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/buffer_head.h> /* __set_page_dirty_buffers */
36 : #include <linux/pagevec.h>
37 : #include <linux/timer.h>
38 : #include <linux/sched/rt.h>
39 : #include <linux/sched/signal.h>
40 : #include <linux/mm_inline.h>
41 : #include <trace/events/writeback.h>
42 :
43 : #include "internal.h"
44 :
45 : /*
46 : * Sleep at most 200ms at a time in balance_dirty_pages().
47 : */
48 : #define MAX_PAUSE max(HZ/5, 1)
49 :
50 : /*
51 : * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 : * by raising pause time to max_pause when falls below it.
53 : */
54 : #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55 :
56 : /*
57 : * Estimate write bandwidth at 200ms intervals.
58 : */
59 : #define BANDWIDTH_INTERVAL max(HZ/5, 1)
60 :
61 : #define RATELIMIT_CALC_SHIFT 10
62 :
63 : /*
64 : * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 : * will look to see if it needs to force writeback or throttling.
66 : */
67 : static long ratelimit_pages = 32;
68 :
69 : /* The following parameters are exported via /proc/sys/vm */
70 :
71 : /*
72 : * Start background writeback (via writeback threads) at this percentage
73 : */
74 : int dirty_background_ratio = 10;
75 :
76 : /*
77 : * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 : * dirty_background_ratio * the amount of dirtyable memory
79 : */
80 : unsigned long dirty_background_bytes;
81 :
82 : /*
83 : * free highmem will not be subtracted from the total free memory
84 : * for calculating free ratios if vm_highmem_is_dirtyable is true
85 : */
86 : int vm_highmem_is_dirtyable;
87 :
88 : /*
89 : * The generator of dirty data starts writeback at this percentage
90 : */
91 : int vm_dirty_ratio = 20;
92 :
93 : /*
94 : * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 : * vm_dirty_ratio * the amount of dirtyable memory
96 : */
97 : unsigned long vm_dirty_bytes;
98 :
99 : /*
100 : * The interval between `kupdate'-style writebacks
101 : */
102 : unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103 :
104 : EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105 :
106 : /*
107 : * The longest time for which data is allowed to remain dirty
108 : */
109 : unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110 :
111 : /*
112 : * Flag that makes the machine dump writes/reads and block dirtyings.
113 : */
114 : int block_dump;
115 :
116 : /*
117 : * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 : * a full sync is triggered after this time elapses without any disk activity.
119 : */
120 : int laptop_mode;
121 :
122 : EXPORT_SYMBOL(laptop_mode);
123 :
124 : /* End of sysctl-exported parameters */
125 :
126 : struct wb_domain global_wb_domain;
127 :
128 : /* consolidated parameters for balance_dirty_pages() and its subroutines */
129 : struct dirty_throttle_control {
130 : #ifdef CONFIG_CGROUP_WRITEBACK
131 : struct wb_domain *dom;
132 : struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
133 : #endif
134 : struct bdi_writeback *wb;
135 : struct fprop_local_percpu *wb_completions;
136 :
137 : unsigned long avail; /* dirtyable */
138 : unsigned long dirty; /* file_dirty + write + nfs */
139 : unsigned long thresh; /* dirty threshold */
140 : unsigned long bg_thresh; /* dirty background threshold */
141 :
142 : unsigned long wb_dirty; /* per-wb counterparts */
143 : unsigned long wb_thresh;
144 : unsigned long wb_bg_thresh;
145 :
146 : unsigned long pos_ratio;
147 : };
148 :
149 : /*
150 : * Length of period for aging writeout fractions of bdis. This is an
151 : * arbitrarily chosen number. The longer the period, the slower fractions will
152 : * reflect changes in current writeout rate.
153 : */
154 : #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155 :
156 : #ifdef CONFIG_CGROUP_WRITEBACK
157 :
158 : #define GDTC_INIT(__wb) .wb = (__wb), \
159 : .dom = &global_wb_domain, \
160 : .wb_completions = &(__wb)->completions
161 :
162 : #define GDTC_INIT_NO_WB .dom = &global_wb_domain
163 :
164 : #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
165 : .dom = mem_cgroup_wb_domain(__wb), \
166 : .wb_completions = &(__wb)->memcg_completions, \
167 : .gdtc = __gdtc
168 :
169 : static bool mdtc_valid(struct dirty_throttle_control *dtc)
170 : {
171 : return dtc->dom;
172 : }
173 :
174 : static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175 : {
176 : return dtc->dom;
177 : }
178 :
179 : static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180 : {
181 : return mdtc->gdtc;
182 : }
183 :
184 : static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185 : {
186 : return &wb->memcg_completions;
187 : }
188 :
189 : static void wb_min_max_ratio(struct bdi_writeback *wb,
190 : unsigned long *minp, unsigned long *maxp)
191 : {
192 : unsigned long this_bw = wb->avg_write_bandwidth;
193 : unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194 : unsigned long long min = wb->bdi->min_ratio;
195 : unsigned long long max = wb->bdi->max_ratio;
196 :
197 : /*
198 : * @wb may already be clean by the time control reaches here and
199 : * the total may not include its bw.
200 : */
201 : if (this_bw < tot_bw) {
202 : if (min) {
203 : min *= this_bw;
204 : min = div64_ul(min, tot_bw);
205 : }
206 : if (max < 100) {
207 : max *= this_bw;
208 : max = div64_ul(max, tot_bw);
209 : }
210 : }
211 :
212 : *minp = min;
213 : *maxp = max;
214 : }
215 :
216 : #else /* CONFIG_CGROUP_WRITEBACK */
217 :
218 : #define GDTC_INIT(__wb) .wb = (__wb), \
219 : .wb_completions = &(__wb)->completions
220 : #define GDTC_INIT_NO_WB
221 : #define MDTC_INIT(__wb, __gdtc)
222 :
223 32 : static bool mdtc_valid(struct dirty_throttle_control *dtc)
224 : {
225 32 : return false;
226 : }
227 :
228 7 : static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229 : {
230 7 : return &global_wb_domain;
231 : }
232 :
233 36 : static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234 : {
235 36 : return NULL;
236 : }
237 :
238 : static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239 : {
240 : return NULL;
241 : }
242 :
243 7 : static void wb_min_max_ratio(struct bdi_writeback *wb,
244 : unsigned long *minp, unsigned long *maxp)
245 : {
246 7 : *minp = wb->bdi->min_ratio;
247 7 : *maxp = wb->bdi->max_ratio;
248 : }
249 :
250 : #endif /* CONFIG_CGROUP_WRITEBACK */
251 :
252 : /*
253 : * In a memory zone, there is a certain amount of pages we consider
254 : * available for the page cache, which is essentially the number of
255 : * free and reclaimable pages, minus some zone reserves to protect
256 : * lowmem and the ability to uphold the zone's watermarks without
257 : * requiring writeback.
258 : *
259 : * This number of dirtyable pages is the base value of which the
260 : * user-configurable dirty ratio is the effective number of pages that
261 : * are allowed to be actually dirtied. Per individual zone, or
262 : * globally by using the sum of dirtyable pages over all zones.
263 : *
264 : * Because the user is allowed to specify the dirty limit globally as
265 : * absolute number of bytes, calculating the per-zone dirty limit can
266 : * require translating the configured limit into a percentage of
267 : * global dirtyable memory first.
268 : */
269 :
270 : /**
271 : * node_dirtyable_memory - number of dirtyable pages in a node
272 : * @pgdat: the node
273 : *
274 : * Return: the node's number of pages potentially available for dirty
275 : * page cache. This is the base value for the per-node dirty limits.
276 : */
277 1509 : static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278 : {
279 1509 : unsigned long nr_pages = 0;
280 1509 : int z;
281 :
282 6036 : for (z = 0; z < MAX_NR_ZONES; z++) {
283 4527 : struct zone *zone = pgdat->node_zones + z;
284 :
285 4527 : if (!populated_zone(zone))
286 3018 : continue;
287 :
288 1509 : nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289 : }
290 :
291 : /*
292 : * Pages reserved for the kernel should not be considered
293 : * dirtyable, to prevent a situation where reclaim has to
294 : * clean pages in order to balance the zones.
295 : */
296 1509 : nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297 :
298 1509 : nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299 1509 : nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
300 :
301 1509 : return nr_pages;
302 : }
303 :
304 : static unsigned long highmem_dirtyable_memory(unsigned long total)
305 : {
306 : #ifdef CONFIG_HIGHMEM
307 : int node;
308 : unsigned long x = 0;
309 : int i;
310 :
311 : for_each_node_state(node, N_HIGH_MEMORY) {
312 : for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313 : struct zone *z;
314 : unsigned long nr_pages;
315 :
316 : if (!is_highmem_idx(i))
317 : continue;
318 :
319 : z = &NODE_DATA(node)->node_zones[i];
320 : if (!populated_zone(z))
321 : continue;
322 :
323 : nr_pages = zone_page_state(z, NR_FREE_PAGES);
324 : /* watch for underflows */
325 : nr_pages -= min(nr_pages, high_wmark_pages(z));
326 : nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327 : nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328 : x += nr_pages;
329 : }
330 : }
331 :
332 : /*
333 : * Unreclaimable memory (kernel memory or anonymous memory
334 : * without swap) can bring down the dirtyable pages below
335 : * the zone's dirty balance reserve and the above calculation
336 : * will underflow. However we still want to add in nodes
337 : * which are below threshold (negative values) to get a more
338 : * accurate calculation but make sure that the total never
339 : * underflows.
340 : */
341 : if ((long)x < 0)
342 : x = 0;
343 :
344 : /*
345 : * Make sure that the number of highmem pages is never larger
346 : * than the number of the total dirtyable memory. This can only
347 : * occur in very strange VM situations but we want to make sure
348 : * that this does not occur.
349 : */
350 : return min(x, total);
351 : #else
352 : return 0;
353 : #endif
354 : }
355 :
356 : /**
357 : * global_dirtyable_memory - number of globally dirtyable pages
358 : *
359 : * Return: the global number of pages potentially available for dirty
360 : * page cache. This is the base value for the global dirty limits.
361 : */
362 36 : static unsigned long global_dirtyable_memory(void)
363 : {
364 36 : unsigned long x;
365 :
366 36 : x = global_zone_page_state(NR_FREE_PAGES);
367 : /*
368 : * Pages reserved for the kernel should not be considered
369 : * dirtyable, to prevent a situation where reclaim has to
370 : * clean pages in order to balance the zones.
371 : */
372 36 : x -= min(x, totalreserve_pages);
373 :
374 36 : x += global_node_page_state(NR_INACTIVE_FILE);
375 36 : x += global_node_page_state(NR_ACTIVE_FILE);
376 :
377 36 : if (!vm_highmem_is_dirtyable)
378 36 : x -= highmem_dirtyable_memory(x);
379 :
380 36 : return x + 1; /* Ensure that we never return 0 */
381 : }
382 :
383 : /**
384 : * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385 : * @dtc: dirty_throttle_control of interest
386 : *
387 : * Calculate @dtc->thresh and ->bg_thresh considering
388 : * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
389 : * must ensure that @dtc->avail is set before calling this function. The
390 : * dirty limits will be lifted by 1/4 for real-time tasks.
391 : */
392 36 : static void domain_dirty_limits(struct dirty_throttle_control *dtc)
393 : {
394 36 : const unsigned long available_memory = dtc->avail;
395 36 : struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
396 36 : unsigned long bytes = vm_dirty_bytes;
397 36 : unsigned long bg_bytes = dirty_background_bytes;
398 : /* convert ratios to per-PAGE_SIZE for higher precision */
399 36 : unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
400 36 : unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
401 36 : unsigned long thresh;
402 36 : unsigned long bg_thresh;
403 36 : struct task_struct *tsk;
404 :
405 : /* gdtc is !NULL iff @dtc is for memcg domain */
406 36 : if (gdtc) {
407 : unsigned long global_avail = gdtc->avail;
408 :
409 : /*
410 : * The byte settings can't be applied directly to memcg
411 : * domains. Convert them to ratios by scaling against
412 : * globally available memory. As the ratios are in
413 : * per-PAGE_SIZE, they can be obtained by dividing bytes by
414 : * number of pages.
415 : */
416 : if (bytes)
417 : ratio = min(DIV_ROUND_UP(bytes, global_avail),
418 : PAGE_SIZE);
419 : if (bg_bytes)
420 : bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
421 : PAGE_SIZE);
422 : bytes = bg_bytes = 0;
423 : }
424 :
425 36 : if (bytes)
426 0 : thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
427 : else
428 36 : thresh = (ratio * available_memory) / PAGE_SIZE;
429 :
430 36 : if (bg_bytes)
431 0 : bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
432 : else
433 36 : bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
434 :
435 36 : if (bg_thresh >= thresh)
436 0 : bg_thresh = thresh / 2;
437 36 : tsk = current;
438 36 : if (rt_task(tsk)) {
439 0 : bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
440 0 : thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
441 : }
442 36 : dtc->thresh = thresh;
443 36 : dtc->bg_thresh = bg_thresh;
444 :
445 : /* we should eventually report the domain in the TP */
446 36 : if (!gdtc)
447 36 : trace_global_dirty_state(bg_thresh, thresh);
448 36 : }
449 :
450 : /**
451 : * global_dirty_limits - background-writeback and dirty-throttling thresholds
452 : * @pbackground: out parameter for bg_thresh
453 : * @pdirty: out parameter for thresh
454 : *
455 : * Calculate bg_thresh and thresh for global_wb_domain. See
456 : * domain_dirty_limits() for details.
457 : */
458 4 : void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
459 : {
460 4 : struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
461 :
462 4 : gdtc.avail = global_dirtyable_memory();
463 4 : domain_dirty_limits(&gdtc);
464 :
465 4 : *pbackground = gdtc.bg_thresh;
466 4 : *pdirty = gdtc.thresh;
467 4 : }
468 :
469 : /**
470 : * node_dirty_limit - maximum number of dirty pages allowed in a node
471 : * @pgdat: the node
472 : *
473 : * Return: the maximum number of dirty pages allowed in a node, based
474 : * on the node's dirtyable memory.
475 : */
476 1509 : static unsigned long node_dirty_limit(struct pglist_data *pgdat)
477 : {
478 1509 : unsigned long node_memory = node_dirtyable_memory(pgdat);
479 1509 : struct task_struct *tsk = current;
480 1509 : unsigned long dirty;
481 :
482 1509 : if (vm_dirty_bytes)
483 0 : dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
484 0 : node_memory / global_dirtyable_memory();
485 : else
486 1509 : dirty = vm_dirty_ratio * node_memory / 100;
487 :
488 1509 : if (rt_task(tsk))
489 0 : dirty += dirty / 4;
490 :
491 1509 : return dirty;
492 : }
493 :
494 : /**
495 : * node_dirty_ok - tells whether a node is within its dirty limits
496 : * @pgdat: the node to check
497 : *
498 : * Return: %true when the dirty pages in @pgdat are within the node's
499 : * dirty limit, %false if the limit is exceeded.
500 : */
501 1509 : bool node_dirty_ok(struct pglist_data *pgdat)
502 : {
503 1509 : unsigned long limit = node_dirty_limit(pgdat);
504 1509 : unsigned long nr_pages = 0;
505 :
506 1509 : nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
507 1509 : nr_pages += node_page_state(pgdat, NR_WRITEBACK);
508 :
509 1509 : return nr_pages <= limit;
510 : }
511 :
512 0 : int dirty_background_ratio_handler(struct ctl_table *table, int write,
513 : void *buffer, size_t *lenp, loff_t *ppos)
514 : {
515 0 : int ret;
516 :
517 0 : ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
518 0 : if (ret == 0 && write)
519 0 : dirty_background_bytes = 0;
520 0 : return ret;
521 : }
522 :
523 0 : int dirty_background_bytes_handler(struct ctl_table *table, int write,
524 : void *buffer, size_t *lenp, loff_t *ppos)
525 : {
526 0 : int ret;
527 :
528 0 : ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
529 0 : if (ret == 0 && write)
530 0 : dirty_background_ratio = 0;
531 0 : return ret;
532 : }
533 :
534 0 : int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
535 : size_t *lenp, loff_t *ppos)
536 : {
537 0 : int old_ratio = vm_dirty_ratio;
538 0 : int ret;
539 :
540 0 : ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
541 0 : if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
542 0 : writeback_set_ratelimit();
543 0 : vm_dirty_bytes = 0;
544 : }
545 0 : return ret;
546 : }
547 :
548 0 : int dirty_bytes_handler(struct ctl_table *table, int write,
549 : void *buffer, size_t *lenp, loff_t *ppos)
550 : {
551 0 : unsigned long old_bytes = vm_dirty_bytes;
552 0 : int ret;
553 :
554 0 : ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
555 0 : if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
556 0 : writeback_set_ratelimit();
557 0 : vm_dirty_ratio = 0;
558 : }
559 0 : return ret;
560 : }
561 :
562 11 : static unsigned long wp_next_time(unsigned long cur_time)
563 : {
564 11 : cur_time += VM_COMPLETIONS_PERIOD_LEN;
565 : /* 0 has a special meaning... */
566 11 : if (!cur_time)
567 : return 1;
568 : return cur_time;
569 : }
570 :
571 1284 : static void wb_domain_writeout_inc(struct wb_domain *dom,
572 : struct fprop_local_percpu *completions,
573 : unsigned int max_prop_frac)
574 : {
575 1284 : __fprop_inc_percpu_max(&dom->completions, completions,
576 : max_prop_frac);
577 : /* First event after period switching was turned off? */
578 1284 : if (unlikely(!dom->period_time)) {
579 : /*
580 : * We can race with other __bdi_writeout_inc calls here but
581 : * it does not cause any harm since the resulting time when
582 : * timer will fire and what is in writeout_period_time will be
583 : * roughly the same.
584 : */
585 1 : dom->period_time = wp_next_time(jiffies);
586 1 : mod_timer(&dom->period_timer, dom->period_time);
587 : }
588 1284 : }
589 :
590 : /*
591 : * Increment @wb's writeout completion count and the global writeout
592 : * completion count. Called from test_clear_page_writeback().
593 : */
594 1284 : static inline void __wb_writeout_inc(struct bdi_writeback *wb)
595 : {
596 1284 : struct wb_domain *cgdom;
597 :
598 1284 : inc_wb_stat(wb, WB_WRITTEN);
599 1284 : wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
600 1284 : wb->bdi->max_prop_frac);
601 :
602 1284 : cgdom = mem_cgroup_wb_domain(wb);
603 1284 : if (cgdom)
604 : wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
605 : wb->bdi->max_prop_frac);
606 1284 : }
607 :
608 0 : void wb_writeout_inc(struct bdi_writeback *wb)
609 : {
610 0 : unsigned long flags;
611 :
612 0 : local_irq_save(flags);
613 0 : __wb_writeout_inc(wb);
614 0 : local_irq_restore(flags);
615 0 : }
616 : EXPORT_SYMBOL_GPL(wb_writeout_inc);
617 :
618 : /*
619 : * On idle system, we can be called long after we scheduled because we use
620 : * deferred timers so count with missed periods.
621 : */
622 10 : static void writeout_period(struct timer_list *t)
623 : {
624 10 : struct wb_domain *dom = from_timer(dom, t, period_timer);
625 10 : int miss_periods = (jiffies - dom->period_time) /
626 : VM_COMPLETIONS_PERIOD_LEN;
627 :
628 10 : if (fprop_new_period(&dom->completions, miss_periods + 1)) {
629 10 : dom->period_time = wp_next_time(dom->period_time +
630 10 : miss_periods * VM_COMPLETIONS_PERIOD_LEN);
631 10 : mod_timer(&dom->period_timer, dom->period_time);
632 : } else {
633 : /*
634 : * Aging has zeroed all fractions. Stop wasting CPU on period
635 : * updates.
636 : */
637 0 : dom->period_time = 0;
638 : }
639 10 : }
640 :
641 1 : int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
642 : {
643 1 : memset(dom, 0, sizeof(*dom));
644 :
645 1 : spin_lock_init(&dom->lock);
646 :
647 1 : timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
648 :
649 1 : dom->dirty_limit_tstamp = jiffies;
650 :
651 1 : return fprop_global_init(&dom->completions, gfp);
652 : }
653 :
654 : #ifdef CONFIG_CGROUP_WRITEBACK
655 : void wb_domain_exit(struct wb_domain *dom)
656 : {
657 : del_timer_sync(&dom->period_timer);
658 : fprop_global_destroy(&dom->completions);
659 : }
660 : #endif
661 :
662 : /*
663 : * bdi_min_ratio keeps the sum of the minimum dirty shares of all
664 : * registered backing devices, which, for obvious reasons, can not
665 : * exceed 100%.
666 : */
667 : static unsigned int bdi_min_ratio;
668 :
669 0 : int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
670 : {
671 0 : int ret = 0;
672 :
673 0 : spin_lock_bh(&bdi_lock);
674 0 : if (min_ratio > bdi->max_ratio) {
675 : ret = -EINVAL;
676 : } else {
677 0 : min_ratio -= bdi->min_ratio;
678 0 : if (bdi_min_ratio + min_ratio < 100) {
679 0 : bdi_min_ratio += min_ratio;
680 0 : bdi->min_ratio += min_ratio;
681 : } else {
682 : ret = -EINVAL;
683 : }
684 : }
685 0 : spin_unlock_bh(&bdi_lock);
686 :
687 0 : return ret;
688 : }
689 :
690 0 : int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
691 : {
692 0 : int ret = 0;
693 :
694 0 : if (max_ratio > 100)
695 : return -EINVAL;
696 :
697 0 : spin_lock_bh(&bdi_lock);
698 0 : if (bdi->min_ratio > max_ratio) {
699 : ret = -EINVAL;
700 : } else {
701 0 : bdi->max_ratio = max_ratio;
702 0 : bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
703 : }
704 0 : spin_unlock_bh(&bdi_lock);
705 :
706 0 : return ret;
707 : }
708 : EXPORT_SYMBOL(bdi_set_max_ratio);
709 :
710 25 : static unsigned long dirty_freerun_ceiling(unsigned long thresh,
711 : unsigned long bg_thresh)
712 : {
713 25 : return (thresh + bg_thresh) / 2;
714 : }
715 :
716 0 : static unsigned long hard_dirty_limit(struct wb_domain *dom,
717 : unsigned long thresh)
718 : {
719 0 : return max(thresh, dom->dirty_limit);
720 : }
721 :
722 : /*
723 : * Memory which can be further allocated to a memcg domain is capped by
724 : * system-wide clean memory excluding the amount being used in the domain.
725 : */
726 : static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
727 : unsigned long filepages, unsigned long headroom)
728 : {
729 : struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
730 : unsigned long clean = filepages - min(filepages, mdtc->dirty);
731 : unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
732 : unsigned long other_clean = global_clean - min(global_clean, clean);
733 :
734 : mdtc->avail = filepages + min(headroom, other_clean);
735 : }
736 :
737 : /**
738 : * __wb_calc_thresh - @wb's share of dirty throttling threshold
739 : * @dtc: dirty_throttle_context of interest
740 : *
741 : * Note that balance_dirty_pages() will only seriously take it as a hard limit
742 : * when sleeping max_pause per page is not enough to keep the dirty pages under
743 : * control. For example, when the device is completely stalled due to some error
744 : * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
745 : * In the other normal situations, it acts more gently by throttling the tasks
746 : * more (rather than completely block them) when the wb dirty pages go high.
747 : *
748 : * It allocates high/low dirty limits to fast/slow devices, in order to prevent
749 : * - starving fast devices
750 : * - piling up dirty pages (that will take long time to sync) on slow devices
751 : *
752 : * The wb's share of dirty limit will be adapting to its throughput and
753 : * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
754 : *
755 : * Return: @wb's dirty limit in pages. The term "dirty" in the context of
756 : * dirty balancing includes all PG_dirty and PG_writeback pages.
757 : */
758 7 : static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
759 : {
760 7 : struct wb_domain *dom = dtc_dom(dtc);
761 7 : unsigned long thresh = dtc->thresh;
762 7 : u64 wb_thresh;
763 7 : unsigned long numerator, denominator;
764 7 : unsigned long wb_min_ratio, wb_max_ratio;
765 :
766 : /*
767 : * Calculate this BDI's share of the thresh ratio.
768 : */
769 7 : fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
770 : &numerator, &denominator);
771 :
772 7 : wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
773 7 : wb_thresh *= numerator;
774 7 : wb_thresh = div64_ul(wb_thresh, denominator);
775 :
776 7 : wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
777 :
778 7 : wb_thresh += (thresh * wb_min_ratio) / 100;
779 7 : if (wb_thresh > (thresh * wb_max_ratio) / 100)
780 : wb_thresh = thresh * wb_max_ratio / 100;
781 :
782 7 : return wb_thresh;
783 : }
784 :
785 7 : unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
786 : {
787 7 : struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
788 : .thresh = thresh };
789 7 : return __wb_calc_thresh(&gdtc);
790 : }
791 :
792 : /*
793 : * setpoint - dirty 3
794 : * f(dirty) := 1.0 + (----------------)
795 : * limit - setpoint
796 : *
797 : * it's a 3rd order polynomial that subjects to
798 : *
799 : * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
800 : * (2) f(setpoint) = 1.0 => the balance point
801 : * (3) f(limit) = 0 => the hard limit
802 : * (4) df/dx <= 0 => negative feedback control
803 : * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
804 : * => fast response on large errors; small oscillation near setpoint
805 : */
806 0 : static long long pos_ratio_polynom(unsigned long setpoint,
807 : unsigned long dirty,
808 : unsigned long limit)
809 : {
810 0 : long long pos_ratio;
811 0 : long x;
812 :
813 0 : x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
814 0 : (limit - setpoint) | 1);
815 0 : pos_ratio = x;
816 0 : pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
817 0 : pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
818 0 : pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
819 :
820 0 : return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
821 : }
822 :
823 : /*
824 : * Dirty position control.
825 : *
826 : * (o) global/bdi setpoints
827 : *
828 : * We want the dirty pages be balanced around the global/wb setpoints.
829 : * When the number of dirty pages is higher/lower than the setpoint, the
830 : * dirty position control ratio (and hence task dirty ratelimit) will be
831 : * decreased/increased to bring the dirty pages back to the setpoint.
832 : *
833 : * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
834 : *
835 : * if (dirty < setpoint) scale up pos_ratio
836 : * if (dirty > setpoint) scale down pos_ratio
837 : *
838 : * if (wb_dirty < wb_setpoint) scale up pos_ratio
839 : * if (wb_dirty > wb_setpoint) scale down pos_ratio
840 : *
841 : * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
842 : *
843 : * (o) global control line
844 : *
845 : * ^ pos_ratio
846 : * |
847 : * | |<===== global dirty control scope ======>|
848 : * 2.0 .............*
849 : * | .*
850 : * | . *
851 : * | . *
852 : * | . *
853 : * | . *
854 : * | . *
855 : * 1.0 ................................*
856 : * | . . *
857 : * | . . *
858 : * | . . *
859 : * | . . *
860 : * | . . *
861 : * 0 +------------.------------------.----------------------*------------->
862 : * freerun^ setpoint^ limit^ dirty pages
863 : *
864 : * (o) wb control line
865 : *
866 : * ^ pos_ratio
867 : * |
868 : * | *
869 : * | *
870 : * | *
871 : * | *
872 : * | * |<=========== span ============>|
873 : * 1.0 .......................*
874 : * | . *
875 : * | . *
876 : * | . *
877 : * | . *
878 : * | . *
879 : * | . *
880 : * | . *
881 : * | . *
882 : * | . *
883 : * | . *
884 : * | . *
885 : * 1/4 ...............................................* * * * * * * * * * * *
886 : * | . .
887 : * | . .
888 : * | . .
889 : * 0 +----------------------.-------------------------------.------------->
890 : * wb_setpoint^ x_intercept^
891 : *
892 : * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
893 : * be smoothly throttled down to normal if it starts high in situations like
894 : * - start writing to a slow SD card and a fast disk at the same time. The SD
895 : * card's wb_dirty may rush to many times higher than wb_setpoint.
896 : * - the wb dirty thresh drops quickly due to change of JBOD workload
897 : */
898 0 : static void wb_position_ratio(struct dirty_throttle_control *dtc)
899 : {
900 0 : struct bdi_writeback *wb = dtc->wb;
901 0 : unsigned long write_bw = wb->avg_write_bandwidth;
902 0 : unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
903 0 : unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
904 0 : unsigned long wb_thresh = dtc->wb_thresh;
905 0 : unsigned long x_intercept;
906 0 : unsigned long setpoint; /* dirty pages' target balance point */
907 0 : unsigned long wb_setpoint;
908 0 : unsigned long span;
909 0 : long long pos_ratio; /* for scaling up/down the rate limit */
910 0 : long x;
911 :
912 0 : dtc->pos_ratio = 0;
913 :
914 0 : if (unlikely(dtc->dirty >= limit))
915 : return;
916 :
917 : /*
918 : * global setpoint
919 : *
920 : * See comment for pos_ratio_polynom().
921 : */
922 0 : setpoint = (freerun + limit) / 2;
923 0 : pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
924 :
925 : /*
926 : * The strictlimit feature is a tool preventing mistrusted filesystems
927 : * from growing a large number of dirty pages before throttling. For
928 : * such filesystems balance_dirty_pages always checks wb counters
929 : * against wb limits. Even if global "nr_dirty" is under "freerun".
930 : * This is especially important for fuse which sets bdi->max_ratio to
931 : * 1% by default. Without strictlimit feature, fuse writeback may
932 : * consume arbitrary amount of RAM because it is accounted in
933 : * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
934 : *
935 : * Here, in wb_position_ratio(), we calculate pos_ratio based on
936 : * two values: wb_dirty and wb_thresh. Let's consider an example:
937 : * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
938 : * limits are set by default to 10% and 20% (background and throttle).
939 : * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
940 : * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
941 : * about ~6K pages (as the average of background and throttle wb
942 : * limits). The 3rd order polynomial will provide positive feedback if
943 : * wb_dirty is under wb_setpoint and vice versa.
944 : *
945 : * Note, that we cannot use global counters in these calculations
946 : * because we want to throttle process writing to a strictlimit wb
947 : * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
948 : * in the example above).
949 : */
950 0 : if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
951 0 : long long wb_pos_ratio;
952 :
953 0 : if (dtc->wb_dirty < 8) {
954 0 : dtc->pos_ratio = min_t(long long, pos_ratio * 2,
955 : 2 << RATELIMIT_CALC_SHIFT);
956 0 : return;
957 : }
958 :
959 0 : if (dtc->wb_dirty >= wb_thresh)
960 : return;
961 :
962 0 : wb_setpoint = dirty_freerun_ceiling(wb_thresh,
963 : dtc->wb_bg_thresh);
964 :
965 0 : if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
966 : return;
967 :
968 0 : wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
969 : wb_thresh);
970 :
971 : /*
972 : * Typically, for strictlimit case, wb_setpoint << setpoint
973 : * and pos_ratio >> wb_pos_ratio. In the other words global
974 : * state ("dirty") is not limiting factor and we have to
975 : * make decision based on wb counters. But there is an
976 : * important case when global pos_ratio should get precedence:
977 : * global limits are exceeded (e.g. due to activities on other
978 : * wb's) while given strictlimit wb is below limit.
979 : *
980 : * "pos_ratio * wb_pos_ratio" would work for the case above,
981 : * but it would look too non-natural for the case of all
982 : * activity in the system coming from a single strictlimit wb
983 : * with bdi->max_ratio == 100%.
984 : *
985 : * Note that min() below somewhat changes the dynamics of the
986 : * control system. Normally, pos_ratio value can be well over 3
987 : * (when globally we are at freerun and wb is well below wb
988 : * setpoint). Now the maximum pos_ratio in the same situation
989 : * is 2. We might want to tweak this if we observe the control
990 : * system is too slow to adapt.
991 : */
992 0 : dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
993 0 : return;
994 : }
995 :
996 : /*
997 : * We have computed basic pos_ratio above based on global situation. If
998 : * the wb is over/under its share of dirty pages, we want to scale
999 : * pos_ratio further down/up. That is done by the following mechanism.
1000 : */
1001 :
1002 : /*
1003 : * wb setpoint
1004 : *
1005 : * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1006 : *
1007 : * x_intercept - wb_dirty
1008 : * := --------------------------
1009 : * x_intercept - wb_setpoint
1010 : *
1011 : * The main wb control line is a linear function that subjects to
1012 : *
1013 : * (1) f(wb_setpoint) = 1.0
1014 : * (2) k = - 1 / (8 * write_bw) (in single wb case)
1015 : * or equally: x_intercept = wb_setpoint + 8 * write_bw
1016 : *
1017 : * For single wb case, the dirty pages are observed to fluctuate
1018 : * regularly within range
1019 : * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1020 : * for various filesystems, where (2) can yield in a reasonable 12.5%
1021 : * fluctuation range for pos_ratio.
1022 : *
1023 : * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1024 : * own size, so move the slope over accordingly and choose a slope that
1025 : * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1026 : */
1027 0 : if (unlikely(wb_thresh > dtc->thresh))
1028 0 : wb_thresh = dtc->thresh;
1029 : /*
1030 : * It's very possible that wb_thresh is close to 0 not because the
1031 : * device is slow, but that it has remained inactive for long time.
1032 : * Honour such devices a reasonable good (hopefully IO efficient)
1033 : * threshold, so that the occasional writes won't be blocked and active
1034 : * writes can rampup the threshold quickly.
1035 : */
1036 0 : wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1037 : /*
1038 : * scale global setpoint to wb's:
1039 : * wb_setpoint = setpoint * wb_thresh / thresh
1040 : */
1041 0 : x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1042 0 : wb_setpoint = setpoint * (u64)x >> 16;
1043 : /*
1044 : * Use span=(8*write_bw) in single wb case as indicated by
1045 : * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1046 : *
1047 : * wb_thresh thresh - wb_thresh
1048 : * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1049 : * thresh thresh
1050 : */
1051 0 : span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1052 0 : x_intercept = wb_setpoint + span;
1053 :
1054 0 : if (dtc->wb_dirty < x_intercept - span / 4) {
1055 0 : pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1056 0 : (x_intercept - wb_setpoint) | 1);
1057 : } else
1058 0 : pos_ratio /= 4;
1059 :
1060 : /*
1061 : * wb reserve area, safeguard against dirty pool underrun and disk idle
1062 : * It may push the desired control point of global dirty pages higher
1063 : * than setpoint.
1064 : */
1065 0 : x_intercept = wb_thresh / 2;
1066 0 : if (dtc->wb_dirty < x_intercept) {
1067 0 : if (dtc->wb_dirty > x_intercept / 8)
1068 0 : pos_ratio = div_u64(pos_ratio * x_intercept,
1069 : dtc->wb_dirty);
1070 : else
1071 0 : pos_ratio *= 8;
1072 : }
1073 :
1074 0 : dtc->pos_ratio = pos_ratio;
1075 : }
1076 :
1077 0 : static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1078 : unsigned long elapsed,
1079 : unsigned long written)
1080 : {
1081 0 : const unsigned long period = roundup_pow_of_two(3 * HZ);
1082 0 : unsigned long avg = wb->avg_write_bandwidth;
1083 0 : unsigned long old = wb->write_bandwidth;
1084 0 : u64 bw;
1085 :
1086 : /*
1087 : * bw = written * HZ / elapsed
1088 : *
1089 : * bw * elapsed + write_bandwidth * (period - elapsed)
1090 : * write_bandwidth = ---------------------------------------------------
1091 : * period
1092 : *
1093 : * @written may have decreased due to account_page_redirty().
1094 : * Avoid underflowing @bw calculation.
1095 : */
1096 0 : bw = written - min(written, wb->written_stamp);
1097 0 : bw *= HZ;
1098 0 : if (unlikely(elapsed > period)) {
1099 0 : bw = div64_ul(bw, elapsed);
1100 0 : avg = bw;
1101 0 : goto out;
1102 : }
1103 0 : bw += (u64)wb->write_bandwidth * (period - elapsed);
1104 0 : bw >>= ilog2(period);
1105 :
1106 : /*
1107 : * one more level of smoothing, for filtering out sudden spikes
1108 : */
1109 0 : if (avg > old && old >= (unsigned long)bw)
1110 0 : avg -= (avg - old) >> 3;
1111 :
1112 0 : if (avg < old && old <= (unsigned long)bw)
1113 0 : avg += (old - avg) >> 3;
1114 :
1115 0 : out:
1116 : /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1117 0 : avg = max(avg, 1LU);
1118 0 : if (wb_has_dirty_io(wb)) {
1119 0 : long delta = avg - wb->avg_write_bandwidth;
1120 0 : WARN_ON_ONCE(atomic_long_add_return(delta,
1121 : &wb->bdi->tot_write_bandwidth) <= 0);
1122 : }
1123 0 : wb->write_bandwidth = bw;
1124 0 : wb->avg_write_bandwidth = avg;
1125 0 : }
1126 :
1127 0 : static void update_dirty_limit(struct dirty_throttle_control *dtc)
1128 : {
1129 0 : struct wb_domain *dom = dtc_dom(dtc);
1130 0 : unsigned long thresh = dtc->thresh;
1131 0 : unsigned long limit = dom->dirty_limit;
1132 :
1133 : /*
1134 : * Follow up in one step.
1135 : */
1136 0 : if (limit < thresh) {
1137 0 : limit = thresh;
1138 0 : goto update;
1139 : }
1140 :
1141 : /*
1142 : * Follow down slowly. Use the higher one as the target, because thresh
1143 : * may drop below dirty. This is exactly the reason to introduce
1144 : * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1145 : */
1146 0 : thresh = max(thresh, dtc->dirty);
1147 0 : if (limit > thresh) {
1148 0 : limit -= (limit - thresh) >> 5;
1149 0 : goto update;
1150 : }
1151 : return;
1152 0 : update:
1153 0 : dom->dirty_limit = limit;
1154 : }
1155 :
1156 0 : static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1157 : unsigned long now)
1158 : {
1159 0 : struct wb_domain *dom = dtc_dom(dtc);
1160 :
1161 : /*
1162 : * check locklessly first to optimize away locking for the most time
1163 : */
1164 0 : if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1165 : return;
1166 :
1167 0 : spin_lock(&dom->lock);
1168 0 : if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1169 0 : update_dirty_limit(dtc);
1170 0 : dom->dirty_limit_tstamp = now;
1171 : }
1172 0 : spin_unlock(&dom->lock);
1173 : }
1174 :
1175 : /*
1176 : * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1177 : *
1178 : * Normal wb tasks will be curbed at or below it in long term.
1179 : * Obviously it should be around (write_bw / N) when there are N dd tasks.
1180 : */
1181 0 : static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1182 : unsigned long dirtied,
1183 : unsigned long elapsed)
1184 : {
1185 0 : struct bdi_writeback *wb = dtc->wb;
1186 0 : unsigned long dirty = dtc->dirty;
1187 0 : unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1188 0 : unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1189 0 : unsigned long setpoint = (freerun + limit) / 2;
1190 0 : unsigned long write_bw = wb->avg_write_bandwidth;
1191 0 : unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1192 0 : unsigned long dirty_rate;
1193 0 : unsigned long task_ratelimit;
1194 0 : unsigned long balanced_dirty_ratelimit;
1195 0 : unsigned long step;
1196 0 : unsigned long x;
1197 0 : unsigned long shift;
1198 :
1199 : /*
1200 : * The dirty rate will match the writeout rate in long term, except
1201 : * when dirty pages are truncated by userspace or re-dirtied by FS.
1202 : */
1203 0 : dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1204 :
1205 : /*
1206 : * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1207 : */
1208 0 : task_ratelimit = (u64)dirty_ratelimit *
1209 0 : dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1210 0 : task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1211 :
1212 : /*
1213 : * A linear estimation of the "balanced" throttle rate. The theory is,
1214 : * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1215 : * dirty_rate will be measured to be (N * task_ratelimit). So the below
1216 : * formula will yield the balanced rate limit (write_bw / N).
1217 : *
1218 : * Note that the expanded form is not a pure rate feedback:
1219 : * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1220 : * but also takes pos_ratio into account:
1221 : * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1222 : *
1223 : * (1) is not realistic because pos_ratio also takes part in balancing
1224 : * the dirty rate. Consider the state
1225 : * pos_ratio = 0.5 (3)
1226 : * rate = 2 * (write_bw / N) (4)
1227 : * If (1) is used, it will stuck in that state! Because each dd will
1228 : * be throttled at
1229 : * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1230 : * yielding
1231 : * dirty_rate = N * task_ratelimit = write_bw (6)
1232 : * put (6) into (1) we get
1233 : * rate_(i+1) = rate_(i) (7)
1234 : *
1235 : * So we end up using (2) to always keep
1236 : * rate_(i+1) ~= (write_bw / N) (8)
1237 : * regardless of the value of pos_ratio. As long as (8) is satisfied,
1238 : * pos_ratio is able to drive itself to 1.0, which is not only where
1239 : * the dirty count meet the setpoint, but also where the slope of
1240 : * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1241 : */
1242 0 : balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1243 : dirty_rate | 1);
1244 : /*
1245 : * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1246 : */
1247 0 : if (unlikely(balanced_dirty_ratelimit > write_bw))
1248 0 : balanced_dirty_ratelimit = write_bw;
1249 :
1250 : /*
1251 : * We could safely do this and return immediately:
1252 : *
1253 : * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1254 : *
1255 : * However to get a more stable dirty_ratelimit, the below elaborated
1256 : * code makes use of task_ratelimit to filter out singular points and
1257 : * limit the step size.
1258 : *
1259 : * The below code essentially only uses the relative value of
1260 : *
1261 : * task_ratelimit - dirty_ratelimit
1262 : * = (pos_ratio - 1) * dirty_ratelimit
1263 : *
1264 : * which reflects the direction and size of dirty position error.
1265 : */
1266 :
1267 : /*
1268 : * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1269 : * task_ratelimit is on the same side of dirty_ratelimit, too.
1270 : * For example, when
1271 : * - dirty_ratelimit > balanced_dirty_ratelimit
1272 : * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1273 : * lowering dirty_ratelimit will help meet both the position and rate
1274 : * control targets. Otherwise, don't update dirty_ratelimit if it will
1275 : * only help meet the rate target. After all, what the users ultimately
1276 : * feel and care are stable dirty rate and small position error.
1277 : *
1278 : * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1279 : * and filter out the singular points of balanced_dirty_ratelimit. Which
1280 : * keeps jumping around randomly and can even leap far away at times
1281 : * due to the small 200ms estimation period of dirty_rate (we want to
1282 : * keep that period small to reduce time lags).
1283 : */
1284 0 : step = 0;
1285 :
1286 : /*
1287 : * For strictlimit case, calculations above were based on wb counters
1288 : * and limits (starting from pos_ratio = wb_position_ratio() and up to
1289 : * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1290 : * Hence, to calculate "step" properly, we have to use wb_dirty as
1291 : * "dirty" and wb_setpoint as "setpoint".
1292 : *
1293 : * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1294 : * it's possible that wb_thresh is close to zero due to inactivity
1295 : * of backing device.
1296 : */
1297 0 : if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1298 0 : dirty = dtc->wb_dirty;
1299 0 : if (dtc->wb_dirty < 8)
1300 0 : setpoint = dtc->wb_dirty + 1;
1301 : else
1302 0 : setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1303 : }
1304 :
1305 0 : if (dirty < setpoint) {
1306 0 : x = min3(wb->balanced_dirty_ratelimit,
1307 : balanced_dirty_ratelimit, task_ratelimit);
1308 0 : if (dirty_ratelimit < x)
1309 0 : step = x - dirty_ratelimit;
1310 : } else {
1311 0 : x = max3(wb->balanced_dirty_ratelimit,
1312 : balanced_dirty_ratelimit, task_ratelimit);
1313 0 : if (dirty_ratelimit > x)
1314 0 : step = dirty_ratelimit - x;
1315 : }
1316 :
1317 : /*
1318 : * Don't pursue 100% rate matching. It's impossible since the balanced
1319 : * rate itself is constantly fluctuating. So decrease the track speed
1320 : * when it gets close to the target. Helps eliminate pointless tremors.
1321 : */
1322 0 : shift = dirty_ratelimit / (2 * step + 1);
1323 0 : if (shift < BITS_PER_LONG)
1324 0 : step = DIV_ROUND_UP(step >> shift, 8);
1325 : else
1326 : step = 0;
1327 :
1328 0 : if (dirty_ratelimit < balanced_dirty_ratelimit)
1329 0 : dirty_ratelimit += step;
1330 : else
1331 0 : dirty_ratelimit -= step;
1332 :
1333 0 : wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1334 0 : wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1335 :
1336 0 : trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1337 0 : }
1338 :
1339 13 : static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1340 : struct dirty_throttle_control *mdtc,
1341 : unsigned long start_time,
1342 : bool update_ratelimit)
1343 : {
1344 13 : struct bdi_writeback *wb = gdtc->wb;
1345 13 : unsigned long now = jiffies;
1346 13 : unsigned long elapsed = now - wb->bw_time_stamp;
1347 13 : unsigned long dirtied;
1348 13 : unsigned long written;
1349 :
1350 26 : lockdep_assert_held(&wb->list_lock);
1351 :
1352 : /*
1353 : * rate-limit, only update once every 200ms.
1354 : */
1355 13 : if (elapsed < BANDWIDTH_INTERVAL)
1356 : return;
1357 :
1358 7 : dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1359 7 : written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1360 :
1361 : /*
1362 : * Skip quiet periods when disk bandwidth is under-utilized.
1363 : * (at least 1s idle time between two flusher runs)
1364 : */
1365 7 : if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1366 7 : goto snapshot;
1367 :
1368 0 : if (update_ratelimit) {
1369 0 : domain_update_bandwidth(gdtc, now);
1370 0 : wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1371 :
1372 : /*
1373 : * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1374 : * compiler has no way to figure that out. Help it.
1375 : */
1376 0 : if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1377 : domain_update_bandwidth(mdtc, now);
1378 : wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1379 : }
1380 : }
1381 0 : wb_update_write_bandwidth(wb, elapsed, written);
1382 :
1383 7 : snapshot:
1384 7 : wb->dirtied_stamp = dirtied;
1385 7 : wb->written_stamp = written;
1386 7 : wb->bw_time_stamp = now;
1387 : }
1388 :
1389 13 : void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1390 : {
1391 13 : struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1392 :
1393 13 : __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1394 13 : }
1395 :
1396 : /*
1397 : * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1398 : * will look to see if it needs to start dirty throttling.
1399 : *
1400 : * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1401 : * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1402 : * (the number of pages we may dirty without exceeding the dirty limits).
1403 : */
1404 25 : static unsigned long dirty_poll_interval(unsigned long dirty,
1405 : unsigned long thresh)
1406 : {
1407 25 : if (thresh > dirty)
1408 25 : return 1UL << (ilog2(thresh - dirty) >> 1);
1409 :
1410 : return 1;
1411 : }
1412 :
1413 0 : static unsigned long wb_max_pause(struct bdi_writeback *wb,
1414 : unsigned long wb_dirty)
1415 : {
1416 0 : unsigned long bw = wb->avg_write_bandwidth;
1417 0 : unsigned long t;
1418 :
1419 : /*
1420 : * Limit pause time for small memory systems. If sleeping for too long
1421 : * time, a small pool of dirty/writeback pages may go empty and disk go
1422 : * idle.
1423 : *
1424 : * 8 serves as the safety ratio.
1425 : */
1426 0 : t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1427 0 : t++;
1428 :
1429 0 : return min_t(unsigned long, t, MAX_PAUSE);
1430 : }
1431 :
1432 0 : static long wb_min_pause(struct bdi_writeback *wb,
1433 : long max_pause,
1434 : unsigned long task_ratelimit,
1435 : unsigned long dirty_ratelimit,
1436 : int *nr_dirtied_pause)
1437 : {
1438 0 : long hi = ilog2(wb->avg_write_bandwidth);
1439 0 : long lo = ilog2(wb->dirty_ratelimit);
1440 0 : long t; /* target pause */
1441 0 : long pause; /* estimated next pause */
1442 0 : int pages; /* target nr_dirtied_pause */
1443 :
1444 : /* target for 10ms pause on 1-dd case */
1445 0 : t = max(1, HZ / 100);
1446 :
1447 : /*
1448 : * Scale up pause time for concurrent dirtiers in order to reduce CPU
1449 : * overheads.
1450 : *
1451 : * (N * 10ms) on 2^N concurrent tasks.
1452 : */
1453 0 : if (hi > lo)
1454 0 : t += (hi - lo) * (10 * HZ) / 1024;
1455 :
1456 : /*
1457 : * This is a bit convoluted. We try to base the next nr_dirtied_pause
1458 : * on the much more stable dirty_ratelimit. However the next pause time
1459 : * will be computed based on task_ratelimit and the two rate limits may
1460 : * depart considerably at some time. Especially if task_ratelimit goes
1461 : * below dirty_ratelimit/2 and the target pause is max_pause, the next
1462 : * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1463 : * result task_ratelimit won't be executed faithfully, which could
1464 : * eventually bring down dirty_ratelimit.
1465 : *
1466 : * We apply two rules to fix it up:
1467 : * 1) try to estimate the next pause time and if necessary, use a lower
1468 : * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1469 : * nr_dirtied_pause will be "dancing" with task_ratelimit.
1470 : * 2) limit the target pause time to max_pause/2, so that the normal
1471 : * small fluctuations of task_ratelimit won't trigger rule (1) and
1472 : * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1473 : */
1474 0 : t = min(t, 1 + max_pause / 2);
1475 0 : pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1476 :
1477 : /*
1478 : * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1479 : * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1480 : * When the 16 consecutive reads are often interrupted by some dirty
1481 : * throttling pause during the async writes, cfq will go into idles
1482 : * (deadline is fine). So push nr_dirtied_pause as high as possible
1483 : * until reaches DIRTY_POLL_THRESH=32 pages.
1484 : */
1485 0 : if (pages < DIRTY_POLL_THRESH) {
1486 0 : t = max_pause;
1487 0 : pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1488 0 : if (pages > DIRTY_POLL_THRESH) {
1489 0 : pages = DIRTY_POLL_THRESH;
1490 0 : t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1491 : }
1492 : }
1493 :
1494 0 : pause = HZ * pages / (task_ratelimit + 1);
1495 0 : if (pause > max_pause) {
1496 0 : t = max_pause;
1497 0 : pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1498 : }
1499 :
1500 0 : *nr_dirtied_pause = pages;
1501 : /*
1502 : * The minimal pause time will normally be half the target pause time.
1503 : */
1504 0 : return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1505 : }
1506 :
1507 0 : static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1508 : {
1509 0 : struct bdi_writeback *wb = dtc->wb;
1510 0 : unsigned long wb_reclaimable;
1511 :
1512 : /*
1513 : * wb_thresh is not treated as some limiting factor as
1514 : * dirty_thresh, due to reasons
1515 : * - in JBOD setup, wb_thresh can fluctuate a lot
1516 : * - in a system with HDD and USB key, the USB key may somehow
1517 : * go into state (wb_dirty >> wb_thresh) either because
1518 : * wb_dirty starts high, or because wb_thresh drops low.
1519 : * In this case we don't want to hard throttle the USB key
1520 : * dirtiers for 100 seconds until wb_dirty drops under
1521 : * wb_thresh. Instead the auxiliary wb control line in
1522 : * wb_position_ratio() will let the dirtier task progress
1523 : * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1524 : */
1525 0 : dtc->wb_thresh = __wb_calc_thresh(dtc);
1526 0 : dtc->wb_bg_thresh = dtc->thresh ?
1527 0 : div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1528 :
1529 : /*
1530 : * In order to avoid the stacked BDI deadlock we need
1531 : * to ensure we accurately count the 'dirty' pages when
1532 : * the threshold is low.
1533 : *
1534 : * Otherwise it would be possible to get thresh+n pages
1535 : * reported dirty, even though there are thresh-m pages
1536 : * actually dirty; with m+n sitting in the percpu
1537 : * deltas.
1538 : */
1539 0 : if (dtc->wb_thresh < 2 * wb_stat_error()) {
1540 0 : wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1541 0 : dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1542 : } else {
1543 0 : wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1544 0 : dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1545 : }
1546 0 : }
1547 :
1548 : /*
1549 : * balance_dirty_pages() must be called by processes which are generating dirty
1550 : * data. It looks at the number of dirty pages in the machine and will force
1551 : * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1552 : * If we're over `background_thresh' then the writeback threads are woken to
1553 : * perform some writeout.
1554 : */
1555 25 : static void balance_dirty_pages(struct bdi_writeback *wb,
1556 : unsigned long pages_dirtied)
1557 : {
1558 25 : struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1559 25 : struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1560 25 : struct dirty_throttle_control * const gdtc = &gdtc_stor;
1561 25 : struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1562 : &mdtc_stor : NULL;
1563 25 : struct dirty_throttle_control *sdtc;
1564 25 : unsigned long nr_reclaimable; /* = file_dirty */
1565 25 : long period;
1566 25 : long pause;
1567 25 : long max_pause;
1568 25 : long min_pause;
1569 25 : int nr_dirtied_pause;
1570 25 : bool dirty_exceeded = false;
1571 25 : unsigned long task_ratelimit;
1572 25 : unsigned long dirty_ratelimit;
1573 25 : struct backing_dev_info *bdi = wb->bdi;
1574 25 : bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1575 25 : unsigned long start_time = jiffies;
1576 :
1577 25 : for (;;) {
1578 25 : unsigned long now = jiffies;
1579 25 : unsigned long dirty, thresh, bg_thresh;
1580 25 : unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1581 25 : unsigned long m_thresh = 0;
1582 25 : unsigned long m_bg_thresh = 0;
1583 :
1584 25 : nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1585 25 : gdtc->avail = global_dirtyable_memory();
1586 25 : gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1587 :
1588 25 : domain_dirty_limits(gdtc);
1589 :
1590 25 : if (unlikely(strictlimit)) {
1591 0 : wb_dirty_limits(gdtc);
1592 :
1593 0 : dirty = gdtc->wb_dirty;
1594 0 : thresh = gdtc->wb_thresh;
1595 0 : bg_thresh = gdtc->wb_bg_thresh;
1596 : } else {
1597 25 : dirty = gdtc->dirty;
1598 25 : thresh = gdtc->thresh;
1599 25 : bg_thresh = gdtc->bg_thresh;
1600 : }
1601 :
1602 25 : if (mdtc) {
1603 : unsigned long filepages, headroom, writeback;
1604 :
1605 : /*
1606 : * If @wb belongs to !root memcg, repeat the same
1607 : * basic calculations for the memcg domain.
1608 : */
1609 : mem_cgroup_wb_stats(wb, &filepages, &headroom,
1610 : &mdtc->dirty, &writeback);
1611 : mdtc->dirty += writeback;
1612 : mdtc_calc_avail(mdtc, filepages, headroom);
1613 :
1614 : domain_dirty_limits(mdtc);
1615 :
1616 : if (unlikely(strictlimit)) {
1617 : wb_dirty_limits(mdtc);
1618 : m_dirty = mdtc->wb_dirty;
1619 : m_thresh = mdtc->wb_thresh;
1620 : m_bg_thresh = mdtc->wb_bg_thresh;
1621 : } else {
1622 : m_dirty = mdtc->dirty;
1623 : m_thresh = mdtc->thresh;
1624 : m_bg_thresh = mdtc->bg_thresh;
1625 : }
1626 : }
1627 :
1628 : /*
1629 : * Throttle it only when the background writeback cannot
1630 : * catch-up. This avoids (excessively) small writeouts
1631 : * when the wb limits are ramping up in case of !strictlimit.
1632 : *
1633 : * In strictlimit case make decision based on the wb counters
1634 : * and limits. Small writeouts when the wb limits are ramping
1635 : * up are the price we consciously pay for strictlimit-ing.
1636 : *
1637 : * If memcg domain is in effect, @dirty should be under
1638 : * both global and memcg freerun ceilings.
1639 : */
1640 25 : if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1641 : (!mdtc ||
1642 : m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1643 25 : unsigned long intv;
1644 25 : unsigned long m_intv;
1645 :
1646 25 : free_running:
1647 25 : intv = dirty_poll_interval(dirty, thresh);
1648 25 : m_intv = ULONG_MAX;
1649 :
1650 25 : current->dirty_paused_when = now;
1651 25 : current->nr_dirtied = 0;
1652 25 : if (mdtc)
1653 : m_intv = dirty_poll_interval(m_dirty, m_thresh);
1654 25 : current->nr_dirtied_pause = min(intv, m_intv);
1655 25 : break;
1656 : }
1657 :
1658 0 : if (unlikely(!writeback_in_progress(wb)))
1659 0 : wb_start_background_writeback(wb);
1660 :
1661 0 : mem_cgroup_flush_foreign(wb);
1662 :
1663 : /*
1664 : * Calculate global domain's pos_ratio and select the
1665 : * global dtc by default.
1666 : */
1667 0 : if (!strictlimit) {
1668 0 : wb_dirty_limits(gdtc);
1669 :
1670 0 : if ((current->flags & PF_LOCAL_THROTTLE) &&
1671 0 : gdtc->wb_dirty <
1672 0 : dirty_freerun_ceiling(gdtc->wb_thresh,
1673 : gdtc->wb_bg_thresh))
1674 : /*
1675 : * LOCAL_THROTTLE tasks must not be throttled
1676 : * when below the per-wb freerun ceiling.
1677 : */
1678 0 : goto free_running;
1679 : }
1680 :
1681 0 : dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1682 0 : ((gdtc->dirty > gdtc->thresh) || strictlimit);
1683 :
1684 0 : wb_position_ratio(gdtc);
1685 0 : sdtc = gdtc;
1686 :
1687 0 : if (mdtc) {
1688 : /*
1689 : * If memcg domain is in effect, calculate its
1690 : * pos_ratio. @wb should satisfy constraints from
1691 : * both global and memcg domains. Choose the one
1692 : * w/ lower pos_ratio.
1693 : */
1694 : if (!strictlimit) {
1695 : wb_dirty_limits(mdtc);
1696 :
1697 : if ((current->flags & PF_LOCAL_THROTTLE) &&
1698 : mdtc->wb_dirty <
1699 : dirty_freerun_ceiling(mdtc->wb_thresh,
1700 : mdtc->wb_bg_thresh))
1701 : /*
1702 : * LOCAL_THROTTLE tasks must not be
1703 : * throttled when below the per-wb
1704 : * freerun ceiling.
1705 : */
1706 : goto free_running;
1707 : }
1708 : dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1709 : ((mdtc->dirty > mdtc->thresh) || strictlimit);
1710 :
1711 : wb_position_ratio(mdtc);
1712 : if (mdtc->pos_ratio < gdtc->pos_ratio)
1713 : sdtc = mdtc;
1714 : }
1715 :
1716 0 : if (dirty_exceeded && !wb->dirty_exceeded)
1717 0 : wb->dirty_exceeded = 1;
1718 :
1719 0 : if (time_is_before_jiffies(wb->bw_time_stamp +
1720 : BANDWIDTH_INTERVAL)) {
1721 0 : spin_lock(&wb->list_lock);
1722 0 : __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1723 0 : spin_unlock(&wb->list_lock);
1724 : }
1725 :
1726 : /* throttle according to the chosen dtc */
1727 0 : dirty_ratelimit = wb->dirty_ratelimit;
1728 0 : task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1729 : RATELIMIT_CALC_SHIFT;
1730 0 : max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1731 0 : min_pause = wb_min_pause(wb, max_pause,
1732 : task_ratelimit, dirty_ratelimit,
1733 : &nr_dirtied_pause);
1734 :
1735 0 : if (unlikely(task_ratelimit == 0)) {
1736 0 : period = max_pause;
1737 0 : pause = max_pause;
1738 0 : goto pause;
1739 : }
1740 0 : period = HZ * pages_dirtied / task_ratelimit;
1741 0 : pause = period;
1742 0 : if (current->dirty_paused_when)
1743 0 : pause -= now - current->dirty_paused_when;
1744 : /*
1745 : * For less than 1s think time (ext3/4 may block the dirtier
1746 : * for up to 800ms from time to time on 1-HDD; so does xfs,
1747 : * however at much less frequency), try to compensate it in
1748 : * future periods by updating the virtual time; otherwise just
1749 : * do a reset, as it may be a light dirtier.
1750 : */
1751 0 : if (pause < min_pause) {
1752 0 : trace_balance_dirty_pages(wb,
1753 : sdtc->thresh,
1754 : sdtc->bg_thresh,
1755 : sdtc->dirty,
1756 : sdtc->wb_thresh,
1757 : sdtc->wb_dirty,
1758 : dirty_ratelimit,
1759 : task_ratelimit,
1760 : pages_dirtied,
1761 : period,
1762 0 : min(pause, 0L),
1763 : start_time);
1764 0 : if (pause < -HZ) {
1765 0 : current->dirty_paused_when = now;
1766 0 : current->nr_dirtied = 0;
1767 0 : } else if (period) {
1768 0 : current->dirty_paused_when += period;
1769 0 : current->nr_dirtied = 0;
1770 0 : } else if (current->nr_dirtied_pause <= pages_dirtied)
1771 0 : current->nr_dirtied_pause += pages_dirtied;
1772 : break;
1773 : }
1774 0 : if (unlikely(pause > max_pause)) {
1775 : /* for occasional dropped task_ratelimit */
1776 0 : now += min(pause - max_pause, max_pause);
1777 0 : pause = max_pause;
1778 : }
1779 :
1780 0 : pause:
1781 0 : trace_balance_dirty_pages(wb,
1782 : sdtc->thresh,
1783 : sdtc->bg_thresh,
1784 : sdtc->dirty,
1785 : sdtc->wb_thresh,
1786 : sdtc->wb_dirty,
1787 : dirty_ratelimit,
1788 : task_ratelimit,
1789 : pages_dirtied,
1790 : period,
1791 : pause,
1792 : start_time);
1793 0 : __set_current_state(TASK_KILLABLE);
1794 0 : wb->dirty_sleep = now;
1795 0 : io_schedule_timeout(pause);
1796 :
1797 0 : current->dirty_paused_when = now + pause;
1798 0 : current->nr_dirtied = 0;
1799 0 : current->nr_dirtied_pause = nr_dirtied_pause;
1800 :
1801 : /*
1802 : * This is typically equal to (dirty < thresh) and can also
1803 : * keep "1000+ dd on a slow USB stick" under control.
1804 : */
1805 0 : if (task_ratelimit)
1806 : break;
1807 :
1808 : /*
1809 : * In the case of an unresponding NFS server and the NFS dirty
1810 : * pages exceeds dirty_thresh, give the other good wb's a pipe
1811 : * to go through, so that tasks on them still remain responsive.
1812 : *
1813 : * In theory 1 page is enough to keep the consumer-producer
1814 : * pipe going: the flusher cleans 1 page => the task dirties 1
1815 : * more page. However wb_dirty has accounting errors. So use
1816 : * the larger and more IO friendly wb_stat_error.
1817 : */
1818 0 : if (sdtc->wb_dirty <= wb_stat_error())
1819 : break;
1820 :
1821 0 : if (fatal_signal_pending(current))
1822 : break;
1823 : }
1824 :
1825 25 : if (!dirty_exceeded && wb->dirty_exceeded)
1826 0 : wb->dirty_exceeded = 0;
1827 :
1828 25 : if (writeback_in_progress(wb))
1829 0 : return;
1830 :
1831 : /*
1832 : * In laptop mode, we wait until hitting the higher threshold before
1833 : * starting background writeout, and then write out all the way down
1834 : * to the lower threshold. So slow writers cause minimal disk activity.
1835 : *
1836 : * In normal mode, we start background writeout at the lower
1837 : * background_thresh, to keep the amount of dirty memory low.
1838 : */
1839 25 : if (laptop_mode)
1840 : return;
1841 :
1842 25 : if (nr_reclaimable > gdtc->bg_thresh)
1843 0 : wb_start_background_writeback(wb);
1844 : }
1845 :
1846 : static DEFINE_PER_CPU(int, bdp_ratelimits);
1847 :
1848 : /*
1849 : * Normal tasks are throttled by
1850 : * loop {
1851 : * dirty tsk->nr_dirtied_pause pages;
1852 : * take a snap in balance_dirty_pages();
1853 : * }
1854 : * However there is a worst case. If every task exit immediately when dirtied
1855 : * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1856 : * called to throttle the page dirties. The solution is to save the not yet
1857 : * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1858 : * randomly into the running tasks. This works well for the above worst case,
1859 : * as the new task will pick up and accumulate the old task's leaked dirty
1860 : * count and eventually get throttled.
1861 : */
1862 : DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1863 :
1864 : /**
1865 : * balance_dirty_pages_ratelimited - balance dirty memory state
1866 : * @mapping: address_space which was dirtied
1867 : *
1868 : * Processes which are dirtying memory should call in here once for each page
1869 : * which was newly dirtied. The function will periodically check the system's
1870 : * dirty state and will initiate writeback if needed.
1871 : *
1872 : * On really big machines, get_writeback_state is expensive, so try to avoid
1873 : * calling it too often (ratelimiting). But once we're over the dirty memory
1874 : * limit we decrease the ratelimiting by a lot, to prevent individual processes
1875 : * from overshooting the limit by (ratelimit_pages) each.
1876 : */
1877 2896 : void balance_dirty_pages_ratelimited(struct address_space *mapping)
1878 : {
1879 2896 : struct inode *inode = mapping->host;
1880 2896 : struct backing_dev_info *bdi = inode_to_bdi(inode);
1881 2896 : struct bdi_writeback *wb = NULL;
1882 2896 : int ratelimit;
1883 2896 : int *p;
1884 :
1885 2896 : if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1886 : return;
1887 :
1888 2181 : if (inode_cgwb_enabled(inode))
1889 : wb = wb_get_create_current(bdi, GFP_KERNEL);
1890 2181 : if (!wb)
1891 2181 : wb = &bdi->wb;
1892 :
1893 2181 : ratelimit = current->nr_dirtied_pause;
1894 2181 : if (wb->dirty_exceeded)
1895 0 : ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1896 :
1897 2181 : preempt_disable();
1898 : /*
1899 : * This prevents one CPU to accumulate too many dirtied pages without
1900 : * calling into balance_dirty_pages(), which can happen when there are
1901 : * 1000+ tasks, all of them start dirtying pages at exactly the same
1902 : * time, hence all honoured too large initial task->nr_dirtied_pause.
1903 : */
1904 2181 : p = this_cpu_ptr(&bdp_ratelimits);
1905 2181 : if (unlikely(current->nr_dirtied >= ratelimit))
1906 14 : *p = 0;
1907 2167 : else if (unlikely(*p >= ratelimit_pages)) {
1908 1 : *p = 0;
1909 1 : ratelimit = 0;
1910 : }
1911 : /*
1912 : * Pick up the dirtied pages by the exited tasks. This avoids lots of
1913 : * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1914 : * the dirty throttling and livelock other long-run dirtiers.
1915 : */
1916 2181 : p = this_cpu_ptr(&dirty_throttle_leaks);
1917 2181 : if (*p > 0 && current->nr_dirtied < ratelimit) {
1918 192 : unsigned long nr_pages_dirtied;
1919 192 : nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1920 192 : *p -= nr_pages_dirtied;
1921 192 : current->nr_dirtied += nr_pages_dirtied;
1922 : }
1923 2181 : preempt_enable();
1924 :
1925 2181 : if (unlikely(current->nr_dirtied >= ratelimit))
1926 25 : balance_dirty_pages(wb, current->nr_dirtied);
1927 :
1928 2896 : wb_put(wb);
1929 : }
1930 : EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1931 :
1932 : /**
1933 : * wb_over_bg_thresh - does @wb need to be written back?
1934 : * @wb: bdi_writeback of interest
1935 : *
1936 : * Determines whether background writeback should keep writing @wb or it's
1937 : * clean enough.
1938 : *
1939 : * Return: %true if writeback should continue.
1940 : */
1941 7 : bool wb_over_bg_thresh(struct bdi_writeback *wb)
1942 : {
1943 7 : struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1944 7 : struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1945 7 : struct dirty_throttle_control * const gdtc = &gdtc_stor;
1946 7 : struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1947 : &mdtc_stor : NULL;
1948 :
1949 : /*
1950 : * Similar to balance_dirty_pages() but ignores pages being written
1951 : * as we're trying to decide whether to put more under writeback.
1952 : */
1953 7 : gdtc->avail = global_dirtyable_memory();
1954 7 : gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1955 7 : domain_dirty_limits(gdtc);
1956 :
1957 7 : if (gdtc->dirty > gdtc->bg_thresh)
1958 : return true;
1959 :
1960 7 : if (wb_stat(wb, WB_RECLAIMABLE) >
1961 7 : wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1962 0 : return true;
1963 :
1964 : if (mdtc) {
1965 : unsigned long filepages, headroom, writeback;
1966 :
1967 : mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1968 : &writeback);
1969 : mdtc_calc_avail(mdtc, filepages, headroom);
1970 : domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1971 :
1972 : if (mdtc->dirty > mdtc->bg_thresh)
1973 : return true;
1974 :
1975 : if (wb_stat(wb, WB_RECLAIMABLE) >
1976 : wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1977 : return true;
1978 : }
1979 :
1980 : return false;
1981 : }
1982 :
1983 : /*
1984 : * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1985 : */
1986 0 : int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1987 : void *buffer, size_t *length, loff_t *ppos)
1988 : {
1989 0 : unsigned int old_interval = dirty_writeback_interval;
1990 0 : int ret;
1991 :
1992 0 : ret = proc_dointvec(table, write, buffer, length, ppos);
1993 :
1994 : /*
1995 : * Writing 0 to dirty_writeback_interval will disable periodic writeback
1996 : * and a different non-zero value will wakeup the writeback threads.
1997 : * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1998 : * iterate over all bdis and wbs.
1999 : * The reason we do this is to make the change take effect immediately.
2000 : */
2001 0 : if (!ret && write && dirty_writeback_interval &&
2002 : dirty_writeback_interval != old_interval)
2003 0 : wakeup_flusher_threads(WB_REASON_PERIODIC);
2004 :
2005 0 : return ret;
2006 : }
2007 :
2008 : #ifdef CONFIG_BLOCK
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 0 : struct backing_dev_info *bdi;
2035 :
2036 0 : 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 0 : rcu_read_unlock();
2042 0 : }
2043 : #endif
2044 :
2045 : /*
2046 : * If ratelimit_pages is too high then we can get into dirty-data overload
2047 : * if a large number of processes all perform writes at the same time.
2048 : * If it is too low then SMP machines will call the (expensive)
2049 : * get_writeback_state too often.
2050 : *
2051 : * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2052 : * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2053 : * thresholds.
2054 : */
2055 :
2056 4 : void writeback_set_ratelimit(void)
2057 : {
2058 4 : struct wb_domain *dom = &global_wb_domain;
2059 4 : unsigned long background_thresh;
2060 4 : unsigned long dirty_thresh;
2061 :
2062 4 : global_dirty_limits(&background_thresh, &dirty_thresh);
2063 4 : dom->dirty_limit = dirty_thresh;
2064 4 : ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2065 4 : if (ratelimit_pages < 16)
2066 0 : ratelimit_pages = 16;
2067 4 : }
2068 :
2069 4 : static int page_writeback_cpu_online(unsigned int cpu)
2070 : {
2071 4 : writeback_set_ratelimit();
2072 4 : return 0;
2073 : }
2074 :
2075 : /*
2076 : * Called early on to tune the page writeback dirty limits.
2077 : *
2078 : * We used to scale dirty pages according to how total memory
2079 : * related to pages that could be allocated for buffers.
2080 : *
2081 : * However, that was when we used "dirty_ratio" to scale with
2082 : * all memory, and we don't do that any more. "dirty_ratio"
2083 : * is now applied to total non-HIGHPAGE memory, and as such we can't
2084 : * get into the old insane situation any more where we had
2085 : * large amounts of dirty pages compared to a small amount of
2086 : * non-HIGHMEM memory.
2087 : *
2088 : * But we might still want to scale the dirty_ratio by how
2089 : * much memory the box has..
2090 : */
2091 1 : void __init page_writeback_init(void)
2092 : {
2093 1 : BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2094 :
2095 1 : cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2096 : page_writeback_cpu_online, NULL);
2097 1 : cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2098 : page_writeback_cpu_online);
2099 1 : }
2100 :
2101 : /**
2102 : * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2103 : * @mapping: address space structure to write
2104 : * @start: starting page index
2105 : * @end: ending page index (inclusive)
2106 : *
2107 : * This function scans the page range from @start to @end (inclusive) and tags
2108 : * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2109 : * that write_cache_pages (or whoever calls this function) will then use
2110 : * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2111 : * used to avoid livelocking of writeback by a process steadily creating new
2112 : * dirty pages in the file (thus it is important for this function to be quick
2113 : * so that it can tag pages faster than a dirtying process can create them).
2114 : */
2115 69 : void tag_pages_for_writeback(struct address_space *mapping,
2116 : pgoff_t start, pgoff_t end)
2117 : {
2118 69 : XA_STATE(xas, &mapping->i_pages, start);
2119 69 : unsigned int tagged = 0;
2120 69 : void *page;
2121 :
2122 69 : xas_lock_irq(&xas);
2123 433 : xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2124 364 : xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2125 364 : if (++tagged % XA_CHECK_SCHED)
2126 364 : continue;
2127 :
2128 0 : xas_pause(&xas);
2129 0 : xas_unlock_irq(&xas);
2130 0 : cond_resched();
2131 364 : xas_lock_irq(&xas);
2132 : }
2133 69 : xas_unlock_irq(&xas);
2134 69 : }
2135 : EXPORT_SYMBOL(tag_pages_for_writeback);
2136 :
2137 : /**
2138 : * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2139 : * @mapping: address space structure to write
2140 : * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2141 : * @writepage: function called for each page
2142 : * @data: data passed to writepage function
2143 : *
2144 : * If a page is already under I/O, write_cache_pages() skips it, even
2145 : * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2146 : * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2147 : * and msync() need to guarantee that all the data which was dirty at the time
2148 : * the call was made get new I/O started against them. If wbc->sync_mode is
2149 : * WB_SYNC_ALL then we were called for data integrity and we must wait for
2150 : * existing IO to complete.
2151 : *
2152 : * To avoid livelocks (when other process dirties new pages), we first tag
2153 : * pages which should be written back with TOWRITE tag and only then start
2154 : * writing them. For data-integrity sync we have to be careful so that we do
2155 : * not miss some pages (e.g., because some other process has cleared TOWRITE
2156 : * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2157 : * by the process clearing the DIRTY tag (and submitting the page for IO).
2158 : *
2159 : * To avoid deadlocks between range_cyclic writeback and callers that hold
2160 : * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2161 : * we do not loop back to the start of the file. Doing so causes a page
2162 : * lock/page writeback access order inversion - we should only ever lock
2163 : * multiple pages in ascending page->index order, and looping back to the start
2164 : * of the file violates that rule and causes deadlocks.
2165 : *
2166 : * Return: %0 on success, negative error code otherwise
2167 : */
2168 3 : int write_cache_pages(struct address_space *mapping,
2169 : struct writeback_control *wbc, writepage_t writepage,
2170 : void *data)
2171 : {
2172 3 : int ret = 0;
2173 3 : int done = 0;
2174 3 : int error;
2175 3 : struct pagevec pvec;
2176 3 : int nr_pages;
2177 3 : pgoff_t index;
2178 3 : pgoff_t end; /* Inclusive */
2179 3 : pgoff_t done_index;
2180 3 : int range_whole = 0;
2181 3 : xa_mark_t tag;
2182 :
2183 3 : pagevec_init(&pvec);
2184 3 : if (wbc->range_cyclic) {
2185 1 : index = mapping->writeback_index; /* prev offset */
2186 1 : end = -1;
2187 : } else {
2188 2 : index = wbc->range_start >> PAGE_SHIFT;
2189 2 : end = wbc->range_end >> PAGE_SHIFT;
2190 2 : if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2191 0 : range_whole = 1;
2192 : }
2193 3 : if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2194 2 : tag_pages_for_writeback(mapping, index, end);
2195 2 : tag = PAGECACHE_TAG_TOWRITE;
2196 : } else {
2197 : tag = PAGECACHE_TAG_DIRTY;
2198 : }
2199 3 : done_index = index;
2200 64 : while (!done && (index <= end)) {
2201 62 : int i;
2202 :
2203 62 : nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2204 : tag);
2205 62 : if (nr_pages == 0)
2206 : break;
2207 :
2208 935 : for (i = 0; i < nr_pages; i++) {
2209 874 : struct page *page = pvec.pages[i];
2210 :
2211 874 : done_index = page->index;
2212 :
2213 874 : lock_page(page);
2214 :
2215 : /*
2216 : * Page truncated or invalidated. We can freely skip it
2217 : * then, even for data integrity operations: the page
2218 : * has disappeared concurrently, so there could be no
2219 : * real expectation of this data interity operation
2220 : * even if there is now a new, dirty page at the same
2221 : * pagecache address.
2222 : */
2223 874 : if (unlikely(page->mapping != mapping)) {
2224 0 : continue_unlock:
2225 0 : unlock_page(page);
2226 0 : continue;
2227 : }
2228 :
2229 1748 : if (!PageDirty(page)) {
2230 : /* someone wrote it for us */
2231 0 : goto continue_unlock;
2232 : }
2233 :
2234 1748 : if (PageWriteback(page)) {
2235 0 : if (wbc->sync_mode != WB_SYNC_NONE)
2236 0 : wait_on_page_writeback(page);
2237 : else
2238 0 : goto continue_unlock;
2239 : }
2240 :
2241 1748 : BUG_ON(PageWriteback(page));
2242 874 : if (!clear_page_dirty_for_io(page))
2243 0 : goto continue_unlock;
2244 :
2245 874 : trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2246 874 : error = (*writepage)(page, wbc, data);
2247 874 : if (unlikely(error)) {
2248 : /*
2249 : * Handle errors according to the type of
2250 : * writeback. There's no need to continue for
2251 : * background writeback. Just push done_index
2252 : * past this page so media errors won't choke
2253 : * writeout for the entire file. For integrity
2254 : * writeback, we must process the entire dirty
2255 : * set regardless of errors because the fs may
2256 : * still have state to clear for each page. In
2257 : * that case we continue processing and return
2258 : * the first error.
2259 : */
2260 0 : if (error == AOP_WRITEPAGE_ACTIVATE) {
2261 0 : unlock_page(page);
2262 0 : error = 0;
2263 0 : } else if (wbc->sync_mode != WB_SYNC_ALL) {
2264 0 : ret = error;
2265 0 : done_index = page->index + 1;
2266 0 : done = 1;
2267 0 : break;
2268 : }
2269 0 : if (!ret)
2270 0 : ret = error;
2271 : }
2272 :
2273 : /*
2274 : * We stop writing back only if we are not doing
2275 : * integrity sync. In case of integrity sync we have to
2276 : * keep going until we have written all the pages
2277 : * we tagged for writeback prior to entering this loop.
2278 : */
2279 874 : if (--wbc->nr_to_write <= 0 &&
2280 0 : wbc->sync_mode == WB_SYNC_NONE) {
2281 : done = 1;
2282 : break;
2283 : }
2284 : }
2285 61 : pagevec_release(&pvec);
2286 61 : cond_resched();
2287 : }
2288 :
2289 : /*
2290 : * If we hit the last page and there is more work to be done: wrap
2291 : * back the index back to the start of the file for the next
2292 : * time we are called.
2293 : */
2294 3 : if (wbc->range_cyclic && !done)
2295 1 : done_index = 0;
2296 3 : if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2297 1 : mapping->writeback_index = done_index;
2298 :
2299 3 : return ret;
2300 : }
2301 : EXPORT_SYMBOL(write_cache_pages);
2302 :
2303 : /*
2304 : * Function used by generic_writepages to call the real writepage
2305 : * function and set the mapping flags on error
2306 : */
2307 874 : static int __writepage(struct page *page, struct writeback_control *wbc,
2308 : void *data)
2309 : {
2310 874 : struct address_space *mapping = data;
2311 874 : int ret = mapping->a_ops->writepage(page, wbc);
2312 874 : mapping_set_error(mapping, ret);
2313 874 : return ret;
2314 : }
2315 :
2316 : /**
2317 : * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2318 : * @mapping: address space structure to write
2319 : * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2320 : *
2321 : * This is a library function, which implements the writepages()
2322 : * address_space_operation.
2323 : *
2324 : * Return: %0 on success, negative error code otherwise
2325 : */
2326 479 : int generic_writepages(struct address_space *mapping,
2327 : struct writeback_control *wbc)
2328 : {
2329 479 : struct blk_plug plug;
2330 479 : int ret;
2331 :
2332 : /* deal with chardevs and other special file */
2333 479 : if (!mapping->a_ops->writepage)
2334 : return 0;
2335 :
2336 3 : blk_start_plug(&plug);
2337 3 : ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2338 3 : blk_finish_plug(&plug);
2339 3 : return ret;
2340 : }
2341 :
2342 : EXPORT_SYMBOL(generic_writepages);
2343 :
2344 878 : int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2345 : {
2346 878 : int ret;
2347 :
2348 878 : if (wbc->nr_to_write <= 0)
2349 : return 0;
2350 878 : while (1) {
2351 878 : if (mapping->a_ops->writepages)
2352 402 : ret = mapping->a_ops->writepages(mapping, wbc);
2353 : else
2354 476 : ret = generic_writepages(mapping, wbc);
2355 878 : if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2356 : break;
2357 0 : cond_resched();
2358 0 : congestion_wait(BLK_RW_ASYNC, HZ/50);
2359 : }
2360 : return ret;
2361 : }
2362 :
2363 : /**
2364 : * write_one_page - write out a single page and wait on I/O
2365 : * @page: the page to write
2366 : *
2367 : * The page must be locked by the caller and will be unlocked upon return.
2368 : *
2369 : * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2370 : * function returns.
2371 : *
2372 : * Return: %0 on success, negative error code otherwise
2373 : */
2374 0 : int write_one_page(struct page *page)
2375 : {
2376 0 : struct address_space *mapping = page->mapping;
2377 0 : int ret = 0;
2378 0 : struct writeback_control wbc = {
2379 : .sync_mode = WB_SYNC_ALL,
2380 : .nr_to_write = 1,
2381 : };
2382 :
2383 0 : BUG_ON(!PageLocked(page));
2384 :
2385 0 : wait_on_page_writeback(page);
2386 :
2387 0 : if (clear_page_dirty_for_io(page)) {
2388 0 : get_page(page);
2389 0 : ret = mapping->a_ops->writepage(page, &wbc);
2390 0 : if (ret == 0)
2391 0 : wait_on_page_writeback(page);
2392 0 : put_page(page);
2393 : } else {
2394 0 : unlock_page(page);
2395 : }
2396 :
2397 0 : if (!ret)
2398 0 : ret = filemap_check_errors(mapping);
2399 0 : return ret;
2400 : }
2401 : EXPORT_SYMBOL(write_one_page);
2402 :
2403 : /*
2404 : * For address_spaces which do not use buffers nor write back.
2405 : */
2406 2149 : int __set_page_dirty_no_writeback(struct page *page)
2407 : {
2408 4298 : if (!PageDirty(page))
2409 3970 : return !TestSetPageDirty(page);
2410 : return 0;
2411 : }
2412 :
2413 : /*
2414 : * Helper function for set_page_dirty family.
2415 : *
2416 : * Caller must hold lock_page_memcg().
2417 : *
2418 : * NOTE: This relies on being atomic wrt interrupts.
2419 : */
2420 2493 : void account_page_dirtied(struct page *page, struct address_space *mapping)
2421 : {
2422 2493 : struct inode *inode = mapping->host;
2423 :
2424 2493 : trace_writeback_dirty_page(page, mapping);
2425 :
2426 2493 : if (mapping_can_writeback(mapping)) {
2427 2493 : struct bdi_writeback *wb;
2428 :
2429 2493 : inode_attach_wb(inode, page);
2430 2493 : wb = inode_to_wb(inode);
2431 :
2432 2493 : __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2433 2493 : __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2434 2493 : __inc_node_page_state(page, NR_DIRTIED);
2435 2493 : inc_wb_stat(wb, WB_RECLAIMABLE);
2436 2493 : inc_wb_stat(wb, WB_DIRTIED);
2437 2493 : task_io_account_write(PAGE_SIZE);
2438 2493 : current->nr_dirtied++;
2439 2493 : this_cpu_inc(bdp_ratelimits);
2440 :
2441 2493 : mem_cgroup_track_foreign_dirty(page, wb);
2442 : }
2443 2493 : }
2444 :
2445 : /*
2446 : * Helper function for deaccounting dirty page without writeback.
2447 : *
2448 : * Caller must hold lock_page_memcg().
2449 : */
2450 496 : void account_page_cleaned(struct page *page, struct address_space *mapping,
2451 : struct bdi_writeback *wb)
2452 : {
2453 496 : if (mapping_can_writeback(mapping)) {
2454 496 : dec_lruvec_page_state(page, NR_FILE_DIRTY);
2455 496 : dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2456 496 : dec_wb_stat(wb, WB_RECLAIMABLE);
2457 496 : task_io_account_cancelled_write(PAGE_SIZE);
2458 : }
2459 496 : }
2460 :
2461 : /*
2462 : * For address_spaces which do not use buffers. Just tag the page as dirty in
2463 : * the xarray.
2464 : *
2465 : * This is also used when a single buffer is being dirtied: we want to set the
2466 : * page dirty in that case, but not all the buffers. This is a "bottom-up"
2467 : * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2468 : *
2469 : * The caller must ensure this doesn't race with truncation. Most will simply
2470 : * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2471 : * the pte lock held, which also locks out truncation.
2472 : */
2473 2 : int __set_page_dirty_nobuffers(struct page *page)
2474 : {
2475 2 : lock_page_memcg(page);
2476 4 : if (!TestSetPageDirty(page)) {
2477 2 : struct address_space *mapping = page_mapping(page);
2478 2 : unsigned long flags;
2479 :
2480 2 : if (!mapping) {
2481 2 : unlock_page_memcg(page);
2482 : return 1;
2483 : }
2484 :
2485 2 : xa_lock_irqsave(&mapping->i_pages, flags);
2486 2 : BUG_ON(page_mapping(page) != mapping);
2487 2 : WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2488 2 : account_page_dirtied(page, mapping);
2489 2 : __xa_set_mark(&mapping->i_pages, page_index(page),
2490 : PAGECACHE_TAG_DIRTY);
2491 2 : xa_unlock_irqrestore(&mapping->i_pages, flags);
2492 2 : unlock_page_memcg(page);
2493 :
2494 2 : if (mapping->host) {
2495 : /* !PageAnon && !swapper_space */
2496 2 : __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2497 : }
2498 2 : return 1;
2499 : }
2500 2 : unlock_page_memcg(page);
2501 : return 0;
2502 : }
2503 : EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2504 :
2505 : /*
2506 : * Call this whenever redirtying a page, to de-account the dirty counters
2507 : * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2508 : * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2509 : * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2510 : * control.
2511 : */
2512 2 : void account_page_redirty(struct page *page)
2513 : {
2514 2 : struct address_space *mapping = page->mapping;
2515 :
2516 4 : if (mapping && mapping_can_writeback(mapping)) {
2517 2 : struct inode *inode = mapping->host;
2518 2 : struct bdi_writeback *wb;
2519 2 : struct wb_lock_cookie cookie = {};
2520 :
2521 2 : wb = unlocked_inode_to_wb_begin(inode, &cookie);
2522 2 : current->nr_dirtied--;
2523 2 : dec_node_page_state(page, NR_DIRTIED);
2524 2 : dec_wb_stat(wb, WB_DIRTIED);
2525 2 : unlocked_inode_to_wb_end(inode, &cookie);
2526 : }
2527 2 : }
2528 : EXPORT_SYMBOL(account_page_redirty);
2529 :
2530 : /*
2531 : * When a writepage implementation decides that it doesn't want to write this
2532 : * page for some reason, it should redirty the locked page via
2533 : * redirty_page_for_writepage() and it should then unlock the page and return 0
2534 : */
2535 2 : int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2536 : {
2537 2 : int ret;
2538 :
2539 2 : wbc->pages_skipped++;
2540 2 : ret = __set_page_dirty_nobuffers(page);
2541 2 : account_page_redirty(page);
2542 2 : return ret;
2543 : }
2544 : EXPORT_SYMBOL(redirty_page_for_writepage);
2545 :
2546 : /*
2547 : * Dirty a page.
2548 : *
2549 : * For pages with a mapping this should be done under the page lock
2550 : * for the benefit of asynchronous memory errors who prefer a consistent
2551 : * dirty state. This rule can be broken in some special cases,
2552 : * but should be better not to.
2553 : *
2554 : * If the mapping doesn't provide a set_page_dirty a_op, then
2555 : * just fall through and assume that it wants buffer_heads.
2556 : */
2557 3061 : int set_page_dirty(struct page *page)
2558 : {
2559 3061 : struct address_space *mapping = page_mapping(page);
2560 :
2561 3061 : page = compound_head(page);
2562 3061 : if (likely(mapping)) {
2563 3061 : int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2564 : /*
2565 : * readahead/lru_deactivate_page could remain
2566 : * PG_readahead/PG_reclaim due to race with end_page_writeback
2567 : * About readahead, if the page is written, the flags would be
2568 : * reset. So no problem.
2569 : * About lru_deactivate_page, if the page is redirty, the flag
2570 : * will be reset. So no problem. but if the page is used by readahead
2571 : * it will confuse readahead and make it restart the size rampup
2572 : * process. But it's a trivial problem.
2573 : */
2574 6122 : if (PageReclaim(page))
2575 0 : ClearPageReclaim(page);
2576 : #ifdef CONFIG_BLOCK
2577 3061 : if (!spd)
2578 0 : spd = __set_page_dirty_buffers;
2579 : #endif
2580 3061 : return (*spd)(page);
2581 : }
2582 0 : if (!PageDirty(page)) {
2583 0 : if (!TestSetPageDirty(page))
2584 0 : return 1;
2585 : }
2586 : return 0;
2587 : }
2588 : EXPORT_SYMBOL(set_page_dirty);
2589 :
2590 : /*
2591 : * set_page_dirty() is racy if the caller has no reference against
2592 : * page->mapping->host, and if the page is unlocked. This is because another
2593 : * CPU could truncate the page off the mapping and then free the mapping.
2594 : *
2595 : * Usually, the page _is_ locked, or the caller is a user-space process which
2596 : * holds a reference on the inode by having an open file.
2597 : *
2598 : * In other cases, the page should be locked before running set_page_dirty().
2599 : */
2600 0 : int set_page_dirty_lock(struct page *page)
2601 : {
2602 0 : int ret;
2603 :
2604 0 : lock_page(page);
2605 0 : ret = set_page_dirty(page);
2606 0 : unlock_page(page);
2607 0 : return ret;
2608 : }
2609 : EXPORT_SYMBOL(set_page_dirty_lock);
2610 :
2611 : /*
2612 : * This cancels just the dirty bit on the kernel page itself, it does NOT
2613 : * actually remove dirty bits on any mmap's that may be around. It also
2614 : * leaves the page tagged dirty, so any sync activity will still find it on
2615 : * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2616 : * look at the dirty bits in the VM.
2617 : *
2618 : * Doing this should *normally* only ever be done when a page is truncated,
2619 : * and is not actually mapped anywhere at all. However, fs/buffer.c does
2620 : * this when it notices that somebody has cleaned out all the buffers on a
2621 : * page without actually doing it through the VM. Can you say "ext3 is
2622 : * horribly ugly"? Thought you could.
2623 : */
2624 1164 : void __cancel_dirty_page(struct page *page)
2625 : {
2626 1164 : struct address_space *mapping = page_mapping(page);
2627 :
2628 1164 : if (mapping_can_writeback(mapping)) {
2629 496 : struct inode *inode = mapping->host;
2630 496 : struct bdi_writeback *wb;
2631 496 : struct wb_lock_cookie cookie = {};
2632 :
2633 496 : lock_page_memcg(page);
2634 496 : wb = unlocked_inode_to_wb_begin(inode, &cookie);
2635 :
2636 992 : if (TestClearPageDirty(page))
2637 496 : account_page_cleaned(page, mapping, wb);
2638 :
2639 496 : unlocked_inode_to_wb_end(inode, &cookie);
2640 496 : unlock_page_memcg(page);
2641 : } else {
2642 668 : ClearPageDirty(page);
2643 : }
2644 1164 : }
2645 : EXPORT_SYMBOL(__cancel_dirty_page);
2646 :
2647 : /*
2648 : * Clear a page's dirty flag, while caring for dirty memory accounting.
2649 : * Returns true if the page was previously dirty.
2650 : *
2651 : * This is for preparing to put the page under writeout. We leave the page
2652 : * tagged as dirty in the xarray so that a concurrent write-for-sync
2653 : * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2654 : * implementation will run either set_page_writeback() or set_page_dirty(),
2655 : * at which stage we bring the page's dirty flag and xarray dirty tag
2656 : * back into sync.
2657 : *
2658 : * This incoherency between the page's dirty flag and xarray tag is
2659 : * unfortunate, but it only exists while the page is locked.
2660 : */
2661 1286 : int clear_page_dirty_for_io(struct page *page)
2662 : {
2663 1286 : struct address_space *mapping = page_mapping(page);
2664 1286 : int ret = 0;
2665 :
2666 2572 : VM_BUG_ON_PAGE(!PageLocked(page), page);
2667 :
2668 2572 : if (mapping && mapping_can_writeback(mapping)) {
2669 1286 : struct inode *inode = mapping->host;
2670 1286 : struct bdi_writeback *wb;
2671 1286 : struct wb_lock_cookie cookie = {};
2672 :
2673 : /*
2674 : * Yes, Virginia, this is indeed insane.
2675 : *
2676 : * We use this sequence to make sure that
2677 : * (a) we account for dirty stats properly
2678 : * (b) we tell the low-level filesystem to
2679 : * mark the whole page dirty if it was
2680 : * dirty in a pagetable. Only to then
2681 : * (c) clean the page again and return 1 to
2682 : * cause the writeback.
2683 : *
2684 : * This way we avoid all nasty races with the
2685 : * dirty bit in multiple places and clearing
2686 : * them concurrently from different threads.
2687 : *
2688 : * Note! Normally the "set_page_dirty(page)"
2689 : * has no effect on the actual dirty bit - since
2690 : * that will already usually be set. But we
2691 : * need the side effects, and it can help us
2692 : * avoid races.
2693 : *
2694 : * We basically use the page "master dirty bit"
2695 : * as a serialization point for all the different
2696 : * threads doing their things.
2697 : */
2698 1286 : if (page_mkclean(page))
2699 68 : set_page_dirty(page);
2700 : /*
2701 : * We carefully synchronise fault handlers against
2702 : * installing a dirty pte and marking the page dirty
2703 : * at this point. We do this by having them hold the
2704 : * page lock while dirtying the page, and pages are
2705 : * always locked coming in here, so we get the desired
2706 : * exclusion.
2707 : */
2708 1286 : wb = unlocked_inode_to_wb_begin(inode, &cookie);
2709 2572 : if (TestClearPageDirty(page)) {
2710 1286 : dec_lruvec_page_state(page, NR_FILE_DIRTY);
2711 1286 : dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2712 1286 : dec_wb_stat(wb, WB_RECLAIMABLE);
2713 1286 : ret = 1;
2714 : }
2715 1286 : unlocked_inode_to_wb_end(inode, &cookie);
2716 1286 : return ret;
2717 : }
2718 0 : return TestClearPageDirty(page);
2719 : }
2720 : EXPORT_SYMBOL(clear_page_dirty_for_io);
2721 :
2722 1284 : int test_clear_page_writeback(struct page *page)
2723 : {
2724 1284 : struct address_space *mapping = page_mapping(page);
2725 1284 : struct mem_cgroup *memcg;
2726 1284 : struct lruvec *lruvec;
2727 1284 : int ret;
2728 :
2729 1284 : memcg = lock_page_memcg(page);
2730 1284 : lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2731 2568 : if (mapping && mapping_use_writeback_tags(mapping)) {
2732 1284 : struct inode *inode = mapping->host;
2733 1284 : struct backing_dev_info *bdi = inode_to_bdi(inode);
2734 1284 : unsigned long flags;
2735 :
2736 1284 : xa_lock_irqsave(&mapping->i_pages, flags);
2737 1284 : ret = TestClearPageWriteback(page);
2738 1284 : if (ret) {
2739 1284 : __xa_clear_mark(&mapping->i_pages, page_index(page),
2740 : PAGECACHE_TAG_WRITEBACK);
2741 1284 : if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2742 1284 : struct bdi_writeback *wb = inode_to_wb(inode);
2743 :
2744 1284 : dec_wb_stat(wb, WB_WRITEBACK);
2745 1284 : __wb_writeout_inc(wb);
2746 : }
2747 : }
2748 :
2749 1284 : if (mapping->host && !mapping_tagged(mapping,
2750 : PAGECACHE_TAG_WRITEBACK))
2751 83 : sb_clear_inode_writeback(mapping->host);
2752 :
2753 1284 : xa_unlock_irqrestore(&mapping->i_pages, flags);
2754 : } else {
2755 0 : ret = TestClearPageWriteback(page);
2756 : }
2757 1284 : if (ret) {
2758 1284 : dec_lruvec_state(lruvec, NR_WRITEBACK);
2759 1284 : dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2760 1284 : inc_node_page_state(page, NR_WRITTEN);
2761 : }
2762 1284 : __unlock_page_memcg(memcg);
2763 1284 : return ret;
2764 : }
2765 :
2766 1284 : int __test_set_page_writeback(struct page *page, bool keep_write)
2767 : {
2768 1284 : struct address_space *mapping = page_mapping(page);
2769 1284 : int ret, access_ret;
2770 :
2771 1284 : lock_page_memcg(page);
2772 2568 : if (mapping && mapping_use_writeback_tags(mapping)) {
2773 1284 : XA_STATE(xas, &mapping->i_pages, page_index(page));
2774 1284 : struct inode *inode = mapping->host;
2775 1284 : struct backing_dev_info *bdi = inode_to_bdi(inode);
2776 1284 : unsigned long flags;
2777 :
2778 1284 : xas_lock_irqsave(&xas, flags);
2779 1284 : xas_load(&xas);
2780 1284 : ret = TestSetPageWriteback(page);
2781 1284 : if (!ret) {
2782 1284 : bool on_wblist;
2783 :
2784 1284 : on_wblist = mapping_tagged(mapping,
2785 : PAGECACHE_TAG_WRITEBACK);
2786 :
2787 1284 : xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2788 1284 : if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT)
2789 1284 : inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2790 :
2791 : /*
2792 : * We can come through here when swapping anonymous
2793 : * pages, so we don't necessarily have an inode to track
2794 : * for sync.
2795 : */
2796 1284 : if (mapping->host && !on_wblist)
2797 83 : sb_mark_inode_writeback(mapping->host);
2798 : }
2799 2568 : if (!PageDirty(page))
2800 1284 : xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2801 1284 : if (!keep_write)
2802 1284 : xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2803 1284 : xas_unlock_irqrestore(&xas, flags);
2804 : } else {
2805 0 : ret = TestSetPageWriteback(page);
2806 : }
2807 1284 : if (!ret) {
2808 1284 : inc_lruvec_page_state(page, NR_WRITEBACK);
2809 1284 : inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2810 : }
2811 1284 : unlock_page_memcg(page);
2812 1284 : access_ret = arch_make_page_accessible(page);
2813 : /*
2814 : * If writeback has been triggered on a page that cannot be made
2815 : * accessible, it is too late to recover here.
2816 : */
2817 1284 : VM_BUG_ON_PAGE(access_ret != 0, page);
2818 :
2819 1284 : return ret;
2820 :
2821 : }
2822 : EXPORT_SYMBOL(__test_set_page_writeback);
2823 :
2824 : /*
2825 : * Wait for a page to complete writeback
2826 : */
2827 1311 : void wait_on_page_writeback(struct page *page)
2828 : {
2829 2788 : while (PageWriteback(page)) {
2830 83 : trace_wait_on_page_writeback(page, page_mapping(page));
2831 83 : wait_on_page_bit(page, PG_writeback);
2832 : }
2833 1311 : }
2834 : EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2835 :
2836 : /**
2837 : * wait_for_stable_page() - wait for writeback to finish, if necessary.
2838 : * @page: The page to wait on.
2839 : *
2840 : * This function determines if the given page is related to a backing device
2841 : * that requires page contents to be held stable during writeback. If so, then
2842 : * it will wait for any pending writeback to complete.
2843 : */
2844 4058 : void wait_for_stable_page(struct page *page)
2845 : {
2846 4058 : page = thp_head(page);
2847 4058 : if (page->mapping->host->i_sb->s_iflags & SB_I_STABLE_WRITES)
2848 0 : wait_on_page_writeback(page);
2849 4058 : }
2850 : EXPORT_SYMBOL_GPL(wait_for_stable_page);
|
