1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * mm/page-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
11 * 10Apr2002 Andrew Morton
12 * Initial version
13 */
14
15 #include <linux/kernel.h>
16 #include <linux/math64.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/fs.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/slab.h>
23 #include <linux/pagemap.h>
24 #include <linux/writeback.h>
25 #include <linux/init.h>
26 #include <linux/backing-dev.h>
27 #include <linux/task_io_accounting_ops.h>
28 #include <linux/blkdev.h>
29 #include <linux/mpage.h>
30 #include <linux/rmap.h>
31 #include <linux/percpu.h>
32 #include <linux/smp.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/syscalls.h>
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/sched/signal.h>
40 #include <linux/mm_inline.h>
41 #include <trace/events/writeback.h>
42
43 #include "internal.h"
44
45 /*
46 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48 #define MAX_PAUSE max(HZ/5, 1)
49
50 /*
51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56 /*
57 * Estimate write bandwidth or update dirty limit at 200ms intervals.
58 */
59 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
61 #define RATELIMIT_CALC_SHIFT 10
62
63 /*
64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67 static long ratelimit_pages = 32;
68
69 /* The following parameters are exported via /proc/sys/vm */
70
71 /*
72 * Start background writeback (via writeback threads) at this percentage
73 */
74 static int dirty_background_ratio = 10;
75
76 /*
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
79 */
80 static unsigned long dirty_background_bytes;
81
82 /*
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 */
86 static int vm_highmem_is_dirtyable;
87
88 /*
89 * The generator of dirty data starts writeback at this percentage
90 */
91 static int vm_dirty_ratio = 20;
92
93 /*
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
96 */
97 static unsigned long vm_dirty_bytes;
98
99 /*
100 * The interval between `kupdate'-style writebacks
101 */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106 /*
107 * The longest time for which data is allowed to remain dirty
108 */
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111 /*
112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113 * a full sync is triggered after this time elapses without any disk activity.
114 */
115 int laptop_mode;
116
117 EXPORT_SYMBOL(laptop_mode);
118
119 /* End of sysctl-exported parameters */
120
121 struct wb_domain global_wb_domain;
122
123 /* consolidated parameters for balance_dirty_pages() and its subroutines */
124 struct dirty_throttle_control {
125 #ifdef CONFIG_CGROUP_WRITEBACK
126 struct wb_domain *dom;
127 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
128 #endif
129 struct bdi_writeback *wb;
130 struct fprop_local_percpu *wb_completions;
131
132 unsigned long avail; /* dirtyable */
133 unsigned long dirty; /* file_dirty + write + nfs */
134 unsigned long thresh; /* dirty threshold */
135 unsigned long bg_thresh; /* dirty background threshold */
136
137 unsigned long wb_dirty; /* per-wb counterparts */
138 unsigned long wb_thresh;
139 unsigned long wb_bg_thresh;
140
141 unsigned long pos_ratio;
142 bool freerun;
143 bool dirty_exceeded;
144 };
145
146 /*
147 * Length of period for aging writeout fractions of bdis. This is an
148 * arbitrarily chosen number. The longer the period, the slower fractions will
149 * reflect changes in current writeout rate.
150 */
151 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
152
153 #ifdef CONFIG_CGROUP_WRITEBACK
154
155 #define GDTC_INIT(__wb) .wb = (__wb), \
156 .dom = &global_wb_domain, \
157 .wb_completions = &(__wb)->completions
158
159 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
160
161 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
162 .dom = mem_cgroup_wb_domain(__wb), \
163 .wb_completions = &(__wb)->memcg_completions, \
164 .gdtc = __gdtc
165
mdtc_valid(struct dirty_throttle_control * dtc)166 static bool mdtc_valid(struct dirty_throttle_control *dtc)
167 {
168 return dtc->dom;
169 }
170
dtc_dom(struct dirty_throttle_control * dtc)171 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
172 {
173 return dtc->dom;
174 }
175
mdtc_gdtc(struct dirty_throttle_control * mdtc)176 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
177 {
178 return mdtc->gdtc;
179 }
180
wb_memcg_completions(struct bdi_writeback * wb)181 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
182 {
183 return &wb->memcg_completions;
184 }
185
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)186 static void wb_min_max_ratio(struct bdi_writeback *wb,
187 unsigned long *minp, unsigned long *maxp)
188 {
189 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
190 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
191 unsigned long long min = wb->bdi->min_ratio;
192 unsigned long long max = wb->bdi->max_ratio;
193
194 /*
195 * @wb may already be clean by the time control reaches here and
196 * the total may not include its bw.
197 */
198 if (this_bw < tot_bw) {
199 if (min) {
200 min *= this_bw;
201 min = div64_ul(min, tot_bw);
202 }
203 if (max < 100 * BDI_RATIO_SCALE) {
204 max *= this_bw;
205 max = div64_ul(max, tot_bw);
206 }
207 }
208
209 *minp = min;
210 *maxp = max;
211 }
212
213 #else /* CONFIG_CGROUP_WRITEBACK */
214
215 #define GDTC_INIT(__wb) .wb = (__wb), \
216 .wb_completions = &(__wb)->completions
217 #define GDTC_INIT_NO_WB
218 #define MDTC_INIT(__wb, __gdtc)
219
mdtc_valid(struct dirty_throttle_control * dtc)220 static bool mdtc_valid(struct dirty_throttle_control *dtc)
221 {
222 return false;
223 }
224
dtc_dom(struct dirty_throttle_control * dtc)225 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
226 {
227 return &global_wb_domain;
228 }
229
mdtc_gdtc(struct dirty_throttle_control * mdtc)230 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
231 {
232 return NULL;
233 }
234
wb_memcg_completions(struct bdi_writeback * wb)235 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
236 {
237 return NULL;
238 }
239
wb_min_max_ratio(struct bdi_writeback * wb,unsigned long * minp,unsigned long * maxp)240 static void wb_min_max_ratio(struct bdi_writeback *wb,
241 unsigned long *minp, unsigned long *maxp)
242 {
243 *minp = wb->bdi->min_ratio;
244 *maxp = wb->bdi->max_ratio;
245 }
246
247 #endif /* CONFIG_CGROUP_WRITEBACK */
248
249 /*
250 * In a memory zone, there is a certain amount of pages we consider
251 * available for the page cache, which is essentially the number of
252 * free and reclaimable pages, minus some zone reserves to protect
253 * lowmem and the ability to uphold the zone's watermarks without
254 * requiring writeback.
255 *
256 * This number of dirtyable pages is the base value of which the
257 * user-configurable dirty ratio is the effective number of pages that
258 * are allowed to be actually dirtied. Per individual zone, or
259 * globally by using the sum of dirtyable pages over all zones.
260 *
261 * Because the user is allowed to specify the dirty limit globally as
262 * absolute number of bytes, calculating the per-zone dirty limit can
263 * require translating the configured limit into a percentage of
264 * global dirtyable memory first.
265 */
266
267 /**
268 * node_dirtyable_memory - number of dirtyable pages in a node
269 * @pgdat: the node
270 *
271 * Return: the node's number of pages potentially available for dirty
272 * page cache. This is the base value for the per-node dirty limits.
273 */
node_dirtyable_memory(struct pglist_data * pgdat)274 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
275 {
276 unsigned long nr_pages = 0;
277 int z;
278
279 for (z = 0; z < MAX_NR_ZONES; z++) {
280 struct zone *zone = pgdat->node_zones + z;
281
282 if (!populated_zone(zone))
283 continue;
284
285 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
286 }
287
288 /*
289 * Pages reserved for the kernel should not be considered
290 * dirtyable, to prevent a situation where reclaim has to
291 * clean pages in order to balance the zones.
292 */
293 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
294
295 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
296 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
297
298 return nr_pages;
299 }
300
highmem_dirtyable_memory(unsigned long total)301 static unsigned long highmem_dirtyable_memory(unsigned long total)
302 {
303 #ifdef CONFIG_HIGHMEM
304 int node;
305 unsigned long x = 0;
306 int i;
307
308 for_each_node_state(node, N_HIGH_MEMORY) {
309 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
310 struct zone *z;
311 unsigned long nr_pages;
312
313 if (!is_highmem_idx(i))
314 continue;
315
316 z = &NODE_DATA(node)->node_zones[i];
317 if (!populated_zone(z))
318 continue;
319
320 nr_pages = zone_page_state(z, NR_FREE_PAGES);
321 /* watch for underflows */
322 nr_pages -= min(nr_pages, high_wmark_pages(z));
323 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
324 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
325 x += nr_pages;
326 }
327 }
328
329 /*
330 * Make sure that the number of highmem pages is never larger
331 * than the number of the total dirtyable memory. This can only
332 * occur in very strange VM situations but we want to make sure
333 * that this does not occur.
334 */
335 return min(x, total);
336 #else
337 return 0;
338 #endif
339 }
340
341 /**
342 * global_dirtyable_memory - number of globally dirtyable pages
343 *
344 * Return: the global number of pages potentially available for dirty
345 * page cache. This is the base value for the global dirty limits.
346 */
global_dirtyable_memory(void)347 static unsigned long global_dirtyable_memory(void)
348 {
349 unsigned long x;
350
351 x = global_zone_page_state(NR_FREE_PAGES);
352 /*
353 * Pages reserved for the kernel should not be considered
354 * dirtyable, to prevent a situation where reclaim has to
355 * clean pages in order to balance the zones.
356 */
357 x -= min(x, totalreserve_pages);
358
359 x += global_node_page_state(NR_INACTIVE_FILE);
360 x += global_node_page_state(NR_ACTIVE_FILE);
361
362 if (!vm_highmem_is_dirtyable)
363 x -= highmem_dirtyable_memory(x);
364
365 return x + 1; /* Ensure that we never return 0 */
366 }
367
368 /**
369 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
370 * @dtc: dirty_throttle_control of interest
371 *
372 * Calculate @dtc->thresh and ->bg_thresh considering
373 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
374 * must ensure that @dtc->avail is set before calling this function. The
375 * dirty limits will be lifted by 1/4 for real-time tasks.
376 */
domain_dirty_limits(struct dirty_throttle_control * dtc)377 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
378 {
379 const unsigned long available_memory = dtc->avail;
380 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
381 unsigned long bytes = vm_dirty_bytes;
382 unsigned long bg_bytes = dirty_background_bytes;
383 /* convert ratios to per-PAGE_SIZE for higher precision */
384 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
385 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
386 unsigned long thresh;
387 unsigned long bg_thresh;
388 struct task_struct *tsk;
389
390 /* gdtc is !NULL iff @dtc is for memcg domain */
391 if (gdtc) {
392 unsigned long global_avail = gdtc->avail;
393
394 /*
395 * The byte settings can't be applied directly to memcg
396 * domains. Convert them to ratios by scaling against
397 * globally available memory. As the ratios are in
398 * per-PAGE_SIZE, they can be obtained by dividing bytes by
399 * number of pages.
400 */
401 if (bytes)
402 ratio = min(DIV_ROUND_UP(bytes, global_avail),
403 PAGE_SIZE);
404 if (bg_bytes)
405 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
406 PAGE_SIZE);
407 bytes = bg_bytes = 0;
408 }
409
410 if (bytes)
411 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
412 else
413 thresh = (ratio * available_memory) / PAGE_SIZE;
414
415 if (bg_bytes)
416 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
417 else
418 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
419
420 tsk = current;
421 if (rt_or_dl_task(tsk)) {
422 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
423 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
424 }
425 /*
426 * Dirty throttling logic assumes the limits in page units fit into
427 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
428 */
429 if (thresh > UINT_MAX)
430 thresh = UINT_MAX;
431 /* This makes sure bg_thresh is within 32-bits as well */
432 if (bg_thresh >= thresh)
433 bg_thresh = thresh / 2;
434 dtc->thresh = thresh;
435 dtc->bg_thresh = bg_thresh;
436
437 /* we should eventually report the domain in the TP */
438 if (!gdtc)
439 trace_global_dirty_state(bg_thresh, thresh);
440 }
441
442 /**
443 * global_dirty_limits - background-writeback and dirty-throttling thresholds
444 * @pbackground: out parameter for bg_thresh
445 * @pdirty: out parameter for thresh
446 *
447 * Calculate bg_thresh and thresh for global_wb_domain. See
448 * domain_dirty_limits() for details.
449 */
global_dirty_limits(unsigned long * pbackground,unsigned long * pdirty)450 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
451 {
452 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
453
454 gdtc.avail = global_dirtyable_memory();
455 domain_dirty_limits(&gdtc);
456
457 *pbackground = gdtc.bg_thresh;
458 *pdirty = gdtc.thresh;
459 }
460
461 /**
462 * node_dirty_limit - maximum number of dirty pages allowed in a node
463 * @pgdat: the node
464 *
465 * Return: the maximum number of dirty pages allowed in a node, based
466 * on the node's dirtyable memory.
467 */
node_dirty_limit(struct pglist_data * pgdat)468 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
469 {
470 unsigned long node_memory = node_dirtyable_memory(pgdat);
471 struct task_struct *tsk = current;
472 unsigned long dirty;
473
474 if (vm_dirty_bytes)
475 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
476 node_memory / global_dirtyable_memory();
477 else
478 dirty = vm_dirty_ratio * node_memory / 100;
479
480 if (rt_or_dl_task(tsk))
481 dirty += dirty / 4;
482
483 /*
484 * Dirty throttling logic assumes the limits in page units fit into
485 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
486 */
487 return min_t(unsigned long, dirty, UINT_MAX);
488 }
489
490 /**
491 * node_dirty_ok - tells whether a node is within its dirty limits
492 * @pgdat: the node to check
493 *
494 * Return: %true when the dirty pages in @pgdat are within the node's
495 * dirty limit, %false if the limit is exceeded.
496 */
node_dirty_ok(struct pglist_data * pgdat)497 bool node_dirty_ok(struct pglist_data *pgdat)
498 {
499 unsigned long limit = node_dirty_limit(pgdat);
500 unsigned long nr_pages = 0;
501
502 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
503 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
504
505 return nr_pages <= limit;
506 }
507
508 #ifdef CONFIG_SYSCTL
dirty_background_ratio_handler(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)509 static int dirty_background_ratio_handler(const struct ctl_table *table, int write,
510 void *buffer, size_t *lenp, loff_t *ppos)
511 {
512 int ret;
513
514 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
515 if (ret == 0 && write)
516 dirty_background_bytes = 0;
517 return ret;
518 }
519
dirty_background_bytes_handler(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)520 static int dirty_background_bytes_handler(const struct ctl_table *table, int write,
521 void *buffer, size_t *lenp, loff_t *ppos)
522 {
523 int ret;
524 unsigned long old_bytes = dirty_background_bytes;
525
526 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
527 if (ret == 0 && write) {
528 if (DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE) >
529 UINT_MAX) {
530 dirty_background_bytes = old_bytes;
531 return -ERANGE;
532 }
533 dirty_background_ratio = 0;
534 }
535 return ret;
536 }
537
dirty_ratio_handler(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)538 static int dirty_ratio_handler(const struct ctl_table *table, int write, void *buffer,
539 size_t *lenp, loff_t *ppos)
540 {
541 int old_ratio = vm_dirty_ratio;
542 int ret;
543
544 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
545 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
546 writeback_set_ratelimit();
547 vm_dirty_bytes = 0;
548 }
549 return ret;
550 }
551
dirty_bytes_handler(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)552 static int dirty_bytes_handler(const struct ctl_table *table, int write,
553 void *buffer, size_t *lenp, loff_t *ppos)
554 {
555 unsigned long old_bytes = vm_dirty_bytes;
556 int ret;
557
558 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
559 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
560 if (DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) > UINT_MAX) {
561 vm_dirty_bytes = old_bytes;
562 return -ERANGE;
563 }
564 writeback_set_ratelimit();
565 vm_dirty_ratio = 0;
566 }
567 return ret;
568 }
569 #endif
570
wp_next_time(unsigned long cur_time)571 static unsigned long wp_next_time(unsigned long cur_time)
572 {
573 cur_time += VM_COMPLETIONS_PERIOD_LEN;
574 /* 0 has a special meaning... */
575 if (!cur_time)
576 return 1;
577 return cur_time;
578 }
579
wb_domain_writeout_add(struct wb_domain * dom,struct fprop_local_percpu * completions,unsigned int max_prop_frac,long nr)580 static void wb_domain_writeout_add(struct wb_domain *dom,
581 struct fprop_local_percpu *completions,
582 unsigned int max_prop_frac, long nr)
583 {
584 __fprop_add_percpu_max(&dom->completions, completions,
585 max_prop_frac, nr);
586 /* First event after period switching was turned off? */
587 if (unlikely(!dom->period_time)) {
588 /*
589 * We can race with other wb_domain_writeout_add calls here but
590 * it does not cause any harm since the resulting time when
591 * timer will fire and what is in writeout_period_time will be
592 * roughly the same.
593 */
594 dom->period_time = wp_next_time(jiffies);
595 mod_timer(&dom->period_timer, dom->period_time);
596 }
597 }
598
599 /*
600 * Increment @wb's writeout completion count and the global writeout
601 * completion count. Called from __folio_end_writeback().
602 */
__wb_writeout_add(struct bdi_writeback * wb,long nr)603 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
604 {
605 struct wb_domain *cgdom;
606
607 wb_stat_mod(wb, WB_WRITTEN, nr);
608 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
609 wb->bdi->max_prop_frac, nr);
610
611 cgdom = mem_cgroup_wb_domain(wb);
612 if (cgdom)
613 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
614 wb->bdi->max_prop_frac, nr);
615 }
616
wb_writeout_inc(struct bdi_writeback * wb)617 void wb_writeout_inc(struct bdi_writeback *wb)
618 {
619 unsigned long flags;
620
621 local_irq_save(flags);
622 __wb_writeout_add(wb, 1);
623 local_irq_restore(flags);
624 }
625 EXPORT_SYMBOL_GPL(wb_writeout_inc);
626
627 /*
628 * On idle system, we can be called long after we scheduled because we use
629 * deferred timers so count with missed periods.
630 */
writeout_period(struct timer_list * t)631 static void writeout_period(struct timer_list *t)
632 {
633 struct wb_domain *dom = from_timer(dom, t, period_timer);
634 int miss_periods = (jiffies - dom->period_time) /
635 VM_COMPLETIONS_PERIOD_LEN;
636
637 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
638 dom->period_time = wp_next_time(dom->period_time +
639 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
640 mod_timer(&dom->period_timer, dom->period_time);
641 } else {
642 /*
643 * Aging has zeroed all fractions. Stop wasting CPU on period
644 * updates.
645 */
646 dom->period_time = 0;
647 }
648 }
649
wb_domain_init(struct wb_domain * dom,gfp_t gfp)650 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
651 {
652 memset(dom, 0, sizeof(*dom));
653
654 spin_lock_init(&dom->lock);
655
656 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
657
658 dom->dirty_limit_tstamp = jiffies;
659
660 return fprop_global_init(&dom->completions, gfp);
661 }
662
663 #ifdef CONFIG_CGROUP_WRITEBACK
wb_domain_exit(struct wb_domain * dom)664 void wb_domain_exit(struct wb_domain *dom)
665 {
666 del_timer_sync(&dom->period_timer);
667 fprop_global_destroy(&dom->completions);
668 }
669 #endif
670
671 /*
672 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
673 * registered backing devices, which, for obvious reasons, can not
674 * exceed 100%.
675 */
676 static unsigned int bdi_min_ratio;
677
bdi_check_pages_limit(unsigned long pages)678 static int bdi_check_pages_limit(unsigned long pages)
679 {
680 unsigned long max_dirty_pages = global_dirtyable_memory();
681
682 if (pages > max_dirty_pages)
683 return -EINVAL;
684
685 return 0;
686 }
687
bdi_ratio_from_pages(unsigned long pages)688 static unsigned long bdi_ratio_from_pages(unsigned long pages)
689 {
690 unsigned long background_thresh;
691 unsigned long dirty_thresh;
692 unsigned long ratio;
693
694 global_dirty_limits(&background_thresh, &dirty_thresh);
695 if (!dirty_thresh)
696 return -EINVAL;
697 ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
698
699 return ratio;
700 }
701
bdi_get_bytes(unsigned int ratio)702 static u64 bdi_get_bytes(unsigned int ratio)
703 {
704 unsigned long background_thresh;
705 unsigned long dirty_thresh;
706 u64 bytes;
707
708 global_dirty_limits(&background_thresh, &dirty_thresh);
709 bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
710
711 return bytes;
712 }
713
__bdi_set_min_ratio(struct backing_dev_info * bdi,unsigned int min_ratio)714 static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
715 {
716 unsigned int delta;
717 int ret = 0;
718
719 if (min_ratio > 100 * BDI_RATIO_SCALE)
720 return -EINVAL;
721
722 spin_lock_bh(&bdi_lock);
723 if (min_ratio > bdi->max_ratio) {
724 ret = -EINVAL;
725 } else {
726 if (min_ratio < bdi->min_ratio) {
727 delta = bdi->min_ratio - min_ratio;
728 bdi_min_ratio -= delta;
729 bdi->min_ratio = min_ratio;
730 } else {
731 delta = min_ratio - bdi->min_ratio;
732 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
733 bdi_min_ratio += delta;
734 bdi->min_ratio = min_ratio;
735 } else {
736 ret = -EINVAL;
737 }
738 }
739 }
740 spin_unlock_bh(&bdi_lock);
741
742 return ret;
743 }
744
__bdi_set_max_ratio(struct backing_dev_info * bdi,unsigned int max_ratio)745 static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
746 {
747 int ret = 0;
748
749 if (max_ratio > 100 * BDI_RATIO_SCALE)
750 return -EINVAL;
751
752 spin_lock_bh(&bdi_lock);
753 if (bdi->min_ratio > max_ratio) {
754 ret = -EINVAL;
755 } else {
756 bdi->max_ratio = max_ratio;
757 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
758 (100 * BDI_RATIO_SCALE);
759 }
760 spin_unlock_bh(&bdi_lock);
761
762 return ret;
763 }
764
bdi_set_min_ratio_no_scale(struct backing_dev_info * bdi,unsigned int min_ratio)765 int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
766 {
767 return __bdi_set_min_ratio(bdi, min_ratio);
768 }
769
bdi_set_max_ratio_no_scale(struct backing_dev_info * bdi,unsigned int max_ratio)770 int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
771 {
772 return __bdi_set_max_ratio(bdi, max_ratio);
773 }
774
bdi_set_min_ratio(struct backing_dev_info * bdi,unsigned int min_ratio)775 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
776 {
777 return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
778 }
779
bdi_set_max_ratio(struct backing_dev_info * bdi,unsigned int max_ratio)780 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
781 {
782 return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
783 }
784 EXPORT_SYMBOL(bdi_set_max_ratio);
785
bdi_get_min_bytes(struct backing_dev_info * bdi)786 u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
787 {
788 return bdi_get_bytes(bdi->min_ratio);
789 }
790
bdi_set_min_bytes(struct backing_dev_info * bdi,u64 min_bytes)791 int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
792 {
793 int ret;
794 unsigned long pages = min_bytes >> PAGE_SHIFT;
795 long min_ratio;
796
797 ret = bdi_check_pages_limit(pages);
798 if (ret)
799 return ret;
800
801 min_ratio = bdi_ratio_from_pages(pages);
802 if (min_ratio < 0)
803 return min_ratio;
804 return __bdi_set_min_ratio(bdi, min_ratio);
805 }
806
bdi_get_max_bytes(struct backing_dev_info * bdi)807 u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
808 {
809 return bdi_get_bytes(bdi->max_ratio);
810 }
811
bdi_set_max_bytes(struct backing_dev_info * bdi,u64 max_bytes)812 int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
813 {
814 int ret;
815 unsigned long pages = max_bytes >> PAGE_SHIFT;
816 long max_ratio;
817
818 ret = bdi_check_pages_limit(pages);
819 if (ret)
820 return ret;
821
822 max_ratio = bdi_ratio_from_pages(pages);
823 if (max_ratio < 0)
824 return max_ratio;
825 return __bdi_set_max_ratio(bdi, max_ratio);
826 }
827
bdi_set_strict_limit(struct backing_dev_info * bdi,unsigned int strict_limit)828 int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
829 {
830 if (strict_limit > 1)
831 return -EINVAL;
832
833 spin_lock_bh(&bdi_lock);
834 if (strict_limit)
835 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
836 else
837 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
838 spin_unlock_bh(&bdi_lock);
839
840 return 0;
841 }
842
dirty_freerun_ceiling(unsigned long thresh,unsigned long bg_thresh)843 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
844 unsigned long bg_thresh)
845 {
846 return (thresh + bg_thresh) / 2;
847 }
848
hard_dirty_limit(struct wb_domain * dom,unsigned long thresh)849 static unsigned long hard_dirty_limit(struct wb_domain *dom,
850 unsigned long thresh)
851 {
852 return max(thresh, dom->dirty_limit);
853 }
854
855 /*
856 * Memory which can be further allocated to a memcg domain is capped by
857 * system-wide clean memory excluding the amount being used in the domain.
858 */
mdtc_calc_avail(struct dirty_throttle_control * mdtc,unsigned long filepages,unsigned long headroom)859 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
860 unsigned long filepages, unsigned long headroom)
861 {
862 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
863 unsigned long clean = filepages - min(filepages, mdtc->dirty);
864 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
865 unsigned long other_clean = global_clean - min(global_clean, clean);
866
867 mdtc->avail = filepages + min(headroom, other_clean);
868 }
869
dtc_is_global(struct dirty_throttle_control * dtc)870 static inline bool dtc_is_global(struct dirty_throttle_control *dtc)
871 {
872 return mdtc_gdtc(dtc) == NULL;
873 }
874
875 /*
876 * Dirty background will ignore pages being written as we're trying to
877 * decide whether to put more under writeback.
878 */
domain_dirty_avail(struct dirty_throttle_control * dtc,bool include_writeback)879 static void domain_dirty_avail(struct dirty_throttle_control *dtc,
880 bool include_writeback)
881 {
882 if (dtc_is_global(dtc)) {
883 dtc->avail = global_dirtyable_memory();
884 dtc->dirty = global_node_page_state(NR_FILE_DIRTY);
885 if (include_writeback)
886 dtc->dirty += global_node_page_state(NR_WRITEBACK);
887 } else {
888 unsigned long filepages = 0, headroom = 0, writeback = 0;
889
890 mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty,
891 &writeback);
892 if (include_writeback)
893 dtc->dirty += writeback;
894 mdtc_calc_avail(dtc, filepages, headroom);
895 }
896 }
897
898 /**
899 * __wb_calc_thresh - @wb's share of dirty threshold
900 * @dtc: dirty_throttle_context of interest
901 * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
902 *
903 * Note that balance_dirty_pages() will only seriously take dirty throttling
904 * threshold as a hard limit when sleeping max_pause per page is not enough
905 * to keep the dirty pages under control. For example, when the device is
906 * completely stalled due to some error conditions, or when there are 1000
907 * dd tasks writing to a slow 10MB/s USB key.
908 * In the other normal situations, it acts more gently by throttling the tasks
909 * more (rather than completely block them) when the wb dirty pages go high.
910 *
911 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
912 * - starving fast devices
913 * - piling up dirty pages (that will take long time to sync) on slow devices
914 *
915 * The wb's share of dirty limit will be adapting to its throughput and
916 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
917 *
918 * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
919 * "dirty" in the context of dirty balancing includes all PG_dirty and
920 * PG_writeback pages.
921 */
__wb_calc_thresh(struct dirty_throttle_control * dtc,unsigned long thresh)922 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
923 unsigned long thresh)
924 {
925 struct wb_domain *dom = dtc_dom(dtc);
926 struct bdi_writeback *wb = dtc->wb;
927 u64 wb_thresh;
928 u64 wb_max_thresh;
929 unsigned long numerator, denominator;
930 unsigned long wb_min_ratio, wb_max_ratio;
931
932 /*
933 * Calculate this wb's share of the thresh ratio.
934 */
935 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
936 &numerator, &denominator);
937
938 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
939 wb_thresh *= numerator;
940 wb_thresh = div64_ul(wb_thresh, denominator);
941
942 wb_min_max_ratio(wb, &wb_min_ratio, &wb_max_ratio);
943
944 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
945
946 /*
947 * It's very possible that wb_thresh is close to 0 not because the
948 * device is slow, but that it has remained inactive for long time.
949 * Honour such devices a reasonable good (hopefully IO efficient)
950 * threshold, so that the occasional writes won't be blocked and active
951 * writes can rampup the threshold quickly.
952 */
953 if (thresh > dtc->dirty) {
954 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT))
955 wb_thresh = max(wb_thresh, (thresh - dtc->dirty) / 100);
956 else
957 wb_thresh = max(wb_thresh, (thresh - dtc->dirty) / 8);
958 }
959
960 wb_max_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
961 if (wb_thresh > wb_max_thresh)
962 wb_thresh = wb_max_thresh;
963
964 return wb_thresh;
965 }
966
wb_calc_thresh(struct bdi_writeback * wb,unsigned long thresh)967 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
968 {
969 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
970
971 domain_dirty_avail(&gdtc, true);
972 return __wb_calc_thresh(&gdtc, thresh);
973 }
974
cgwb_calc_thresh(struct bdi_writeback * wb)975 unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
976 {
977 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
978 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
979
980 domain_dirty_avail(&gdtc, true);
981 domain_dirty_avail(&mdtc, true);
982 domain_dirty_limits(&mdtc);
983
984 return __wb_calc_thresh(&mdtc, mdtc.thresh);
985 }
986
987 /*
988 * setpoint - dirty 3
989 * f(dirty) := 1.0 + (----------------)
990 * limit - setpoint
991 *
992 * it's a 3rd order polynomial that subjects to
993 *
994 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
995 * (2) f(setpoint) = 1.0 => the balance point
996 * (3) f(limit) = 0 => the hard limit
997 * (4) df/dx <= 0 => negative feedback control
998 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
999 * => fast response on large errors; small oscillation near setpoint
1000 */
pos_ratio_polynom(unsigned long setpoint,unsigned long dirty,unsigned long limit)1001 static long long pos_ratio_polynom(unsigned long setpoint,
1002 unsigned long dirty,
1003 unsigned long limit)
1004 {
1005 long long pos_ratio;
1006 long x;
1007
1008 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
1009 (limit - setpoint) | 1);
1010 pos_ratio = x;
1011 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1012 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1013 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
1014
1015 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
1016 }
1017
1018 /*
1019 * Dirty position control.
1020 *
1021 * (o) global/bdi setpoints
1022 *
1023 * We want the dirty pages be balanced around the global/wb setpoints.
1024 * When the number of dirty pages is higher/lower than the setpoint, the
1025 * dirty position control ratio (and hence task dirty ratelimit) will be
1026 * decreased/increased to bring the dirty pages back to the setpoint.
1027 *
1028 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
1029 *
1030 * if (dirty < setpoint) scale up pos_ratio
1031 * if (dirty > setpoint) scale down pos_ratio
1032 *
1033 * if (wb_dirty < wb_setpoint) scale up pos_ratio
1034 * if (wb_dirty > wb_setpoint) scale down pos_ratio
1035 *
1036 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
1037 *
1038 * (o) global control line
1039 *
1040 * ^ pos_ratio
1041 * |
1042 * | |<===== global dirty control scope ======>|
1043 * 2.0 * * * * * * *
1044 * | .*
1045 * | . *
1046 * | . *
1047 * | . *
1048 * | . *
1049 * | . *
1050 * 1.0 ................................*
1051 * | . . *
1052 * | . . *
1053 * | . . *
1054 * | . . *
1055 * | . . *
1056 * 0 +------------.------------------.----------------------*------------->
1057 * freerun^ setpoint^ limit^ dirty pages
1058 *
1059 * (o) wb control line
1060 *
1061 * ^ pos_ratio
1062 * |
1063 * | *
1064 * | *
1065 * | *
1066 * | *
1067 * | * |<=========== span ============>|
1068 * 1.0 .......................*
1069 * | . *
1070 * | . *
1071 * | . *
1072 * | . *
1073 * | . *
1074 * | . *
1075 * | . *
1076 * | . *
1077 * | . *
1078 * | . *
1079 * | . *
1080 * 1/4 ...............................................* * * * * * * * * * * *
1081 * | . .
1082 * | . .
1083 * | . .
1084 * 0 +----------------------.-------------------------------.------------->
1085 * wb_setpoint^ x_intercept^
1086 *
1087 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
1088 * be smoothly throttled down to normal if it starts high in situations like
1089 * - start writing to a slow SD card and a fast disk at the same time. The SD
1090 * card's wb_dirty may rush to many times higher than wb_setpoint.
1091 * - the wb dirty thresh drops quickly due to change of JBOD workload
1092 */
wb_position_ratio(struct dirty_throttle_control * dtc)1093 static void wb_position_ratio(struct dirty_throttle_control *dtc)
1094 {
1095 struct bdi_writeback *wb = dtc->wb;
1096 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
1097 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1098 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1099 unsigned long wb_thresh = dtc->wb_thresh;
1100 unsigned long x_intercept;
1101 unsigned long setpoint; /* dirty pages' target balance point */
1102 unsigned long wb_setpoint;
1103 unsigned long span;
1104 long long pos_ratio; /* for scaling up/down the rate limit */
1105 long x;
1106
1107 dtc->pos_ratio = 0;
1108
1109 if (unlikely(dtc->dirty >= limit))
1110 return;
1111
1112 /*
1113 * global setpoint
1114 *
1115 * See comment for pos_ratio_polynom().
1116 */
1117 setpoint = (freerun + limit) / 2;
1118 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
1119
1120 /*
1121 * The strictlimit feature is a tool preventing mistrusted filesystems
1122 * from growing a large number of dirty pages before throttling. For
1123 * such filesystems balance_dirty_pages always checks wb counters
1124 * against wb limits. Even if global "nr_dirty" is under "freerun".
1125 * This is especially important for fuse which sets bdi->max_ratio to
1126 * 1% by default. Without strictlimit feature, fuse writeback may
1127 * consume arbitrary amount of RAM because it is accounted in
1128 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1129 *
1130 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1131 * two values: wb_dirty and wb_thresh. Let's consider an example:
1132 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1133 * limits are set by default to 10% and 20% (background and throttle).
1134 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1135 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1136 * about ~6K pages (as the average of background and throttle wb
1137 * limits). The 3rd order polynomial will provide positive feedback if
1138 * wb_dirty is under wb_setpoint and vice versa.
1139 *
1140 * Note, that we cannot use global counters in these calculations
1141 * because we want to throttle process writing to a strictlimit wb
1142 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1143 * in the example above).
1144 */
1145 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1146 long long wb_pos_ratio;
1147
1148 if (dtc->wb_dirty >= wb_thresh)
1149 return;
1150
1151 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1152 dtc->wb_bg_thresh);
1153
1154 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1155 return;
1156
1157 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1158 wb_thresh);
1159
1160 /*
1161 * Typically, for strictlimit case, wb_setpoint << setpoint
1162 * and pos_ratio >> wb_pos_ratio. In the other words global
1163 * state ("dirty") is not limiting factor and we have to
1164 * make decision based on wb counters. But there is an
1165 * important case when global pos_ratio should get precedence:
1166 * global limits are exceeded (e.g. due to activities on other
1167 * wb's) while given strictlimit wb is below limit.
1168 *
1169 * "pos_ratio * wb_pos_ratio" would work for the case above,
1170 * but it would look too non-natural for the case of all
1171 * activity in the system coming from a single strictlimit wb
1172 * with bdi->max_ratio == 100%.
1173 *
1174 * Note that min() below somewhat changes the dynamics of the
1175 * control system. Normally, pos_ratio value can be well over 3
1176 * (when globally we are at freerun and wb is well below wb
1177 * setpoint). Now the maximum pos_ratio in the same situation
1178 * is 2. We might want to tweak this if we observe the control
1179 * system is too slow to adapt.
1180 */
1181 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1182 return;
1183 }
1184
1185 /*
1186 * We have computed basic pos_ratio above based on global situation. If
1187 * the wb is over/under its share of dirty pages, we want to scale
1188 * pos_ratio further down/up. That is done by the following mechanism.
1189 */
1190
1191 /*
1192 * wb setpoint
1193 *
1194 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1195 *
1196 * x_intercept - wb_dirty
1197 * := --------------------------
1198 * x_intercept - wb_setpoint
1199 *
1200 * The main wb control line is a linear function that subjects to
1201 *
1202 * (1) f(wb_setpoint) = 1.0
1203 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1204 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1205 *
1206 * For single wb case, the dirty pages are observed to fluctuate
1207 * regularly within range
1208 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1209 * for various filesystems, where (2) can yield in a reasonable 12.5%
1210 * fluctuation range for pos_ratio.
1211 *
1212 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1213 * own size, so move the slope over accordingly and choose a slope that
1214 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1215 */
1216 if (unlikely(wb_thresh > dtc->thresh))
1217 wb_thresh = dtc->thresh;
1218 /*
1219 * scale global setpoint to wb's:
1220 * wb_setpoint = setpoint * wb_thresh / thresh
1221 */
1222 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1223 wb_setpoint = setpoint * (u64)x >> 16;
1224 /*
1225 * Use span=(8*write_bw) in single wb case as indicated by
1226 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1227 *
1228 * wb_thresh thresh - wb_thresh
1229 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1230 * thresh thresh
1231 */
1232 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1233 x_intercept = wb_setpoint + span;
1234
1235 if (dtc->wb_dirty < x_intercept - span / 4) {
1236 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1237 (x_intercept - wb_setpoint) | 1);
1238 } else
1239 pos_ratio /= 4;
1240
1241 /*
1242 * wb reserve area, safeguard against dirty pool underrun and disk idle
1243 * It may push the desired control point of global dirty pages higher
1244 * than setpoint.
1245 */
1246 x_intercept = wb_thresh / 2;
1247 if (dtc->wb_dirty < x_intercept) {
1248 if (dtc->wb_dirty > x_intercept / 8)
1249 pos_ratio = div_u64(pos_ratio * x_intercept,
1250 dtc->wb_dirty);
1251 else
1252 pos_ratio *= 8;
1253 }
1254
1255 dtc->pos_ratio = pos_ratio;
1256 }
1257
wb_update_write_bandwidth(struct bdi_writeback * wb,unsigned long elapsed,unsigned long written)1258 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1259 unsigned long elapsed,
1260 unsigned long written)
1261 {
1262 const unsigned long period = roundup_pow_of_two(3 * HZ);
1263 unsigned long avg = wb->avg_write_bandwidth;
1264 unsigned long old = wb->write_bandwidth;
1265 u64 bw;
1266
1267 /*
1268 * bw = written * HZ / elapsed
1269 *
1270 * bw * elapsed + write_bandwidth * (period - elapsed)
1271 * write_bandwidth = ---------------------------------------------------
1272 * period
1273 *
1274 * @written may have decreased due to folio_redirty_for_writepage().
1275 * Avoid underflowing @bw calculation.
1276 */
1277 bw = written - min(written, wb->written_stamp);
1278 bw *= HZ;
1279 if (unlikely(elapsed > period)) {
1280 bw = div64_ul(bw, elapsed);
1281 avg = bw;
1282 goto out;
1283 }
1284 bw += (u64)wb->write_bandwidth * (period - elapsed);
1285 bw >>= ilog2(period);
1286
1287 /*
1288 * one more level of smoothing, for filtering out sudden spikes
1289 */
1290 if (avg > old && old >= (unsigned long)bw)
1291 avg -= (avg - old) >> 3;
1292
1293 if (avg < old && old <= (unsigned long)bw)
1294 avg += (old - avg) >> 3;
1295
1296 out:
1297 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1298 avg = max(avg, 1LU);
1299 if (wb_has_dirty_io(wb)) {
1300 long delta = avg - wb->avg_write_bandwidth;
1301 WARN_ON_ONCE(atomic_long_add_return(delta,
1302 &wb->bdi->tot_write_bandwidth) <= 0);
1303 }
1304 wb->write_bandwidth = bw;
1305 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1306 }
1307
update_dirty_limit(struct dirty_throttle_control * dtc)1308 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1309 {
1310 struct wb_domain *dom = dtc_dom(dtc);
1311 unsigned long thresh = dtc->thresh;
1312 unsigned long limit = dom->dirty_limit;
1313
1314 /*
1315 * Follow up in one step.
1316 */
1317 if (limit < thresh) {
1318 limit = thresh;
1319 goto update;
1320 }
1321
1322 /*
1323 * Follow down slowly. Use the higher one as the target, because thresh
1324 * may drop below dirty. This is exactly the reason to introduce
1325 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1326 */
1327 thresh = max(thresh, dtc->dirty);
1328 if (limit > thresh) {
1329 limit -= (limit - thresh) >> 5;
1330 goto update;
1331 }
1332 return;
1333 update:
1334 dom->dirty_limit = limit;
1335 }
1336
domain_update_dirty_limit(struct dirty_throttle_control * dtc,unsigned long now)1337 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1338 unsigned long now)
1339 {
1340 struct wb_domain *dom = dtc_dom(dtc);
1341
1342 /*
1343 * check locklessly first to optimize away locking for the most time
1344 */
1345 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1346 return;
1347
1348 spin_lock(&dom->lock);
1349 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1350 update_dirty_limit(dtc);
1351 dom->dirty_limit_tstamp = now;
1352 }
1353 spin_unlock(&dom->lock);
1354 }
1355
1356 /*
1357 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1358 *
1359 * Normal wb tasks will be curbed at or below it in long term.
1360 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1361 */
wb_update_dirty_ratelimit(struct dirty_throttle_control * dtc,unsigned long dirtied,unsigned long elapsed)1362 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1363 unsigned long dirtied,
1364 unsigned long elapsed)
1365 {
1366 struct bdi_writeback *wb = dtc->wb;
1367 unsigned long dirty = dtc->dirty;
1368 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1369 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1370 unsigned long setpoint = (freerun + limit) / 2;
1371 unsigned long write_bw = wb->avg_write_bandwidth;
1372 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1373 unsigned long dirty_rate;
1374 unsigned long task_ratelimit;
1375 unsigned long balanced_dirty_ratelimit;
1376 unsigned long step;
1377 unsigned long x;
1378 unsigned long shift;
1379
1380 /*
1381 * The dirty rate will match the writeout rate in long term, except
1382 * when dirty pages are truncated by userspace or re-dirtied by FS.
1383 */
1384 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1385
1386 /*
1387 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1388 */
1389 task_ratelimit = (u64)dirty_ratelimit *
1390 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1391 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1392
1393 /*
1394 * A linear estimation of the "balanced" throttle rate. The theory is,
1395 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1396 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1397 * formula will yield the balanced rate limit (write_bw / N).
1398 *
1399 * Note that the expanded form is not a pure rate feedback:
1400 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1401 * but also takes pos_ratio into account:
1402 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1403 *
1404 * (1) is not realistic because pos_ratio also takes part in balancing
1405 * the dirty rate. Consider the state
1406 * pos_ratio = 0.5 (3)
1407 * rate = 2 * (write_bw / N) (4)
1408 * If (1) is used, it will stuck in that state! Because each dd will
1409 * be throttled at
1410 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1411 * yielding
1412 * dirty_rate = N * task_ratelimit = write_bw (6)
1413 * put (6) into (1) we get
1414 * rate_(i+1) = rate_(i) (7)
1415 *
1416 * So we end up using (2) to always keep
1417 * rate_(i+1) ~= (write_bw / N) (8)
1418 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1419 * pos_ratio is able to drive itself to 1.0, which is not only where
1420 * the dirty count meet the setpoint, but also where the slope of
1421 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1422 */
1423 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1424 dirty_rate | 1);
1425 /*
1426 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1427 */
1428 if (unlikely(balanced_dirty_ratelimit > write_bw))
1429 balanced_dirty_ratelimit = write_bw;
1430
1431 /*
1432 * We could safely do this and return immediately:
1433 *
1434 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1435 *
1436 * However to get a more stable dirty_ratelimit, the below elaborated
1437 * code makes use of task_ratelimit to filter out singular points and
1438 * limit the step size.
1439 *
1440 * The below code essentially only uses the relative value of
1441 *
1442 * task_ratelimit - dirty_ratelimit
1443 * = (pos_ratio - 1) * dirty_ratelimit
1444 *
1445 * which reflects the direction and size of dirty position error.
1446 */
1447
1448 /*
1449 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1450 * task_ratelimit is on the same side of dirty_ratelimit, too.
1451 * For example, when
1452 * - dirty_ratelimit > balanced_dirty_ratelimit
1453 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1454 * lowering dirty_ratelimit will help meet both the position and rate
1455 * control targets. Otherwise, don't update dirty_ratelimit if it will
1456 * only help meet the rate target. After all, what the users ultimately
1457 * feel and care are stable dirty rate and small position error.
1458 *
1459 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1460 * and filter out the singular points of balanced_dirty_ratelimit. Which
1461 * keeps jumping around randomly and can even leap far away at times
1462 * due to the small 200ms estimation period of dirty_rate (we want to
1463 * keep that period small to reduce time lags).
1464 */
1465 step = 0;
1466
1467 /*
1468 * For strictlimit case, calculations above were based on wb counters
1469 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1470 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1471 * Hence, to calculate "step" properly, we have to use wb_dirty as
1472 * "dirty" and wb_setpoint as "setpoint".
1473 */
1474 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1475 dirty = dtc->wb_dirty;
1476 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1477 }
1478
1479 if (dirty < setpoint) {
1480 x = min3(wb->balanced_dirty_ratelimit,
1481 balanced_dirty_ratelimit, task_ratelimit);
1482 if (dirty_ratelimit < x)
1483 step = x - dirty_ratelimit;
1484 } else {
1485 x = max3(wb->balanced_dirty_ratelimit,
1486 balanced_dirty_ratelimit, task_ratelimit);
1487 if (dirty_ratelimit > x)
1488 step = dirty_ratelimit - x;
1489 }
1490
1491 /*
1492 * Don't pursue 100% rate matching. It's impossible since the balanced
1493 * rate itself is constantly fluctuating. So decrease the track speed
1494 * when it gets close to the target. Helps eliminate pointless tremors.
1495 */
1496 shift = dirty_ratelimit / (2 * step + 1);
1497 if (shift < BITS_PER_LONG)
1498 step = DIV_ROUND_UP(step >> shift, 8);
1499 else
1500 step = 0;
1501
1502 if (dirty_ratelimit < balanced_dirty_ratelimit)
1503 dirty_ratelimit += step;
1504 else
1505 dirty_ratelimit -= step;
1506
1507 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1508 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1509
1510 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1511 }
1512
__wb_update_bandwidth(struct dirty_throttle_control * gdtc,struct dirty_throttle_control * mdtc,bool update_ratelimit)1513 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1514 struct dirty_throttle_control *mdtc,
1515 bool update_ratelimit)
1516 {
1517 struct bdi_writeback *wb = gdtc->wb;
1518 unsigned long now = jiffies;
1519 unsigned long elapsed;
1520 unsigned long dirtied;
1521 unsigned long written;
1522
1523 spin_lock(&wb->list_lock);
1524
1525 /*
1526 * Lockless checks for elapsed time are racy and delayed update after
1527 * IO completion doesn't do it at all (to make sure written pages are
1528 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1529 * division errors.
1530 */
1531 elapsed = max(now - wb->bw_time_stamp, 1UL);
1532 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1533 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1534
1535 if (update_ratelimit) {
1536 domain_update_dirty_limit(gdtc, now);
1537 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1538
1539 /*
1540 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1541 * compiler has no way to figure that out. Help it.
1542 */
1543 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1544 domain_update_dirty_limit(mdtc, now);
1545 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1546 }
1547 }
1548 wb_update_write_bandwidth(wb, elapsed, written);
1549
1550 wb->dirtied_stamp = dirtied;
1551 wb->written_stamp = written;
1552 WRITE_ONCE(wb->bw_time_stamp, now);
1553 spin_unlock(&wb->list_lock);
1554 }
1555
wb_update_bandwidth(struct bdi_writeback * wb)1556 void wb_update_bandwidth(struct bdi_writeback *wb)
1557 {
1558 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1559
1560 __wb_update_bandwidth(&gdtc, NULL, false);
1561 }
1562
1563 /* Interval after which we consider wb idle and don't estimate bandwidth */
1564 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1565
wb_bandwidth_estimate_start(struct bdi_writeback * wb)1566 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1567 {
1568 unsigned long now = jiffies;
1569 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1570
1571 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1572 !atomic_read(&wb->writeback_inodes)) {
1573 spin_lock(&wb->list_lock);
1574 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1575 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1576 WRITE_ONCE(wb->bw_time_stamp, now);
1577 spin_unlock(&wb->list_lock);
1578 }
1579 }
1580
1581 /*
1582 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1583 * will look to see if it needs to start dirty throttling.
1584 *
1585 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1586 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1587 * (the number of pages we may dirty without exceeding the dirty limits).
1588 */
dirty_poll_interval(unsigned long dirty,unsigned long thresh)1589 static unsigned long dirty_poll_interval(unsigned long dirty,
1590 unsigned long thresh)
1591 {
1592 if (thresh > dirty)
1593 return 1UL << (ilog2(thresh - dirty) >> 1);
1594
1595 return 1;
1596 }
1597
wb_max_pause(struct bdi_writeback * wb,unsigned long wb_dirty)1598 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1599 unsigned long wb_dirty)
1600 {
1601 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1602 unsigned long t;
1603
1604 /*
1605 * Limit pause time for small memory systems. If sleeping for too long
1606 * time, a small pool of dirty/writeback pages may go empty and disk go
1607 * idle.
1608 *
1609 * 8 serves as the safety ratio.
1610 */
1611 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1612 t++;
1613
1614 return min_t(unsigned long, t, MAX_PAUSE);
1615 }
1616
wb_min_pause(struct bdi_writeback * wb,long max_pause,unsigned long task_ratelimit,unsigned long dirty_ratelimit,int * nr_dirtied_pause)1617 static long wb_min_pause(struct bdi_writeback *wb,
1618 long max_pause,
1619 unsigned long task_ratelimit,
1620 unsigned long dirty_ratelimit,
1621 int *nr_dirtied_pause)
1622 {
1623 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1624 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1625 long t; /* target pause */
1626 long pause; /* estimated next pause */
1627 int pages; /* target nr_dirtied_pause */
1628
1629 /* target for 10ms pause on 1-dd case */
1630 t = max(1, HZ / 100);
1631
1632 /*
1633 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1634 * overheads.
1635 *
1636 * (N * 10ms) on 2^N concurrent tasks.
1637 */
1638 if (hi > lo)
1639 t += (hi - lo) * (10 * HZ) / 1024;
1640
1641 /*
1642 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1643 * on the much more stable dirty_ratelimit. However the next pause time
1644 * will be computed based on task_ratelimit and the two rate limits may
1645 * depart considerably at some time. Especially if task_ratelimit goes
1646 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1647 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1648 * result task_ratelimit won't be executed faithfully, which could
1649 * eventually bring down dirty_ratelimit.
1650 *
1651 * We apply two rules to fix it up:
1652 * 1) try to estimate the next pause time and if necessary, use a lower
1653 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1654 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1655 * 2) limit the target pause time to max_pause/2, so that the normal
1656 * small fluctuations of task_ratelimit won't trigger rule (1) and
1657 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1658 */
1659 t = min(t, 1 + max_pause / 2);
1660 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1661
1662 /*
1663 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1664 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1665 * When the 16 consecutive reads are often interrupted by some dirty
1666 * throttling pause during the async writes, cfq will go into idles
1667 * (deadline is fine). So push nr_dirtied_pause as high as possible
1668 * until reaches DIRTY_POLL_THRESH=32 pages.
1669 */
1670 if (pages < DIRTY_POLL_THRESH) {
1671 t = max_pause;
1672 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1673 if (pages > DIRTY_POLL_THRESH) {
1674 pages = DIRTY_POLL_THRESH;
1675 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1676 }
1677 }
1678
1679 pause = HZ * pages / (task_ratelimit + 1);
1680 if (pause > max_pause) {
1681 t = max_pause;
1682 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1683 }
1684
1685 *nr_dirtied_pause = pages;
1686 /*
1687 * The minimal pause time will normally be half the target pause time.
1688 */
1689 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1690 }
1691
wb_dirty_limits(struct dirty_throttle_control * dtc)1692 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1693 {
1694 struct bdi_writeback *wb = dtc->wb;
1695 unsigned long wb_reclaimable;
1696
1697 /*
1698 * wb_thresh is not treated as some limiting factor as
1699 * dirty_thresh, due to reasons
1700 * - in JBOD setup, wb_thresh can fluctuate a lot
1701 * - in a system with HDD and USB key, the USB key may somehow
1702 * go into state (wb_dirty >> wb_thresh) either because
1703 * wb_dirty starts high, or because wb_thresh drops low.
1704 * In this case we don't want to hard throttle the USB key
1705 * dirtiers for 100 seconds until wb_dirty drops under
1706 * wb_thresh. Instead the auxiliary wb control line in
1707 * wb_position_ratio() will let the dirtier task progress
1708 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1709 */
1710 dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
1711 dtc->wb_bg_thresh = dtc->thresh ?
1712 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1713
1714 /*
1715 * In order to avoid the stacked BDI deadlock we need
1716 * to ensure we accurately count the 'dirty' pages when
1717 * the threshold is low.
1718 *
1719 * Otherwise it would be possible to get thresh+n pages
1720 * reported dirty, even though there are thresh-m pages
1721 * actually dirty; with m+n sitting in the percpu
1722 * deltas.
1723 */
1724 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1725 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1726 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1727 } else {
1728 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1729 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1730 }
1731 }
1732
domain_poll_intv(struct dirty_throttle_control * dtc,bool strictlimit)1733 static unsigned long domain_poll_intv(struct dirty_throttle_control *dtc,
1734 bool strictlimit)
1735 {
1736 unsigned long dirty, thresh;
1737
1738 if (strictlimit) {
1739 dirty = dtc->wb_dirty;
1740 thresh = dtc->wb_thresh;
1741 } else {
1742 dirty = dtc->dirty;
1743 thresh = dtc->thresh;
1744 }
1745
1746 return dirty_poll_interval(dirty, thresh);
1747 }
1748
1749 /*
1750 * Throttle it only when the background writeback cannot catch-up. This avoids
1751 * (excessively) small writeouts when the wb limits are ramping up in case of
1752 * !strictlimit.
1753 *
1754 * In strictlimit case make decision based on the wb counters and limits. Small
1755 * writeouts when the wb limits are ramping up are the price we consciously pay
1756 * for strictlimit-ing.
1757 */
domain_dirty_freerun(struct dirty_throttle_control * dtc,bool strictlimit)1758 static void domain_dirty_freerun(struct dirty_throttle_control *dtc,
1759 bool strictlimit)
1760 {
1761 unsigned long dirty, thresh, bg_thresh;
1762
1763 if (unlikely(strictlimit)) {
1764 wb_dirty_limits(dtc);
1765 dirty = dtc->wb_dirty;
1766 thresh = dtc->wb_thresh;
1767 bg_thresh = dtc->wb_bg_thresh;
1768 } else {
1769 dirty = dtc->dirty;
1770 thresh = dtc->thresh;
1771 bg_thresh = dtc->bg_thresh;
1772 }
1773 dtc->freerun = dirty <= dirty_freerun_ceiling(thresh, bg_thresh);
1774 }
1775
balance_domain_limits(struct dirty_throttle_control * dtc,bool strictlimit)1776 static void balance_domain_limits(struct dirty_throttle_control *dtc,
1777 bool strictlimit)
1778 {
1779 domain_dirty_avail(dtc, true);
1780 domain_dirty_limits(dtc);
1781 domain_dirty_freerun(dtc, strictlimit);
1782 }
1783
wb_dirty_freerun(struct dirty_throttle_control * dtc,bool strictlimit)1784 static void wb_dirty_freerun(struct dirty_throttle_control *dtc,
1785 bool strictlimit)
1786 {
1787 dtc->freerun = false;
1788
1789 /* was already handled in domain_dirty_freerun */
1790 if (strictlimit)
1791 return;
1792
1793 wb_dirty_limits(dtc);
1794 /*
1795 * LOCAL_THROTTLE tasks must not be throttled when below the per-wb
1796 * freerun ceiling.
1797 */
1798 if (!(current->flags & PF_LOCAL_THROTTLE))
1799 return;
1800
1801 dtc->freerun = dtc->wb_dirty <
1802 dirty_freerun_ceiling(dtc->wb_thresh, dtc->wb_bg_thresh);
1803 }
1804
wb_dirty_exceeded(struct dirty_throttle_control * dtc,bool strictlimit)1805 static inline void wb_dirty_exceeded(struct dirty_throttle_control *dtc,
1806 bool strictlimit)
1807 {
1808 dtc->dirty_exceeded = (dtc->wb_dirty > dtc->wb_thresh) &&
1809 ((dtc->dirty > dtc->thresh) || strictlimit);
1810 }
1811
1812 /*
1813 * The limits fields dirty_exceeded and pos_ratio won't be updated if wb is
1814 * in freerun state. Please don't use these invalid fields in freerun case.
1815 */
balance_wb_limits(struct dirty_throttle_control * dtc,bool strictlimit)1816 static void balance_wb_limits(struct dirty_throttle_control *dtc,
1817 bool strictlimit)
1818 {
1819 wb_dirty_freerun(dtc, strictlimit);
1820 if (dtc->freerun)
1821 return;
1822
1823 wb_dirty_exceeded(dtc, strictlimit);
1824 wb_position_ratio(dtc);
1825 }
1826
1827 /*
1828 * balance_dirty_pages() must be called by processes which are generating dirty
1829 * data. It looks at the number of dirty pages in the machine and will force
1830 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1831 * If we're over `background_thresh' then the writeback threads are woken to
1832 * perform some writeout.
1833 */
balance_dirty_pages(struct bdi_writeback * wb,unsigned long pages_dirtied,unsigned int flags)1834 static int balance_dirty_pages(struct bdi_writeback *wb,
1835 unsigned long pages_dirtied, unsigned int flags)
1836 {
1837 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1838 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1839 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1840 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1841 &mdtc_stor : NULL;
1842 struct dirty_throttle_control *sdtc;
1843 unsigned long nr_dirty;
1844 long period;
1845 long pause;
1846 long max_pause;
1847 long min_pause;
1848 int nr_dirtied_pause;
1849 unsigned long task_ratelimit;
1850 unsigned long dirty_ratelimit;
1851 struct backing_dev_info *bdi = wb->bdi;
1852 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1853 unsigned long start_time = jiffies;
1854 int ret = 0;
1855
1856 for (;;) {
1857 unsigned long now = jiffies;
1858
1859 nr_dirty = global_node_page_state(NR_FILE_DIRTY);
1860
1861 balance_domain_limits(gdtc, strictlimit);
1862 if (mdtc) {
1863 /*
1864 * If @wb belongs to !root memcg, repeat the same
1865 * basic calculations for the memcg domain.
1866 */
1867 balance_domain_limits(mdtc, strictlimit);
1868 }
1869
1870 /*
1871 * In laptop mode, we wait until hitting the higher threshold
1872 * before starting background writeout, and then write out all
1873 * the way down to the lower threshold. So slow writers cause
1874 * minimal disk activity.
1875 *
1876 * In normal mode, we start background writeout at the lower
1877 * background_thresh, to keep the amount of dirty memory low.
1878 */
1879 if (!laptop_mode && nr_dirty > gdtc->bg_thresh &&
1880 !writeback_in_progress(wb))
1881 wb_start_background_writeback(wb);
1882
1883 /*
1884 * If memcg domain is in effect, @dirty should be under
1885 * both global and memcg freerun ceilings.
1886 */
1887 if (gdtc->freerun && (!mdtc || mdtc->freerun)) {
1888 unsigned long intv;
1889 unsigned long m_intv;
1890
1891 free_running:
1892 intv = domain_poll_intv(gdtc, strictlimit);
1893 m_intv = ULONG_MAX;
1894
1895 current->dirty_paused_when = now;
1896 current->nr_dirtied = 0;
1897 if (mdtc)
1898 m_intv = domain_poll_intv(mdtc, strictlimit);
1899 current->nr_dirtied_pause = min(intv, m_intv);
1900 break;
1901 }
1902
1903 /* Start writeback even when in laptop mode */
1904 if (unlikely(!writeback_in_progress(wb)))
1905 wb_start_background_writeback(wb);
1906
1907 mem_cgroup_flush_foreign(wb);
1908
1909 /*
1910 * Calculate global domain's pos_ratio and select the
1911 * global dtc by default.
1912 */
1913 balance_wb_limits(gdtc, strictlimit);
1914 if (gdtc->freerun)
1915 goto free_running;
1916 sdtc = gdtc;
1917
1918 if (mdtc) {
1919 /*
1920 * If memcg domain is in effect, calculate its
1921 * pos_ratio. @wb should satisfy constraints from
1922 * both global and memcg domains. Choose the one
1923 * w/ lower pos_ratio.
1924 */
1925 balance_wb_limits(mdtc, strictlimit);
1926 if (mdtc->freerun)
1927 goto free_running;
1928 if (mdtc->pos_ratio < gdtc->pos_ratio)
1929 sdtc = mdtc;
1930 }
1931
1932 wb->dirty_exceeded = gdtc->dirty_exceeded ||
1933 (mdtc && mdtc->dirty_exceeded);
1934 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1935 BANDWIDTH_INTERVAL))
1936 __wb_update_bandwidth(gdtc, mdtc, true);
1937
1938 /* throttle according to the chosen dtc */
1939 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1940 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1941 RATELIMIT_CALC_SHIFT;
1942 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1943 min_pause = wb_min_pause(wb, max_pause,
1944 task_ratelimit, dirty_ratelimit,
1945 &nr_dirtied_pause);
1946
1947 if (unlikely(task_ratelimit == 0)) {
1948 period = max_pause;
1949 pause = max_pause;
1950 goto pause;
1951 }
1952 period = HZ * pages_dirtied / task_ratelimit;
1953 pause = period;
1954 if (current->dirty_paused_when)
1955 pause -= now - current->dirty_paused_when;
1956 /*
1957 * For less than 1s think time (ext3/4 may block the dirtier
1958 * for up to 800ms from time to time on 1-HDD; so does xfs,
1959 * however at much less frequency), try to compensate it in
1960 * future periods by updating the virtual time; otherwise just
1961 * do a reset, as it may be a light dirtier.
1962 */
1963 if (pause < min_pause) {
1964 trace_balance_dirty_pages(wb,
1965 sdtc->thresh,
1966 sdtc->bg_thresh,
1967 sdtc->dirty,
1968 sdtc->wb_thresh,
1969 sdtc->wb_dirty,
1970 dirty_ratelimit,
1971 task_ratelimit,
1972 pages_dirtied,
1973 period,
1974 min(pause, 0L),
1975 start_time);
1976 if (pause < -HZ) {
1977 current->dirty_paused_when = now;
1978 current->nr_dirtied = 0;
1979 } else if (period) {
1980 current->dirty_paused_when += period;
1981 current->nr_dirtied = 0;
1982 } else if (current->nr_dirtied_pause <= pages_dirtied)
1983 current->nr_dirtied_pause += pages_dirtied;
1984 break;
1985 }
1986 if (unlikely(pause > max_pause)) {
1987 /* for occasional dropped task_ratelimit */
1988 now += min(pause - max_pause, max_pause);
1989 pause = max_pause;
1990 }
1991
1992 pause:
1993 trace_balance_dirty_pages(wb,
1994 sdtc->thresh,
1995 sdtc->bg_thresh,
1996 sdtc->dirty,
1997 sdtc->wb_thresh,
1998 sdtc->wb_dirty,
1999 dirty_ratelimit,
2000 task_ratelimit,
2001 pages_dirtied,
2002 period,
2003 pause,
2004 start_time);
2005 if (flags & BDP_ASYNC) {
2006 ret = -EAGAIN;
2007 break;
2008 }
2009 __set_current_state(TASK_KILLABLE);
2010 bdi->last_bdp_sleep = jiffies;
2011 io_schedule_timeout(pause);
2012
2013 current->dirty_paused_when = now + pause;
2014 current->nr_dirtied = 0;
2015 current->nr_dirtied_pause = nr_dirtied_pause;
2016
2017 /*
2018 * This is typically equal to (dirty < thresh) and can also
2019 * keep "1000+ dd on a slow USB stick" under control.
2020 */
2021 if (task_ratelimit)
2022 break;
2023
2024 /*
2025 * In the case of an unresponsive NFS server and the NFS dirty
2026 * pages exceeds dirty_thresh, give the other good wb's a pipe
2027 * to go through, so that tasks on them still remain responsive.
2028 *
2029 * In theory 1 page is enough to keep the consumer-producer
2030 * pipe going: the flusher cleans 1 page => the task dirties 1
2031 * more page. However wb_dirty has accounting errors. So use
2032 * the larger and more IO friendly wb_stat_error.
2033 */
2034 if (sdtc->wb_dirty <= wb_stat_error())
2035 break;
2036
2037 if (fatal_signal_pending(current))
2038 break;
2039 }
2040 return ret;
2041 }
2042
2043 static DEFINE_PER_CPU(int, bdp_ratelimits);
2044
2045 /*
2046 * Normal tasks are throttled by
2047 * loop {
2048 * dirty tsk->nr_dirtied_pause pages;
2049 * take a snap in balance_dirty_pages();
2050 * }
2051 * However there is a worst case. If every task exit immediately when dirtied
2052 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
2053 * called to throttle the page dirties. The solution is to save the not yet
2054 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
2055 * randomly into the running tasks. This works well for the above worst case,
2056 * as the new task will pick up and accumulate the old task's leaked dirty
2057 * count and eventually get throttled.
2058 */
2059 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
2060
2061 /**
2062 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
2063 * @mapping: address_space which was dirtied.
2064 * @flags: BDP flags.
2065 *
2066 * Processes which are dirtying memory should call in here once for each page
2067 * which was newly dirtied. The function will periodically check the system's
2068 * dirty state and will initiate writeback if needed.
2069 *
2070 * See balance_dirty_pages_ratelimited() for details.
2071 *
2072 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
2073 * indicate that memory is out of balance and the caller must wait
2074 * for I/O to complete. Otherwise, it will return 0 to indicate
2075 * that either memory was already in balance, or it was able to sleep
2076 * until the amount of dirty memory returned to balance.
2077 */
balance_dirty_pages_ratelimited_flags(struct address_space * mapping,unsigned int flags)2078 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
2079 unsigned int flags)
2080 {
2081 struct inode *inode = mapping->host;
2082 struct backing_dev_info *bdi = inode_to_bdi(inode);
2083 struct bdi_writeback *wb = NULL;
2084 int ratelimit;
2085 int ret = 0;
2086 int *p;
2087
2088 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
2089 return ret;
2090
2091 if (inode_cgwb_enabled(inode))
2092 wb = wb_get_create_current(bdi, GFP_KERNEL);
2093 if (!wb)
2094 wb = &bdi->wb;
2095
2096 ratelimit = current->nr_dirtied_pause;
2097 if (wb->dirty_exceeded)
2098 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
2099
2100 preempt_disable();
2101 /*
2102 * This prevents one CPU to accumulate too many dirtied pages without
2103 * calling into balance_dirty_pages(), which can happen when there are
2104 * 1000+ tasks, all of them start dirtying pages at exactly the same
2105 * time, hence all honoured too large initial task->nr_dirtied_pause.
2106 */
2107 p = this_cpu_ptr(&bdp_ratelimits);
2108 if (unlikely(current->nr_dirtied >= ratelimit))
2109 *p = 0;
2110 else if (unlikely(*p >= ratelimit_pages)) {
2111 *p = 0;
2112 ratelimit = 0;
2113 }
2114 /*
2115 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2116 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2117 * the dirty throttling and livelock other long-run dirtiers.
2118 */
2119 p = this_cpu_ptr(&dirty_throttle_leaks);
2120 if (*p > 0 && current->nr_dirtied < ratelimit) {
2121 unsigned long nr_pages_dirtied;
2122 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2123 *p -= nr_pages_dirtied;
2124 current->nr_dirtied += nr_pages_dirtied;
2125 }
2126 preempt_enable();
2127
2128 if (unlikely(current->nr_dirtied >= ratelimit))
2129 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2130
2131 wb_put(wb);
2132 return ret;
2133 }
2134 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2135
2136 /**
2137 * balance_dirty_pages_ratelimited - balance dirty memory state.
2138 * @mapping: address_space which was dirtied.
2139 *
2140 * Processes which are dirtying memory should call in here once for each page
2141 * which was newly dirtied. The function will periodically check the system's
2142 * dirty state and will initiate writeback if needed.
2143 *
2144 * Once we're over the dirty memory limit we decrease the ratelimiting
2145 * by a lot, to prevent individual processes from overshooting the limit
2146 * by (ratelimit_pages) each.
2147 */
balance_dirty_pages_ratelimited(struct address_space * mapping)2148 void balance_dirty_pages_ratelimited(struct address_space *mapping)
2149 {
2150 balance_dirty_pages_ratelimited_flags(mapping, 0);
2151 }
2152 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
2153
2154 /*
2155 * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty
2156 * and thresh, but it's for background writeback.
2157 */
wb_bg_dirty_limits(struct dirty_throttle_control * dtc)2158 static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc)
2159 {
2160 struct bdi_writeback *wb = dtc->wb;
2161
2162 dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh);
2163 if (dtc->wb_bg_thresh < 2 * wb_stat_error())
2164 dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE);
2165 else
2166 dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE);
2167 }
2168
domain_over_bg_thresh(struct dirty_throttle_control * dtc)2169 static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc)
2170 {
2171 domain_dirty_avail(dtc, false);
2172 domain_dirty_limits(dtc);
2173 if (dtc->dirty > dtc->bg_thresh)
2174 return true;
2175
2176 wb_bg_dirty_limits(dtc);
2177 if (dtc->wb_dirty > dtc->wb_bg_thresh)
2178 return true;
2179
2180 return false;
2181 }
2182
2183 /**
2184 * wb_over_bg_thresh - does @wb need to be written back?
2185 * @wb: bdi_writeback of interest
2186 *
2187 * Determines whether background writeback should keep writing @wb or it's
2188 * clean enough.
2189 *
2190 * Return: %true if writeback should continue.
2191 */
wb_over_bg_thresh(struct bdi_writeback * wb)2192 bool wb_over_bg_thresh(struct bdi_writeback *wb)
2193 {
2194 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
2195 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
2196
2197 if (domain_over_bg_thresh(&gdtc))
2198 return true;
2199
2200 if (mdtc_valid(&mdtc))
2201 return domain_over_bg_thresh(&mdtc);
2202
2203 return false;
2204 }
2205
2206 #ifdef CONFIG_SYSCTL
2207 /*
2208 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2209 */
dirty_writeback_centisecs_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)2210 static int dirty_writeback_centisecs_handler(const struct ctl_table *table, int write,
2211 void *buffer, size_t *length, loff_t *ppos)
2212 {
2213 unsigned int old_interval = dirty_writeback_interval;
2214 int ret;
2215
2216 ret = proc_dointvec(table, write, buffer, length, ppos);
2217
2218 /*
2219 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2220 * and a different non-zero value will wakeup the writeback threads.
2221 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2222 * iterate over all bdis and wbs.
2223 * The reason we do this is to make the change take effect immediately.
2224 */
2225 if (!ret && write && dirty_writeback_interval &&
2226 dirty_writeback_interval != old_interval)
2227 wakeup_flusher_threads(WB_REASON_PERIODIC);
2228
2229 return ret;
2230 }
2231 #endif
2232
laptop_mode_timer_fn(struct timer_list * t)2233 void laptop_mode_timer_fn(struct timer_list *t)
2234 {
2235 struct backing_dev_info *backing_dev_info =
2236 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2237
2238 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2239 }
2240
2241 /*
2242 * We've spun up the disk and we're in laptop mode: schedule writeback
2243 * of all dirty data a few seconds from now. If the flush is already scheduled
2244 * then push it back - the user is still using the disk.
2245 */
laptop_io_completion(struct backing_dev_info * info)2246 void laptop_io_completion(struct backing_dev_info *info)
2247 {
2248 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2249 }
2250
2251 /*
2252 * We're in laptop mode and we've just synced. The sync's writes will have
2253 * caused another writeback to be scheduled by laptop_io_completion.
2254 * Nothing needs to be written back anymore, so we unschedule the writeback.
2255 */
laptop_sync_completion(void)2256 void laptop_sync_completion(void)
2257 {
2258 struct backing_dev_info *bdi;
2259
2260 rcu_read_lock();
2261
2262 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2263 del_timer(&bdi->laptop_mode_wb_timer);
2264
2265 rcu_read_unlock();
2266 }
2267
2268 /*
2269 * If ratelimit_pages is too high then we can get into dirty-data overload
2270 * if a large number of processes all perform writes at the same time.
2271 *
2272 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2273 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2274 * thresholds.
2275 */
2276
writeback_set_ratelimit(void)2277 void writeback_set_ratelimit(void)
2278 {
2279 struct wb_domain *dom = &global_wb_domain;
2280 unsigned long background_thresh;
2281 unsigned long dirty_thresh;
2282
2283 global_dirty_limits(&background_thresh, &dirty_thresh);
2284 dom->dirty_limit = dirty_thresh;
2285 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2286 if (ratelimit_pages < 16)
2287 ratelimit_pages = 16;
2288 }
2289
page_writeback_cpu_online(unsigned int cpu)2290 static int page_writeback_cpu_online(unsigned int cpu)
2291 {
2292 writeback_set_ratelimit();
2293 return 0;
2294 }
2295
2296 #ifdef CONFIG_SYSCTL
2297
2298 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2299 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2300
2301 static const struct ctl_table vm_page_writeback_sysctls[] = {
2302 {
2303 .procname = "dirty_background_ratio",
2304 .data = &dirty_background_ratio,
2305 .maxlen = sizeof(dirty_background_ratio),
2306 .mode = 0644,
2307 .proc_handler = dirty_background_ratio_handler,
2308 .extra1 = SYSCTL_ZERO,
2309 .extra2 = SYSCTL_ONE_HUNDRED,
2310 },
2311 {
2312 .procname = "dirty_background_bytes",
2313 .data = &dirty_background_bytes,
2314 .maxlen = sizeof(dirty_background_bytes),
2315 .mode = 0644,
2316 .proc_handler = dirty_background_bytes_handler,
2317 .extra1 = SYSCTL_LONG_ONE,
2318 },
2319 {
2320 .procname = "dirty_ratio",
2321 .data = &vm_dirty_ratio,
2322 .maxlen = sizeof(vm_dirty_ratio),
2323 .mode = 0644,
2324 .proc_handler = dirty_ratio_handler,
2325 .extra1 = SYSCTL_ZERO,
2326 .extra2 = SYSCTL_ONE_HUNDRED,
2327 },
2328 {
2329 .procname = "dirty_bytes",
2330 .data = &vm_dirty_bytes,
2331 .maxlen = sizeof(vm_dirty_bytes),
2332 .mode = 0644,
2333 .proc_handler = dirty_bytes_handler,
2334 .extra1 = (void *)&dirty_bytes_min,
2335 },
2336 {
2337 .procname = "dirty_writeback_centisecs",
2338 .data = &dirty_writeback_interval,
2339 .maxlen = sizeof(dirty_writeback_interval),
2340 .mode = 0644,
2341 .proc_handler = dirty_writeback_centisecs_handler,
2342 },
2343 {
2344 .procname = "dirty_expire_centisecs",
2345 .data = &dirty_expire_interval,
2346 .maxlen = sizeof(dirty_expire_interval),
2347 .mode = 0644,
2348 .proc_handler = proc_dointvec_minmax,
2349 .extra1 = SYSCTL_ZERO,
2350 },
2351 #ifdef CONFIG_HIGHMEM
2352 {
2353 .procname = "highmem_is_dirtyable",
2354 .data = &vm_highmem_is_dirtyable,
2355 .maxlen = sizeof(vm_highmem_is_dirtyable),
2356 .mode = 0644,
2357 .proc_handler = proc_dointvec_minmax,
2358 .extra1 = SYSCTL_ZERO,
2359 .extra2 = SYSCTL_ONE,
2360 },
2361 #endif
2362 {
2363 .procname = "laptop_mode",
2364 .data = &laptop_mode,
2365 .maxlen = sizeof(laptop_mode),
2366 .mode = 0644,
2367 .proc_handler = proc_dointvec_jiffies,
2368 },
2369 };
2370 #endif
2371
2372 /*
2373 * Called early on to tune the page writeback dirty limits.
2374 *
2375 * We used to scale dirty pages according to how total memory
2376 * related to pages that could be allocated for buffers.
2377 *
2378 * However, that was when we used "dirty_ratio" to scale with
2379 * all memory, and we don't do that any more. "dirty_ratio"
2380 * is now applied to total non-HIGHPAGE memory, and as such we can't
2381 * get into the old insane situation any more where we had
2382 * large amounts of dirty pages compared to a small amount of
2383 * non-HIGHMEM memory.
2384 *
2385 * But we might still want to scale the dirty_ratio by how
2386 * much memory the box has..
2387 */
page_writeback_init(void)2388 void __init page_writeback_init(void)
2389 {
2390 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2391
2392 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2393 page_writeback_cpu_online, NULL);
2394 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2395 page_writeback_cpu_online);
2396 #ifdef CONFIG_SYSCTL
2397 register_sysctl_init("vm", vm_page_writeback_sysctls);
2398 #endif
2399 }
2400
2401 /**
2402 * tag_pages_for_writeback - tag pages to be written by writeback
2403 * @mapping: address space structure to write
2404 * @start: starting page index
2405 * @end: ending page index (inclusive)
2406 *
2407 * This function scans the page range from @start to @end (inclusive) and tags
2408 * all pages that have DIRTY tag set with a special TOWRITE tag. The caller
2409 * can then use the TOWRITE tag to identify pages eligible for writeback.
2410 * This mechanism is used to avoid livelocking of writeback by a process
2411 * steadily creating new dirty pages in the file (thus it is important for this
2412 * function to be quick so that it can tag pages faster than a dirtying process
2413 * can create them).
2414 */
tag_pages_for_writeback(struct address_space * mapping,pgoff_t start,pgoff_t end)2415 void tag_pages_for_writeback(struct address_space *mapping,
2416 pgoff_t start, pgoff_t end)
2417 {
2418 XA_STATE(xas, &mapping->i_pages, start);
2419 unsigned int tagged = 0;
2420 void *page;
2421
2422 xas_lock_irq(&xas);
2423 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2424 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2425 if (++tagged % XA_CHECK_SCHED)
2426 continue;
2427
2428 xas_pause(&xas);
2429 xas_unlock_irq(&xas);
2430 cond_resched();
2431 xas_lock_irq(&xas);
2432 }
2433 xas_unlock_irq(&xas);
2434 }
2435 EXPORT_SYMBOL(tag_pages_for_writeback);
2436
folio_prepare_writeback(struct address_space * mapping,struct writeback_control * wbc,struct folio * folio)2437 static bool folio_prepare_writeback(struct address_space *mapping,
2438 struct writeback_control *wbc, struct folio *folio)
2439 {
2440 /*
2441 * Folio truncated or invalidated. We can freely skip it then,
2442 * even for data integrity operations: the folio has disappeared
2443 * concurrently, so there could be no real expectation of this
2444 * data integrity operation even if there is now a new, dirty
2445 * folio at the same pagecache index.
2446 */
2447 if (unlikely(folio->mapping != mapping))
2448 return false;
2449
2450 /*
2451 * Did somebody else write it for us?
2452 */
2453 if (!folio_test_dirty(folio))
2454 return false;
2455
2456 if (folio_test_writeback(folio)) {
2457 if (wbc->sync_mode == WB_SYNC_NONE)
2458 return false;
2459 folio_wait_writeback(folio);
2460 }
2461 BUG_ON(folio_test_writeback(folio));
2462
2463 if (!folio_clear_dirty_for_io(folio))
2464 return false;
2465
2466 return true;
2467 }
2468
wbc_to_tag(struct writeback_control * wbc)2469 static xa_mark_t wbc_to_tag(struct writeback_control *wbc)
2470 {
2471 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2472 return PAGECACHE_TAG_TOWRITE;
2473 return PAGECACHE_TAG_DIRTY;
2474 }
2475
wbc_end(struct writeback_control * wbc)2476 static pgoff_t wbc_end(struct writeback_control *wbc)
2477 {
2478 if (wbc->range_cyclic)
2479 return -1;
2480 return wbc->range_end >> PAGE_SHIFT;
2481 }
2482
writeback_get_folio(struct address_space * mapping,struct writeback_control * wbc)2483 static struct folio *writeback_get_folio(struct address_space *mapping,
2484 struct writeback_control *wbc)
2485 {
2486 struct folio *folio;
2487
2488 retry:
2489 folio = folio_batch_next(&wbc->fbatch);
2490 if (!folio) {
2491 folio_batch_release(&wbc->fbatch);
2492 cond_resched();
2493 filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
2494 wbc_to_tag(wbc), &wbc->fbatch);
2495 folio = folio_batch_next(&wbc->fbatch);
2496 if (!folio)
2497 return NULL;
2498 }
2499
2500 folio_lock(folio);
2501 if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
2502 folio_unlock(folio);
2503 goto retry;
2504 }
2505
2506 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2507 return folio;
2508 }
2509
2510 /**
2511 * writeback_iter - iterate folio of a mapping for writeback
2512 * @mapping: address space structure to write
2513 * @wbc: writeback context
2514 * @folio: previously iterated folio (%NULL to start)
2515 * @error: in-out pointer for writeback errors (see below)
2516 *
2517 * This function returns the next folio for the writeback operation described by
2518 * @wbc on @mapping and should be called in a while loop in the ->writepages
2519 * implementation.
2520 *
2521 * To start the writeback operation, %NULL is passed in the @folio argument, and
2522 * for every subsequent iteration the folio returned previously should be passed
2523 * back in.
2524 *
2525 * If there was an error in the per-folio writeback inside the writeback_iter()
2526 * loop, @error should be set to the error value.
2527 *
2528 * Once the writeback described in @wbc has finished, this function will return
2529 * %NULL and if there was an error in any iteration restore it to @error.
2530 *
2531 * Note: callers should not manually break out of the loop using break or goto
2532 * but must keep calling writeback_iter() until it returns %NULL.
2533 *
2534 * Return: the folio to write or %NULL if the loop is done.
2535 */
writeback_iter(struct address_space * mapping,struct writeback_control * wbc,struct folio * folio,int * error)2536 struct folio *writeback_iter(struct address_space *mapping,
2537 struct writeback_control *wbc, struct folio *folio, int *error)
2538 {
2539 if (!folio) {
2540 folio_batch_init(&wbc->fbatch);
2541 wbc->saved_err = *error = 0;
2542
2543 /*
2544 * For range cyclic writeback we remember where we stopped so
2545 * that we can continue where we stopped.
2546 *
2547 * For non-cyclic writeback we always start at the beginning of
2548 * the passed in range.
2549 */
2550 if (wbc->range_cyclic)
2551 wbc->index = mapping->writeback_index;
2552 else
2553 wbc->index = wbc->range_start >> PAGE_SHIFT;
2554
2555 /*
2556 * To avoid livelocks when other processes dirty new pages, we
2557 * first tag pages which should be written back and only then
2558 * start writing them.
2559 *
2560 * For data-integrity writeback we have to be careful so that we
2561 * do not miss some pages (e.g., because some other process has
2562 * cleared the TOWRITE tag we set). The rule we follow is that
2563 * TOWRITE tag can be cleared only by the process clearing the
2564 * DIRTY tag (and submitting the page for I/O).
2565 */
2566 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2567 tag_pages_for_writeback(mapping, wbc->index,
2568 wbc_end(wbc));
2569 } else {
2570 wbc->nr_to_write -= folio_nr_pages(folio);
2571
2572 WARN_ON_ONCE(*error > 0);
2573
2574 /*
2575 * For integrity writeback we have to keep going until we have
2576 * written all the folios we tagged for writeback above, even if
2577 * we run past wbc->nr_to_write or encounter errors.
2578 * We stash away the first error we encounter in wbc->saved_err
2579 * so that it can be retrieved when we're done. This is because
2580 * the file system may still have state to clear for each folio.
2581 *
2582 * For background writeback we exit as soon as we run past
2583 * wbc->nr_to_write or encounter the first error.
2584 */
2585 if (wbc->sync_mode == WB_SYNC_ALL) {
2586 if (*error && !wbc->saved_err)
2587 wbc->saved_err = *error;
2588 } else {
2589 if (*error || wbc->nr_to_write <= 0)
2590 goto done;
2591 }
2592 }
2593
2594 folio = writeback_get_folio(mapping, wbc);
2595 if (!folio) {
2596 /*
2597 * To avoid deadlocks between range_cyclic writeback and callers
2598 * that hold pages in PageWriteback to aggregate I/O until
2599 * the writeback iteration finishes, we do not loop back to the
2600 * start of the file. Doing so causes a page lock/page
2601 * writeback access order inversion - we should only ever lock
2602 * multiple pages in ascending page->index order, and looping
2603 * back to the start of the file violates that rule and causes
2604 * deadlocks.
2605 */
2606 if (wbc->range_cyclic)
2607 mapping->writeback_index = 0;
2608
2609 /*
2610 * Return the first error we encountered (if there was any) to
2611 * the caller.
2612 */
2613 *error = wbc->saved_err;
2614 }
2615 return folio;
2616
2617 done:
2618 if (wbc->range_cyclic)
2619 mapping->writeback_index = folio_next_index(folio);
2620 folio_batch_release(&wbc->fbatch);
2621 return NULL;
2622 }
2623 EXPORT_SYMBOL_GPL(writeback_iter);
2624
2625 /**
2626 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2627 * @mapping: address space structure to write
2628 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2629 * @writepage: function called for each page
2630 * @data: data passed to writepage function
2631 *
2632 * Return: %0 on success, negative error code otherwise
2633 *
2634 * Note: please use writeback_iter() instead.
2635 */
write_cache_pages(struct address_space * mapping,struct writeback_control * wbc,writepage_t writepage,void * data)2636 int write_cache_pages(struct address_space *mapping,
2637 struct writeback_control *wbc, writepage_t writepage,
2638 void *data)
2639 {
2640 struct folio *folio = NULL;
2641 int error;
2642
2643 while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
2644 error = writepage(folio, wbc, data);
2645 if (error == AOP_WRITEPAGE_ACTIVATE) {
2646 folio_unlock(folio);
2647 error = 0;
2648 }
2649 }
2650
2651 return error;
2652 }
2653 EXPORT_SYMBOL(write_cache_pages);
2654
writeback_use_writepage(struct address_space * mapping,struct writeback_control * wbc)2655 static int writeback_use_writepage(struct address_space *mapping,
2656 struct writeback_control *wbc)
2657 {
2658 struct folio *folio = NULL;
2659 struct blk_plug plug;
2660 int err;
2661
2662 blk_start_plug(&plug);
2663 while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
2664 err = mapping->a_ops->writepage(&folio->page, wbc);
2665 if (err == AOP_WRITEPAGE_ACTIVATE) {
2666 folio_unlock(folio);
2667 err = 0;
2668 }
2669 mapping_set_error(mapping, err);
2670 }
2671 blk_finish_plug(&plug);
2672
2673 return err;
2674 }
2675
do_writepages(struct address_space * mapping,struct writeback_control * wbc)2676 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2677 {
2678 int ret;
2679 struct bdi_writeback *wb;
2680
2681 if (wbc->nr_to_write <= 0)
2682 return 0;
2683 wb = inode_to_wb_wbc(mapping->host, wbc);
2684 wb_bandwidth_estimate_start(wb);
2685 while (1) {
2686 if (mapping->a_ops->writepages) {
2687 ret = mapping->a_ops->writepages(mapping, wbc);
2688 } else if (mapping->a_ops->writepage) {
2689 ret = writeback_use_writepage(mapping, wbc);
2690 } else {
2691 /* deal with chardevs and other special files */
2692 ret = 0;
2693 }
2694 if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
2695 break;
2696
2697 /*
2698 * Lacking an allocation context or the locality or writeback
2699 * state of any of the inode's pages, throttle based on
2700 * writeback activity on the local node. It's as good a
2701 * guess as any.
2702 */
2703 reclaim_throttle(NODE_DATA(numa_node_id()),
2704 VMSCAN_THROTTLE_WRITEBACK);
2705 }
2706 /*
2707 * Usually few pages are written by now from those we've just submitted
2708 * but if there's constant writeback being submitted, this makes sure
2709 * writeback bandwidth is updated once in a while.
2710 */
2711 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2712 BANDWIDTH_INTERVAL))
2713 wb_update_bandwidth(wb);
2714 return ret;
2715 }
2716
2717 /*
2718 * For address_spaces which do not use buffers nor write back.
2719 */
noop_dirty_folio(struct address_space * mapping,struct folio * folio)2720 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2721 {
2722 if (!folio_test_dirty(folio))
2723 return !folio_test_set_dirty(folio);
2724 return false;
2725 }
2726 EXPORT_SYMBOL(noop_dirty_folio);
2727
2728 /*
2729 * Helper function for set_page_dirty family.
2730 *
2731 * NOTE: This relies on being atomic wrt interrupts.
2732 */
folio_account_dirtied(struct folio * folio,struct address_space * mapping)2733 static void folio_account_dirtied(struct folio *folio,
2734 struct address_space *mapping)
2735 {
2736 struct inode *inode = mapping->host;
2737
2738 trace_writeback_dirty_folio(folio, mapping);
2739
2740 if (mapping_can_writeback(mapping)) {
2741 struct bdi_writeback *wb;
2742 long nr = folio_nr_pages(folio);
2743
2744 inode_attach_wb(inode, folio);
2745 wb = inode_to_wb(inode);
2746
2747 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2748 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2749 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2750 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2751 wb_stat_mod(wb, WB_DIRTIED, nr);
2752 task_io_account_write(nr * PAGE_SIZE);
2753 current->nr_dirtied += nr;
2754 __this_cpu_add(bdp_ratelimits, nr);
2755
2756 mem_cgroup_track_foreign_dirty(folio, wb);
2757 }
2758 }
2759
2760 /*
2761 * Helper function for deaccounting dirty page without writeback.
2762 *
2763 */
folio_account_cleaned(struct folio * folio,struct bdi_writeback * wb)2764 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2765 {
2766 long nr = folio_nr_pages(folio);
2767
2768 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2769 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2770 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2771 task_io_account_cancelled_write(nr * PAGE_SIZE);
2772 }
2773
2774 /*
2775 * Mark the folio dirty, and set it dirty in the page cache.
2776 *
2777 * If warn is true, then emit a warning if the folio is not uptodate and has
2778 * not been truncated.
2779 *
2780 * It is the caller's responsibility to prevent the folio from being truncated
2781 * while this function is in progress, although it may have been truncated
2782 * before this function is called. Most callers have the folio locked.
2783 * A few have the folio blocked from truncation through other means (e.g.
2784 * zap_vma_pages() has it mapped and is holding the page table lock).
2785 * When called from mark_buffer_dirty(), the filesystem should hold a
2786 * reference to the buffer_head that is being marked dirty, which causes
2787 * try_to_free_buffers() to fail.
2788 */
__folio_mark_dirty(struct folio * folio,struct address_space * mapping,int warn)2789 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2790 int warn)
2791 {
2792 unsigned long flags;
2793
2794 xa_lock_irqsave(&mapping->i_pages, flags);
2795 if (folio->mapping) { /* Race with truncate? */
2796 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2797 folio_account_dirtied(folio, mapping);
2798 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2799 PAGECACHE_TAG_DIRTY);
2800 }
2801 xa_unlock_irqrestore(&mapping->i_pages, flags);
2802 }
2803
2804 /**
2805 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2806 * @mapping: Address space this folio belongs to.
2807 * @folio: Folio to be marked as dirty.
2808 *
2809 * Filesystems which do not use buffer heads should call this function
2810 * from their dirty_folio address space operation. It ignores the
2811 * contents of folio_get_private(), so if the filesystem marks individual
2812 * blocks as dirty, the filesystem should handle that itself.
2813 *
2814 * This is also sometimes used by filesystems which use buffer_heads when
2815 * a single buffer is being dirtied: we want to set the folio dirty in
2816 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2817 * whereas block_dirty_folio() is a "top-down" dirtying.
2818 *
2819 * The caller must ensure this doesn't race with truncation. Most will
2820 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2821 * folio mapped and the pte lock held, which also locks out truncation.
2822 */
filemap_dirty_folio(struct address_space * mapping,struct folio * folio)2823 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2824 {
2825 if (folio_test_set_dirty(folio))
2826 return false;
2827
2828 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2829
2830 if (mapping->host) {
2831 /* !PageAnon && !swapper_space */
2832 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2833 }
2834 return true;
2835 }
2836 EXPORT_SYMBOL(filemap_dirty_folio);
2837
2838 /**
2839 * folio_redirty_for_writepage - Decline to write a dirty folio.
2840 * @wbc: The writeback control.
2841 * @folio: The folio.
2842 *
2843 * When a writepage implementation decides that it doesn't want to write
2844 * @folio for some reason, it should call this function, unlock @folio and
2845 * return 0.
2846 *
2847 * Return: True if we redirtied the folio. False if someone else dirtied
2848 * it first.
2849 */
folio_redirty_for_writepage(struct writeback_control * wbc,struct folio * folio)2850 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2851 struct folio *folio)
2852 {
2853 struct address_space *mapping = folio->mapping;
2854 long nr = folio_nr_pages(folio);
2855 bool ret;
2856
2857 wbc->pages_skipped += nr;
2858 ret = filemap_dirty_folio(mapping, folio);
2859 if (mapping && mapping_can_writeback(mapping)) {
2860 struct inode *inode = mapping->host;
2861 struct bdi_writeback *wb;
2862 struct wb_lock_cookie cookie = {};
2863
2864 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2865 current->nr_dirtied -= nr;
2866 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2867 wb_stat_mod(wb, WB_DIRTIED, -nr);
2868 unlocked_inode_to_wb_end(inode, &cookie);
2869 }
2870 return ret;
2871 }
2872 EXPORT_SYMBOL(folio_redirty_for_writepage);
2873
2874 /**
2875 * folio_mark_dirty - Mark a folio as being modified.
2876 * @folio: The folio.
2877 *
2878 * The folio may not be truncated while this function is running.
2879 * Holding the folio lock is sufficient to prevent truncation, but some
2880 * callers cannot acquire a sleeping lock. These callers instead hold
2881 * the page table lock for a page table which contains at least one page
2882 * in this folio. Truncation will block on the page table lock as it
2883 * unmaps pages before removing the folio from its mapping.
2884 *
2885 * Return: True if the folio was newly dirtied, false if it was already dirty.
2886 */
folio_mark_dirty(struct folio * folio)2887 bool folio_mark_dirty(struct folio *folio)
2888 {
2889 struct address_space *mapping = folio_mapping(folio);
2890
2891 if (likely(mapping)) {
2892 /*
2893 * readahead/folio_deactivate could remain
2894 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2895 * About readahead, if the folio is written, the flags would be
2896 * reset. So no problem.
2897 * About folio_deactivate, if the folio is redirtied,
2898 * the flag will be reset. So no problem. but if the
2899 * folio is used by readahead it will confuse readahead
2900 * and make it restart the size rampup process. But it's
2901 * a trivial problem.
2902 */
2903 if (folio_test_reclaim(folio))
2904 folio_clear_reclaim(folio);
2905 return mapping->a_ops->dirty_folio(mapping, folio);
2906 }
2907
2908 return noop_dirty_folio(mapping, folio);
2909 }
2910 EXPORT_SYMBOL(folio_mark_dirty);
2911
2912 /*
2913 * folio_mark_dirty() is racy if the caller has no reference against
2914 * folio->mapping->host, and if the folio is unlocked. This is because another
2915 * CPU could truncate the folio off the mapping and then free the mapping.
2916 *
2917 * Usually, the folio _is_ locked, or the caller is a user-space process which
2918 * holds a reference on the inode by having an open file.
2919 *
2920 * In other cases, the folio should be locked before running folio_mark_dirty().
2921 */
folio_mark_dirty_lock(struct folio * folio)2922 bool folio_mark_dirty_lock(struct folio *folio)
2923 {
2924 bool ret;
2925
2926 folio_lock(folio);
2927 ret = folio_mark_dirty(folio);
2928 folio_unlock(folio);
2929 return ret;
2930 }
2931 EXPORT_SYMBOL(folio_mark_dirty_lock);
2932
2933 /*
2934 * This cancels just the dirty bit on the kernel page itself, it does NOT
2935 * actually remove dirty bits on any mmap's that may be around. It also
2936 * leaves the page tagged dirty, so any sync activity will still find it on
2937 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2938 * look at the dirty bits in the VM.
2939 *
2940 * Doing this should *normally* only ever be done when a page is truncated,
2941 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2942 * this when it notices that somebody has cleaned out all the buffers on a
2943 * page without actually doing it through the VM. Can you say "ext3 is
2944 * horribly ugly"? Thought you could.
2945 */
__folio_cancel_dirty(struct folio * folio)2946 void __folio_cancel_dirty(struct folio *folio)
2947 {
2948 struct address_space *mapping = folio_mapping(folio);
2949
2950 if (mapping_can_writeback(mapping)) {
2951 struct inode *inode = mapping->host;
2952 struct bdi_writeback *wb;
2953 struct wb_lock_cookie cookie = {};
2954
2955 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2956
2957 if (folio_test_clear_dirty(folio))
2958 folio_account_cleaned(folio, wb);
2959
2960 unlocked_inode_to_wb_end(inode, &cookie);
2961 } else {
2962 folio_clear_dirty(folio);
2963 }
2964 }
2965 EXPORT_SYMBOL(__folio_cancel_dirty);
2966
2967 /*
2968 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2969 * Returns true if the folio was previously dirty.
2970 *
2971 * This is for preparing to put the folio under writeout. We leave
2972 * the folio tagged as dirty in the xarray so that a concurrent
2973 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2974 * The ->writepage implementation will run either folio_start_writeback()
2975 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2976 * and xarray dirty tag back into sync.
2977 *
2978 * This incoherency between the folio's dirty flag and xarray tag is
2979 * unfortunate, but it only exists while the folio is locked.
2980 */
folio_clear_dirty_for_io(struct folio * folio)2981 bool folio_clear_dirty_for_io(struct folio *folio)
2982 {
2983 struct address_space *mapping = folio_mapping(folio);
2984 bool ret = false;
2985
2986 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2987
2988 if (mapping && mapping_can_writeback(mapping)) {
2989 struct inode *inode = mapping->host;
2990 struct bdi_writeback *wb;
2991 struct wb_lock_cookie cookie = {};
2992
2993 /*
2994 * Yes, Virginia, this is indeed insane.
2995 *
2996 * We use this sequence to make sure that
2997 * (a) we account for dirty stats properly
2998 * (b) we tell the low-level filesystem to
2999 * mark the whole folio dirty if it was
3000 * dirty in a pagetable. Only to then
3001 * (c) clean the folio again and return 1 to
3002 * cause the writeback.
3003 *
3004 * This way we avoid all nasty races with the
3005 * dirty bit in multiple places and clearing
3006 * them concurrently from different threads.
3007 *
3008 * Note! Normally the "folio_mark_dirty(folio)"
3009 * has no effect on the actual dirty bit - since
3010 * that will already usually be set. But we
3011 * need the side effects, and it can help us
3012 * avoid races.
3013 *
3014 * We basically use the folio "master dirty bit"
3015 * as a serialization point for all the different
3016 * threads doing their things.
3017 */
3018 if (folio_mkclean(folio))
3019 folio_mark_dirty(folio);
3020 /*
3021 * We carefully synchronise fault handlers against
3022 * installing a dirty pte and marking the folio dirty
3023 * at this point. We do this by having them hold the
3024 * page lock while dirtying the folio, and folios are
3025 * always locked coming in here, so we get the desired
3026 * exclusion.
3027 */
3028 wb = unlocked_inode_to_wb_begin(inode, &cookie);
3029 if (folio_test_clear_dirty(folio)) {
3030 long nr = folio_nr_pages(folio);
3031 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
3032 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3033 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
3034 ret = true;
3035 }
3036 unlocked_inode_to_wb_end(inode, &cookie);
3037 return ret;
3038 }
3039 return folio_test_clear_dirty(folio);
3040 }
3041 EXPORT_SYMBOL(folio_clear_dirty_for_io);
3042
wb_inode_writeback_start(struct bdi_writeback * wb)3043 static void wb_inode_writeback_start(struct bdi_writeback *wb)
3044 {
3045 atomic_inc(&wb->writeback_inodes);
3046 }
3047
wb_inode_writeback_end(struct bdi_writeback * wb)3048 static void wb_inode_writeback_end(struct bdi_writeback *wb)
3049 {
3050 unsigned long flags;
3051 atomic_dec(&wb->writeback_inodes);
3052 /*
3053 * Make sure estimate of writeback throughput gets updated after
3054 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
3055 * (which is the interval other bandwidth updates use for batching) so
3056 * that if multiple inodes end writeback at a similar time, they get
3057 * batched into one bandwidth update.
3058 */
3059 spin_lock_irqsave(&wb->work_lock, flags);
3060 if (test_bit(WB_registered, &wb->state))
3061 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
3062 spin_unlock_irqrestore(&wb->work_lock, flags);
3063 }
3064
__folio_end_writeback(struct folio * folio)3065 bool __folio_end_writeback(struct folio *folio)
3066 {
3067 long nr = folio_nr_pages(folio);
3068 struct address_space *mapping = folio_mapping(folio);
3069 bool ret;
3070
3071 if (mapping && mapping_use_writeback_tags(mapping)) {
3072 struct inode *inode = mapping->host;
3073 struct backing_dev_info *bdi = inode_to_bdi(inode);
3074 unsigned long flags;
3075
3076 xa_lock_irqsave(&mapping->i_pages, flags);
3077 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3078 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
3079 PAGECACHE_TAG_WRITEBACK);
3080 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3081 struct bdi_writeback *wb = inode_to_wb(inode);
3082
3083 wb_stat_mod(wb, WB_WRITEBACK, -nr);
3084 __wb_writeout_add(wb, nr);
3085 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
3086 wb_inode_writeback_end(wb);
3087 }
3088
3089 if (mapping->host && !mapping_tagged(mapping,
3090 PAGECACHE_TAG_WRITEBACK))
3091 sb_clear_inode_writeback(mapping->host);
3092
3093 xa_unlock_irqrestore(&mapping->i_pages, flags);
3094 } else {
3095 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3096 }
3097
3098 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3099 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3100 node_stat_mod_folio(folio, NR_WRITTEN, nr);
3101
3102 return ret;
3103 }
3104
__folio_start_writeback(struct folio * folio,bool keep_write)3105 void __folio_start_writeback(struct folio *folio, bool keep_write)
3106 {
3107 long nr = folio_nr_pages(folio);
3108 struct address_space *mapping = folio_mapping(folio);
3109 int access_ret;
3110
3111 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
3112
3113 if (mapping && mapping_use_writeback_tags(mapping)) {
3114 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3115 struct inode *inode = mapping->host;
3116 struct backing_dev_info *bdi = inode_to_bdi(inode);
3117 unsigned long flags;
3118 bool on_wblist;
3119
3120 xas_lock_irqsave(&xas, flags);
3121 xas_load(&xas);
3122 folio_test_set_writeback(folio);
3123
3124 on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
3125
3126 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3127 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3128 struct bdi_writeback *wb = inode_to_wb(inode);
3129
3130 wb_stat_mod(wb, WB_WRITEBACK, nr);
3131 if (!on_wblist)
3132 wb_inode_writeback_start(wb);
3133 }
3134
3135 /*
3136 * We can come through here when swapping anonymous
3137 * folios, so we don't necessarily have an inode to
3138 * track for sync.
3139 */
3140 if (mapping->host && !on_wblist)
3141 sb_mark_inode_writeback(mapping->host);
3142 if (!folio_test_dirty(folio))
3143 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3144 if (!keep_write)
3145 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3146 xas_unlock_irqrestore(&xas, flags);
3147 } else {
3148 folio_test_set_writeback(folio);
3149 }
3150
3151 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3152 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3153
3154 access_ret = arch_make_folio_accessible(folio);
3155 /*
3156 * If writeback has been triggered on a page that cannot be made
3157 * accessible, it is too late to recover here.
3158 */
3159 VM_BUG_ON_FOLIO(access_ret != 0, folio);
3160 }
3161 EXPORT_SYMBOL(__folio_start_writeback);
3162
3163 /**
3164 * folio_wait_writeback - Wait for a folio to finish writeback.
3165 * @folio: The folio to wait for.
3166 *
3167 * If the folio is currently being written back to storage, wait for the
3168 * I/O to complete.
3169 *
3170 * Context: Sleeps. Must be called in process context and with
3171 * no spinlocks held. Caller should hold a reference on the folio.
3172 * If the folio is not locked, writeback may start again after writeback
3173 * has finished.
3174 */
folio_wait_writeback(struct folio * folio)3175 void folio_wait_writeback(struct folio *folio)
3176 {
3177 while (folio_test_writeback(folio)) {
3178 trace_folio_wait_writeback(folio, folio_mapping(folio));
3179 folio_wait_bit(folio, PG_writeback);
3180 }
3181 }
3182 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3183
3184 /**
3185 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3186 * @folio: The folio to wait for.
3187 *
3188 * If the folio is currently being written back to storage, wait for the
3189 * I/O to complete or a fatal signal to arrive.
3190 *
3191 * Context: Sleeps. Must be called in process context and with
3192 * no spinlocks held. Caller should hold a reference on the folio.
3193 * If the folio is not locked, writeback may start again after writeback
3194 * has finished.
3195 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3196 */
folio_wait_writeback_killable(struct folio * folio)3197 int folio_wait_writeback_killable(struct folio *folio)
3198 {
3199 while (folio_test_writeback(folio)) {
3200 trace_folio_wait_writeback(folio, folio_mapping(folio));
3201 if (folio_wait_bit_killable(folio, PG_writeback))
3202 return -EINTR;
3203 }
3204
3205 return 0;
3206 }
3207 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3208
3209 /**
3210 * folio_wait_stable() - wait for writeback to finish, if necessary.
3211 * @folio: The folio to wait on.
3212 *
3213 * This function determines if the given folio is related to a backing
3214 * device that requires folio contents to be held stable during writeback.
3215 * If so, then it will wait for any pending writeback to complete.
3216 *
3217 * Context: Sleeps. Must be called in process context and with
3218 * no spinlocks held. Caller should hold a reference on the folio.
3219 * If the folio is not locked, writeback may start again after writeback
3220 * has finished.
3221 */
folio_wait_stable(struct folio * folio)3222 void folio_wait_stable(struct folio *folio)
3223 {
3224 if (mapping_stable_writes(folio_mapping(folio)))
3225 folio_wait_writeback(folio);
3226 }
3227 EXPORT_SYMBOL_GPL(folio_wait_stable);
3228