Loading...
1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
25#include <linux/task_io_accounting_ops.h>
26#include <linux/blkdev.h>
27#include <linux/mpage.h>
28#include <linux/rmap.h>
29#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h>
36#include <linux/pagevec.h>
37#include <trace/events/writeback.h>
38
39/*
40 * Sleep at most 200ms at a time in balance_dirty_pages().
41 */
42#define MAX_PAUSE max(HZ/5, 1)
43
44/*
45 * Estimate write bandwidth at 200ms intervals.
46 */
47#define BANDWIDTH_INTERVAL max(HZ/5, 1)
48
49/*
50 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
51 * will look to see if it needs to force writeback or throttling.
52 */
53static long ratelimit_pages = 32;
54
55/*
56 * When balance_dirty_pages decides that the caller needs to perform some
57 * non-background writeback, this is how many pages it will attempt to write.
58 * It should be somewhat larger than dirtied pages to ensure that reasonably
59 * large amounts of I/O are submitted.
60 */
61static inline long sync_writeback_pages(unsigned long dirtied)
62{
63 if (dirtied < ratelimit_pages)
64 dirtied = ratelimit_pages;
65
66 return dirtied + dirtied / 2;
67}
68
69/* The following parameters are exported via /proc/sys/vm */
70
71/*
72 * Start background writeback (via writeback threads) at this percentage
73 */
74int 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 */
80unsigned 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 */
86int vm_highmem_is_dirtyable;
87
88/*
89 * The generator of dirty data starts writeback at this percentage
90 */
91int 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 */
97unsigned long vm_dirty_bytes;
98
99/*
100 * The interval between `kupdate'-style writebacks
101 */
102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104/*
105 * The longest time for which data is allowed to remain dirty
106 */
107unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
108
109/*
110 * Flag that makes the machine dump writes/reads and block dirtyings.
111 */
112int block_dump;
113
114/*
115 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
116 * a full sync is triggered after this time elapses without any disk activity.
117 */
118int laptop_mode;
119
120EXPORT_SYMBOL(laptop_mode);
121
122/* End of sysctl-exported parameters */
123
124unsigned long global_dirty_limit;
125
126/*
127 * Scale the writeback cache size proportional to the relative writeout speeds.
128 *
129 * We do this by keeping a floating proportion between BDIs, based on page
130 * writeback completions [end_page_writeback()]. Those devices that write out
131 * pages fastest will get the larger share, while the slower will get a smaller
132 * share.
133 *
134 * We use page writeout completions because we are interested in getting rid of
135 * dirty pages. Having them written out is the primary goal.
136 *
137 * We introduce a concept of time, a period over which we measure these events,
138 * because demand can/will vary over time. The length of this period itself is
139 * measured in page writeback completions.
140 *
141 */
142static struct prop_descriptor vm_completions;
143static struct prop_descriptor vm_dirties;
144
145/*
146 * couple the period to the dirty_ratio:
147 *
148 * period/2 ~ roundup_pow_of_two(dirty limit)
149 */
150static int calc_period_shift(void)
151{
152 unsigned long dirty_total;
153
154 if (vm_dirty_bytes)
155 dirty_total = vm_dirty_bytes / PAGE_SIZE;
156 else
157 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
158 100;
159 return 2 + ilog2(dirty_total - 1);
160}
161
162/*
163 * update the period when the dirty threshold changes.
164 */
165static void update_completion_period(void)
166{
167 int shift = calc_period_shift();
168 prop_change_shift(&vm_completions, shift);
169 prop_change_shift(&vm_dirties, shift);
170}
171
172int dirty_background_ratio_handler(struct ctl_table *table, int write,
173 void __user *buffer, size_t *lenp,
174 loff_t *ppos)
175{
176 int ret;
177
178 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
179 if (ret == 0 && write)
180 dirty_background_bytes = 0;
181 return ret;
182}
183
184int dirty_background_bytes_handler(struct ctl_table *table, int write,
185 void __user *buffer, size_t *lenp,
186 loff_t *ppos)
187{
188 int ret;
189
190 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
191 if (ret == 0 && write)
192 dirty_background_ratio = 0;
193 return ret;
194}
195
196int dirty_ratio_handler(struct ctl_table *table, int write,
197 void __user *buffer, size_t *lenp,
198 loff_t *ppos)
199{
200 int old_ratio = vm_dirty_ratio;
201 int ret;
202
203 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
204 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
205 update_completion_period();
206 vm_dirty_bytes = 0;
207 }
208 return ret;
209}
210
211
212int dirty_bytes_handler(struct ctl_table *table, int write,
213 void __user *buffer, size_t *lenp,
214 loff_t *ppos)
215{
216 unsigned long old_bytes = vm_dirty_bytes;
217 int ret;
218
219 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
220 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
221 update_completion_period();
222 vm_dirty_ratio = 0;
223 }
224 return ret;
225}
226
227/*
228 * Increment the BDI's writeout completion count and the global writeout
229 * completion count. Called from test_clear_page_writeback().
230 */
231static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
232{
233 __inc_bdi_stat(bdi, BDI_WRITTEN);
234 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
235 bdi->max_prop_frac);
236}
237
238void bdi_writeout_inc(struct backing_dev_info *bdi)
239{
240 unsigned long flags;
241
242 local_irq_save(flags);
243 __bdi_writeout_inc(bdi);
244 local_irq_restore(flags);
245}
246EXPORT_SYMBOL_GPL(bdi_writeout_inc);
247
248void task_dirty_inc(struct task_struct *tsk)
249{
250 prop_inc_single(&vm_dirties, &tsk->dirties);
251}
252
253/*
254 * Obtain an accurate fraction of the BDI's portion.
255 */
256static void bdi_writeout_fraction(struct backing_dev_info *bdi,
257 long *numerator, long *denominator)
258{
259 prop_fraction_percpu(&vm_completions, &bdi->completions,
260 numerator, denominator);
261}
262
263static inline void task_dirties_fraction(struct task_struct *tsk,
264 long *numerator, long *denominator)
265{
266 prop_fraction_single(&vm_dirties, &tsk->dirties,
267 numerator, denominator);
268}
269
270/*
271 * task_dirty_limit - scale down dirty throttling threshold for one task
272 *
273 * task specific dirty limit:
274 *
275 * dirty -= (dirty/8) * p_{t}
276 *
277 * To protect light/slow dirtying tasks from heavier/fast ones, we start
278 * throttling individual tasks before reaching the bdi dirty limit.
279 * Relatively low thresholds will be allocated to heavy dirtiers. So when
280 * dirty pages grow large, heavy dirtiers will be throttled first, which will
281 * effectively curb the growth of dirty pages. Light dirtiers with high enough
282 * dirty threshold may never get throttled.
283 */
284#define TASK_LIMIT_FRACTION 8
285static unsigned long task_dirty_limit(struct task_struct *tsk,
286 unsigned long bdi_dirty)
287{
288 long numerator, denominator;
289 unsigned long dirty = bdi_dirty;
290 u64 inv = dirty / TASK_LIMIT_FRACTION;
291
292 task_dirties_fraction(tsk, &numerator, &denominator);
293 inv *= numerator;
294 do_div(inv, denominator);
295
296 dirty -= inv;
297
298 return max(dirty, bdi_dirty/2);
299}
300
301/* Minimum limit for any task */
302static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
303{
304 return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
305}
306
307/*
308 *
309 */
310static unsigned int bdi_min_ratio;
311
312int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
313{
314 int ret = 0;
315
316 spin_lock_bh(&bdi_lock);
317 if (min_ratio > bdi->max_ratio) {
318 ret = -EINVAL;
319 } else {
320 min_ratio -= bdi->min_ratio;
321 if (bdi_min_ratio + min_ratio < 100) {
322 bdi_min_ratio += min_ratio;
323 bdi->min_ratio += min_ratio;
324 } else {
325 ret = -EINVAL;
326 }
327 }
328 spin_unlock_bh(&bdi_lock);
329
330 return ret;
331}
332
333int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
334{
335 int ret = 0;
336
337 if (max_ratio > 100)
338 return -EINVAL;
339
340 spin_lock_bh(&bdi_lock);
341 if (bdi->min_ratio > max_ratio) {
342 ret = -EINVAL;
343 } else {
344 bdi->max_ratio = max_ratio;
345 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
346 }
347 spin_unlock_bh(&bdi_lock);
348
349 return ret;
350}
351EXPORT_SYMBOL(bdi_set_max_ratio);
352
353/*
354 * Work out the current dirty-memory clamping and background writeout
355 * thresholds.
356 *
357 * The main aim here is to lower them aggressively if there is a lot of mapped
358 * memory around. To avoid stressing page reclaim with lots of unreclaimable
359 * pages. It is better to clamp down on writers than to start swapping, and
360 * performing lots of scanning.
361 *
362 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
363 *
364 * We don't permit the clamping level to fall below 5% - that is getting rather
365 * excessive.
366 *
367 * We make sure that the background writeout level is below the adjusted
368 * clamping level.
369 */
370
371static unsigned long highmem_dirtyable_memory(unsigned long total)
372{
373#ifdef CONFIG_HIGHMEM
374 int node;
375 unsigned long x = 0;
376
377 for_each_node_state(node, N_HIGH_MEMORY) {
378 struct zone *z =
379 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
380
381 x += zone_page_state(z, NR_FREE_PAGES) +
382 zone_reclaimable_pages(z);
383 }
384 /*
385 * Make sure that the number of highmem pages is never larger
386 * than the number of the total dirtyable memory. This can only
387 * occur in very strange VM situations but we want to make sure
388 * that this does not occur.
389 */
390 return min(x, total);
391#else
392 return 0;
393#endif
394}
395
396/**
397 * determine_dirtyable_memory - amount of memory that may be used
398 *
399 * Returns the numebr of pages that can currently be freed and used
400 * by the kernel for direct mappings.
401 */
402unsigned long determine_dirtyable_memory(void)
403{
404 unsigned long x;
405
406 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
407
408 if (!vm_highmem_is_dirtyable)
409 x -= highmem_dirtyable_memory(x);
410
411 return x + 1; /* Ensure that we never return 0 */
412}
413
414static unsigned long hard_dirty_limit(unsigned long thresh)
415{
416 return max(thresh, global_dirty_limit);
417}
418
419/*
420 * global_dirty_limits - background-writeback and dirty-throttling thresholds
421 *
422 * Calculate the dirty thresholds based on sysctl parameters
423 * - vm.dirty_background_ratio or vm.dirty_background_bytes
424 * - vm.dirty_ratio or vm.dirty_bytes
425 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
426 * real-time tasks.
427 */
428void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
429{
430 unsigned long background;
431 unsigned long dirty;
432 unsigned long uninitialized_var(available_memory);
433 struct task_struct *tsk;
434
435 if (!vm_dirty_bytes || !dirty_background_bytes)
436 available_memory = determine_dirtyable_memory();
437
438 if (vm_dirty_bytes)
439 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
440 else
441 dirty = (vm_dirty_ratio * available_memory) / 100;
442
443 if (dirty_background_bytes)
444 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
445 else
446 background = (dirty_background_ratio * available_memory) / 100;
447
448 if (background >= dirty)
449 background = dirty / 2;
450 tsk = current;
451 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
452 background += background / 4;
453 dirty += dirty / 4;
454 }
455 *pbackground = background;
456 *pdirty = dirty;
457 trace_global_dirty_state(background, dirty);
458}
459
460/**
461 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
462 * @bdi: the backing_dev_info to query
463 * @dirty: global dirty limit in pages
464 *
465 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
466 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
467 * And the "limit" in the name is not seriously taken as hard limit in
468 * balance_dirty_pages().
469 *
470 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
471 * - starving fast devices
472 * - piling up dirty pages (that will take long time to sync) on slow devices
473 *
474 * The bdi's share of dirty limit will be adapting to its throughput and
475 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
476 */
477unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
478{
479 u64 bdi_dirty;
480 long numerator, denominator;
481
482 /*
483 * Calculate this BDI's share of the dirty ratio.
484 */
485 bdi_writeout_fraction(bdi, &numerator, &denominator);
486
487 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
488 bdi_dirty *= numerator;
489 do_div(bdi_dirty, denominator);
490
491 bdi_dirty += (dirty * bdi->min_ratio) / 100;
492 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
493 bdi_dirty = dirty * bdi->max_ratio / 100;
494
495 return bdi_dirty;
496}
497
498static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
499 unsigned long elapsed,
500 unsigned long written)
501{
502 const unsigned long period = roundup_pow_of_two(3 * HZ);
503 unsigned long avg = bdi->avg_write_bandwidth;
504 unsigned long old = bdi->write_bandwidth;
505 u64 bw;
506
507 /*
508 * bw = written * HZ / elapsed
509 *
510 * bw * elapsed + write_bandwidth * (period - elapsed)
511 * write_bandwidth = ---------------------------------------------------
512 * period
513 */
514 bw = written - bdi->written_stamp;
515 bw *= HZ;
516 if (unlikely(elapsed > period)) {
517 do_div(bw, elapsed);
518 avg = bw;
519 goto out;
520 }
521 bw += (u64)bdi->write_bandwidth * (period - elapsed);
522 bw >>= ilog2(period);
523
524 /*
525 * one more level of smoothing, for filtering out sudden spikes
526 */
527 if (avg > old && old >= (unsigned long)bw)
528 avg -= (avg - old) >> 3;
529
530 if (avg < old && old <= (unsigned long)bw)
531 avg += (old - avg) >> 3;
532
533out:
534 bdi->write_bandwidth = bw;
535 bdi->avg_write_bandwidth = avg;
536}
537
538/*
539 * The global dirtyable memory and dirty threshold could be suddenly knocked
540 * down by a large amount (eg. on the startup of KVM in a swapless system).
541 * This may throw the system into deep dirty exceeded state and throttle
542 * heavy/light dirtiers alike. To retain good responsiveness, maintain
543 * global_dirty_limit for tracking slowly down to the knocked down dirty
544 * threshold.
545 */
546static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
547{
548 unsigned long limit = global_dirty_limit;
549
550 /*
551 * Follow up in one step.
552 */
553 if (limit < thresh) {
554 limit = thresh;
555 goto update;
556 }
557
558 /*
559 * Follow down slowly. Use the higher one as the target, because thresh
560 * may drop below dirty. This is exactly the reason to introduce
561 * global_dirty_limit which is guaranteed to lie above the dirty pages.
562 */
563 thresh = max(thresh, dirty);
564 if (limit > thresh) {
565 limit -= (limit - thresh) >> 5;
566 goto update;
567 }
568 return;
569update:
570 global_dirty_limit = limit;
571}
572
573static void global_update_bandwidth(unsigned long thresh,
574 unsigned long dirty,
575 unsigned long now)
576{
577 static DEFINE_SPINLOCK(dirty_lock);
578 static unsigned long update_time;
579
580 /*
581 * check locklessly first to optimize away locking for the most time
582 */
583 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
584 return;
585
586 spin_lock(&dirty_lock);
587 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
588 update_dirty_limit(thresh, dirty);
589 update_time = now;
590 }
591 spin_unlock(&dirty_lock);
592}
593
594void __bdi_update_bandwidth(struct backing_dev_info *bdi,
595 unsigned long thresh,
596 unsigned long dirty,
597 unsigned long bdi_thresh,
598 unsigned long bdi_dirty,
599 unsigned long start_time)
600{
601 unsigned long now = jiffies;
602 unsigned long elapsed = now - bdi->bw_time_stamp;
603 unsigned long written;
604
605 /*
606 * rate-limit, only update once every 200ms.
607 */
608 if (elapsed < BANDWIDTH_INTERVAL)
609 return;
610
611 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
612
613 /*
614 * Skip quiet periods when disk bandwidth is under-utilized.
615 * (at least 1s idle time between two flusher runs)
616 */
617 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
618 goto snapshot;
619
620 if (thresh)
621 global_update_bandwidth(thresh, dirty, now);
622
623 bdi_update_write_bandwidth(bdi, elapsed, written);
624
625snapshot:
626 bdi->written_stamp = written;
627 bdi->bw_time_stamp = now;
628}
629
630static void bdi_update_bandwidth(struct backing_dev_info *bdi,
631 unsigned long thresh,
632 unsigned long dirty,
633 unsigned long bdi_thresh,
634 unsigned long bdi_dirty,
635 unsigned long start_time)
636{
637 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
638 return;
639 spin_lock(&bdi->wb.list_lock);
640 __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
641 start_time);
642 spin_unlock(&bdi->wb.list_lock);
643}
644
645/*
646 * balance_dirty_pages() must be called by processes which are generating dirty
647 * data. It looks at the number of dirty pages in the machine and will force
648 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
649 * If we're over `background_thresh' then the writeback threads are woken to
650 * perform some writeout.
651 */
652static void balance_dirty_pages(struct address_space *mapping,
653 unsigned long write_chunk)
654{
655 unsigned long nr_reclaimable, bdi_nr_reclaimable;
656 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
657 unsigned long bdi_dirty;
658 unsigned long background_thresh;
659 unsigned long dirty_thresh;
660 unsigned long bdi_thresh;
661 unsigned long task_bdi_thresh;
662 unsigned long min_task_bdi_thresh;
663 unsigned long pages_written = 0;
664 unsigned long pause = 1;
665 bool dirty_exceeded = false;
666 bool clear_dirty_exceeded = true;
667 struct backing_dev_info *bdi = mapping->backing_dev_info;
668 unsigned long start_time = jiffies;
669
670 for (;;) {
671 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
672 global_page_state(NR_UNSTABLE_NFS);
673 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
674
675 global_dirty_limits(&background_thresh, &dirty_thresh);
676
677 /*
678 * Throttle it only when the background writeback cannot
679 * catch-up. This avoids (excessively) small writeouts
680 * when the bdi limits are ramping up.
681 */
682 if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
683 break;
684
685 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
686 min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
687 task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
688
689 /*
690 * In order to avoid the stacked BDI deadlock we need
691 * to ensure we accurately count the 'dirty' pages when
692 * the threshold is low.
693 *
694 * Otherwise it would be possible to get thresh+n pages
695 * reported dirty, even though there are thresh-m pages
696 * actually dirty; with m+n sitting in the percpu
697 * deltas.
698 */
699 if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) {
700 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
701 bdi_dirty = bdi_nr_reclaimable +
702 bdi_stat_sum(bdi, BDI_WRITEBACK);
703 } else {
704 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
705 bdi_dirty = bdi_nr_reclaimable +
706 bdi_stat(bdi, BDI_WRITEBACK);
707 }
708
709 /*
710 * The bdi thresh is somehow "soft" limit derived from the
711 * global "hard" limit. The former helps to prevent heavy IO
712 * bdi or process from holding back light ones; The latter is
713 * the last resort safeguard.
714 */
715 dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
716 (nr_dirty > dirty_thresh);
717 clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) &&
718 (nr_dirty <= dirty_thresh);
719
720 if (!dirty_exceeded)
721 break;
722
723 if (!bdi->dirty_exceeded)
724 bdi->dirty_exceeded = 1;
725
726 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
727 bdi_thresh, bdi_dirty, start_time);
728
729 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
730 * Unstable writes are a feature of certain networked
731 * filesystems (i.e. NFS) in which data may have been
732 * written to the server's write cache, but has not yet
733 * been flushed to permanent storage.
734 * Only move pages to writeback if this bdi is over its
735 * threshold otherwise wait until the disk writes catch
736 * up.
737 */
738 trace_balance_dirty_start(bdi);
739 if (bdi_nr_reclaimable > task_bdi_thresh) {
740 pages_written += writeback_inodes_wb(&bdi->wb,
741 write_chunk);
742 trace_balance_dirty_written(bdi, pages_written);
743 if (pages_written >= write_chunk)
744 break; /* We've done our duty */
745 }
746 __set_current_state(TASK_UNINTERRUPTIBLE);
747 io_schedule_timeout(pause);
748 trace_balance_dirty_wait(bdi);
749
750 dirty_thresh = hard_dirty_limit(dirty_thresh);
751 /*
752 * max-pause area. If dirty exceeded but still within this
753 * area, no need to sleep for more than 200ms: (a) 8 pages per
754 * 200ms is typically more than enough to curb heavy dirtiers;
755 * (b) the pause time limit makes the dirtiers more responsive.
756 */
757 if (nr_dirty < dirty_thresh &&
758 bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
759 time_after(jiffies, start_time + MAX_PAUSE))
760 break;
761
762 /*
763 * Increase the delay for each loop, up to our previous
764 * default of taking a 100ms nap.
765 */
766 pause <<= 1;
767 if (pause > HZ / 10)
768 pause = HZ / 10;
769 }
770
771 /* Clear dirty_exceeded flag only when no task can exceed the limit */
772 if (clear_dirty_exceeded && bdi->dirty_exceeded)
773 bdi->dirty_exceeded = 0;
774
775 if (writeback_in_progress(bdi))
776 return;
777
778 /*
779 * In laptop mode, we wait until hitting the higher threshold before
780 * starting background writeout, and then write out all the way down
781 * to the lower threshold. So slow writers cause minimal disk activity.
782 *
783 * In normal mode, we start background writeout at the lower
784 * background_thresh, to keep the amount of dirty memory low.
785 */
786 if ((laptop_mode && pages_written) ||
787 (!laptop_mode && (nr_reclaimable > background_thresh)))
788 bdi_start_background_writeback(bdi);
789}
790
791void set_page_dirty_balance(struct page *page, int page_mkwrite)
792{
793 if (set_page_dirty(page) || page_mkwrite) {
794 struct address_space *mapping = page_mapping(page);
795
796 if (mapping)
797 balance_dirty_pages_ratelimited(mapping);
798 }
799}
800
801static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
802
803/**
804 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
805 * @mapping: address_space which was dirtied
806 * @nr_pages_dirtied: number of pages which the caller has just dirtied
807 *
808 * Processes which are dirtying memory should call in here once for each page
809 * which was newly dirtied. The function will periodically check the system's
810 * dirty state and will initiate writeback if needed.
811 *
812 * On really big machines, get_writeback_state is expensive, so try to avoid
813 * calling it too often (ratelimiting). But once we're over the dirty memory
814 * limit we decrease the ratelimiting by a lot, to prevent individual processes
815 * from overshooting the limit by (ratelimit_pages) each.
816 */
817void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
818 unsigned long nr_pages_dirtied)
819{
820 struct backing_dev_info *bdi = mapping->backing_dev_info;
821 unsigned long ratelimit;
822 unsigned long *p;
823
824 if (!bdi_cap_account_dirty(bdi))
825 return;
826
827 ratelimit = ratelimit_pages;
828 if (mapping->backing_dev_info->dirty_exceeded)
829 ratelimit = 8;
830
831 /*
832 * Check the rate limiting. Also, we do not want to throttle real-time
833 * tasks in balance_dirty_pages(). Period.
834 */
835 preempt_disable();
836 p = &__get_cpu_var(bdp_ratelimits);
837 *p += nr_pages_dirtied;
838 if (unlikely(*p >= ratelimit)) {
839 ratelimit = sync_writeback_pages(*p);
840 *p = 0;
841 preempt_enable();
842 balance_dirty_pages(mapping, ratelimit);
843 return;
844 }
845 preempt_enable();
846}
847EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
848
849void throttle_vm_writeout(gfp_t gfp_mask)
850{
851 unsigned long background_thresh;
852 unsigned long dirty_thresh;
853
854 for ( ; ; ) {
855 global_dirty_limits(&background_thresh, &dirty_thresh);
856
857 /*
858 * Boost the allowable dirty threshold a bit for page
859 * allocators so they don't get DoS'ed by heavy writers
860 */
861 dirty_thresh += dirty_thresh / 10; /* wheeee... */
862
863 if (global_page_state(NR_UNSTABLE_NFS) +
864 global_page_state(NR_WRITEBACK) <= dirty_thresh)
865 break;
866 congestion_wait(BLK_RW_ASYNC, HZ/10);
867
868 /*
869 * The caller might hold locks which can prevent IO completion
870 * or progress in the filesystem. So we cannot just sit here
871 * waiting for IO to complete.
872 */
873 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
874 break;
875 }
876}
877
878/*
879 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
880 */
881int dirty_writeback_centisecs_handler(ctl_table *table, int write,
882 void __user *buffer, size_t *length, loff_t *ppos)
883{
884 proc_dointvec(table, write, buffer, length, ppos);
885 bdi_arm_supers_timer();
886 return 0;
887}
888
889#ifdef CONFIG_BLOCK
890void laptop_mode_timer_fn(unsigned long data)
891{
892 struct request_queue *q = (struct request_queue *)data;
893 int nr_pages = global_page_state(NR_FILE_DIRTY) +
894 global_page_state(NR_UNSTABLE_NFS);
895
896 /*
897 * We want to write everything out, not just down to the dirty
898 * threshold
899 */
900 if (bdi_has_dirty_io(&q->backing_dev_info))
901 bdi_start_writeback(&q->backing_dev_info, nr_pages);
902}
903
904/*
905 * We've spun up the disk and we're in laptop mode: schedule writeback
906 * of all dirty data a few seconds from now. If the flush is already scheduled
907 * then push it back - the user is still using the disk.
908 */
909void laptop_io_completion(struct backing_dev_info *info)
910{
911 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
912}
913
914/*
915 * We're in laptop mode and we've just synced. The sync's writes will have
916 * caused another writeback to be scheduled by laptop_io_completion.
917 * Nothing needs to be written back anymore, so we unschedule the writeback.
918 */
919void laptop_sync_completion(void)
920{
921 struct backing_dev_info *bdi;
922
923 rcu_read_lock();
924
925 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
926 del_timer(&bdi->laptop_mode_wb_timer);
927
928 rcu_read_unlock();
929}
930#endif
931
932/*
933 * If ratelimit_pages is too high then we can get into dirty-data overload
934 * if a large number of processes all perform writes at the same time.
935 * If it is too low then SMP machines will call the (expensive)
936 * get_writeback_state too often.
937 *
938 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
939 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
940 * thresholds before writeback cuts in.
941 *
942 * But the limit should not be set too high. Because it also controls the
943 * amount of memory which the balance_dirty_pages() caller has to write back.
944 * If this is too large then the caller will block on the IO queue all the
945 * time. So limit it to four megabytes - the balance_dirty_pages() caller
946 * will write six megabyte chunks, max.
947 */
948
949void writeback_set_ratelimit(void)
950{
951 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
952 if (ratelimit_pages < 16)
953 ratelimit_pages = 16;
954 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
955 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
956}
957
958static int __cpuinit
959ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
960{
961 writeback_set_ratelimit();
962 return NOTIFY_DONE;
963}
964
965static struct notifier_block __cpuinitdata ratelimit_nb = {
966 .notifier_call = ratelimit_handler,
967 .next = NULL,
968};
969
970/*
971 * Called early on to tune the page writeback dirty limits.
972 *
973 * We used to scale dirty pages according to how total memory
974 * related to pages that could be allocated for buffers (by
975 * comparing nr_free_buffer_pages() to vm_total_pages.
976 *
977 * However, that was when we used "dirty_ratio" to scale with
978 * all memory, and we don't do that any more. "dirty_ratio"
979 * is now applied to total non-HIGHPAGE memory (by subtracting
980 * totalhigh_pages from vm_total_pages), and as such we can't
981 * get into the old insane situation any more where we had
982 * large amounts of dirty pages compared to a small amount of
983 * non-HIGHMEM memory.
984 *
985 * But we might still want to scale the dirty_ratio by how
986 * much memory the box has..
987 */
988void __init page_writeback_init(void)
989{
990 int shift;
991
992 writeback_set_ratelimit();
993 register_cpu_notifier(&ratelimit_nb);
994
995 shift = calc_period_shift();
996 prop_descriptor_init(&vm_completions, shift);
997 prop_descriptor_init(&vm_dirties, shift);
998}
999
1000/**
1001 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1002 * @mapping: address space structure to write
1003 * @start: starting page index
1004 * @end: ending page index (inclusive)
1005 *
1006 * This function scans the page range from @start to @end (inclusive) and tags
1007 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1008 * that write_cache_pages (or whoever calls this function) will then use
1009 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1010 * used to avoid livelocking of writeback by a process steadily creating new
1011 * dirty pages in the file (thus it is important for this function to be quick
1012 * so that it can tag pages faster than a dirtying process can create them).
1013 */
1014/*
1015 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1016 */
1017void tag_pages_for_writeback(struct address_space *mapping,
1018 pgoff_t start, pgoff_t end)
1019{
1020#define WRITEBACK_TAG_BATCH 4096
1021 unsigned long tagged;
1022
1023 do {
1024 spin_lock_irq(&mapping->tree_lock);
1025 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1026 &start, end, WRITEBACK_TAG_BATCH,
1027 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1028 spin_unlock_irq(&mapping->tree_lock);
1029 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1030 cond_resched();
1031 /* We check 'start' to handle wrapping when end == ~0UL */
1032 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1033}
1034EXPORT_SYMBOL(tag_pages_for_writeback);
1035
1036/**
1037 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1038 * @mapping: address space structure to write
1039 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1040 * @writepage: function called for each page
1041 * @data: data passed to writepage function
1042 *
1043 * If a page is already under I/O, write_cache_pages() skips it, even
1044 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1045 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1046 * and msync() need to guarantee that all the data which was dirty at the time
1047 * the call was made get new I/O started against them. If wbc->sync_mode is
1048 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1049 * existing IO to complete.
1050 *
1051 * To avoid livelocks (when other process dirties new pages), we first tag
1052 * pages which should be written back with TOWRITE tag and only then start
1053 * writing them. For data-integrity sync we have to be careful so that we do
1054 * not miss some pages (e.g., because some other process has cleared TOWRITE
1055 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1056 * by the process clearing the DIRTY tag (and submitting the page for IO).
1057 */
1058int write_cache_pages(struct address_space *mapping,
1059 struct writeback_control *wbc, writepage_t writepage,
1060 void *data)
1061{
1062 int ret = 0;
1063 int done = 0;
1064 struct pagevec pvec;
1065 int nr_pages;
1066 pgoff_t uninitialized_var(writeback_index);
1067 pgoff_t index;
1068 pgoff_t end; /* Inclusive */
1069 pgoff_t done_index;
1070 int cycled;
1071 int range_whole = 0;
1072 int tag;
1073
1074 pagevec_init(&pvec, 0);
1075 if (wbc->range_cyclic) {
1076 writeback_index = mapping->writeback_index; /* prev offset */
1077 index = writeback_index;
1078 if (index == 0)
1079 cycled = 1;
1080 else
1081 cycled = 0;
1082 end = -1;
1083 } else {
1084 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1085 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1086 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1087 range_whole = 1;
1088 cycled = 1; /* ignore range_cyclic tests */
1089 }
1090 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1091 tag = PAGECACHE_TAG_TOWRITE;
1092 else
1093 tag = PAGECACHE_TAG_DIRTY;
1094retry:
1095 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1096 tag_pages_for_writeback(mapping, index, end);
1097 done_index = index;
1098 while (!done && (index <= end)) {
1099 int i;
1100
1101 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1102 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1103 if (nr_pages == 0)
1104 break;
1105
1106 for (i = 0; i < nr_pages; i++) {
1107 struct page *page = pvec.pages[i];
1108
1109 /*
1110 * At this point, the page may be truncated or
1111 * invalidated (changing page->mapping to NULL), or
1112 * even swizzled back from swapper_space to tmpfs file
1113 * mapping. However, page->index will not change
1114 * because we have a reference on the page.
1115 */
1116 if (page->index > end) {
1117 /*
1118 * can't be range_cyclic (1st pass) because
1119 * end == -1 in that case.
1120 */
1121 done = 1;
1122 break;
1123 }
1124
1125 done_index = page->index;
1126
1127 lock_page(page);
1128
1129 /*
1130 * Page truncated or invalidated. We can freely skip it
1131 * then, even for data integrity operations: the page
1132 * has disappeared concurrently, so there could be no
1133 * real expectation of this data interity operation
1134 * even if there is now a new, dirty page at the same
1135 * pagecache address.
1136 */
1137 if (unlikely(page->mapping != mapping)) {
1138continue_unlock:
1139 unlock_page(page);
1140 continue;
1141 }
1142
1143 if (!PageDirty(page)) {
1144 /* someone wrote it for us */
1145 goto continue_unlock;
1146 }
1147
1148 if (PageWriteback(page)) {
1149 if (wbc->sync_mode != WB_SYNC_NONE)
1150 wait_on_page_writeback(page);
1151 else
1152 goto continue_unlock;
1153 }
1154
1155 BUG_ON(PageWriteback(page));
1156 if (!clear_page_dirty_for_io(page))
1157 goto continue_unlock;
1158
1159 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1160 ret = (*writepage)(page, wbc, data);
1161 if (unlikely(ret)) {
1162 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1163 unlock_page(page);
1164 ret = 0;
1165 } else {
1166 /*
1167 * done_index is set past this page,
1168 * so media errors will not choke
1169 * background writeout for the entire
1170 * file. This has consequences for
1171 * range_cyclic semantics (ie. it may
1172 * not be suitable for data integrity
1173 * writeout).
1174 */
1175 done_index = page->index + 1;
1176 done = 1;
1177 break;
1178 }
1179 }
1180
1181 /*
1182 * We stop writing back only if we are not doing
1183 * integrity sync. In case of integrity sync we have to
1184 * keep going until we have written all the pages
1185 * we tagged for writeback prior to entering this loop.
1186 */
1187 if (--wbc->nr_to_write <= 0 &&
1188 wbc->sync_mode == WB_SYNC_NONE) {
1189 done = 1;
1190 break;
1191 }
1192 }
1193 pagevec_release(&pvec);
1194 cond_resched();
1195 }
1196 if (!cycled && !done) {
1197 /*
1198 * range_cyclic:
1199 * We hit the last page and there is more work to be done: wrap
1200 * back to the start of the file
1201 */
1202 cycled = 1;
1203 index = 0;
1204 end = writeback_index - 1;
1205 goto retry;
1206 }
1207 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1208 mapping->writeback_index = done_index;
1209
1210 return ret;
1211}
1212EXPORT_SYMBOL(write_cache_pages);
1213
1214/*
1215 * Function used by generic_writepages to call the real writepage
1216 * function and set the mapping flags on error
1217 */
1218static int __writepage(struct page *page, struct writeback_control *wbc,
1219 void *data)
1220{
1221 struct address_space *mapping = data;
1222 int ret = mapping->a_ops->writepage(page, wbc);
1223 mapping_set_error(mapping, ret);
1224 return ret;
1225}
1226
1227/**
1228 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1229 * @mapping: address space structure to write
1230 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1231 *
1232 * This is a library function, which implements the writepages()
1233 * address_space_operation.
1234 */
1235int generic_writepages(struct address_space *mapping,
1236 struct writeback_control *wbc)
1237{
1238 struct blk_plug plug;
1239 int ret;
1240
1241 /* deal with chardevs and other special file */
1242 if (!mapping->a_ops->writepage)
1243 return 0;
1244
1245 blk_start_plug(&plug);
1246 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1247 blk_finish_plug(&plug);
1248 return ret;
1249}
1250
1251EXPORT_SYMBOL(generic_writepages);
1252
1253int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1254{
1255 int ret;
1256
1257 if (wbc->nr_to_write <= 0)
1258 return 0;
1259 if (mapping->a_ops->writepages)
1260 ret = mapping->a_ops->writepages(mapping, wbc);
1261 else
1262 ret = generic_writepages(mapping, wbc);
1263 return ret;
1264}
1265
1266/**
1267 * write_one_page - write out a single page and optionally wait on I/O
1268 * @page: the page to write
1269 * @wait: if true, wait on writeout
1270 *
1271 * The page must be locked by the caller and will be unlocked upon return.
1272 *
1273 * write_one_page() returns a negative error code if I/O failed.
1274 */
1275int write_one_page(struct page *page, int wait)
1276{
1277 struct address_space *mapping = page->mapping;
1278 int ret = 0;
1279 struct writeback_control wbc = {
1280 .sync_mode = WB_SYNC_ALL,
1281 .nr_to_write = 1,
1282 };
1283
1284 BUG_ON(!PageLocked(page));
1285
1286 if (wait)
1287 wait_on_page_writeback(page);
1288
1289 if (clear_page_dirty_for_io(page)) {
1290 page_cache_get(page);
1291 ret = mapping->a_ops->writepage(page, &wbc);
1292 if (ret == 0 && wait) {
1293 wait_on_page_writeback(page);
1294 if (PageError(page))
1295 ret = -EIO;
1296 }
1297 page_cache_release(page);
1298 } else {
1299 unlock_page(page);
1300 }
1301 return ret;
1302}
1303EXPORT_SYMBOL(write_one_page);
1304
1305/*
1306 * For address_spaces which do not use buffers nor write back.
1307 */
1308int __set_page_dirty_no_writeback(struct page *page)
1309{
1310 if (!PageDirty(page))
1311 return !TestSetPageDirty(page);
1312 return 0;
1313}
1314
1315/*
1316 * Helper function for set_page_dirty family.
1317 * NOTE: This relies on being atomic wrt interrupts.
1318 */
1319void account_page_dirtied(struct page *page, struct address_space *mapping)
1320{
1321 if (mapping_cap_account_dirty(mapping)) {
1322 __inc_zone_page_state(page, NR_FILE_DIRTY);
1323 __inc_zone_page_state(page, NR_DIRTIED);
1324 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1325 task_dirty_inc(current);
1326 task_io_account_write(PAGE_CACHE_SIZE);
1327 }
1328}
1329EXPORT_SYMBOL(account_page_dirtied);
1330
1331/*
1332 * Helper function for set_page_writeback family.
1333 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1334 * wrt interrupts.
1335 */
1336void account_page_writeback(struct page *page)
1337{
1338 inc_zone_page_state(page, NR_WRITEBACK);
1339}
1340EXPORT_SYMBOL(account_page_writeback);
1341
1342/*
1343 * For address_spaces which do not use buffers. Just tag the page as dirty in
1344 * its radix tree.
1345 *
1346 * This is also used when a single buffer is being dirtied: we want to set the
1347 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1348 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1349 *
1350 * Most callers have locked the page, which pins the address_space in memory.
1351 * But zap_pte_range() does not lock the page, however in that case the
1352 * mapping is pinned by the vma's ->vm_file reference.
1353 *
1354 * We take care to handle the case where the page was truncated from the
1355 * mapping by re-checking page_mapping() inside tree_lock.
1356 */
1357int __set_page_dirty_nobuffers(struct page *page)
1358{
1359 if (!TestSetPageDirty(page)) {
1360 struct address_space *mapping = page_mapping(page);
1361 struct address_space *mapping2;
1362
1363 if (!mapping)
1364 return 1;
1365
1366 spin_lock_irq(&mapping->tree_lock);
1367 mapping2 = page_mapping(page);
1368 if (mapping2) { /* Race with truncate? */
1369 BUG_ON(mapping2 != mapping);
1370 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1371 account_page_dirtied(page, mapping);
1372 radix_tree_tag_set(&mapping->page_tree,
1373 page_index(page), PAGECACHE_TAG_DIRTY);
1374 }
1375 spin_unlock_irq(&mapping->tree_lock);
1376 if (mapping->host) {
1377 /* !PageAnon && !swapper_space */
1378 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1379 }
1380 return 1;
1381 }
1382 return 0;
1383}
1384EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1385
1386/*
1387 * When a writepage implementation decides that it doesn't want to write this
1388 * page for some reason, it should redirty the locked page via
1389 * redirty_page_for_writepage() and it should then unlock the page and return 0
1390 */
1391int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1392{
1393 wbc->pages_skipped++;
1394 return __set_page_dirty_nobuffers(page);
1395}
1396EXPORT_SYMBOL(redirty_page_for_writepage);
1397
1398/*
1399 * Dirty a page.
1400 *
1401 * For pages with a mapping this should be done under the page lock
1402 * for the benefit of asynchronous memory errors who prefer a consistent
1403 * dirty state. This rule can be broken in some special cases,
1404 * but should be better not to.
1405 *
1406 * If the mapping doesn't provide a set_page_dirty a_op, then
1407 * just fall through and assume that it wants buffer_heads.
1408 */
1409int set_page_dirty(struct page *page)
1410{
1411 struct address_space *mapping = page_mapping(page);
1412
1413 if (likely(mapping)) {
1414 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1415 /*
1416 * readahead/lru_deactivate_page could remain
1417 * PG_readahead/PG_reclaim due to race with end_page_writeback
1418 * About readahead, if the page is written, the flags would be
1419 * reset. So no problem.
1420 * About lru_deactivate_page, if the page is redirty, the flag
1421 * will be reset. So no problem. but if the page is used by readahead
1422 * it will confuse readahead and make it restart the size rampup
1423 * process. But it's a trivial problem.
1424 */
1425 ClearPageReclaim(page);
1426#ifdef CONFIG_BLOCK
1427 if (!spd)
1428 spd = __set_page_dirty_buffers;
1429#endif
1430 return (*spd)(page);
1431 }
1432 if (!PageDirty(page)) {
1433 if (!TestSetPageDirty(page))
1434 return 1;
1435 }
1436 return 0;
1437}
1438EXPORT_SYMBOL(set_page_dirty);
1439
1440/*
1441 * set_page_dirty() is racy if the caller has no reference against
1442 * page->mapping->host, and if the page is unlocked. This is because another
1443 * CPU could truncate the page off the mapping and then free the mapping.
1444 *
1445 * Usually, the page _is_ locked, or the caller is a user-space process which
1446 * holds a reference on the inode by having an open file.
1447 *
1448 * In other cases, the page should be locked before running set_page_dirty().
1449 */
1450int set_page_dirty_lock(struct page *page)
1451{
1452 int ret;
1453
1454 lock_page(page);
1455 ret = set_page_dirty(page);
1456 unlock_page(page);
1457 return ret;
1458}
1459EXPORT_SYMBOL(set_page_dirty_lock);
1460
1461/*
1462 * Clear a page's dirty flag, while caring for dirty memory accounting.
1463 * Returns true if the page was previously dirty.
1464 *
1465 * This is for preparing to put the page under writeout. We leave the page
1466 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1467 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1468 * implementation will run either set_page_writeback() or set_page_dirty(),
1469 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1470 * back into sync.
1471 *
1472 * This incoherency between the page's dirty flag and radix-tree tag is
1473 * unfortunate, but it only exists while the page is locked.
1474 */
1475int clear_page_dirty_for_io(struct page *page)
1476{
1477 struct address_space *mapping = page_mapping(page);
1478
1479 BUG_ON(!PageLocked(page));
1480
1481 if (mapping && mapping_cap_account_dirty(mapping)) {
1482 /*
1483 * Yes, Virginia, this is indeed insane.
1484 *
1485 * We use this sequence to make sure that
1486 * (a) we account for dirty stats properly
1487 * (b) we tell the low-level filesystem to
1488 * mark the whole page dirty if it was
1489 * dirty in a pagetable. Only to then
1490 * (c) clean the page again and return 1 to
1491 * cause the writeback.
1492 *
1493 * This way we avoid all nasty races with the
1494 * dirty bit in multiple places and clearing
1495 * them concurrently from different threads.
1496 *
1497 * Note! Normally the "set_page_dirty(page)"
1498 * has no effect on the actual dirty bit - since
1499 * that will already usually be set. But we
1500 * need the side effects, and it can help us
1501 * avoid races.
1502 *
1503 * We basically use the page "master dirty bit"
1504 * as a serialization point for all the different
1505 * threads doing their things.
1506 */
1507 if (page_mkclean(page))
1508 set_page_dirty(page);
1509 /*
1510 * We carefully synchronise fault handlers against
1511 * installing a dirty pte and marking the page dirty
1512 * at this point. We do this by having them hold the
1513 * page lock at some point after installing their
1514 * pte, but before marking the page dirty.
1515 * Pages are always locked coming in here, so we get
1516 * the desired exclusion. See mm/memory.c:do_wp_page()
1517 * for more comments.
1518 */
1519 if (TestClearPageDirty(page)) {
1520 dec_zone_page_state(page, NR_FILE_DIRTY);
1521 dec_bdi_stat(mapping->backing_dev_info,
1522 BDI_RECLAIMABLE);
1523 return 1;
1524 }
1525 return 0;
1526 }
1527 return TestClearPageDirty(page);
1528}
1529EXPORT_SYMBOL(clear_page_dirty_for_io);
1530
1531int test_clear_page_writeback(struct page *page)
1532{
1533 struct address_space *mapping = page_mapping(page);
1534 int ret;
1535
1536 if (mapping) {
1537 struct backing_dev_info *bdi = mapping->backing_dev_info;
1538 unsigned long flags;
1539
1540 spin_lock_irqsave(&mapping->tree_lock, flags);
1541 ret = TestClearPageWriteback(page);
1542 if (ret) {
1543 radix_tree_tag_clear(&mapping->page_tree,
1544 page_index(page),
1545 PAGECACHE_TAG_WRITEBACK);
1546 if (bdi_cap_account_writeback(bdi)) {
1547 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1548 __bdi_writeout_inc(bdi);
1549 }
1550 }
1551 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1552 } else {
1553 ret = TestClearPageWriteback(page);
1554 }
1555 if (ret) {
1556 dec_zone_page_state(page, NR_WRITEBACK);
1557 inc_zone_page_state(page, NR_WRITTEN);
1558 }
1559 return ret;
1560}
1561
1562int test_set_page_writeback(struct page *page)
1563{
1564 struct address_space *mapping = page_mapping(page);
1565 int ret;
1566
1567 if (mapping) {
1568 struct backing_dev_info *bdi = mapping->backing_dev_info;
1569 unsigned long flags;
1570
1571 spin_lock_irqsave(&mapping->tree_lock, flags);
1572 ret = TestSetPageWriteback(page);
1573 if (!ret) {
1574 radix_tree_tag_set(&mapping->page_tree,
1575 page_index(page),
1576 PAGECACHE_TAG_WRITEBACK);
1577 if (bdi_cap_account_writeback(bdi))
1578 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1579 }
1580 if (!PageDirty(page))
1581 radix_tree_tag_clear(&mapping->page_tree,
1582 page_index(page),
1583 PAGECACHE_TAG_DIRTY);
1584 radix_tree_tag_clear(&mapping->page_tree,
1585 page_index(page),
1586 PAGECACHE_TAG_TOWRITE);
1587 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1588 } else {
1589 ret = TestSetPageWriteback(page);
1590 }
1591 if (!ret)
1592 account_page_writeback(page);
1593 return ret;
1594
1595}
1596EXPORT_SYMBOL(test_set_page_writeback);
1597
1598/*
1599 * Return true if any of the pages in the mapping are marked with the
1600 * passed tag.
1601 */
1602int mapping_tagged(struct address_space *mapping, int tag)
1603{
1604 return radix_tree_tagged(&mapping->page_tree, tag);
1605}
1606EXPORT_SYMBOL(mapping_tagged);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/page-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
11 * 10Apr2002 Andrew Morton
12 * Initial version
13 */
14
15#include <linux/kernel.h>
16#include <linux/math64.h>
17#include <linux/export.h>
18#include <linux/spinlock.h>
19#include <linux/fs.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/slab.h>
23#include <linux/pagemap.h>
24#include <linux/writeback.h>
25#include <linux/init.h>
26#include <linux/backing-dev.h>
27#include <linux/task_io_accounting_ops.h>
28#include <linux/blkdev.h>
29#include <linux/mpage.h>
30#include <linux/rmap.h>
31#include <linux/percpu.h>
32#include <linux/smp.h>
33#include <linux/sysctl.h>
34#include <linux/cpu.h>
35#include <linux/syscalls.h>
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/sched/signal.h>
40#include <linux/mm_inline.h>
41#include <trace/events/writeback.h>
42
43#include "internal.h"
44
45/*
46 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48#define MAX_PAUSE max(HZ/5, 1)
49
50/*
51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56/*
57 * Estimate write bandwidth or update dirty limit at 200ms intervals.
58 */
59#define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
61#define RATELIMIT_CALC_SHIFT 10
62
63/*
64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67static 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 */
74static 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 */
80static 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 */
86static int vm_highmem_is_dirtyable;
87
88/*
89 * The generator of dirty data starts writeback at this percentage
90 */
91static 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 */
97static unsigned long vm_dirty_bytes;
98
99/*
100 * The interval between `kupdate'-style writebacks
101 */
102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106/*
107 * The longest time for which data is allowed to remain dirty
108 */
109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111/*
112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113 * a full sync is triggered after this time elapses without any disk activity.
114 */
115int laptop_mode;
116
117EXPORT_SYMBOL(laptop_mode);
118
119/* End of sysctl-exported parameters */
120
121struct wb_domain global_wb_domain;
122
123/* consolidated parameters for balance_dirty_pages() and its subroutines */
124struct dirty_throttle_control {
125#ifdef CONFIG_CGROUP_WRITEBACK
126 struct wb_domain *dom;
127 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
128#endif
129 struct bdi_writeback *wb;
130 struct fprop_local_percpu *wb_completions;
131
132 unsigned long avail; /* dirtyable */
133 unsigned long dirty; /* file_dirty + write + nfs */
134 unsigned long thresh; /* dirty threshold */
135 unsigned long bg_thresh; /* dirty background threshold */
136
137 unsigned long wb_dirty; /* per-wb counterparts */
138 unsigned long wb_thresh;
139 unsigned long wb_bg_thresh;
140
141 unsigned long pos_ratio;
142 bool freerun;
143 bool dirty_exceeded;
144};
145
146/*
147 * Length of period for aging writeout fractions of bdis. This is an
148 * arbitrarily chosen number. The longer the period, the slower fractions will
149 * reflect changes in current writeout rate.
150 */
151#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
152
153#ifdef CONFIG_CGROUP_WRITEBACK
154
155#define GDTC_INIT(__wb) .wb = (__wb), \
156 .dom = &global_wb_domain, \
157 .wb_completions = &(__wb)->completions
158
159#define GDTC_INIT_NO_WB .dom = &global_wb_domain
160
161#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
162 .dom = mem_cgroup_wb_domain(__wb), \
163 .wb_completions = &(__wb)->memcg_completions, \
164 .gdtc = __gdtc
165
166static bool mdtc_valid(struct dirty_throttle_control *dtc)
167{
168 return dtc->dom;
169}
170
171static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
172{
173 return dtc->dom;
174}
175
176static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
177{
178 return mdtc->gdtc;
179}
180
181static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
182{
183 return &wb->memcg_completions;
184}
185
186static void wb_min_max_ratio(struct bdi_writeback *wb,
187 unsigned long *minp, unsigned long *maxp)
188{
189 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
190 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
191 unsigned long long min = wb->bdi->min_ratio;
192 unsigned long long max = wb->bdi->max_ratio;
193
194 /*
195 * @wb may already be clean by the time control reaches here and
196 * the total may not include its bw.
197 */
198 if (this_bw < tot_bw) {
199 if (min) {
200 min *= this_bw;
201 min = div64_ul(min, tot_bw);
202 }
203 if (max < 100 * BDI_RATIO_SCALE) {
204 max *= this_bw;
205 max = div64_ul(max, tot_bw);
206 }
207 }
208
209 *minp = min;
210 *maxp = max;
211}
212
213#else /* CONFIG_CGROUP_WRITEBACK */
214
215#define GDTC_INIT(__wb) .wb = (__wb), \
216 .wb_completions = &(__wb)->completions
217#define GDTC_INIT_NO_WB
218#define MDTC_INIT(__wb, __gdtc)
219
220static bool mdtc_valid(struct dirty_throttle_control *dtc)
221{
222 return false;
223}
224
225static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
226{
227 return &global_wb_domain;
228}
229
230static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
231{
232 return NULL;
233}
234
235static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
236{
237 return NULL;
238}
239
240static void wb_min_max_ratio(struct bdi_writeback *wb,
241 unsigned long *minp, unsigned long *maxp)
242{
243 *minp = wb->bdi->min_ratio;
244 *maxp = wb->bdi->max_ratio;
245}
246
247#endif /* CONFIG_CGROUP_WRITEBACK */
248
249/*
250 * In a memory zone, there is a certain amount of pages we consider
251 * available for the page cache, which is essentially the number of
252 * free and reclaimable pages, minus some zone reserves to protect
253 * lowmem and the ability to uphold the zone's watermarks without
254 * requiring writeback.
255 *
256 * This number of dirtyable pages is the base value of which the
257 * user-configurable dirty ratio is the effective number of pages that
258 * are allowed to be actually dirtied. Per individual zone, or
259 * globally by using the sum of dirtyable pages over all zones.
260 *
261 * Because the user is allowed to specify the dirty limit globally as
262 * absolute number of bytes, calculating the per-zone dirty limit can
263 * require translating the configured limit into a percentage of
264 * global dirtyable memory first.
265 */
266
267/**
268 * node_dirtyable_memory - number of dirtyable pages in a node
269 * @pgdat: the node
270 *
271 * Return: the node's number of pages potentially available for dirty
272 * page cache. This is the base value for the per-node dirty limits.
273 */
274static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
275{
276 unsigned long nr_pages = 0;
277 int z;
278
279 for (z = 0; z < MAX_NR_ZONES; z++) {
280 struct zone *zone = pgdat->node_zones + z;
281
282 if (!populated_zone(zone))
283 continue;
284
285 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
286 }
287
288 /*
289 * Pages reserved for the kernel should not be considered
290 * dirtyable, to prevent a situation where reclaim has to
291 * clean pages in order to balance the zones.
292 */
293 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
294
295 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
296 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
297
298 return nr_pages;
299}
300
301static unsigned long highmem_dirtyable_memory(unsigned long total)
302{
303#ifdef CONFIG_HIGHMEM
304 int node;
305 unsigned long x = 0;
306 int i;
307
308 for_each_node_state(node, N_HIGH_MEMORY) {
309 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
310 struct zone *z;
311 unsigned long nr_pages;
312
313 if (!is_highmem_idx(i))
314 continue;
315
316 z = &NODE_DATA(node)->node_zones[i];
317 if (!populated_zone(z))
318 continue;
319
320 nr_pages = zone_page_state(z, NR_FREE_PAGES);
321 /* watch for underflows */
322 nr_pages -= min(nr_pages, high_wmark_pages(z));
323 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
324 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
325 x += nr_pages;
326 }
327 }
328
329 /*
330 * Make sure that the number of highmem pages is never larger
331 * than the number of the total dirtyable memory. This can only
332 * occur in very strange VM situations but we want to make sure
333 * that this does not occur.
334 */
335 return min(x, total);
336#else
337 return 0;
338#endif
339}
340
341/**
342 * global_dirtyable_memory - number of globally dirtyable pages
343 *
344 * Return: the global number of pages potentially available for dirty
345 * page cache. This is the base value for the global dirty limits.
346 */
347static unsigned long global_dirtyable_memory(void)
348{
349 unsigned long x;
350
351 x = global_zone_page_state(NR_FREE_PAGES);
352 /*
353 * Pages reserved for the kernel should not be considered
354 * dirtyable, to prevent a situation where reclaim has to
355 * clean pages in order to balance the zones.
356 */
357 x -= min(x, totalreserve_pages);
358
359 x += global_node_page_state(NR_INACTIVE_FILE);
360 x += global_node_page_state(NR_ACTIVE_FILE);
361
362 if (!vm_highmem_is_dirtyable)
363 x -= highmem_dirtyable_memory(x);
364
365 return x + 1; /* Ensure that we never return 0 */
366}
367
368/**
369 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
370 * @dtc: dirty_throttle_control of interest
371 *
372 * Calculate @dtc->thresh and ->bg_thresh considering
373 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
374 * must ensure that @dtc->avail is set before calling this function. The
375 * dirty limits will be lifted by 1/4 for real-time tasks.
376 */
377static void domain_dirty_limits(struct dirty_throttle_control *dtc)
378{
379 const unsigned long available_memory = dtc->avail;
380 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
381 unsigned long bytes = vm_dirty_bytes;
382 unsigned long bg_bytes = dirty_background_bytes;
383 /* convert ratios to per-PAGE_SIZE for higher precision */
384 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
385 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
386 unsigned long thresh;
387 unsigned long bg_thresh;
388 struct task_struct *tsk;
389
390 /* gdtc is !NULL iff @dtc is for memcg domain */
391 if (gdtc) {
392 unsigned long global_avail = gdtc->avail;
393
394 /*
395 * The byte settings can't be applied directly to memcg
396 * domains. Convert them to ratios by scaling against
397 * globally available memory. As the ratios are in
398 * per-PAGE_SIZE, they can be obtained by dividing bytes by
399 * number of pages.
400 */
401 if (bytes)
402 ratio = min(DIV_ROUND_UP(bytes, global_avail),
403 PAGE_SIZE);
404 if (bg_bytes)
405 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
406 PAGE_SIZE);
407 bytes = bg_bytes = 0;
408 }
409
410 if (bytes)
411 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
412 else
413 thresh = (ratio * available_memory) / PAGE_SIZE;
414
415 if (bg_bytes)
416 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
417 else
418 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
419
420 tsk = current;
421 if (rt_or_dl_task(tsk)) {
422 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
423 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
424 }
425 /*
426 * Dirty throttling logic assumes the limits in page units fit into
427 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
428 */
429 if (thresh > UINT_MAX)
430 thresh = UINT_MAX;
431 /* This makes sure bg_thresh is within 32-bits as well */
432 if (bg_thresh >= thresh)
433 bg_thresh = thresh / 2;
434 dtc->thresh = thresh;
435 dtc->bg_thresh = bg_thresh;
436
437 /* we should eventually report the domain in the TP */
438 if (!gdtc)
439 trace_global_dirty_state(bg_thresh, thresh);
440}
441
442/**
443 * global_dirty_limits - background-writeback and dirty-throttling thresholds
444 * @pbackground: out parameter for bg_thresh
445 * @pdirty: out parameter for thresh
446 *
447 * Calculate bg_thresh and thresh for global_wb_domain. See
448 * domain_dirty_limits() for details.
449 */
450void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
451{
452 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
453
454 gdtc.avail = global_dirtyable_memory();
455 domain_dirty_limits(&gdtc);
456
457 *pbackground = gdtc.bg_thresh;
458 *pdirty = gdtc.thresh;
459}
460
461/**
462 * node_dirty_limit - maximum number of dirty pages allowed in a node
463 * @pgdat: the node
464 *
465 * Return: the maximum number of dirty pages allowed in a node, based
466 * on the node's dirtyable memory.
467 */
468static unsigned long node_dirty_limit(struct pglist_data *pgdat)
469{
470 unsigned long node_memory = node_dirtyable_memory(pgdat);
471 struct task_struct *tsk = current;
472 unsigned long dirty;
473
474 if (vm_dirty_bytes)
475 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
476 node_memory / global_dirtyable_memory();
477 else
478 dirty = vm_dirty_ratio * node_memory / 100;
479
480 if (rt_or_dl_task(tsk))
481 dirty += dirty / 4;
482
483 /*
484 * Dirty throttling logic assumes the limits in page units fit into
485 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
486 */
487 return min_t(unsigned long, dirty, UINT_MAX);
488}
489
490/**
491 * node_dirty_ok - tells whether a node is within its dirty limits
492 * @pgdat: the node to check
493 *
494 * Return: %true when the dirty pages in @pgdat are within the node's
495 * dirty limit, %false if the limit is exceeded.
496 */
497bool node_dirty_ok(struct pglist_data *pgdat)
498{
499 unsigned long limit = node_dirty_limit(pgdat);
500 unsigned long nr_pages = 0;
501
502 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
503 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
504
505 return nr_pages <= limit;
506}
507
508#ifdef CONFIG_SYSCTL
509static int dirty_background_ratio_handler(const struct ctl_table *table, int write,
510 void *buffer, size_t *lenp, loff_t *ppos)
511{
512 int ret;
513
514 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
515 if (ret == 0 && write)
516 dirty_background_bytes = 0;
517 return ret;
518}
519
520static int dirty_background_bytes_handler(const struct ctl_table *table, int write,
521 void *buffer, size_t *lenp, loff_t *ppos)
522{
523 int ret;
524 unsigned long old_bytes = dirty_background_bytes;
525
526 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
527 if (ret == 0 && write) {
528 if (DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE) >
529 UINT_MAX) {
530 dirty_background_bytes = old_bytes;
531 return -ERANGE;
532 }
533 dirty_background_ratio = 0;
534 }
535 return ret;
536}
537
538static int dirty_ratio_handler(const struct ctl_table *table, int write, void *buffer,
539 size_t *lenp, loff_t *ppos)
540{
541 int old_ratio = vm_dirty_ratio;
542 int ret;
543
544 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
545 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
546 writeback_set_ratelimit();
547 vm_dirty_bytes = 0;
548 }
549 return ret;
550}
551
552static int dirty_bytes_handler(const struct ctl_table *table, int write,
553 void *buffer, size_t *lenp, loff_t *ppos)
554{
555 unsigned long old_bytes = vm_dirty_bytes;
556 int ret;
557
558 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
559 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
560 if (DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) > UINT_MAX) {
561 vm_dirty_bytes = old_bytes;
562 return -ERANGE;
563 }
564 writeback_set_ratelimit();
565 vm_dirty_ratio = 0;
566 }
567 return ret;
568}
569#endif
570
571static unsigned long wp_next_time(unsigned long cur_time)
572{
573 cur_time += VM_COMPLETIONS_PERIOD_LEN;
574 /* 0 has a special meaning... */
575 if (!cur_time)
576 return 1;
577 return cur_time;
578}
579
580static void wb_domain_writeout_add(struct wb_domain *dom,
581 struct fprop_local_percpu *completions,
582 unsigned int max_prop_frac, long nr)
583{
584 __fprop_add_percpu_max(&dom->completions, completions,
585 max_prop_frac, nr);
586 /* First event after period switching was turned off? */
587 if (unlikely(!dom->period_time)) {
588 /*
589 * We can race with other wb_domain_writeout_add calls here but
590 * it does not cause any harm since the resulting time when
591 * timer will fire and what is in writeout_period_time will be
592 * roughly the same.
593 */
594 dom->period_time = wp_next_time(jiffies);
595 mod_timer(&dom->period_timer, dom->period_time);
596 }
597}
598
599/*
600 * Increment @wb's writeout completion count and the global writeout
601 * completion count. Called from __folio_end_writeback().
602 */
603static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
604{
605 struct wb_domain *cgdom;
606
607 wb_stat_mod(wb, WB_WRITTEN, nr);
608 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
609 wb->bdi->max_prop_frac, nr);
610
611 cgdom = mem_cgroup_wb_domain(wb);
612 if (cgdom)
613 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
614 wb->bdi->max_prop_frac, nr);
615}
616
617void wb_writeout_inc(struct bdi_writeback *wb)
618{
619 unsigned long flags;
620
621 local_irq_save(flags);
622 __wb_writeout_add(wb, 1);
623 local_irq_restore(flags);
624}
625EXPORT_SYMBOL_GPL(wb_writeout_inc);
626
627/*
628 * On idle system, we can be called long after we scheduled because we use
629 * deferred timers so count with missed periods.
630 */
631static void writeout_period(struct timer_list *t)
632{
633 struct wb_domain *dom = from_timer(dom, t, period_timer);
634 int miss_periods = (jiffies - dom->period_time) /
635 VM_COMPLETIONS_PERIOD_LEN;
636
637 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
638 dom->period_time = wp_next_time(dom->period_time +
639 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
640 mod_timer(&dom->period_timer, dom->period_time);
641 } else {
642 /*
643 * Aging has zeroed all fractions. Stop wasting CPU on period
644 * updates.
645 */
646 dom->period_time = 0;
647 }
648}
649
650int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
651{
652 memset(dom, 0, sizeof(*dom));
653
654 spin_lock_init(&dom->lock);
655
656 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
657
658 dom->dirty_limit_tstamp = jiffies;
659
660 return fprop_global_init(&dom->completions, gfp);
661}
662
663#ifdef CONFIG_CGROUP_WRITEBACK
664void wb_domain_exit(struct wb_domain *dom)
665{
666 del_timer_sync(&dom->period_timer);
667 fprop_global_destroy(&dom->completions);
668}
669#endif
670
671/*
672 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
673 * registered backing devices, which, for obvious reasons, can not
674 * exceed 100%.
675 */
676static unsigned int bdi_min_ratio;
677
678static int bdi_check_pages_limit(unsigned long pages)
679{
680 unsigned long max_dirty_pages = global_dirtyable_memory();
681
682 if (pages > max_dirty_pages)
683 return -EINVAL;
684
685 return 0;
686}
687
688static unsigned long bdi_ratio_from_pages(unsigned long pages)
689{
690 unsigned long background_thresh;
691 unsigned long dirty_thresh;
692 unsigned long ratio;
693
694 global_dirty_limits(&background_thresh, &dirty_thresh);
695 if (!dirty_thresh)
696 return -EINVAL;
697 ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
698
699 return ratio;
700}
701
702static u64 bdi_get_bytes(unsigned int ratio)
703{
704 unsigned long background_thresh;
705 unsigned long dirty_thresh;
706 u64 bytes;
707
708 global_dirty_limits(&background_thresh, &dirty_thresh);
709 bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
710
711 return bytes;
712}
713
714static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
715{
716 unsigned int delta;
717 int ret = 0;
718
719 if (min_ratio > 100 * BDI_RATIO_SCALE)
720 return -EINVAL;
721
722 spin_lock_bh(&bdi_lock);
723 if (min_ratio > bdi->max_ratio) {
724 ret = -EINVAL;
725 } else {
726 if (min_ratio < bdi->min_ratio) {
727 delta = bdi->min_ratio - min_ratio;
728 bdi_min_ratio -= delta;
729 bdi->min_ratio = min_ratio;
730 } else {
731 delta = min_ratio - bdi->min_ratio;
732 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
733 bdi_min_ratio += delta;
734 bdi->min_ratio = min_ratio;
735 } else {
736 ret = -EINVAL;
737 }
738 }
739 }
740 spin_unlock_bh(&bdi_lock);
741
742 return ret;
743}
744
745static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
746{
747 int ret = 0;
748
749 if (max_ratio > 100 * BDI_RATIO_SCALE)
750 return -EINVAL;
751
752 spin_lock_bh(&bdi_lock);
753 if (bdi->min_ratio > max_ratio) {
754 ret = -EINVAL;
755 } else {
756 bdi->max_ratio = max_ratio;
757 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
758 (100 * BDI_RATIO_SCALE);
759 }
760 spin_unlock_bh(&bdi_lock);
761
762 return ret;
763}
764
765int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
766{
767 return __bdi_set_min_ratio(bdi, min_ratio);
768}
769
770int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
771{
772 return __bdi_set_max_ratio(bdi, max_ratio);
773}
774
775int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
776{
777 return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
778}
779
780int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
781{
782 return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
783}
784EXPORT_SYMBOL(bdi_set_max_ratio);
785
786u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
787{
788 return bdi_get_bytes(bdi->min_ratio);
789}
790
791int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
792{
793 int ret;
794 unsigned long pages = min_bytes >> PAGE_SHIFT;
795 long min_ratio;
796
797 ret = bdi_check_pages_limit(pages);
798 if (ret)
799 return ret;
800
801 min_ratio = bdi_ratio_from_pages(pages);
802 if (min_ratio < 0)
803 return min_ratio;
804 return __bdi_set_min_ratio(bdi, min_ratio);
805}
806
807u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
808{
809 return bdi_get_bytes(bdi->max_ratio);
810}
811
812int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
813{
814 int ret;
815 unsigned long pages = max_bytes >> PAGE_SHIFT;
816 long max_ratio;
817
818 ret = bdi_check_pages_limit(pages);
819 if (ret)
820 return ret;
821
822 max_ratio = bdi_ratio_from_pages(pages);
823 if (max_ratio < 0)
824 return max_ratio;
825 return __bdi_set_max_ratio(bdi, max_ratio);
826}
827
828int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
829{
830 if (strict_limit > 1)
831 return -EINVAL;
832
833 spin_lock_bh(&bdi_lock);
834 if (strict_limit)
835 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
836 else
837 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
838 spin_unlock_bh(&bdi_lock);
839
840 return 0;
841}
842
843static unsigned long dirty_freerun_ceiling(unsigned long thresh,
844 unsigned long bg_thresh)
845{
846 return (thresh + bg_thresh) / 2;
847}
848
849static unsigned long hard_dirty_limit(struct wb_domain *dom,
850 unsigned long thresh)
851{
852 return max(thresh, dom->dirty_limit);
853}
854
855/*
856 * Memory which can be further allocated to a memcg domain is capped by
857 * system-wide clean memory excluding the amount being used in the domain.
858 */
859static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
860 unsigned long filepages, unsigned long headroom)
861{
862 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
863 unsigned long clean = filepages - min(filepages, mdtc->dirty);
864 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
865 unsigned long other_clean = global_clean - min(global_clean, clean);
866
867 mdtc->avail = filepages + min(headroom, other_clean);
868}
869
870static inline bool dtc_is_global(struct dirty_throttle_control *dtc)
871{
872 return mdtc_gdtc(dtc) == NULL;
873}
874
875/*
876 * Dirty background will ignore pages being written as we're trying to
877 * decide whether to put more under writeback.
878 */
879static void domain_dirty_avail(struct dirty_throttle_control *dtc,
880 bool include_writeback)
881{
882 if (dtc_is_global(dtc)) {
883 dtc->avail = global_dirtyable_memory();
884 dtc->dirty = global_node_page_state(NR_FILE_DIRTY);
885 if (include_writeback)
886 dtc->dirty += global_node_page_state(NR_WRITEBACK);
887 } else {
888 unsigned long filepages = 0, headroom = 0, writeback = 0;
889
890 mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty,
891 &writeback);
892 if (include_writeback)
893 dtc->dirty += writeback;
894 mdtc_calc_avail(dtc, filepages, headroom);
895 }
896}
897
898/**
899 * __wb_calc_thresh - @wb's share of dirty threshold
900 * @dtc: dirty_throttle_context of interest
901 * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
902 *
903 * Note that balance_dirty_pages() will only seriously take dirty throttling
904 * threshold as a hard limit when sleeping max_pause per page is not enough
905 * to keep the dirty pages under control. For example, when the device is
906 * completely stalled due to some error conditions, or when there are 1000
907 * dd tasks writing to a slow 10MB/s USB key.
908 * In the other normal situations, it acts more gently by throttling the tasks
909 * more (rather than completely block them) when the wb dirty pages go high.
910 *
911 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
912 * - starving fast devices
913 * - piling up dirty pages (that will take long time to sync) on slow devices
914 *
915 * The wb's share of dirty limit will be adapting to its throughput and
916 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
917 *
918 * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
919 * "dirty" in the context of dirty balancing includes all PG_dirty and
920 * PG_writeback pages.
921 */
922static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
923 unsigned long thresh)
924{
925 struct wb_domain *dom = dtc_dom(dtc);
926 struct bdi_writeback *wb = dtc->wb;
927 u64 wb_thresh;
928 u64 wb_max_thresh;
929 unsigned long numerator, denominator;
930 unsigned long wb_min_ratio, wb_max_ratio;
931
932 /*
933 * Calculate this wb's share of the thresh ratio.
934 */
935 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
936 &numerator, &denominator);
937
938 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
939 wb_thresh *= numerator;
940 wb_thresh = div64_ul(wb_thresh, denominator);
941
942 wb_min_max_ratio(wb, &wb_min_ratio, &wb_max_ratio);
943
944 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
945 wb_max_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
946 if (wb_thresh > wb_max_thresh)
947 wb_thresh = wb_max_thresh;
948
949 /*
950 * With strictlimit flag, the wb_thresh is treated as
951 * a hard limit in balance_dirty_pages() and wb_position_ratio().
952 * It's possible that wb_thresh is close to zero, not because
953 * the device is slow, but because it has been inactive.
954 * To prevent occasional writes from being blocked, we raise wb_thresh.
955 */
956 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
957 unsigned long limit = hard_dirty_limit(dom, dtc->thresh);
958 u64 wb_scale_thresh = 0;
959
960 if (limit > dtc->dirty)
961 wb_scale_thresh = (limit - dtc->dirty) / 100;
962 wb_thresh = max(wb_thresh, min(wb_scale_thresh, wb_max_thresh / 4));
963 }
964
965 return wb_thresh;
966}
967
968unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
969{
970 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
971
972 return __wb_calc_thresh(&gdtc, thresh);
973}
974
975unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
976{
977 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
978 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
979
980 domain_dirty_avail(&gdtc, true);
981 domain_dirty_avail(&mdtc, true);
982 domain_dirty_limits(&mdtc);
983
984 return __wb_calc_thresh(&mdtc, mdtc.thresh);
985}
986
987/*
988 * setpoint - dirty 3
989 * f(dirty) := 1.0 + (----------------)
990 * limit - setpoint
991 *
992 * it's a 3rd order polynomial that subjects to
993 *
994 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
995 * (2) f(setpoint) = 1.0 => the balance point
996 * (3) f(limit) = 0 => the hard limit
997 * (4) df/dx <= 0 => negative feedback control
998 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
999 * => fast response on large errors; small oscillation near setpoint
1000 */
1001static long long pos_ratio_polynom(unsigned long setpoint,
1002 unsigned long dirty,
1003 unsigned long limit)
1004{
1005 long long pos_ratio;
1006 long x;
1007
1008 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
1009 (limit - setpoint) | 1);
1010 pos_ratio = x;
1011 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1012 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1013 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
1014
1015 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
1016}
1017
1018/*
1019 * Dirty position control.
1020 *
1021 * (o) global/bdi setpoints
1022 *
1023 * We want the dirty pages be balanced around the global/wb setpoints.
1024 * When the number of dirty pages is higher/lower than the setpoint, the
1025 * dirty position control ratio (and hence task dirty ratelimit) will be
1026 * decreased/increased to bring the dirty pages back to the setpoint.
1027 *
1028 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
1029 *
1030 * if (dirty < setpoint) scale up pos_ratio
1031 * if (dirty > setpoint) scale down pos_ratio
1032 *
1033 * if (wb_dirty < wb_setpoint) scale up pos_ratio
1034 * if (wb_dirty > wb_setpoint) scale down pos_ratio
1035 *
1036 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
1037 *
1038 * (o) global control line
1039 *
1040 * ^ pos_ratio
1041 * |
1042 * | |<===== global dirty control scope ======>|
1043 * 2.0 * * * * * * *
1044 * | .*
1045 * | . *
1046 * | . *
1047 * | . *
1048 * | . *
1049 * | . *
1050 * 1.0 ................................*
1051 * | . . *
1052 * | . . *
1053 * | . . *
1054 * | . . *
1055 * | . . *
1056 * 0 +------------.------------------.----------------------*------------->
1057 * freerun^ setpoint^ limit^ dirty pages
1058 *
1059 * (o) wb control line
1060 *
1061 * ^ pos_ratio
1062 * |
1063 * | *
1064 * | *
1065 * | *
1066 * | *
1067 * | * |<=========== span ============>|
1068 * 1.0 .......................*
1069 * | . *
1070 * | . *
1071 * | . *
1072 * | . *
1073 * | . *
1074 * | . *
1075 * | . *
1076 * | . *
1077 * | . *
1078 * | . *
1079 * | . *
1080 * 1/4 ...............................................* * * * * * * * * * * *
1081 * | . .
1082 * | . .
1083 * | . .
1084 * 0 +----------------------.-------------------------------.------------->
1085 * wb_setpoint^ x_intercept^
1086 *
1087 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
1088 * be smoothly throttled down to normal if it starts high in situations like
1089 * - start writing to a slow SD card and a fast disk at the same time. The SD
1090 * card's wb_dirty may rush to many times higher than wb_setpoint.
1091 * - the wb dirty thresh drops quickly due to change of JBOD workload
1092 */
1093static void wb_position_ratio(struct dirty_throttle_control *dtc)
1094{
1095 struct bdi_writeback *wb = dtc->wb;
1096 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
1097 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1098 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1099 unsigned long wb_thresh = dtc->wb_thresh;
1100 unsigned long x_intercept;
1101 unsigned long setpoint; /* dirty pages' target balance point */
1102 unsigned long wb_setpoint;
1103 unsigned long span;
1104 long long pos_ratio; /* for scaling up/down the rate limit */
1105 long x;
1106
1107 dtc->pos_ratio = 0;
1108
1109 if (unlikely(dtc->dirty >= limit))
1110 return;
1111
1112 /*
1113 * global setpoint
1114 *
1115 * See comment for pos_ratio_polynom().
1116 */
1117 setpoint = (freerun + limit) / 2;
1118 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
1119
1120 /*
1121 * The strictlimit feature is a tool preventing mistrusted filesystems
1122 * from growing a large number of dirty pages before throttling. For
1123 * such filesystems balance_dirty_pages always checks wb counters
1124 * against wb limits. Even if global "nr_dirty" is under "freerun".
1125 * This is especially important for fuse which sets bdi->max_ratio to
1126 * 1% by default. Without strictlimit feature, fuse writeback may
1127 * consume arbitrary amount of RAM because it is accounted in
1128 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1129 *
1130 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1131 * two values: wb_dirty and wb_thresh. Let's consider an example:
1132 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1133 * limits are set by default to 10% and 20% (background and throttle).
1134 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1135 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1136 * about ~6K pages (as the average of background and throttle wb
1137 * limits). The 3rd order polynomial will provide positive feedback if
1138 * wb_dirty is under wb_setpoint and vice versa.
1139 *
1140 * Note, that we cannot use global counters in these calculations
1141 * because we want to throttle process writing to a strictlimit wb
1142 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1143 * in the example above).
1144 */
1145 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1146 long long wb_pos_ratio;
1147
1148 if (dtc->wb_dirty < 8) {
1149 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1150 2 << RATELIMIT_CALC_SHIFT);
1151 return;
1152 }
1153
1154 if (dtc->wb_dirty >= wb_thresh)
1155 return;
1156
1157 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1158 dtc->wb_bg_thresh);
1159
1160 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1161 return;
1162
1163 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1164 wb_thresh);
1165
1166 /*
1167 * Typically, for strictlimit case, wb_setpoint << setpoint
1168 * and pos_ratio >> wb_pos_ratio. In the other words global
1169 * state ("dirty") is not limiting factor and we have to
1170 * make decision based on wb counters. But there is an
1171 * important case when global pos_ratio should get precedence:
1172 * global limits are exceeded (e.g. due to activities on other
1173 * wb's) while given strictlimit wb is below limit.
1174 *
1175 * "pos_ratio * wb_pos_ratio" would work for the case above,
1176 * but it would look too non-natural for the case of all
1177 * activity in the system coming from a single strictlimit wb
1178 * with bdi->max_ratio == 100%.
1179 *
1180 * Note that min() below somewhat changes the dynamics of the
1181 * control system. Normally, pos_ratio value can be well over 3
1182 * (when globally we are at freerun and wb is well below wb
1183 * setpoint). Now the maximum pos_ratio in the same situation
1184 * is 2. We might want to tweak this if we observe the control
1185 * system is too slow to adapt.
1186 */
1187 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1188 return;
1189 }
1190
1191 /*
1192 * We have computed basic pos_ratio above based on global situation. If
1193 * the wb is over/under its share of dirty pages, we want to scale
1194 * pos_ratio further down/up. That is done by the following mechanism.
1195 */
1196
1197 /*
1198 * wb setpoint
1199 *
1200 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1201 *
1202 * x_intercept - wb_dirty
1203 * := --------------------------
1204 * x_intercept - wb_setpoint
1205 *
1206 * The main wb control line is a linear function that subjects to
1207 *
1208 * (1) f(wb_setpoint) = 1.0
1209 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1210 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1211 *
1212 * For single wb case, the dirty pages are observed to fluctuate
1213 * regularly within range
1214 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1215 * for various filesystems, where (2) can yield in a reasonable 12.5%
1216 * fluctuation range for pos_ratio.
1217 *
1218 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1219 * own size, so move the slope over accordingly and choose a slope that
1220 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1221 */
1222 if (unlikely(wb_thresh > dtc->thresh))
1223 wb_thresh = dtc->thresh;
1224 /*
1225 * It's very possible that wb_thresh is close to 0 not because the
1226 * device is slow, but that it has remained inactive for long time.
1227 * Honour such devices a reasonable good (hopefully IO efficient)
1228 * threshold, so that the occasional writes won't be blocked and active
1229 * writes can rampup the threshold quickly.
1230 */
1231 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1232 /*
1233 * scale global setpoint to wb's:
1234 * wb_setpoint = setpoint * wb_thresh / thresh
1235 */
1236 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1237 wb_setpoint = setpoint * (u64)x >> 16;
1238 /*
1239 * Use span=(8*write_bw) in single wb case as indicated by
1240 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1241 *
1242 * wb_thresh thresh - wb_thresh
1243 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1244 * thresh thresh
1245 */
1246 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1247 x_intercept = wb_setpoint + span;
1248
1249 if (dtc->wb_dirty < x_intercept - span / 4) {
1250 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1251 (x_intercept - wb_setpoint) | 1);
1252 } else
1253 pos_ratio /= 4;
1254
1255 /*
1256 * wb reserve area, safeguard against dirty pool underrun and disk idle
1257 * It may push the desired control point of global dirty pages higher
1258 * than setpoint.
1259 */
1260 x_intercept = wb_thresh / 2;
1261 if (dtc->wb_dirty < x_intercept) {
1262 if (dtc->wb_dirty > x_intercept / 8)
1263 pos_ratio = div_u64(pos_ratio * x_intercept,
1264 dtc->wb_dirty);
1265 else
1266 pos_ratio *= 8;
1267 }
1268
1269 dtc->pos_ratio = pos_ratio;
1270}
1271
1272static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1273 unsigned long elapsed,
1274 unsigned long written)
1275{
1276 const unsigned long period = roundup_pow_of_two(3 * HZ);
1277 unsigned long avg = wb->avg_write_bandwidth;
1278 unsigned long old = wb->write_bandwidth;
1279 u64 bw;
1280
1281 /*
1282 * bw = written * HZ / elapsed
1283 *
1284 * bw * elapsed + write_bandwidth * (period - elapsed)
1285 * write_bandwidth = ---------------------------------------------------
1286 * period
1287 *
1288 * @written may have decreased due to folio_redirty_for_writepage().
1289 * Avoid underflowing @bw calculation.
1290 */
1291 bw = written - min(written, wb->written_stamp);
1292 bw *= HZ;
1293 if (unlikely(elapsed > period)) {
1294 bw = div64_ul(bw, elapsed);
1295 avg = bw;
1296 goto out;
1297 }
1298 bw += (u64)wb->write_bandwidth * (period - elapsed);
1299 bw >>= ilog2(period);
1300
1301 /*
1302 * one more level of smoothing, for filtering out sudden spikes
1303 */
1304 if (avg > old && old >= (unsigned long)bw)
1305 avg -= (avg - old) >> 3;
1306
1307 if (avg < old && old <= (unsigned long)bw)
1308 avg += (old - avg) >> 3;
1309
1310out:
1311 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1312 avg = max(avg, 1LU);
1313 if (wb_has_dirty_io(wb)) {
1314 long delta = avg - wb->avg_write_bandwidth;
1315 WARN_ON_ONCE(atomic_long_add_return(delta,
1316 &wb->bdi->tot_write_bandwidth) <= 0);
1317 }
1318 wb->write_bandwidth = bw;
1319 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1320}
1321
1322static void update_dirty_limit(struct dirty_throttle_control *dtc)
1323{
1324 struct wb_domain *dom = dtc_dom(dtc);
1325 unsigned long thresh = dtc->thresh;
1326 unsigned long limit = dom->dirty_limit;
1327
1328 /*
1329 * Follow up in one step.
1330 */
1331 if (limit < thresh) {
1332 limit = thresh;
1333 goto update;
1334 }
1335
1336 /*
1337 * Follow down slowly. Use the higher one as the target, because thresh
1338 * may drop below dirty. This is exactly the reason to introduce
1339 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1340 */
1341 thresh = max(thresh, dtc->dirty);
1342 if (limit > thresh) {
1343 limit -= (limit - thresh) >> 5;
1344 goto update;
1345 }
1346 return;
1347update:
1348 dom->dirty_limit = limit;
1349}
1350
1351static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1352 unsigned long now)
1353{
1354 struct wb_domain *dom = dtc_dom(dtc);
1355
1356 /*
1357 * check locklessly first to optimize away locking for the most time
1358 */
1359 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1360 return;
1361
1362 spin_lock(&dom->lock);
1363 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1364 update_dirty_limit(dtc);
1365 dom->dirty_limit_tstamp = now;
1366 }
1367 spin_unlock(&dom->lock);
1368}
1369
1370/*
1371 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1372 *
1373 * Normal wb tasks will be curbed at or below it in long term.
1374 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1375 */
1376static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1377 unsigned long dirtied,
1378 unsigned long elapsed)
1379{
1380 struct bdi_writeback *wb = dtc->wb;
1381 unsigned long dirty = dtc->dirty;
1382 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1383 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1384 unsigned long setpoint = (freerun + limit) / 2;
1385 unsigned long write_bw = wb->avg_write_bandwidth;
1386 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1387 unsigned long dirty_rate;
1388 unsigned long task_ratelimit;
1389 unsigned long balanced_dirty_ratelimit;
1390 unsigned long step;
1391 unsigned long x;
1392 unsigned long shift;
1393
1394 /*
1395 * The dirty rate will match the writeout rate in long term, except
1396 * when dirty pages are truncated by userspace or re-dirtied by FS.
1397 */
1398 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1399
1400 /*
1401 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1402 */
1403 task_ratelimit = (u64)dirty_ratelimit *
1404 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1405 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1406
1407 /*
1408 * A linear estimation of the "balanced" throttle rate. The theory is,
1409 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1410 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1411 * formula will yield the balanced rate limit (write_bw / N).
1412 *
1413 * Note that the expanded form is not a pure rate feedback:
1414 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1415 * but also takes pos_ratio into account:
1416 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1417 *
1418 * (1) is not realistic because pos_ratio also takes part in balancing
1419 * the dirty rate. Consider the state
1420 * pos_ratio = 0.5 (3)
1421 * rate = 2 * (write_bw / N) (4)
1422 * If (1) is used, it will stuck in that state! Because each dd will
1423 * be throttled at
1424 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1425 * yielding
1426 * dirty_rate = N * task_ratelimit = write_bw (6)
1427 * put (6) into (1) we get
1428 * rate_(i+1) = rate_(i) (7)
1429 *
1430 * So we end up using (2) to always keep
1431 * rate_(i+1) ~= (write_bw / N) (8)
1432 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1433 * pos_ratio is able to drive itself to 1.0, which is not only where
1434 * the dirty count meet the setpoint, but also where the slope of
1435 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1436 */
1437 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1438 dirty_rate | 1);
1439 /*
1440 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1441 */
1442 if (unlikely(balanced_dirty_ratelimit > write_bw))
1443 balanced_dirty_ratelimit = write_bw;
1444
1445 /*
1446 * We could safely do this and return immediately:
1447 *
1448 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1449 *
1450 * However to get a more stable dirty_ratelimit, the below elaborated
1451 * code makes use of task_ratelimit to filter out singular points and
1452 * limit the step size.
1453 *
1454 * The below code essentially only uses the relative value of
1455 *
1456 * task_ratelimit - dirty_ratelimit
1457 * = (pos_ratio - 1) * dirty_ratelimit
1458 *
1459 * which reflects the direction and size of dirty position error.
1460 */
1461
1462 /*
1463 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1464 * task_ratelimit is on the same side of dirty_ratelimit, too.
1465 * For example, when
1466 * - dirty_ratelimit > balanced_dirty_ratelimit
1467 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1468 * lowering dirty_ratelimit will help meet both the position and rate
1469 * control targets. Otherwise, don't update dirty_ratelimit if it will
1470 * only help meet the rate target. After all, what the users ultimately
1471 * feel and care are stable dirty rate and small position error.
1472 *
1473 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1474 * and filter out the singular points of balanced_dirty_ratelimit. Which
1475 * keeps jumping around randomly and can even leap far away at times
1476 * due to the small 200ms estimation period of dirty_rate (we want to
1477 * keep that period small to reduce time lags).
1478 */
1479 step = 0;
1480
1481 /*
1482 * For strictlimit case, calculations above were based on wb counters
1483 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1484 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1485 * Hence, to calculate "step" properly, we have to use wb_dirty as
1486 * "dirty" and wb_setpoint as "setpoint".
1487 *
1488 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1489 * it's possible that wb_thresh is close to zero due to inactivity
1490 * of backing device.
1491 */
1492 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1493 dirty = dtc->wb_dirty;
1494 if (dtc->wb_dirty < 8)
1495 setpoint = dtc->wb_dirty + 1;
1496 else
1497 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1498 }
1499
1500 if (dirty < setpoint) {
1501 x = min3(wb->balanced_dirty_ratelimit,
1502 balanced_dirty_ratelimit, task_ratelimit);
1503 if (dirty_ratelimit < x)
1504 step = x - dirty_ratelimit;
1505 } else {
1506 x = max3(wb->balanced_dirty_ratelimit,
1507 balanced_dirty_ratelimit, task_ratelimit);
1508 if (dirty_ratelimit > x)
1509 step = dirty_ratelimit - x;
1510 }
1511
1512 /*
1513 * Don't pursue 100% rate matching. It's impossible since the balanced
1514 * rate itself is constantly fluctuating. So decrease the track speed
1515 * when it gets close to the target. Helps eliminate pointless tremors.
1516 */
1517 shift = dirty_ratelimit / (2 * step + 1);
1518 if (shift < BITS_PER_LONG)
1519 step = DIV_ROUND_UP(step >> shift, 8);
1520 else
1521 step = 0;
1522
1523 if (dirty_ratelimit < balanced_dirty_ratelimit)
1524 dirty_ratelimit += step;
1525 else
1526 dirty_ratelimit -= step;
1527
1528 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1529 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1530
1531 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1532}
1533
1534static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1535 struct dirty_throttle_control *mdtc,
1536 bool update_ratelimit)
1537{
1538 struct bdi_writeback *wb = gdtc->wb;
1539 unsigned long now = jiffies;
1540 unsigned long elapsed;
1541 unsigned long dirtied;
1542 unsigned long written;
1543
1544 spin_lock(&wb->list_lock);
1545
1546 /*
1547 * Lockless checks for elapsed time are racy and delayed update after
1548 * IO completion doesn't do it at all (to make sure written pages are
1549 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1550 * division errors.
1551 */
1552 elapsed = max(now - wb->bw_time_stamp, 1UL);
1553 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1554 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1555
1556 if (update_ratelimit) {
1557 domain_update_dirty_limit(gdtc, now);
1558 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1559
1560 /*
1561 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1562 * compiler has no way to figure that out. Help it.
1563 */
1564 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1565 domain_update_dirty_limit(mdtc, now);
1566 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1567 }
1568 }
1569 wb_update_write_bandwidth(wb, elapsed, written);
1570
1571 wb->dirtied_stamp = dirtied;
1572 wb->written_stamp = written;
1573 WRITE_ONCE(wb->bw_time_stamp, now);
1574 spin_unlock(&wb->list_lock);
1575}
1576
1577void wb_update_bandwidth(struct bdi_writeback *wb)
1578{
1579 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1580
1581 __wb_update_bandwidth(&gdtc, NULL, false);
1582}
1583
1584/* Interval after which we consider wb idle and don't estimate bandwidth */
1585#define WB_BANDWIDTH_IDLE_JIF (HZ)
1586
1587static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1588{
1589 unsigned long now = jiffies;
1590 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1591
1592 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1593 !atomic_read(&wb->writeback_inodes)) {
1594 spin_lock(&wb->list_lock);
1595 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1596 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1597 WRITE_ONCE(wb->bw_time_stamp, now);
1598 spin_unlock(&wb->list_lock);
1599 }
1600}
1601
1602/*
1603 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1604 * will look to see if it needs to start dirty throttling.
1605 *
1606 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1607 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1608 * (the number of pages we may dirty without exceeding the dirty limits).
1609 */
1610static unsigned long dirty_poll_interval(unsigned long dirty,
1611 unsigned long thresh)
1612{
1613 if (thresh > dirty)
1614 return 1UL << (ilog2(thresh - dirty) >> 1);
1615
1616 return 1;
1617}
1618
1619static unsigned long wb_max_pause(struct bdi_writeback *wb,
1620 unsigned long wb_dirty)
1621{
1622 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1623 unsigned long t;
1624
1625 /*
1626 * Limit pause time for small memory systems. If sleeping for too long
1627 * time, a small pool of dirty/writeback pages may go empty and disk go
1628 * idle.
1629 *
1630 * 8 serves as the safety ratio.
1631 */
1632 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1633 t++;
1634
1635 return min_t(unsigned long, t, MAX_PAUSE);
1636}
1637
1638static long wb_min_pause(struct bdi_writeback *wb,
1639 long max_pause,
1640 unsigned long task_ratelimit,
1641 unsigned long dirty_ratelimit,
1642 int *nr_dirtied_pause)
1643{
1644 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1645 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1646 long t; /* target pause */
1647 long pause; /* estimated next pause */
1648 int pages; /* target nr_dirtied_pause */
1649
1650 /* target for 10ms pause on 1-dd case */
1651 t = max(1, HZ / 100);
1652
1653 /*
1654 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1655 * overheads.
1656 *
1657 * (N * 10ms) on 2^N concurrent tasks.
1658 */
1659 if (hi > lo)
1660 t += (hi - lo) * (10 * HZ) / 1024;
1661
1662 /*
1663 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1664 * on the much more stable dirty_ratelimit. However the next pause time
1665 * will be computed based on task_ratelimit and the two rate limits may
1666 * depart considerably at some time. Especially if task_ratelimit goes
1667 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1668 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1669 * result task_ratelimit won't be executed faithfully, which could
1670 * eventually bring down dirty_ratelimit.
1671 *
1672 * We apply two rules to fix it up:
1673 * 1) try to estimate the next pause time and if necessary, use a lower
1674 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1675 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1676 * 2) limit the target pause time to max_pause/2, so that the normal
1677 * small fluctuations of task_ratelimit won't trigger rule (1) and
1678 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1679 */
1680 t = min(t, 1 + max_pause / 2);
1681 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1682
1683 /*
1684 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1685 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1686 * When the 16 consecutive reads are often interrupted by some dirty
1687 * throttling pause during the async writes, cfq will go into idles
1688 * (deadline is fine). So push nr_dirtied_pause as high as possible
1689 * until reaches DIRTY_POLL_THRESH=32 pages.
1690 */
1691 if (pages < DIRTY_POLL_THRESH) {
1692 t = max_pause;
1693 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1694 if (pages > DIRTY_POLL_THRESH) {
1695 pages = DIRTY_POLL_THRESH;
1696 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1697 }
1698 }
1699
1700 pause = HZ * pages / (task_ratelimit + 1);
1701 if (pause > max_pause) {
1702 t = max_pause;
1703 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1704 }
1705
1706 *nr_dirtied_pause = pages;
1707 /*
1708 * The minimal pause time will normally be half the target pause time.
1709 */
1710 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1711}
1712
1713static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1714{
1715 struct bdi_writeback *wb = dtc->wb;
1716 unsigned long wb_reclaimable;
1717
1718 /*
1719 * wb_thresh is not treated as some limiting factor as
1720 * dirty_thresh, due to reasons
1721 * - in JBOD setup, wb_thresh can fluctuate a lot
1722 * - in a system with HDD and USB key, the USB key may somehow
1723 * go into state (wb_dirty >> wb_thresh) either because
1724 * wb_dirty starts high, or because wb_thresh drops low.
1725 * In this case we don't want to hard throttle the USB key
1726 * dirtiers for 100 seconds until wb_dirty drops under
1727 * wb_thresh. Instead the auxiliary wb control line in
1728 * wb_position_ratio() will let the dirtier task progress
1729 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1730 */
1731 dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
1732 dtc->wb_bg_thresh = dtc->thresh ?
1733 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1734
1735 /*
1736 * In order to avoid the stacked BDI deadlock we need
1737 * to ensure we accurately count the 'dirty' pages when
1738 * the threshold is low.
1739 *
1740 * Otherwise it would be possible to get thresh+n pages
1741 * reported dirty, even though there are thresh-m pages
1742 * actually dirty; with m+n sitting in the percpu
1743 * deltas.
1744 */
1745 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1746 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1747 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1748 } else {
1749 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1750 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1751 }
1752}
1753
1754static unsigned long domain_poll_intv(struct dirty_throttle_control *dtc,
1755 bool strictlimit)
1756{
1757 unsigned long dirty, thresh;
1758
1759 if (strictlimit) {
1760 dirty = dtc->wb_dirty;
1761 thresh = dtc->wb_thresh;
1762 } else {
1763 dirty = dtc->dirty;
1764 thresh = dtc->thresh;
1765 }
1766
1767 return dirty_poll_interval(dirty, thresh);
1768}
1769
1770/*
1771 * Throttle it only when the background writeback cannot catch-up. This avoids
1772 * (excessively) small writeouts when the wb limits are ramping up in case of
1773 * !strictlimit.
1774 *
1775 * In strictlimit case make decision based on the wb counters and limits. Small
1776 * writeouts when the wb limits are ramping up are the price we consciously pay
1777 * for strictlimit-ing.
1778 */
1779static void domain_dirty_freerun(struct dirty_throttle_control *dtc,
1780 bool strictlimit)
1781{
1782 unsigned long dirty, thresh, bg_thresh;
1783
1784 if (unlikely(strictlimit)) {
1785 wb_dirty_limits(dtc);
1786 dirty = dtc->wb_dirty;
1787 thresh = dtc->wb_thresh;
1788 bg_thresh = dtc->wb_bg_thresh;
1789 } else {
1790 dirty = dtc->dirty;
1791 thresh = dtc->thresh;
1792 bg_thresh = dtc->bg_thresh;
1793 }
1794 dtc->freerun = dirty <= dirty_freerun_ceiling(thresh, bg_thresh);
1795}
1796
1797static void balance_domain_limits(struct dirty_throttle_control *dtc,
1798 bool strictlimit)
1799{
1800 domain_dirty_avail(dtc, true);
1801 domain_dirty_limits(dtc);
1802 domain_dirty_freerun(dtc, strictlimit);
1803}
1804
1805static void wb_dirty_freerun(struct dirty_throttle_control *dtc,
1806 bool strictlimit)
1807{
1808 dtc->freerun = false;
1809
1810 /* was already handled in domain_dirty_freerun */
1811 if (strictlimit)
1812 return;
1813
1814 wb_dirty_limits(dtc);
1815 /*
1816 * LOCAL_THROTTLE tasks must not be throttled when below the per-wb
1817 * freerun ceiling.
1818 */
1819 if (!(current->flags & PF_LOCAL_THROTTLE))
1820 return;
1821
1822 dtc->freerun = dtc->wb_dirty <
1823 dirty_freerun_ceiling(dtc->wb_thresh, dtc->wb_bg_thresh);
1824}
1825
1826static inline void wb_dirty_exceeded(struct dirty_throttle_control *dtc,
1827 bool strictlimit)
1828{
1829 dtc->dirty_exceeded = (dtc->wb_dirty > dtc->wb_thresh) &&
1830 ((dtc->dirty > dtc->thresh) || strictlimit);
1831}
1832
1833/*
1834 * The limits fields dirty_exceeded and pos_ratio won't be updated if wb is
1835 * in freerun state. Please don't use these invalid fields in freerun case.
1836 */
1837static void balance_wb_limits(struct dirty_throttle_control *dtc,
1838 bool strictlimit)
1839{
1840 wb_dirty_freerun(dtc, strictlimit);
1841 if (dtc->freerun)
1842 return;
1843
1844 wb_dirty_exceeded(dtc, strictlimit);
1845 wb_position_ratio(dtc);
1846}
1847
1848/*
1849 * balance_dirty_pages() must be called by processes which are generating dirty
1850 * data. It looks at the number of dirty pages in the machine and will force
1851 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1852 * If we're over `background_thresh' then the writeback threads are woken to
1853 * perform some writeout.
1854 */
1855static int balance_dirty_pages(struct bdi_writeback *wb,
1856 unsigned long pages_dirtied, unsigned int flags)
1857{
1858 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1859 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1860 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1861 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1862 &mdtc_stor : NULL;
1863 struct dirty_throttle_control *sdtc;
1864 unsigned long nr_dirty;
1865 long period;
1866 long pause;
1867 long max_pause;
1868 long min_pause;
1869 int nr_dirtied_pause;
1870 unsigned long task_ratelimit;
1871 unsigned long dirty_ratelimit;
1872 struct backing_dev_info *bdi = wb->bdi;
1873 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1874 unsigned long start_time = jiffies;
1875 int ret = 0;
1876
1877 for (;;) {
1878 unsigned long now = jiffies;
1879
1880 nr_dirty = global_node_page_state(NR_FILE_DIRTY);
1881
1882 balance_domain_limits(gdtc, strictlimit);
1883 if (mdtc) {
1884 /*
1885 * If @wb belongs to !root memcg, repeat the same
1886 * basic calculations for the memcg domain.
1887 */
1888 balance_domain_limits(mdtc, strictlimit);
1889 }
1890
1891 /*
1892 * In laptop mode, we wait until hitting the higher threshold
1893 * before starting background writeout, and then write out all
1894 * the way down to the lower threshold. So slow writers cause
1895 * minimal disk activity.
1896 *
1897 * In normal mode, we start background writeout at the lower
1898 * background_thresh, to keep the amount of dirty memory low.
1899 */
1900 if (!laptop_mode && nr_dirty > gdtc->bg_thresh &&
1901 !writeback_in_progress(wb))
1902 wb_start_background_writeback(wb);
1903
1904 /*
1905 * If memcg domain is in effect, @dirty should be under
1906 * both global and memcg freerun ceilings.
1907 */
1908 if (gdtc->freerun && (!mdtc || mdtc->freerun)) {
1909 unsigned long intv;
1910 unsigned long m_intv;
1911
1912free_running:
1913 intv = domain_poll_intv(gdtc, strictlimit);
1914 m_intv = ULONG_MAX;
1915
1916 current->dirty_paused_when = now;
1917 current->nr_dirtied = 0;
1918 if (mdtc)
1919 m_intv = domain_poll_intv(mdtc, strictlimit);
1920 current->nr_dirtied_pause = min(intv, m_intv);
1921 break;
1922 }
1923
1924 /* Start writeback even when in laptop mode */
1925 if (unlikely(!writeback_in_progress(wb)))
1926 wb_start_background_writeback(wb);
1927
1928 mem_cgroup_flush_foreign(wb);
1929
1930 /*
1931 * Calculate global domain's pos_ratio and select the
1932 * global dtc by default.
1933 */
1934 balance_wb_limits(gdtc, strictlimit);
1935 if (gdtc->freerun)
1936 goto free_running;
1937 sdtc = gdtc;
1938
1939 if (mdtc) {
1940 /*
1941 * If memcg domain is in effect, calculate its
1942 * pos_ratio. @wb should satisfy constraints from
1943 * both global and memcg domains. Choose the one
1944 * w/ lower pos_ratio.
1945 */
1946 balance_wb_limits(mdtc, strictlimit);
1947 if (mdtc->freerun)
1948 goto free_running;
1949 if (mdtc->pos_ratio < gdtc->pos_ratio)
1950 sdtc = mdtc;
1951 }
1952
1953 wb->dirty_exceeded = gdtc->dirty_exceeded ||
1954 (mdtc && mdtc->dirty_exceeded);
1955 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1956 BANDWIDTH_INTERVAL))
1957 __wb_update_bandwidth(gdtc, mdtc, true);
1958
1959 /* throttle according to the chosen dtc */
1960 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1961 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1962 RATELIMIT_CALC_SHIFT;
1963 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1964 min_pause = wb_min_pause(wb, max_pause,
1965 task_ratelimit, dirty_ratelimit,
1966 &nr_dirtied_pause);
1967
1968 if (unlikely(task_ratelimit == 0)) {
1969 period = max_pause;
1970 pause = max_pause;
1971 goto pause;
1972 }
1973 period = HZ * pages_dirtied / task_ratelimit;
1974 pause = period;
1975 if (current->dirty_paused_when)
1976 pause -= now - current->dirty_paused_when;
1977 /*
1978 * For less than 1s think time (ext3/4 may block the dirtier
1979 * for up to 800ms from time to time on 1-HDD; so does xfs,
1980 * however at much less frequency), try to compensate it in
1981 * future periods by updating the virtual time; otherwise just
1982 * do a reset, as it may be a light dirtier.
1983 */
1984 if (pause < min_pause) {
1985 trace_balance_dirty_pages(wb,
1986 sdtc->thresh,
1987 sdtc->bg_thresh,
1988 sdtc->dirty,
1989 sdtc->wb_thresh,
1990 sdtc->wb_dirty,
1991 dirty_ratelimit,
1992 task_ratelimit,
1993 pages_dirtied,
1994 period,
1995 min(pause, 0L),
1996 start_time);
1997 if (pause < -HZ) {
1998 current->dirty_paused_when = now;
1999 current->nr_dirtied = 0;
2000 } else if (period) {
2001 current->dirty_paused_when += period;
2002 current->nr_dirtied = 0;
2003 } else if (current->nr_dirtied_pause <= pages_dirtied)
2004 current->nr_dirtied_pause += pages_dirtied;
2005 break;
2006 }
2007 if (unlikely(pause > max_pause)) {
2008 /* for occasional dropped task_ratelimit */
2009 now += min(pause - max_pause, max_pause);
2010 pause = max_pause;
2011 }
2012
2013pause:
2014 trace_balance_dirty_pages(wb,
2015 sdtc->thresh,
2016 sdtc->bg_thresh,
2017 sdtc->dirty,
2018 sdtc->wb_thresh,
2019 sdtc->wb_dirty,
2020 dirty_ratelimit,
2021 task_ratelimit,
2022 pages_dirtied,
2023 period,
2024 pause,
2025 start_time);
2026 if (flags & BDP_ASYNC) {
2027 ret = -EAGAIN;
2028 break;
2029 }
2030 __set_current_state(TASK_KILLABLE);
2031 bdi->last_bdp_sleep = jiffies;
2032 io_schedule_timeout(pause);
2033
2034 current->dirty_paused_when = now + pause;
2035 current->nr_dirtied = 0;
2036 current->nr_dirtied_pause = nr_dirtied_pause;
2037
2038 /*
2039 * This is typically equal to (dirty < thresh) and can also
2040 * keep "1000+ dd on a slow USB stick" under control.
2041 */
2042 if (task_ratelimit)
2043 break;
2044
2045 /*
2046 * In the case of an unresponsive NFS server and the NFS dirty
2047 * pages exceeds dirty_thresh, give the other good wb's a pipe
2048 * to go through, so that tasks on them still remain responsive.
2049 *
2050 * In theory 1 page is enough to keep the consumer-producer
2051 * pipe going: the flusher cleans 1 page => the task dirties 1
2052 * more page. However wb_dirty has accounting errors. So use
2053 * the larger and more IO friendly wb_stat_error.
2054 */
2055 if (sdtc->wb_dirty <= wb_stat_error())
2056 break;
2057
2058 if (fatal_signal_pending(current))
2059 break;
2060 }
2061 return ret;
2062}
2063
2064static DEFINE_PER_CPU(int, bdp_ratelimits);
2065
2066/*
2067 * Normal tasks are throttled by
2068 * loop {
2069 * dirty tsk->nr_dirtied_pause pages;
2070 * take a snap in balance_dirty_pages();
2071 * }
2072 * However there is a worst case. If every task exit immediately when dirtied
2073 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
2074 * called to throttle the page dirties. The solution is to save the not yet
2075 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
2076 * randomly into the running tasks. This works well for the above worst case,
2077 * as the new task will pick up and accumulate the old task's leaked dirty
2078 * count and eventually get throttled.
2079 */
2080DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
2081
2082/**
2083 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
2084 * @mapping: address_space which was dirtied.
2085 * @flags: BDP flags.
2086 *
2087 * Processes which are dirtying memory should call in here once for each page
2088 * which was newly dirtied. The function will periodically check the system's
2089 * dirty state and will initiate writeback if needed.
2090 *
2091 * See balance_dirty_pages_ratelimited() for details.
2092 *
2093 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
2094 * indicate that memory is out of balance and the caller must wait
2095 * for I/O to complete. Otherwise, it will return 0 to indicate
2096 * that either memory was already in balance, or it was able to sleep
2097 * until the amount of dirty memory returned to balance.
2098 */
2099int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
2100 unsigned int flags)
2101{
2102 struct inode *inode = mapping->host;
2103 struct backing_dev_info *bdi = inode_to_bdi(inode);
2104 struct bdi_writeback *wb = NULL;
2105 int ratelimit;
2106 int ret = 0;
2107 int *p;
2108
2109 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
2110 return ret;
2111
2112 if (inode_cgwb_enabled(inode))
2113 wb = wb_get_create_current(bdi, GFP_KERNEL);
2114 if (!wb)
2115 wb = &bdi->wb;
2116
2117 ratelimit = current->nr_dirtied_pause;
2118 if (wb->dirty_exceeded)
2119 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
2120
2121 preempt_disable();
2122 /*
2123 * This prevents one CPU to accumulate too many dirtied pages without
2124 * calling into balance_dirty_pages(), which can happen when there are
2125 * 1000+ tasks, all of them start dirtying pages at exactly the same
2126 * time, hence all honoured too large initial task->nr_dirtied_pause.
2127 */
2128 p = this_cpu_ptr(&bdp_ratelimits);
2129 if (unlikely(current->nr_dirtied >= ratelimit))
2130 *p = 0;
2131 else if (unlikely(*p >= ratelimit_pages)) {
2132 *p = 0;
2133 ratelimit = 0;
2134 }
2135 /*
2136 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2137 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2138 * the dirty throttling and livelock other long-run dirtiers.
2139 */
2140 p = this_cpu_ptr(&dirty_throttle_leaks);
2141 if (*p > 0 && current->nr_dirtied < ratelimit) {
2142 unsigned long nr_pages_dirtied;
2143 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2144 *p -= nr_pages_dirtied;
2145 current->nr_dirtied += nr_pages_dirtied;
2146 }
2147 preempt_enable();
2148
2149 if (unlikely(current->nr_dirtied >= ratelimit))
2150 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2151
2152 wb_put(wb);
2153 return ret;
2154}
2155EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2156
2157/**
2158 * balance_dirty_pages_ratelimited - balance dirty memory state.
2159 * @mapping: address_space which was dirtied.
2160 *
2161 * Processes which are dirtying memory should call in here once for each page
2162 * which was newly dirtied. The function will periodically check the system's
2163 * dirty state and will initiate writeback if needed.
2164 *
2165 * Once we're over the dirty memory limit we decrease the ratelimiting
2166 * by a lot, to prevent individual processes from overshooting the limit
2167 * by (ratelimit_pages) each.
2168 */
2169void balance_dirty_pages_ratelimited(struct address_space *mapping)
2170{
2171 balance_dirty_pages_ratelimited_flags(mapping, 0);
2172}
2173EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
2174
2175/*
2176 * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty
2177 * and thresh, but it's for background writeback.
2178 */
2179static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc)
2180{
2181 struct bdi_writeback *wb = dtc->wb;
2182
2183 dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh);
2184 if (dtc->wb_bg_thresh < 2 * wb_stat_error())
2185 dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE);
2186 else
2187 dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE);
2188}
2189
2190static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc)
2191{
2192 domain_dirty_avail(dtc, false);
2193 domain_dirty_limits(dtc);
2194 if (dtc->dirty > dtc->bg_thresh)
2195 return true;
2196
2197 wb_bg_dirty_limits(dtc);
2198 if (dtc->wb_dirty > dtc->wb_bg_thresh)
2199 return true;
2200
2201 return false;
2202}
2203
2204/**
2205 * wb_over_bg_thresh - does @wb need to be written back?
2206 * @wb: bdi_writeback of interest
2207 *
2208 * Determines whether background writeback should keep writing @wb or it's
2209 * clean enough.
2210 *
2211 * Return: %true if writeback should continue.
2212 */
2213bool wb_over_bg_thresh(struct bdi_writeback *wb)
2214{
2215 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
2216 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
2217
2218 if (domain_over_bg_thresh(&gdtc))
2219 return true;
2220
2221 if (mdtc_valid(&mdtc))
2222 return domain_over_bg_thresh(&mdtc);
2223
2224 return false;
2225}
2226
2227#ifdef CONFIG_SYSCTL
2228/*
2229 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2230 */
2231static int dirty_writeback_centisecs_handler(const struct ctl_table *table, int write,
2232 void *buffer, size_t *length, loff_t *ppos)
2233{
2234 unsigned int old_interval = dirty_writeback_interval;
2235 int ret;
2236
2237 ret = proc_dointvec(table, write, buffer, length, ppos);
2238
2239 /*
2240 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2241 * and a different non-zero value will wakeup the writeback threads.
2242 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2243 * iterate over all bdis and wbs.
2244 * The reason we do this is to make the change take effect immediately.
2245 */
2246 if (!ret && write && dirty_writeback_interval &&
2247 dirty_writeback_interval != old_interval)
2248 wakeup_flusher_threads(WB_REASON_PERIODIC);
2249
2250 return ret;
2251}
2252#endif
2253
2254void laptop_mode_timer_fn(struct timer_list *t)
2255{
2256 struct backing_dev_info *backing_dev_info =
2257 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2258
2259 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2260}
2261
2262/*
2263 * We've spun up the disk and we're in laptop mode: schedule writeback
2264 * of all dirty data a few seconds from now. If the flush is already scheduled
2265 * then push it back - the user is still using the disk.
2266 */
2267void laptop_io_completion(struct backing_dev_info *info)
2268{
2269 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2270}
2271
2272/*
2273 * We're in laptop mode and we've just synced. The sync's writes will have
2274 * caused another writeback to be scheduled by laptop_io_completion.
2275 * Nothing needs to be written back anymore, so we unschedule the writeback.
2276 */
2277void laptop_sync_completion(void)
2278{
2279 struct backing_dev_info *bdi;
2280
2281 rcu_read_lock();
2282
2283 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2284 del_timer(&bdi->laptop_mode_wb_timer);
2285
2286 rcu_read_unlock();
2287}
2288
2289/*
2290 * If ratelimit_pages is too high then we can get into dirty-data overload
2291 * if a large number of processes all perform writes at the same time.
2292 *
2293 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2294 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2295 * thresholds.
2296 */
2297
2298void writeback_set_ratelimit(void)
2299{
2300 struct wb_domain *dom = &global_wb_domain;
2301 unsigned long background_thresh;
2302 unsigned long dirty_thresh;
2303
2304 global_dirty_limits(&background_thresh, &dirty_thresh);
2305 dom->dirty_limit = dirty_thresh;
2306 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2307 if (ratelimit_pages < 16)
2308 ratelimit_pages = 16;
2309}
2310
2311static int page_writeback_cpu_online(unsigned int cpu)
2312{
2313 writeback_set_ratelimit();
2314 return 0;
2315}
2316
2317#ifdef CONFIG_SYSCTL
2318
2319/* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2320static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2321
2322static struct ctl_table vm_page_writeback_sysctls[] = {
2323 {
2324 .procname = "dirty_background_ratio",
2325 .data = &dirty_background_ratio,
2326 .maxlen = sizeof(dirty_background_ratio),
2327 .mode = 0644,
2328 .proc_handler = dirty_background_ratio_handler,
2329 .extra1 = SYSCTL_ZERO,
2330 .extra2 = SYSCTL_ONE_HUNDRED,
2331 },
2332 {
2333 .procname = "dirty_background_bytes",
2334 .data = &dirty_background_bytes,
2335 .maxlen = sizeof(dirty_background_bytes),
2336 .mode = 0644,
2337 .proc_handler = dirty_background_bytes_handler,
2338 .extra1 = SYSCTL_LONG_ONE,
2339 },
2340 {
2341 .procname = "dirty_ratio",
2342 .data = &vm_dirty_ratio,
2343 .maxlen = sizeof(vm_dirty_ratio),
2344 .mode = 0644,
2345 .proc_handler = dirty_ratio_handler,
2346 .extra1 = SYSCTL_ZERO,
2347 .extra2 = SYSCTL_ONE_HUNDRED,
2348 },
2349 {
2350 .procname = "dirty_bytes",
2351 .data = &vm_dirty_bytes,
2352 .maxlen = sizeof(vm_dirty_bytes),
2353 .mode = 0644,
2354 .proc_handler = dirty_bytes_handler,
2355 .extra1 = (void *)&dirty_bytes_min,
2356 },
2357 {
2358 .procname = "dirty_writeback_centisecs",
2359 .data = &dirty_writeback_interval,
2360 .maxlen = sizeof(dirty_writeback_interval),
2361 .mode = 0644,
2362 .proc_handler = dirty_writeback_centisecs_handler,
2363 },
2364 {
2365 .procname = "dirty_expire_centisecs",
2366 .data = &dirty_expire_interval,
2367 .maxlen = sizeof(dirty_expire_interval),
2368 .mode = 0644,
2369 .proc_handler = proc_dointvec_minmax,
2370 .extra1 = SYSCTL_ZERO,
2371 },
2372#ifdef CONFIG_HIGHMEM
2373 {
2374 .procname = "highmem_is_dirtyable",
2375 .data = &vm_highmem_is_dirtyable,
2376 .maxlen = sizeof(vm_highmem_is_dirtyable),
2377 .mode = 0644,
2378 .proc_handler = proc_dointvec_minmax,
2379 .extra1 = SYSCTL_ZERO,
2380 .extra2 = SYSCTL_ONE,
2381 },
2382#endif
2383 {
2384 .procname = "laptop_mode",
2385 .data = &laptop_mode,
2386 .maxlen = sizeof(laptop_mode),
2387 .mode = 0644,
2388 .proc_handler = proc_dointvec_jiffies,
2389 },
2390};
2391#endif
2392
2393/*
2394 * Called early on to tune the page writeback dirty limits.
2395 *
2396 * We used to scale dirty pages according to how total memory
2397 * related to pages that could be allocated for buffers.
2398 *
2399 * However, that was when we used "dirty_ratio" to scale with
2400 * all memory, and we don't do that any more. "dirty_ratio"
2401 * is now applied to total non-HIGHPAGE memory, and as such we can't
2402 * get into the old insane situation any more where we had
2403 * large amounts of dirty pages compared to a small amount of
2404 * non-HIGHMEM memory.
2405 *
2406 * But we might still want to scale the dirty_ratio by how
2407 * much memory the box has..
2408 */
2409void __init page_writeback_init(void)
2410{
2411 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2412
2413 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2414 page_writeback_cpu_online, NULL);
2415 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2416 page_writeback_cpu_online);
2417#ifdef CONFIG_SYSCTL
2418 register_sysctl_init("vm", vm_page_writeback_sysctls);
2419#endif
2420}
2421
2422/**
2423 * tag_pages_for_writeback - tag pages to be written by writeback
2424 * @mapping: address space structure to write
2425 * @start: starting page index
2426 * @end: ending page index (inclusive)
2427 *
2428 * This function scans the page range from @start to @end (inclusive) and tags
2429 * all pages that have DIRTY tag set with a special TOWRITE tag. The caller
2430 * can then use the TOWRITE tag to identify pages eligible for writeback.
2431 * This mechanism is used to avoid livelocking of writeback by a process
2432 * steadily creating new dirty pages in the file (thus it is important for this
2433 * function to be quick so that it can tag pages faster than a dirtying process
2434 * can create them).
2435 */
2436void tag_pages_for_writeback(struct address_space *mapping,
2437 pgoff_t start, pgoff_t end)
2438{
2439 XA_STATE(xas, &mapping->i_pages, start);
2440 unsigned int tagged = 0;
2441 void *page;
2442
2443 xas_lock_irq(&xas);
2444 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2445 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2446 if (++tagged % XA_CHECK_SCHED)
2447 continue;
2448
2449 xas_pause(&xas);
2450 xas_unlock_irq(&xas);
2451 cond_resched();
2452 xas_lock_irq(&xas);
2453 }
2454 xas_unlock_irq(&xas);
2455}
2456EXPORT_SYMBOL(tag_pages_for_writeback);
2457
2458static bool folio_prepare_writeback(struct address_space *mapping,
2459 struct writeback_control *wbc, struct folio *folio)
2460{
2461 /*
2462 * Folio truncated or invalidated. We can freely skip it then,
2463 * even for data integrity operations: the folio has disappeared
2464 * concurrently, so there could be no real expectation of this
2465 * data integrity operation even if there is now a new, dirty
2466 * folio at the same pagecache index.
2467 */
2468 if (unlikely(folio->mapping != mapping))
2469 return false;
2470
2471 /*
2472 * Did somebody else write it for us?
2473 */
2474 if (!folio_test_dirty(folio))
2475 return false;
2476
2477 if (folio_test_writeback(folio)) {
2478 if (wbc->sync_mode == WB_SYNC_NONE)
2479 return false;
2480 folio_wait_writeback(folio);
2481 }
2482 BUG_ON(folio_test_writeback(folio));
2483
2484 if (!folio_clear_dirty_for_io(folio))
2485 return false;
2486
2487 return true;
2488}
2489
2490static xa_mark_t wbc_to_tag(struct writeback_control *wbc)
2491{
2492 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2493 return PAGECACHE_TAG_TOWRITE;
2494 return PAGECACHE_TAG_DIRTY;
2495}
2496
2497static pgoff_t wbc_end(struct writeback_control *wbc)
2498{
2499 if (wbc->range_cyclic)
2500 return -1;
2501 return wbc->range_end >> PAGE_SHIFT;
2502}
2503
2504static struct folio *writeback_get_folio(struct address_space *mapping,
2505 struct writeback_control *wbc)
2506{
2507 struct folio *folio;
2508
2509retry:
2510 folio = folio_batch_next(&wbc->fbatch);
2511 if (!folio) {
2512 folio_batch_release(&wbc->fbatch);
2513 cond_resched();
2514 filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
2515 wbc_to_tag(wbc), &wbc->fbatch);
2516 folio = folio_batch_next(&wbc->fbatch);
2517 if (!folio)
2518 return NULL;
2519 }
2520
2521 folio_lock(folio);
2522 if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
2523 folio_unlock(folio);
2524 goto retry;
2525 }
2526
2527 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2528 return folio;
2529}
2530
2531/**
2532 * writeback_iter - iterate folio of a mapping for writeback
2533 * @mapping: address space structure to write
2534 * @wbc: writeback context
2535 * @folio: previously iterated folio (%NULL to start)
2536 * @error: in-out pointer for writeback errors (see below)
2537 *
2538 * This function returns the next folio for the writeback operation described by
2539 * @wbc on @mapping and should be called in a while loop in the ->writepages
2540 * implementation.
2541 *
2542 * To start the writeback operation, %NULL is passed in the @folio argument, and
2543 * for every subsequent iteration the folio returned previously should be passed
2544 * back in.
2545 *
2546 * If there was an error in the per-folio writeback inside the writeback_iter()
2547 * loop, @error should be set to the error value.
2548 *
2549 * Once the writeback described in @wbc has finished, this function will return
2550 * %NULL and if there was an error in any iteration restore it to @error.
2551 *
2552 * Note: callers should not manually break out of the loop using break or goto
2553 * but must keep calling writeback_iter() until it returns %NULL.
2554 *
2555 * Return: the folio to write or %NULL if the loop is done.
2556 */
2557struct folio *writeback_iter(struct address_space *mapping,
2558 struct writeback_control *wbc, struct folio *folio, int *error)
2559{
2560 if (!folio) {
2561 folio_batch_init(&wbc->fbatch);
2562 wbc->saved_err = *error = 0;
2563
2564 /*
2565 * For range cyclic writeback we remember where we stopped so
2566 * that we can continue where we stopped.
2567 *
2568 * For non-cyclic writeback we always start at the beginning of
2569 * the passed in range.
2570 */
2571 if (wbc->range_cyclic)
2572 wbc->index = mapping->writeback_index;
2573 else
2574 wbc->index = wbc->range_start >> PAGE_SHIFT;
2575
2576 /*
2577 * To avoid livelocks when other processes dirty new pages, we
2578 * first tag pages which should be written back and only then
2579 * start writing them.
2580 *
2581 * For data-integrity writeback we have to be careful so that we
2582 * do not miss some pages (e.g., because some other process has
2583 * cleared the TOWRITE tag we set). The rule we follow is that
2584 * TOWRITE tag can be cleared only by the process clearing the
2585 * DIRTY tag (and submitting the page for I/O).
2586 */
2587 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2588 tag_pages_for_writeback(mapping, wbc->index,
2589 wbc_end(wbc));
2590 } else {
2591 wbc->nr_to_write -= folio_nr_pages(folio);
2592
2593 WARN_ON_ONCE(*error > 0);
2594
2595 /*
2596 * For integrity writeback we have to keep going until we have
2597 * written all the folios we tagged for writeback above, even if
2598 * we run past wbc->nr_to_write or encounter errors.
2599 * We stash away the first error we encounter in wbc->saved_err
2600 * so that it can be retrieved when we're done. This is because
2601 * the file system may still have state to clear for each folio.
2602 *
2603 * For background writeback we exit as soon as we run past
2604 * wbc->nr_to_write or encounter the first error.
2605 */
2606 if (wbc->sync_mode == WB_SYNC_ALL) {
2607 if (*error && !wbc->saved_err)
2608 wbc->saved_err = *error;
2609 } else {
2610 if (*error || wbc->nr_to_write <= 0)
2611 goto done;
2612 }
2613 }
2614
2615 folio = writeback_get_folio(mapping, wbc);
2616 if (!folio) {
2617 /*
2618 * To avoid deadlocks between range_cyclic writeback and callers
2619 * that hold pages in PageWriteback to aggregate I/O until
2620 * the writeback iteration finishes, we do not loop back to the
2621 * start of the file. Doing so causes a page lock/page
2622 * writeback access order inversion - we should only ever lock
2623 * multiple pages in ascending page->index order, and looping
2624 * back to the start of the file violates that rule and causes
2625 * deadlocks.
2626 */
2627 if (wbc->range_cyclic)
2628 mapping->writeback_index = 0;
2629
2630 /*
2631 * Return the first error we encountered (if there was any) to
2632 * the caller.
2633 */
2634 *error = wbc->saved_err;
2635 }
2636 return folio;
2637
2638done:
2639 if (wbc->range_cyclic)
2640 mapping->writeback_index = folio_next_index(folio);
2641 folio_batch_release(&wbc->fbatch);
2642 return NULL;
2643}
2644EXPORT_SYMBOL_GPL(writeback_iter);
2645
2646/**
2647 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2648 * @mapping: address space structure to write
2649 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2650 * @writepage: function called for each page
2651 * @data: data passed to writepage function
2652 *
2653 * Return: %0 on success, negative error code otherwise
2654 *
2655 * Note: please use writeback_iter() instead.
2656 */
2657int write_cache_pages(struct address_space *mapping,
2658 struct writeback_control *wbc, writepage_t writepage,
2659 void *data)
2660{
2661 struct folio *folio = NULL;
2662 int error;
2663
2664 while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
2665 error = writepage(folio, wbc, data);
2666 if (error == AOP_WRITEPAGE_ACTIVATE) {
2667 folio_unlock(folio);
2668 error = 0;
2669 }
2670 }
2671
2672 return error;
2673}
2674EXPORT_SYMBOL(write_cache_pages);
2675
2676static int writeback_use_writepage(struct address_space *mapping,
2677 struct writeback_control *wbc)
2678{
2679 struct folio *folio = NULL;
2680 struct blk_plug plug;
2681 int err;
2682
2683 blk_start_plug(&plug);
2684 while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
2685 err = mapping->a_ops->writepage(&folio->page, wbc);
2686 if (err == AOP_WRITEPAGE_ACTIVATE) {
2687 folio_unlock(folio);
2688 err = 0;
2689 }
2690 mapping_set_error(mapping, err);
2691 }
2692 blk_finish_plug(&plug);
2693
2694 return err;
2695}
2696
2697int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2698{
2699 int ret;
2700 struct bdi_writeback *wb;
2701
2702 if (wbc->nr_to_write <= 0)
2703 return 0;
2704 wb = inode_to_wb_wbc(mapping->host, wbc);
2705 wb_bandwidth_estimate_start(wb);
2706 while (1) {
2707 if (mapping->a_ops->writepages) {
2708 ret = mapping->a_ops->writepages(mapping, wbc);
2709 } else if (mapping->a_ops->writepage) {
2710 ret = writeback_use_writepage(mapping, wbc);
2711 } else {
2712 /* deal with chardevs and other special files */
2713 ret = 0;
2714 }
2715 if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
2716 break;
2717
2718 /*
2719 * Lacking an allocation context or the locality or writeback
2720 * state of any of the inode's pages, throttle based on
2721 * writeback activity on the local node. It's as good a
2722 * guess as any.
2723 */
2724 reclaim_throttle(NODE_DATA(numa_node_id()),
2725 VMSCAN_THROTTLE_WRITEBACK);
2726 }
2727 /*
2728 * Usually few pages are written by now from those we've just submitted
2729 * but if there's constant writeback being submitted, this makes sure
2730 * writeback bandwidth is updated once in a while.
2731 */
2732 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2733 BANDWIDTH_INTERVAL))
2734 wb_update_bandwidth(wb);
2735 return ret;
2736}
2737
2738/*
2739 * For address_spaces which do not use buffers nor write back.
2740 */
2741bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2742{
2743 if (!folio_test_dirty(folio))
2744 return !folio_test_set_dirty(folio);
2745 return false;
2746}
2747EXPORT_SYMBOL(noop_dirty_folio);
2748
2749/*
2750 * Helper function for set_page_dirty family.
2751 *
2752 * NOTE: This relies on being atomic wrt interrupts.
2753 */
2754static void folio_account_dirtied(struct folio *folio,
2755 struct address_space *mapping)
2756{
2757 struct inode *inode = mapping->host;
2758
2759 trace_writeback_dirty_folio(folio, mapping);
2760
2761 if (mapping_can_writeback(mapping)) {
2762 struct bdi_writeback *wb;
2763 long nr = folio_nr_pages(folio);
2764
2765 inode_attach_wb(inode, folio);
2766 wb = inode_to_wb(inode);
2767
2768 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2769 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2770 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2771 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2772 wb_stat_mod(wb, WB_DIRTIED, nr);
2773 task_io_account_write(nr * PAGE_SIZE);
2774 current->nr_dirtied += nr;
2775 __this_cpu_add(bdp_ratelimits, nr);
2776
2777 mem_cgroup_track_foreign_dirty(folio, wb);
2778 }
2779}
2780
2781/*
2782 * Helper function for deaccounting dirty page without writeback.
2783 *
2784 */
2785void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2786{
2787 long nr = folio_nr_pages(folio);
2788
2789 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2790 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2791 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2792 task_io_account_cancelled_write(nr * PAGE_SIZE);
2793}
2794
2795/*
2796 * Mark the folio dirty, and set it dirty in the page cache.
2797 *
2798 * If warn is true, then emit a warning if the folio is not uptodate and has
2799 * not been truncated.
2800 *
2801 * It is the caller's responsibility to prevent the folio from being truncated
2802 * while this function is in progress, although it may have been truncated
2803 * before this function is called. Most callers have the folio locked.
2804 * A few have the folio blocked from truncation through other means (e.g.
2805 * zap_vma_pages() has it mapped and is holding the page table lock).
2806 * When called from mark_buffer_dirty(), the filesystem should hold a
2807 * reference to the buffer_head that is being marked dirty, which causes
2808 * try_to_free_buffers() to fail.
2809 */
2810void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2811 int warn)
2812{
2813 unsigned long flags;
2814
2815 xa_lock_irqsave(&mapping->i_pages, flags);
2816 if (folio->mapping) { /* Race with truncate? */
2817 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2818 folio_account_dirtied(folio, mapping);
2819 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2820 PAGECACHE_TAG_DIRTY);
2821 }
2822 xa_unlock_irqrestore(&mapping->i_pages, flags);
2823}
2824
2825/**
2826 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2827 * @mapping: Address space this folio belongs to.
2828 * @folio: Folio to be marked as dirty.
2829 *
2830 * Filesystems which do not use buffer heads should call this function
2831 * from their dirty_folio address space operation. It ignores the
2832 * contents of folio_get_private(), so if the filesystem marks individual
2833 * blocks as dirty, the filesystem should handle that itself.
2834 *
2835 * This is also sometimes used by filesystems which use buffer_heads when
2836 * a single buffer is being dirtied: we want to set the folio dirty in
2837 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2838 * whereas block_dirty_folio() is a "top-down" dirtying.
2839 *
2840 * The caller must ensure this doesn't race with truncation. Most will
2841 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2842 * folio mapped and the pte lock held, which also locks out truncation.
2843 */
2844bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2845{
2846 if (folio_test_set_dirty(folio))
2847 return false;
2848
2849 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2850
2851 if (mapping->host) {
2852 /* !PageAnon && !swapper_space */
2853 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2854 }
2855 return true;
2856}
2857EXPORT_SYMBOL(filemap_dirty_folio);
2858
2859/**
2860 * folio_redirty_for_writepage - Decline to write a dirty folio.
2861 * @wbc: The writeback control.
2862 * @folio: The folio.
2863 *
2864 * When a writepage implementation decides that it doesn't want to write
2865 * @folio for some reason, it should call this function, unlock @folio and
2866 * return 0.
2867 *
2868 * Return: True if we redirtied the folio. False if someone else dirtied
2869 * it first.
2870 */
2871bool folio_redirty_for_writepage(struct writeback_control *wbc,
2872 struct folio *folio)
2873{
2874 struct address_space *mapping = folio->mapping;
2875 long nr = folio_nr_pages(folio);
2876 bool ret;
2877
2878 wbc->pages_skipped += nr;
2879 ret = filemap_dirty_folio(mapping, folio);
2880 if (mapping && mapping_can_writeback(mapping)) {
2881 struct inode *inode = mapping->host;
2882 struct bdi_writeback *wb;
2883 struct wb_lock_cookie cookie = {};
2884
2885 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2886 current->nr_dirtied -= nr;
2887 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2888 wb_stat_mod(wb, WB_DIRTIED, -nr);
2889 unlocked_inode_to_wb_end(inode, &cookie);
2890 }
2891 return ret;
2892}
2893EXPORT_SYMBOL(folio_redirty_for_writepage);
2894
2895/**
2896 * folio_mark_dirty - Mark a folio as being modified.
2897 * @folio: The folio.
2898 *
2899 * The folio may not be truncated while this function is running.
2900 * Holding the folio lock is sufficient to prevent truncation, but some
2901 * callers cannot acquire a sleeping lock. These callers instead hold
2902 * the page table lock for a page table which contains at least one page
2903 * in this folio. Truncation will block on the page table lock as it
2904 * unmaps pages before removing the folio from its mapping.
2905 *
2906 * Return: True if the folio was newly dirtied, false if it was already dirty.
2907 */
2908bool folio_mark_dirty(struct folio *folio)
2909{
2910 struct address_space *mapping = folio_mapping(folio);
2911
2912 if (likely(mapping)) {
2913 /*
2914 * readahead/folio_deactivate could remain
2915 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2916 * About readahead, if the folio is written, the flags would be
2917 * reset. So no problem.
2918 * About folio_deactivate, if the folio is redirtied,
2919 * the flag will be reset. So no problem. but if the
2920 * folio is used by readahead it will confuse readahead
2921 * and make it restart the size rampup process. But it's
2922 * a trivial problem.
2923 */
2924 if (folio_test_reclaim(folio))
2925 folio_clear_reclaim(folio);
2926 return mapping->a_ops->dirty_folio(mapping, folio);
2927 }
2928
2929 return noop_dirty_folio(mapping, folio);
2930}
2931EXPORT_SYMBOL(folio_mark_dirty);
2932
2933/*
2934 * folio_mark_dirty() is racy if the caller has no reference against
2935 * folio->mapping->host, and if the folio is unlocked. This is because another
2936 * CPU could truncate the folio off the mapping and then free the mapping.
2937 *
2938 * Usually, the folio _is_ locked, or the caller is a user-space process which
2939 * holds a reference on the inode by having an open file.
2940 *
2941 * In other cases, the folio should be locked before running folio_mark_dirty().
2942 */
2943bool folio_mark_dirty_lock(struct folio *folio)
2944{
2945 bool ret;
2946
2947 folio_lock(folio);
2948 ret = folio_mark_dirty(folio);
2949 folio_unlock(folio);
2950 return ret;
2951}
2952EXPORT_SYMBOL(folio_mark_dirty_lock);
2953
2954/*
2955 * This cancels just the dirty bit on the kernel page itself, it does NOT
2956 * actually remove dirty bits on any mmap's that may be around. It also
2957 * leaves the page tagged dirty, so any sync activity will still find it on
2958 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2959 * look at the dirty bits in the VM.
2960 *
2961 * Doing this should *normally* only ever be done when a page is truncated,
2962 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2963 * this when it notices that somebody has cleaned out all the buffers on a
2964 * page without actually doing it through the VM. Can you say "ext3 is
2965 * horribly ugly"? Thought you could.
2966 */
2967void __folio_cancel_dirty(struct folio *folio)
2968{
2969 struct address_space *mapping = folio_mapping(folio);
2970
2971 if (mapping_can_writeback(mapping)) {
2972 struct inode *inode = mapping->host;
2973 struct bdi_writeback *wb;
2974 struct wb_lock_cookie cookie = {};
2975
2976 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2977
2978 if (folio_test_clear_dirty(folio))
2979 folio_account_cleaned(folio, wb);
2980
2981 unlocked_inode_to_wb_end(inode, &cookie);
2982 } else {
2983 folio_clear_dirty(folio);
2984 }
2985}
2986EXPORT_SYMBOL(__folio_cancel_dirty);
2987
2988/*
2989 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2990 * Returns true if the folio was previously dirty.
2991 *
2992 * This is for preparing to put the folio under writeout. We leave
2993 * the folio tagged as dirty in the xarray so that a concurrent
2994 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2995 * The ->writepage implementation will run either folio_start_writeback()
2996 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2997 * and xarray dirty tag back into sync.
2998 *
2999 * This incoherency between the folio's dirty flag and xarray tag is
3000 * unfortunate, but it only exists while the folio is locked.
3001 */
3002bool folio_clear_dirty_for_io(struct folio *folio)
3003{
3004 struct address_space *mapping = folio_mapping(folio);
3005 bool ret = false;
3006
3007 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3008
3009 if (mapping && mapping_can_writeback(mapping)) {
3010 struct inode *inode = mapping->host;
3011 struct bdi_writeback *wb;
3012 struct wb_lock_cookie cookie = {};
3013
3014 /*
3015 * Yes, Virginia, this is indeed insane.
3016 *
3017 * We use this sequence to make sure that
3018 * (a) we account for dirty stats properly
3019 * (b) we tell the low-level filesystem to
3020 * mark the whole folio dirty if it was
3021 * dirty in a pagetable. Only to then
3022 * (c) clean the folio again and return 1 to
3023 * cause the writeback.
3024 *
3025 * This way we avoid all nasty races with the
3026 * dirty bit in multiple places and clearing
3027 * them concurrently from different threads.
3028 *
3029 * Note! Normally the "folio_mark_dirty(folio)"
3030 * has no effect on the actual dirty bit - since
3031 * that will already usually be set. But we
3032 * need the side effects, and it can help us
3033 * avoid races.
3034 *
3035 * We basically use the folio "master dirty bit"
3036 * as a serialization point for all the different
3037 * threads doing their things.
3038 */
3039 if (folio_mkclean(folio))
3040 folio_mark_dirty(folio);
3041 /*
3042 * We carefully synchronise fault handlers against
3043 * installing a dirty pte and marking the folio dirty
3044 * at this point. We do this by having them hold the
3045 * page lock while dirtying the folio, and folios are
3046 * always locked coming in here, so we get the desired
3047 * exclusion.
3048 */
3049 wb = unlocked_inode_to_wb_begin(inode, &cookie);
3050 if (folio_test_clear_dirty(folio)) {
3051 long nr = folio_nr_pages(folio);
3052 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
3053 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3054 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
3055 ret = true;
3056 }
3057 unlocked_inode_to_wb_end(inode, &cookie);
3058 return ret;
3059 }
3060 return folio_test_clear_dirty(folio);
3061}
3062EXPORT_SYMBOL(folio_clear_dirty_for_io);
3063
3064static void wb_inode_writeback_start(struct bdi_writeback *wb)
3065{
3066 atomic_inc(&wb->writeback_inodes);
3067}
3068
3069static void wb_inode_writeback_end(struct bdi_writeback *wb)
3070{
3071 unsigned long flags;
3072 atomic_dec(&wb->writeback_inodes);
3073 /*
3074 * Make sure estimate of writeback throughput gets updated after
3075 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
3076 * (which is the interval other bandwidth updates use for batching) so
3077 * that if multiple inodes end writeback at a similar time, they get
3078 * batched into one bandwidth update.
3079 */
3080 spin_lock_irqsave(&wb->work_lock, flags);
3081 if (test_bit(WB_registered, &wb->state))
3082 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
3083 spin_unlock_irqrestore(&wb->work_lock, flags);
3084}
3085
3086bool __folio_end_writeback(struct folio *folio)
3087{
3088 long nr = folio_nr_pages(folio);
3089 struct address_space *mapping = folio_mapping(folio);
3090 bool ret;
3091
3092 if (mapping && mapping_use_writeback_tags(mapping)) {
3093 struct inode *inode = mapping->host;
3094 struct backing_dev_info *bdi = inode_to_bdi(inode);
3095 unsigned long flags;
3096
3097 xa_lock_irqsave(&mapping->i_pages, flags);
3098 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3099 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
3100 PAGECACHE_TAG_WRITEBACK);
3101 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3102 struct bdi_writeback *wb = inode_to_wb(inode);
3103
3104 wb_stat_mod(wb, WB_WRITEBACK, -nr);
3105 __wb_writeout_add(wb, nr);
3106 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
3107 wb_inode_writeback_end(wb);
3108 }
3109
3110 if (mapping->host && !mapping_tagged(mapping,
3111 PAGECACHE_TAG_WRITEBACK))
3112 sb_clear_inode_writeback(mapping->host);
3113
3114 xa_unlock_irqrestore(&mapping->i_pages, flags);
3115 } else {
3116 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3117 }
3118
3119 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3120 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3121 node_stat_mod_folio(folio, NR_WRITTEN, nr);
3122
3123 return ret;
3124}
3125
3126void __folio_start_writeback(struct folio *folio, bool keep_write)
3127{
3128 long nr = folio_nr_pages(folio);
3129 struct address_space *mapping = folio_mapping(folio);
3130 int access_ret;
3131
3132 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
3133
3134 if (mapping && mapping_use_writeback_tags(mapping)) {
3135 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3136 struct inode *inode = mapping->host;
3137 struct backing_dev_info *bdi = inode_to_bdi(inode);
3138 unsigned long flags;
3139 bool on_wblist;
3140
3141 xas_lock_irqsave(&xas, flags);
3142 xas_load(&xas);
3143 folio_test_set_writeback(folio);
3144
3145 on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
3146
3147 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3148 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3149 struct bdi_writeback *wb = inode_to_wb(inode);
3150
3151 wb_stat_mod(wb, WB_WRITEBACK, nr);
3152 if (!on_wblist)
3153 wb_inode_writeback_start(wb);
3154 }
3155
3156 /*
3157 * We can come through here when swapping anonymous
3158 * folios, so we don't necessarily have an inode to
3159 * track for sync.
3160 */
3161 if (mapping->host && !on_wblist)
3162 sb_mark_inode_writeback(mapping->host);
3163 if (!folio_test_dirty(folio))
3164 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3165 if (!keep_write)
3166 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3167 xas_unlock_irqrestore(&xas, flags);
3168 } else {
3169 folio_test_set_writeback(folio);
3170 }
3171
3172 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3173 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3174
3175 access_ret = arch_make_folio_accessible(folio);
3176 /*
3177 * If writeback has been triggered on a page that cannot be made
3178 * accessible, it is too late to recover here.
3179 */
3180 VM_BUG_ON_FOLIO(access_ret != 0, folio);
3181}
3182EXPORT_SYMBOL(__folio_start_writeback);
3183
3184/**
3185 * folio_wait_writeback - Wait for a folio to finish writeback.
3186 * @folio: The folio to wait for.
3187 *
3188 * If the folio is currently being written back to storage, wait for the
3189 * I/O to complete.
3190 *
3191 * Context: Sleeps. Must be called in process context and with
3192 * no spinlocks held. Caller should hold a reference on the folio.
3193 * If the folio is not locked, writeback may start again after writeback
3194 * has finished.
3195 */
3196void folio_wait_writeback(struct folio *folio)
3197{
3198 while (folio_test_writeback(folio)) {
3199 trace_folio_wait_writeback(folio, folio_mapping(folio));
3200 folio_wait_bit(folio, PG_writeback);
3201 }
3202}
3203EXPORT_SYMBOL_GPL(folio_wait_writeback);
3204
3205/**
3206 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3207 * @folio: The folio to wait for.
3208 *
3209 * If the folio is currently being written back to storage, wait for the
3210 * I/O to complete or a fatal signal to arrive.
3211 *
3212 * Context: Sleeps. Must be called in process context and with
3213 * no spinlocks held. Caller should hold a reference on the folio.
3214 * If the folio is not locked, writeback may start again after writeback
3215 * has finished.
3216 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3217 */
3218int folio_wait_writeback_killable(struct folio *folio)
3219{
3220 while (folio_test_writeback(folio)) {
3221 trace_folio_wait_writeback(folio, folio_mapping(folio));
3222 if (folio_wait_bit_killable(folio, PG_writeback))
3223 return -EINTR;
3224 }
3225
3226 return 0;
3227}
3228EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3229
3230/**
3231 * folio_wait_stable() - wait for writeback to finish, if necessary.
3232 * @folio: The folio to wait on.
3233 *
3234 * This function determines if the given folio is related to a backing
3235 * device that requires folio contents to be held stable during writeback.
3236 * If so, then it will wait for any pending writeback to complete.
3237 *
3238 * Context: Sleeps. Must be called in process context and with
3239 * no spinlocks held. Caller should hold a reference on the folio.
3240 * If the folio is not locked, writeback may start again after writeback
3241 * has finished.
3242 */
3243void folio_wait_stable(struct folio *folio)
3244{
3245 if (mapping_stable_writes(folio_mapping(folio)))
3246 folio_wait_writeback(folio);
3247}
3248EXPORT_SYMBOL_GPL(folio_wait_stable);