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