835 
1836 	ratelimit = current->nr_dirtied_pause;
1837 	if (wb->dirty_exceeded)
1838 		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1839 
1840 	preempt_disable();
1841 	/*
1842 	 * This prevents one CPU to accumulate too many dirtied pages without
1843 	 * calling into balance_dirty_pages(), which can happen when there are
1844 	 * 1000+ tasks, all of them start dirtying pages at exactly the same
1845 	 * time, hence all honoured too large initial task->nr_dirtied_pause.
1846 	 */
1847 	p =  this_cpu_ptr(&bdp_ratelimits);
1848 	if (unlikely(current->nr_dirtied >= ratelimit))
1849 		*p = 0;
1850 	else if (unlikely(*p >= ratelimit_pages)) {
1851 		*p = 0;
1852 		ratelimit = 0;
1853 	}
1854 	/*
1855 	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1856 	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1857 	 * the dirty throttling and livelock other long-run dirtiers.
1858 	 */
1859 	p = this_cpu_ptr(&dirty_throttle_leaks);
1860 	if (*p > 0 && current->nr_dirtied < ratelimit) {
1861 		unsigned long nr_pages_dirtied;
1862 		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1863 		*p -= nr_pages_dirtied;
1864 		current->nr_dirtied += nr_pages_dirtied;
1865 	}
1866 	preempt_enable();
1867 
1868 	if (unlikely(current->nr_dirtied >= ratelimit))
1869 		balance_dirty_pages(mapping, wb, current->nr_dirtied);
1870 
1871 	wb_put(wb);
1872 }
1873 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1874 
1875 /**
1876  * wb_over_bg_thresh - does @wb need to be written back?
1877  * @wb: bdi_writeback of interest
1878  *
1879  * Determines whether background writeback should keep writing @wb or it's
1880  * clean enough.  Returns %true if writeback should continue.
1881  */
1882 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1883 {
1884 	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1885 	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1886 	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1887 	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1888 						     &mdtc_stor : NULL;
1889 
1890 	/*
1891 	 * Similar to balance_dirty_pages() but ignores pages being written
1892 	 * as we're trying to decide whether to put more under writeback.
1893 	 */
1894 	gdtc->avail = global_dirtyable_memory();
1895 	gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1896 		      global_page_state(NR_UNSTABLE_NFS);
1897 	domain_dirty_limits(gdtc);
1898 
1899 	if (gdtc->dirty > gdtc->bg_thresh)
1900 		return true;
1901 
1902 	if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc))
1903 		return true;
1904 
1905 	if (mdtc) {
1906 		unsigned long filepages, headroom, writeback;
1907 
1908 		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1909 				    &writeback);
1910 		mdtc_calc_avail(mdtc, filepages, headroom);
1911 		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
1912 
1913 		if (mdtc->dirty > mdtc->bg_thresh)
1914 			return true;
1915 
1916 		if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(mdtc))
1917 			return true;
1918 	}
1919 
1920 	return false;
1921 }
1922 
1923 void throttle_vm_writeout(gfp_t gfp_mask)
1924 {
1925 	unsigned long background_thresh;
1926 	unsigned long dirty_thresh;
1927 
1928         for ( ; ; ) {
1929 		global_dirty_limits(&background_thresh, &dirty_thresh);
1930 		dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1931 
1932                 /*
1933                  * Boost the allowable dirty threshold a bit for page
1934                  * allocators so they don't get DoS'ed by heavy writers
1935                  */
1936                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1937 
1938                 if (global_page_state(NR_UNSTABLE_NFS) +
1939 			global_page_state(NR_WRITEBACK) <= dirty_thresh)
1940                         	break;
1941                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1942 
1943 		/*
1944 		 * The caller might hold locks which can prevent IO completion
1945 		 * or progress in the filesystem.  So we cannot just sit here
1946 		 * waiting for IO to complete.
1947 		 */
1948 		if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1949 			break;
1950         }
1951 }
1952 
1953 /*
1954  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1955  */
1956 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1957 	void __user *buffer, size_t *length, loff_t *ppos)
1958 {
1959 	proc_dointvec(table, write, buffer, length, ppos);
1960 	return 0;
1961 }
1962 
1963 #ifdef CONFIG_BLOCK
1964 void laptop_mode_timer_fn(unsigned long data)
1965 {
1966 	struct request_queue *q = (struct request_queue *)data;
1967 	int nr_pages = global_page_state(NR_FILE_DIRTY) +
1968 		global_page_state(NR_UNSTABLE_NFS);
1969 	struct bdi_writeback *wb;
1970 
1971 	/*
1972 	 * We want to write everything out, not just down to the dirty
1973 	 * threshold
1974 	 */
1975 	if (!bdi_has_dirty_io(&q->backing_dev_info))
1976 		return;
1977 
1978 	rcu_read_lock();
1979 	list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
1980 		if (wb_has_dirty_io(wb))
1981 			wb_start_writeback(wb, nr_pages, true,
1982 					   WB_REASON_LAPTOP_TIMER);
1983 	rcu_read_unlock();
1984 }
1985 
1986 /*
1987  * We've spun up the disk and we're in laptop mode: schedule writeback
1988  * of all dirty data a few seconds from now.  If the flush is already scheduled
1989  * then push it back - the user is still using the disk.
1990  */
1991 void laptop_io_completion(struct backing_dev_info *info)
1992 {
1993 	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1994 }
1995 
1996 /*
1997  * We're in laptop mode and we've just synced. The sync's writes will have
1998  * caused another writeback to be scheduled by laptop_io_completion.
1999  * Nothing needs to be written back anymore, so we unschedule the writeback.
2000  */
2001 void laptop_sync_completion(void)
2002 {
2003 	struct backing_dev_info *bdi;
2004 
2005 	rcu_read_lock();
2006 
2007 	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2008 		del_timer(&bdi->laptop_mode_wb_timer);
2009 
2010 	rcu_read_unlock();
2011 }
2012 #endif
2013 
2014 /*
2015  * If ratelimit_pages is too high then we can get into dirty-data overload
2016  * if a large number of processes all perform writes at the same time.
2017  * If it is too low then SMP machines will call the (expensive)
2018  * get_writeback_state too often.
2019  *
2020  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2021  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2022  * thresholds.
2023  */
2024 
2025 void writeback_set_ratelimit(void)
2026 {
2027 	struct wb_domain *dom = &global_wb_domain;
2028 	unsigned long background_thresh;
2029 	unsigned long dirty_thresh;
2030 
2031 	global_dirty_limits(&background_thresh, &dirty_thresh);
2032 	dom->dirty_limit = dirty_thresh;
2033 	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2034 	if (ratelimit_pages < 16)
2035 		ratelimit_pages = 16;
2036 }
2037 
2038 static int
2039 ratelimit_handler(struct notifier_block *self, unsigned long action,
2040 		  void *hcpu)
2041 {
2042 
2043 	switch (action & ~CPU_TASKS_FROZEN) {
2044 	case CPU_ONLINE:
2045 	case CPU_DEAD:
2046 		writeback_set_ratelimit();
2047 		return NOTIFY_OK;
2048 	default:
2049 		return NOTIFY_DONE;
2050 	}
2051 }
2052 
2053 static struct notifier_block ratelimit_nb = {
2054 	.notifier_call	= ratelimit_handler,
2055 	.next		= NULL,
2056 };
2057 
2058 /*
2059  * Called early on to tune the page writeback dirty limits.
2060  *
2061  * We used to scale dirty pages according to how total memory
2062  * related to pages that could be allocated for buffers (by
2063  * comparing nr_free_buffer_pages() to vm_total_pages.
2064  *
2065  * However, that was when we used "dirty_ratio" to scale with
2066  * all memory, and we don't do that any more. "dirty_ratio"
2067  * is now applied to total non-HIGHPAGE memory (by subtracting
2068  * totalhigh_pages from vm_total_pages), and as such we can't
2069  * get into the old insane situation any more where we had
2070  * large amounts of dirty pages compared to a small amount of
2071  * non-HIGHMEM memory.
2072  *
2073  * But we might still want to scale the dirty_ratio by how
2074  * much memory the box has..
2075  */
2076 void __init page_writeback_init(void)
2077 {
2078 	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2079 
2080 	writeback_set_ratelimit();
2081 	register_cpu_notifier(&ratelimit_nb);
2082 }
2083 
2084 /**
2085  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2086  * @mapping: address space structure to write
2087  * @start: starting page index
2088  * @end: ending page index (inclusive)
2089  *
2090  * This function scans the page range from @start to @end (inclusive) and tags
2091  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2092  * that write_cache_pages (or whoever calls this function) will then use
2093  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2094  * used to avoid livelocking of writeback by a process steadily creating new
2095  * dirty pages in the file (thus it is important for this function to be quick
2096  * so that it can tag pages faster than a dirtying process can create them).
2097  */
2098 /*
2099  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2100  */
2101 void tag_pages_for_writeback(struct address_space *mapping,
2102 			     pgoff_t start, pgoff_t end)
2103 {
2104 #define WRITEBACK_TAG_BATCH 4096
2105 	unsigned long tagged;
2106 
2107 	do {
2108 		spin_lock_irq(&mapping->tree_lock);
2109 		tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2110 				&start, end, WRITEBACK_TAG_BATCH,
2111 				PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2112 		spin_unlock_irq(&mapping->tree_lock);
2113 		WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2114 		cond_resched();
2115 		/* We check 'start' to handle wrapping when end == ~0UL */
2116 	} while (tagged >= WRITEBACK_TAG_BATCH && start);
2117 }
2118 EXPORT_SYMBOL(tag_pages_for_writeback);
2119 
2120 /**
2121  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2122  * @mapping: address space structure to write
2123  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2124  * @writepage: function called for each page
2125  * @data: data passed to writepage function
2126  *
2127  * If a page is already under I/O, write_cache_pages() skips it, even
2128  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2129  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2130  * and msync() need to guarantee that all the data which was dirty at the time
2131  * the call was made get new I/O started against them.  If wbc->sync_mode is
2132  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2133  * existing IO to complete.
2134  *
2135  * To avoid livelocks (when other process dirties new pages), we first tag
2136  * pages which should be written back with TOWRITE tag and only then start
2137  * writing them. For data-integrity sync we have to be careful so that we do
2138  * not miss some pages (e.g., because some other process has cleared TOWRITE
2139  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2140  * by the process clearing the DIRTY tag (and submitting the page for IO).
2141  */
2142 int write_cache_pages(struct address_space *mapping,
2143 		      struct writeback_control *wbc, writepage_t writepage,
2144 		      void *data)
2145 {
2146 	int ret = 0;
2147 	int done = 0;
2148 	struct pagevec pvec;
2149 	int nr_pages;
2150 	pgoff_t uninitialized_var(writeback_index);
2151 	pgoff_t index;
2152 	pgoff_t end;		/* Inclusive */
2153 	pgoff_t done_index;
2154 	int cycled;
2155 	int range_whole = 0;
2156 	int tag;
2157 
2158 	pagevec_init(&pvec, 0);
2159 	if (wbc->range_cyclic) {
2160 		writeback_index = mapping->writeback_index; /* prev offset */
2161 		index = writeback_index;
2162 		if (index == 0)
2163 			cycled = 1;
2164 		else
2165 			cycled = 0;
2166 		end = -1;
2167 	} else {
2168 		index = wbc->range_start >> PAGE_CACHE_SHIFT;
2169 		end = wbc->range_end >> PAGE_CACHE_SHIFT;
2170 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2171 			range_whole = 1;
2172 		cycled = 1; /* ignore range_cyclic tests */
2173 	}
2174 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2175 		tag = PAGECACHE_TAG_TOWRITE;
2176 	else
2177 		tag = PAGECACHE_TAG_DIRTY;
2178 retry:
2179 	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2180 		tag_pages_for_writeback(mapping, index, end);
2181 	done_index = index;
2182 	while (!done && (index <= end)) {
2183 		int i;
2184 
2185 		nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2186 			      min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2187 		if (nr_pages == 0)
2188 			break;
2189 
2190 		for (i = 0; i < nr_pages; i++) {
2191 			struct page *page = pvec.pages[i];
2192 
2193 			/*
2194 			 * At this point, the page may be truncated or
2195 			 * invalidated (changing page->mapping to NULL), or
2196 			 * even swizzled back from swapper_space to tmpfs file
2197 			 * mapping. However, page->index will not change
2198 			 * because we have a reference on the page.
2199 			 */
2200 			if (page->index > end) {
2201 				/*
2202 				 * can't be range_cyclic (1st pass) because
2203 				 * end == -1 in that case.
2204 				 */
2205 				done = 1;
2206 				break;
2207 			}
2208 
2209 			done_index = page->index;
2210 
2211 			lock_page(page);
2212 
2213 			/*
2214 			 * Page truncated or invalidated. We can freely skip it
2215 			 * then, even for data integrity operations: the page
2216 			 * has disappeared concurrently, so there could be no
2217 			 * real expectation of this data interity operation
2218 			 * even if there is now a new, dirty page at the same
2219 			 * pagecache address.
2220 			 */
2221 			if (unlikely(page->mapping != mapping)) {
2222 continue_unlock:
2223 				unlock_page(page);
2224 				continue;
2225 			}
2226 
2227 			if (!PageDirty(page)) {
2228 				/* someone wrote it for us */
2229 				goto continue_unlock;
2230 			}
2231 
2232 			if (PageWriteback(page)) {
2233 				if (wbc->sync_mode != WB_SYNC_NONE)
2234 					wait_on_page_writeback(page);
2235 				else
2236 					goto continue_unlock;
2237 			}
2238 
2239 			BUG_ON(PageWriteback(page));
2240 			if (!clear_page_dirty_for_io(page))
2241 				goto continue_unlock;
2242 
2243 			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2244 			ret = (*writepage)(page, wbc, data);
2245 			if (unlikely(ret)) {
2246 				if (ret == AOP_WRITEPAGE_ACTIVATE) {
2247 					unlock_page(page);
2248 					ret = 0;
2249 				} else {
2250 					/*
2251 					 * done_index is set past this page,
2252 					 * so media errors will not choke
2253 					 * background writeout for the entire
2254 					 * file. This has consequences for
2255 					 * range_cyclic semantics (ie. it may
2256 					 * not be suitable for data integrity
2257 					 * writeout).
2258 					 */
2259 					done_index = page->index + 1;
2260 					done = 1;
2261 					break;
2262 				}
2263 			}
2264 
2265 			/*
2266 			 * We stop writing back only if we are not doing
2267 			 * integrity sync. In case of integrity sync we have to
2268 			 * keep going until we have written all the pages
2269 			 * we tagged for writeback prior to entering this loop.
2270 			 */
2271 			if (--wbc->nr_to_write <= 0 &&
2272 			    wbc->sync_mode == WB_SYNC_NONE) {
2273 				done = 1;
2274 				break;
2275 			}
2276 		}
2277 		pagevec_release(&pvec);
2278 		cond_resched();
2279 	}
2280 	if (!cycled && !done) {
2281 		/*
2282 		 * range_cyclic:
2283 		 * We hit the last page and there is more work to be done: wrap
2284 		 * back to the start of the file
2285 		 */
2286 		cycled = 1;
2287 		index = 0;
2288 		end = writeback_index - 1;
2289 		goto retry;
2290 	}
2291 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2292 		mapping->writeback_index = done_index;
2293 
2294 	return ret;
2295 }
2296 EXPORT_SYMBOL(write_cache_pages);
2297 
2298 /*
2299  * Function used by generic_writepages to call the real writepage
2300  * function and set the mapping flags on error
2301  */
2302 static int __writepage(struct page *page, struct writeback_control *wbc,
2303 		       void *data)
2304 {
2305 	struct address_space *mapping = data;
2306 	int ret = mapping->a_ops->writepage(page, wbc);
2307 	mapping_set_error(mapping, ret);
2308 	return ret;
2309 }
2310 
2311 /**
2312  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2313  * @mapping: address space structure to write
2314  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2315  *
2316  * This is a library function, which implements the writepages()
2317  * address_space_operation.
2318  */
2319 int generic_writepages(struct address_space *mapping,
2320 		       struct writeback_control *wbc)
2321 {
2322 	struct blk_plug plug;
2323 	int ret;
2324 
2325 	/* deal with chardevs and other special file */
2326 	if (!mapping->a_ops->writepage)
2327 		return 0;
2328 
2329 	blk_start_plug(&plug);
2330 	ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2331 	blk_finish_plug(&plug);
2332 	return ret;
2333 }
2334 
2335 EXPORT_SYMBOL(generic_writepages);
2336 
2337 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2338 {
2339 	int ret;
2340 
2341 	if (wbc->nr_to_write <= 0)
2342 		return 0;
2343 	if (mapping->a_ops->writepages)
2344 		ret = mapping->a_ops->writepages(mapping, wbc);
2345 	else
2346 		ret = generic_writepages(mapping, wbc);
2347 	return ret;
2348 }
2349 
2350 /**
2351  * write_one_page - write out a single page and optionally wait on I/O
2352  * @page: the page to write
2353  * @wait: if true, wait on writeout
2354  *
2355  * The page must be locked by the caller and will be unlocked upon return.
2356  *
2357  * write_one_page() returns a negative error code if I/O failed.
2358  */
2359 int write_one_page(struct page *page, int wait)
2360 {
2361 	struct address_space *mapping = page->mapping;
2362 	int ret = 0;
2363 	struct writeback_control wbc = {
2364 		.sync_mode = WB_SYNC_ALL,
2365 		.nr_to_write = 1,
2366 	};
2367 
2368 	BUG_ON(!PageLocked(page));
2369 
2370 	if (wait)
2371 		wait_on_page_writeback(page);
2372 
2373 	if (clear_page_dirty_for_io(page)) {
2374 		page_cache_get(page);
2375 		ret = mapping->a_ops->writepage(page, &wbc);
2376 		if (ret == 0 && wait) {
2377 			wait_on_page_writeback(page);
2378 			if (PageError(page))
2379 				ret = -EIO;
2380 		}
2381 		page_cache_release(page);
2382 	} else {
2383 		unlock_page(page);
2384 	}
2385 	return ret;
2386 }
2387 EXPORT_SYMBOL(write_one_page);
2388 
2389 /*
2390  * For address_spaces which do not use buffers nor write back.
2391  */
2392 int __set_page_dirty_no_writeback(struct page *page)
2393 {
2394 	if (!PageDirty(page))
2395 		return !TestSetPageDirty(page);
2396 	return 0;
2397 }
2398 
2399 /*
2400  * Helper function for set_page_dirty family.
2401  *
2402  * Caller must hold mem_cgroup_begin_page_stat().
2403  *
2404  * NOTE: This relies on being atomic wrt interrupts.
2405  */
2406 void account_page_dirtied(struct page *page, struct address_space *mapping,
2407 			  struct mem_cgroup *memcg)
2408 {
2409 	struct inode *inode = mapping->host;
2410 
2411 	trace_writeback_dirty_page(page, mapping);
2412 
2413 	if (mapping_cap_account_dirty(mapping)) {
2414 		struct bdi_writeback *wb;
2415 
2416 		inode_attach_wb(inode, page);
2417 		wb = inode_to_wb(inode);
2418 
2419 		mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2420 		__inc_zone_page_state(page, NR_FILE_DIRTY);
2421 		__inc_zone_page_state(page, NR_DIRTIED);
2422 		__inc_wb_stat(wb, WB_RECLAIMABLE);
2423 		__inc_wb_stat(wb, WB_DIRTIED);
2424 		task_io_account_write(PAGE_CACHE_SIZE);
2425 		current->nr_dirtied++;
2426 		this_cpu_inc(bdp_ratelimits);
2427 	}
2428 }
2429 EXPORT_SYMBOL(account_page_dirtied);
2430 
2431 /*
2432  * Helper function for deaccounting dirty page without writeback.
2433  *
2434  * Caller must hold mem_cgroup_begin_page_stat().
2435  */
2436 void account_page_cleaned(struct page *page, struct address_space *mapping,
2437 			  struct mem_cgroup *memcg, struct bdi_writeback *wb)
2438 {
2439 	if (mapping_cap_account_dirty(mapping)) {
2440 		mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2441 		dec_zone_page_state(page, NR_FILE_DIRTY);
2442 		dec_wb_stat(wb, WB_RECLAIMABLE);
2443 		task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2444 	}
2445 }
2446 
2447 /*
2448  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2449  * its radix tree.
2450  *
2451  * This is also used when a single buffer is being dirtied: we want to set the
2452  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2453  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454  *
2455  * The caller must ensure this doesn't race with truncation.  Most will simply
2456  * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457  * the pte lock held, which also locks out truncation.
2458  */
2459 int __set_page_dirty_nobuffers(struct page *page)
2460 {
2461 	struct mem_cgroup *memcg;
2462 
2463 	memcg = mem_cgroup_begin_page_stat(page);
2464 	if (!TestSetPageDirty(page)) {
2465 		struct address_space *mapping = page_mapping(page);
2466 		unsigned long flags;
2467 
2468 		if (!mapping) {
2469 			mem_cgroup_end_page_stat(memcg);
2470 			return 1;
2471 		}
2472 
2473 		spin_lock_irqsave(&mapping->tree_lock, flags);
2474 		BUG_ON(page_mapping(page) != mapping);
2475 		WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2476 		account_page_dirtied(page, mapping, memcg);
2477 		radix_tree_tag_set(&mapping->page_tree, page_index(page),
2478 				   PAGECACHE_TAG_DIRTY);
2479 		spin_unlock_irqrestore(&mapping->tree_lock, flags);
2480 		mem_cgroup_end_page_stat(memcg);
2481 
2482 		if (mapping->host) {
2483 			/* !PageAnon && !swapper_space */
2484 			__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2485 		}
2486 		return 1;
2487 	}
2488 	mem_cgroup_end_page_stat(memcg);
2489 	return 0;
2490 }
2491 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2492 
2493 /*
2494  * Call this whenever redirtying a page, to de-account the dirty counters
2495  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2496  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2497  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2498  * control.
2499  */
2500 void account_page_redirty(struct page *page)
2501 {
2502 	struct address_space *mapping = page->mapping;
2503 
2504 	if (mapping && mapping_cap_account_dirty(mapping)) {
2505 		struct inode *inode = mapping->host;
2506 		struct bdi_writeback *wb;
2507 		bool locked;
2508 
2509 		wb = unlocked_inode_to_wb_begin(inode, &locked);
2510 		current->nr_dirtied--;
2511 		dec_zone_page_state(page, NR_DIRTIED);
2512 		dec_wb_stat(wb, WB_DIRTIED);
2513 		unlocked_inode_to_wb_end(inode, locked);
2514 	}
2515 }
2516 EXPORT_SYMBOL(account_page_redirty);
2517 
2518 /*
2519  * When a writepage implementation decides that it doesn't want to write this
2520  * page for some reason, it should redirty the locked page via
2521  * redirty_page_for_writepage() and it should then unlock the page and return 0
2522  */
2523 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2524 {
2525 	int ret;
2526 
2527 	wbc->pages_skipped++;
2528 	ret = __set_page_dirty_nobuffers(page);
2529 	account_page_redirty(page);
2530 	return ret;
2531 }
2532 EXPORT_SYMBOL(redirty_page_for_writepage);
2533 
2534 /*
2535  * Dirty a page.
2536  *
2537  * For pages with a mapping this should be done under the page lock
2538  * for the benefit of asynchronous memory errors who prefer a consistent
2539  * dirty state. This rule can be broken in some special cases,
2540  * but should be better not to.
2541  *
2542  * If the mapping doesn't provide a set_page_dirty a_op, then
2543  * just fall through and assume that it wants buffer_heads.
2544  */
2545 int set_page_dirty(struct page *page)
2546 {
2547 	struct address_space *mapping = page_mapping(page);
2548 
2549 	if (likely(mapping)) {
2550 		int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2551 		/*
2552 		 * readahead/lru_deactivate_page could remain
2553 		 * PG_readahead/PG_reclaim due to race with end_page_writeback
2554 		 * About readahead, if the page is written, the flags would be
2555 		 * reset. So no problem.
2556 		 * About lru_deactivate_page, if the page is redirty, the flag
2557 		 * will be reset. So no problem. but if the page is used by readahead
2558 		 * it will confuse readahead and make it restart the size rampup
2559 		 * process. But it's a trivial problem.
2560 		 */
2561 		if (PageReclaim(page))
2562 			ClearPageReclaim(page);
2563 #ifdef CONFIG_BLOCK
2564 		if (!spd)
2565 			spd = __set_page_dirty_buffers;
2566 #endif
2567 		return (*spd)(page);
2568 	}
2569 	if (!PageDirty(page)) {
2570 		if (!TestSetPageDirty(page))
2571 			return 1;
2572 	}
2573 	return 0;
2574 }
2575 EXPORT_SYMBOL(set_page_dirty);
2576 
2577 /*
2578  * set_page_dirty() is racy if the caller has no reference against
2579  * page->mapping->host, and if the page is unlocked.  This is because another
2580  * CPU could truncate the page off the mapping and then free the mapping.
2581  *
2582  * Usually, the page _is_ locked, or the caller is a user-space process which
2583  * holds a reference on the inode by having an open file.
2584  *
2585  * In other cases, the page should be locked before running set_page_dirty().
2586  */
2587 int set_page_dirty_lock(struct page *page)
2588 {
2589 	int ret;
2590 
2591 	lock_page(page);
2592 	ret = set_page_dirty(page);
2593 	unlock_page(page);
2594 	return ret;
2595 }
2596 EXPORT_SYMBOL(set_page_dirty_lock);
2597 
2598 /*
2599  * This cancels just the dirty bit on the kernel page itself, it does NOT
2600  * actually remove dirty bits on any mmap's that may be around. It also
2601  * leaves the page tagged dirty, so any sync activity will still find it on
2602  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2603  * look at the dirty bits in the VM.
2604  *
2605  * Doing this should *normally* only ever be done when a page is truncated,
2606  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2607  * this when it notices that somebody has cleaned out all the buffers on a
2608  * page without actually doing it through the VM. Can you say "ext3 is
2609  * horribly ugly"? Thought you could.
2610  */
2611 void cancel_dirty_page(struct page *page)
2612 {
2613 	struct address_space *mapping = page_mapping(page);
2614 
2615 	if (mapping_cap_account_dirty(mapping)) {
2616 		struct inode *inode = mapping->host;
2617 		struct bdi_writeback *wb;
2618 		struct mem_cgroup *memcg;
2619 		bool locked;
2620 
2621 		memcg = mem_cgroup_begin_page_stat(page);
2622 		wb = unlocked_inode_to_wb_begin(inode, &locked);
2623 
2624 		if (TestClearPageDirty(page))
2625 			account_page_cleaned(page, mapping, memcg, wb);
2626 
2627 		unlocked_inode_to_wb_end(inode, locked);
2628 		mem_cgroup_end_page_stat(memcg);
2629 	} else {
2630 		ClearPageDirty(page);
2631 	}
2632 }
2633 EXPORT_SYMBOL(cancel_dirty_page);
2634 
2635 /*
2636  * Clear a page's dirty flag, while caring for dirty memory accounting.
2637  * Returns true if the page was previously dirty.
2638  *
2639  * This is for preparing to put the page under writeout.  We leave the page
2640  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2641  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2642  * implementation will run either set_page_writeback() or set_page_dirty(),
2643  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2644  * back into sync.
2645  *
2646  * This incoherency between the page's dirty flag and radix-tree tag is
2647  * unfortunate, but it only exists while the page is locked.
2648  */
2649 int clear_page_dirty_for_io(struct page *page)
2650 {
2651 	struct address_space *mapping = page_mapping(page);
2652 	int ret = 0;
2653 
2654 	BUG_ON(!PageLocked(page));
2655 
2656 	if (mapping && mapping_cap_account_dirty(mapping)) {
2657 		struct inode *inode = mapping->host;
2658 		struct bdi_writeback *wb;
2659 		struct mem_cgroup *memcg;
2660 		bool locked;
2661 
2662 		/*
2663 		 * Yes, Virginia, this is indeed insane.
2664 		 *
2665 		 * We use this sequence to make sure that
2666 		 *  (a) we account for dirty stats properly
2667 		 *  (b) we tell the low-level filesystem to
2668 		 *      mark the whole page dirty if it was
2669 		 *      dirty in a pagetable. Only to then
2670 		 *  (c) clean the page again and return 1 to
2671 		 *      cause the writeback.
2672 		 *
2673 		 * This way we avoid all nasty races with the
2674 		 * dirty bit in multiple places and clearing
2675 		 * them concurrently from different threads.
2676 		 *
2677 		 * Note! Normally the "set_page_dirty(page)"
2678 		 * has no effect on the actual dirty bit - since
2679 		 * that will already usually be set. But we
2680 		 * need the side effects, and it can help us
2681 		 * avoid races.
2682 		 *
2683 		 * We basically use the page "master dirty bit"
2684 		 * as a serialization point for all the different
2685 		 * threads doing their things.
2686 		 */
2687 		if (page_mkclean(page))
2688 			set_page_dirty(page);
2689 		/*
2690 		 * We carefully synchronise fault handlers against
2691 		 * installing a dirty pte and marking the page dirty
2692 		 * at this point.  We do this by having them hold the
2693 		 * page lock while dirtying the page, and pages are
2694 		 * always locked coming in here, so we get the desired
2695 		 * exclusion.
2696 		 */
2697 		memcg = mem_cgroup_begin_page_stat(page);
2698 		wb = unlocked_inode_to_wb_begin(inode, &locked);
2699 		if (TestClearPageDirty(page)) {
2700 			mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2701 			dec_zone_page_state(page, NR_FILE_DIRTY);
2702 			dec_wb_stat(wb, WB_RECLAIMABLE);
2703 			ret = 1;
2704 		}
2705 		unlocked_inode_to_wb_end(inode, locked);
2706 		mem_cgroup_end_page_stat(memcg);
2707 		return ret;
2708 	}
2709 	return TestClearPageDirty(page);
2710 }
2711 EXPORT_SYMBOL(clear_page_dirty_for_io);
2712 
2713 int test_clear_page_writeback(struct page *page)
2714 {
2715 	struct address_space *mapping = page_mapping(page);
2716 	struct mem_cgroup *memcg;
2717 	int ret;
2718 
2719 	memcg = mem_cgroup_begin_page_stat(page);
2720 	if (mapping) {
2721 		struct inode *inode = mapping->host;
2722 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2723 		unsigned long flags;
2724 
2725 		spin_lock_irqsave(&mapping->tree_lock, flags);
2726 		ret = TestClearPageWriteback(page);
2727 		if (ret) {
2728 			radix_tree_tag_clear(&mapping->page_tree,
2729 						page_index(page),
2730 						PAGECACHE_TAG_WRITEBACK);
2731 			if (bdi_cap_account_writeback(bdi)) {
2732 				struct bdi_writeback *wb = inode_to_wb(inode);
2733 
2734 				__dec_wb_stat(wb, WB_WRITEBACK);
2735 				__wb_writeout_inc(wb);
2736 			}
2737 		}
2738 		spin_unlock_irqrestore(&mapping->tree_lock, flags);
2739 	} else {
2740 		ret = TestClearPageWriteback(page);
2741 	}
2742 	if (ret) {
2743 		mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2744 		dec_zone_page_state(page, NR_WRITEBACK);
2745 		inc_zone_page_state(page, NR_WRITTEN);
2746 	}
2747 	mem_cgroup_end_page_stat(memcg);
2748 	return ret;
2749 }
2750 
2751 int __test_set_page_writeback(struct page *page, bool keep_write)
2752 {
2753 	struct address_space *mapping = page_mapping(page);
2754 	struct mem_cgroup *memcg;
2755 	int ret;
2756 
2757 	memcg = mem_cgroup_begin_page_stat(page);
2758 	if (mapping) {
2759 		struct inode *inode = mapping->host;
2760 		struct backing_dev_info *bdi = inode_to_bdi(inode);
2761 		unsigned long flags;
2762 
2763 		spin_lock_irqsave(&mapping->tree_lock, flags);
2764 		ret = TestSetPageWriteback(page);
2765 		if (!ret) {
2766 			radix_tree_tag_set(&mapping->page_tree,
2767 						page_index(page),
2768 						PAGECACHE_TAG_WRITEBACK);
2769 			if (bdi_cap_account_writeback(bdi))
2770 				__inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2771 		}
2772 		if (!PageDirty(page))
2773 			radix_tree_tag_clear(&mapping->page_tree,
2774 						page_index(page),
2775 						PAGECACHE_TAG_DIRTY);
2776 		if (!keep_write)
2777 			radix_tree_tag_clear(&mapping->page_tree,
2778 						page_index(page),
2779 						PAGECACHE_TAG_TOWRITE);
2780 		spin_unlock_irqrestore(&mapping->tree_lock, flags);
2781 	} else {
2782 		ret = TestSetPageWriteback(page);
2783 	}
2784 	if (!ret) {
2785 		mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2786 		inc_zone_page_state(page, NR_WRITEBACK);
2787 	}
2788 	mem_cgroup_end_page_stat(memcg);
2789 	return ret;
2790 
2791 }
2792 EXPORT_SYMBOL(__test_set_page_writeback);
2793 
2794 /*
2795  * Return true if any of the pages in the mapping are marked with the
2796  * passed tag.
2797  */
2798 int mapping_tagged(struct address_space *mapping, int tag)
2799 {
2800 	return radix_tree_tagged(&mapping->page_tree, tag);
2801 }
2802 EXPORT_SYMBOL(mapping_tagged);
2803 
2804 /**
2805  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2806  * @page:	The page to wait on.
2807  *
2808  * This function determines if the given page is related to a backing device
2809  * that requires page contents to be held stable during writeback.  If so, then
2810  * it will wait for any pending writeback to complete.
2811  */
2812 void wait_for_stable_page(struct page *page)
2813 {
2814 	if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2815 		wait_on_page_writeback(page);
2816 }
2817 EXPORT_SYMBOL_GPL(wait_for_stable_page);
2818