rseq: Use get_user/put_user rather than __get_user/__put_user
[linux-2.6-microblaze.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    Andrew Morton
11  *              Initial version
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/sched/signal.h>
40 #include <linux/mm_inline.h>
41 #include <trace/events/writeback.h>
42
43 #include "internal.h"
44
45 /*
46  * Sleep at most 200ms at a time in balance_dirty_pages().
47  */
48 #define MAX_PAUSE               max(HZ/5, 1)
49
50 /*
51  * Try to keep balance_dirty_pages() call intervals higher than this many pages
52  * by raising pause time to max_pause when falls below it.
53  */
54 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
55
56 /*
57  * Estimate write bandwidth at 200ms intervals.
58  */
59 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
60
61 #define RATELIMIT_CALC_SHIFT    10
62
63 /*
64  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65  * will look to see if it needs to force writeback or throttling.
66  */
67 static long ratelimit_pages = 32;
68
69 /* The following parameters are exported via /proc/sys/vm */
70
71 /*
72  * Start background writeback (via writeback threads) at this percentage
73  */
74 int dirty_background_ratio = 10;
75
76 /*
77  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78  * dirty_background_ratio * the amount of dirtyable memory
79  */
80 unsigned long dirty_background_bytes;
81
82 /*
83  * free highmem will not be subtracted from the total free memory
84  * for calculating free ratios if vm_highmem_is_dirtyable is true
85  */
86 int vm_highmem_is_dirtyable;
87
88 /*
89  * The generator of dirty data starts writeback at this percentage
90  */
91 int vm_dirty_ratio = 20;
92
93 /*
94  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95  * vm_dirty_ratio * the amount of dirtyable memory
96  */
97 unsigned long vm_dirty_bytes;
98
99 /*
100  * The interval between `kupdate'-style writebacks
101  */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106 /*
107  * The longest time for which data is allowed to remain dirty
108  */
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111 /*
112  * Flag that makes the machine dump writes/reads and block dirtyings.
113  */
114 int block_dump;
115
116 /*
117  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118  * a full sync is triggered after this time elapses without any disk activity.
119  */
120 int laptop_mode;
121
122 EXPORT_SYMBOL(laptop_mode);
123
124 /* End of sysctl-exported parameters */
125
126 struct wb_domain global_wb_domain;
127
128 /* consolidated parameters for balance_dirty_pages() and its subroutines */
129 struct dirty_throttle_control {
130 #ifdef CONFIG_CGROUP_WRITEBACK
131         struct wb_domain        *dom;
132         struct dirty_throttle_control *gdtc;    /* only set in memcg dtc's */
133 #endif
134         struct bdi_writeback    *wb;
135         struct fprop_local_percpu *wb_completions;
136
137         unsigned long           avail;          /* dirtyable */
138         unsigned long           dirty;          /* file_dirty + write + nfs */
139         unsigned long           thresh;         /* dirty threshold */
140         unsigned long           bg_thresh;      /* dirty background threshold */
141
142         unsigned long           wb_dirty;       /* per-wb counterparts */
143         unsigned long           wb_thresh;
144         unsigned long           wb_bg_thresh;
145
146         unsigned long           pos_ratio;
147 };
148
149 /*
150  * Length of period for aging writeout fractions of bdis. This is an
151  * arbitrarily chosen number. The longer the period, the slower fractions will
152  * reflect changes in current writeout rate.
153  */
154 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155
156 #ifdef CONFIG_CGROUP_WRITEBACK
157
158 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
159                                 .dom = &global_wb_domain,               \
160                                 .wb_completions = &(__wb)->completions
161
162 #define GDTC_INIT_NO_WB         .dom = &global_wb_domain
163
164 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb),                           \
165                                 .dom = mem_cgroup_wb_domain(__wb),      \
166                                 .wb_completions = &(__wb)->memcg_completions, \
167                                 .gdtc = __gdtc
168
169 static bool mdtc_valid(struct dirty_throttle_control *dtc)
170 {
171         return dtc->dom;
172 }
173
174 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175 {
176         return dtc->dom;
177 }
178
179 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180 {
181         return mdtc->gdtc;
182 }
183
184 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185 {
186         return &wb->memcg_completions;
187 }
188
189 static void wb_min_max_ratio(struct bdi_writeback *wb,
190                              unsigned long *minp, unsigned long *maxp)
191 {
192         unsigned long this_bw = wb->avg_write_bandwidth;
193         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194         unsigned long long min = wb->bdi->min_ratio;
195         unsigned long long max = wb->bdi->max_ratio;
196
197         /*
198          * @wb may already be clean by the time control reaches here and
199          * the total may not include its bw.
200          */
201         if (this_bw < tot_bw) {
202                 if (min) {
203                         min *= this_bw;
204                         do_div(min, tot_bw);
205                 }
206                 if (max < 100) {
207                         max *= this_bw;
208                         do_div(max, tot_bw);
209                 }
210         }
211
212         *minp = min;
213         *maxp = max;
214 }
215
216 #else   /* CONFIG_CGROUP_WRITEBACK */
217
218 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
219                                 .wb_completions = &(__wb)->completions
220 #define GDTC_INIT_NO_WB
221 #define MDTC_INIT(__wb, __gdtc)
222
223 static bool mdtc_valid(struct dirty_throttle_control *dtc)
224 {
225         return false;
226 }
227
228 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229 {
230         return &global_wb_domain;
231 }
232
233 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234 {
235         return NULL;
236 }
237
238 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239 {
240         return NULL;
241 }
242
243 static void wb_min_max_ratio(struct bdi_writeback *wb,
244                              unsigned long *minp, unsigned long *maxp)
245 {
246         *minp = wb->bdi->min_ratio;
247         *maxp = wb->bdi->max_ratio;
248 }
249
250 #endif  /* CONFIG_CGROUP_WRITEBACK */
251
252 /*
253  * In a memory zone, there is a certain amount of pages we consider
254  * available for the page cache, which is essentially the number of
255  * free and reclaimable pages, minus some zone reserves to protect
256  * lowmem and the ability to uphold the zone's watermarks without
257  * requiring writeback.
258  *
259  * This number of dirtyable pages is the base value of which the
260  * user-configurable dirty ratio is the effictive number of pages that
261  * are allowed to be actually dirtied.  Per individual zone, or
262  * globally by using the sum of dirtyable pages over all zones.
263  *
264  * Because the user is allowed to specify the dirty limit globally as
265  * absolute number of bytes, calculating the per-zone dirty limit can
266  * require translating the configured limit into a percentage of
267  * global dirtyable memory first.
268  */
269
270 /**
271  * node_dirtyable_memory - number of dirtyable pages in a node
272  * @pgdat: the node
273  *
274  * Returns the node's number of pages potentially available for dirty
275  * page cache.  This is the base value for the per-node dirty limits.
276  */
277 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278 {
279         unsigned long nr_pages = 0;
280         int z;
281
282         for (z = 0; z < MAX_NR_ZONES; z++) {
283                 struct zone *zone = pgdat->node_zones + z;
284
285                 if (!populated_zone(zone))
286                         continue;
287
288                 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289         }
290
291         /*
292          * Pages reserved for the kernel should not be considered
293          * dirtyable, to prevent a situation where reclaim has to
294          * clean pages in order to balance the zones.
295          */
296         nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297
298         nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299         nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
300
301         return nr_pages;
302 }
303
304 static unsigned long highmem_dirtyable_memory(unsigned long total)
305 {
306 #ifdef CONFIG_HIGHMEM
307         int node;
308         unsigned long x = 0;
309         int i;
310
311         for_each_node_state(node, N_HIGH_MEMORY) {
312                 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313                         struct zone *z;
314                         unsigned long nr_pages;
315
316                         if (!is_highmem_idx(i))
317                                 continue;
318
319                         z = &NODE_DATA(node)->node_zones[i];
320                         if (!populated_zone(z))
321                                 continue;
322
323                         nr_pages = zone_page_state(z, NR_FREE_PAGES);
324                         /* watch for underflows */
325                         nr_pages -= min(nr_pages, high_wmark_pages(z));
326                         nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327                         nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328                         x += nr_pages;
329                 }
330         }
331
332         /*
333          * Unreclaimable memory (kernel memory or anonymous memory
334          * without swap) can bring down the dirtyable pages below
335          * the zone's dirty balance reserve and the above calculation
336          * will underflow.  However we still want to add in nodes
337          * which are below threshold (negative values) to get a more
338          * accurate calculation but make sure that the total never
339          * underflows.
340          */
341         if ((long)x < 0)
342                 x = 0;
343
344         /*
345          * Make sure that the number of highmem pages is never larger
346          * than the number of the total dirtyable memory. This can only
347          * occur in very strange VM situations but we want to make sure
348          * that this does not occur.
349          */
350         return min(x, total);
351 #else
352         return 0;
353 #endif
354 }
355
356 /**
357  * global_dirtyable_memory - number of globally dirtyable pages
358  *
359  * Returns the global number of pages potentially available for dirty
360  * page cache.  This is the base value for the global dirty limits.
361  */
362 static unsigned long global_dirtyable_memory(void)
363 {
364         unsigned long x;
365
366         x = global_zone_page_state(NR_FREE_PAGES);
367         /*
368          * Pages reserved for the kernel should not be considered
369          * dirtyable, to prevent a situation where reclaim has to
370          * clean pages in order to balance the zones.
371          */
372         x -= min(x, totalreserve_pages);
373
374         x += global_node_page_state(NR_INACTIVE_FILE);
375         x += global_node_page_state(NR_ACTIVE_FILE);
376
377         if (!vm_highmem_is_dirtyable)
378                 x -= highmem_dirtyable_memory(x);
379
380         return x + 1;   /* Ensure that we never return 0 */
381 }
382
383 /**
384  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385  * @dtc: dirty_throttle_control of interest
386  *
387  * Calculate @dtc->thresh and ->bg_thresh considering
388  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
389  * must ensure that @dtc->avail is set before calling this function.  The
390  * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
391  * real-time tasks.
392  */
393 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
394 {
395         const unsigned long available_memory = dtc->avail;
396         struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
397         unsigned long bytes = vm_dirty_bytes;
398         unsigned long bg_bytes = dirty_background_bytes;
399         /* convert ratios to per-PAGE_SIZE for higher precision */
400         unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
401         unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
402         unsigned long thresh;
403         unsigned long bg_thresh;
404         struct task_struct *tsk;
405
406         /* gdtc is !NULL iff @dtc is for memcg domain */
407         if (gdtc) {
408                 unsigned long global_avail = gdtc->avail;
409
410                 /*
411                  * The byte settings can't be applied directly to memcg
412                  * domains.  Convert them to ratios by scaling against
413                  * globally available memory.  As the ratios are in
414                  * per-PAGE_SIZE, they can be obtained by dividing bytes by
415                  * number of pages.
416                  */
417                 if (bytes)
418                         ratio = min(DIV_ROUND_UP(bytes, global_avail),
419                                     PAGE_SIZE);
420                 if (bg_bytes)
421                         bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
422                                        PAGE_SIZE);
423                 bytes = bg_bytes = 0;
424         }
425
426         if (bytes)
427                 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
428         else
429                 thresh = (ratio * available_memory) / PAGE_SIZE;
430
431         if (bg_bytes)
432                 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
433         else
434                 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
435
436         if (bg_thresh >= thresh)
437                 bg_thresh = thresh / 2;
438         tsk = current;
439         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
440                 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
441                 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
442         }
443         dtc->thresh = thresh;
444         dtc->bg_thresh = bg_thresh;
445
446         /* we should eventually report the domain in the TP */
447         if (!gdtc)
448                 trace_global_dirty_state(bg_thresh, thresh);
449 }
450
451 /**
452  * global_dirty_limits - background-writeback and dirty-throttling thresholds
453  * @pbackground: out parameter for bg_thresh
454  * @pdirty: out parameter for thresh
455  *
456  * Calculate bg_thresh and thresh for global_wb_domain.  See
457  * domain_dirty_limits() for details.
458  */
459 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
460 {
461         struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
462
463         gdtc.avail = global_dirtyable_memory();
464         domain_dirty_limits(&gdtc);
465
466         *pbackground = gdtc.bg_thresh;
467         *pdirty = gdtc.thresh;
468 }
469
470 /**
471  * node_dirty_limit - maximum number of dirty pages allowed in a node
472  * @pgdat: the node
473  *
474  * Returns the maximum number of dirty pages allowed in a node, based
475  * on the node's dirtyable memory.
476  */
477 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
478 {
479         unsigned long node_memory = node_dirtyable_memory(pgdat);
480         struct task_struct *tsk = current;
481         unsigned long dirty;
482
483         if (vm_dirty_bytes)
484                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
485                         node_memory / global_dirtyable_memory();
486         else
487                 dirty = vm_dirty_ratio * node_memory / 100;
488
489         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
490                 dirty += dirty / 4;
491
492         return dirty;
493 }
494
495 /**
496  * node_dirty_ok - tells whether a node is within its dirty limits
497  * @pgdat: the node to check
498  *
499  * Returns %true when the dirty pages in @pgdat are within the node's
500  * dirty limit, %false if the limit is exceeded.
501  */
502 bool node_dirty_ok(struct pglist_data *pgdat)
503 {
504         unsigned long limit = node_dirty_limit(pgdat);
505         unsigned long nr_pages = 0;
506
507         nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
508         nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
509         nr_pages += node_page_state(pgdat, NR_WRITEBACK);
510
511         return nr_pages <= limit;
512 }
513
514 int dirty_background_ratio_handler(struct ctl_table *table, int write,
515                 void __user *buffer, size_t *lenp,
516                 loff_t *ppos)
517 {
518         int ret;
519
520         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
521         if (ret == 0 && write)
522                 dirty_background_bytes = 0;
523         return ret;
524 }
525
526 int dirty_background_bytes_handler(struct ctl_table *table, int write,
527                 void __user *buffer, size_t *lenp,
528                 loff_t *ppos)
529 {
530         int ret;
531
532         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
533         if (ret == 0 && write)
534                 dirty_background_ratio = 0;
535         return ret;
536 }
537
538 int dirty_ratio_handler(struct ctl_table *table, int write,
539                 void __user *buffer, size_t *lenp,
540                 loff_t *ppos)
541 {
542         int old_ratio = vm_dirty_ratio;
543         int ret;
544
545         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
546         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
547                 writeback_set_ratelimit();
548                 vm_dirty_bytes = 0;
549         }
550         return ret;
551 }
552
553 int dirty_bytes_handler(struct ctl_table *table, int write,
554                 void __user *buffer, size_t *lenp,
555                 loff_t *ppos)
556 {
557         unsigned long old_bytes = vm_dirty_bytes;
558         int ret;
559
560         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
561         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
562                 writeback_set_ratelimit();
563                 vm_dirty_ratio = 0;
564         }
565         return ret;
566 }
567
568 static unsigned long wp_next_time(unsigned long cur_time)
569 {
570         cur_time += VM_COMPLETIONS_PERIOD_LEN;
571         /* 0 has a special meaning... */
572         if (!cur_time)
573                 return 1;
574         return cur_time;
575 }
576
577 static void wb_domain_writeout_inc(struct wb_domain *dom,
578                                    struct fprop_local_percpu *completions,
579                                    unsigned int max_prop_frac)
580 {
581         __fprop_inc_percpu_max(&dom->completions, completions,
582                                max_prop_frac);
583         /* First event after period switching was turned off? */
584         if (unlikely(!dom->period_time)) {
585                 /*
586                  * We can race with other __bdi_writeout_inc calls here but
587                  * it does not cause any harm since the resulting time when
588                  * timer will fire and what is in writeout_period_time will be
589                  * roughly the same.
590                  */
591                 dom->period_time = wp_next_time(jiffies);
592                 mod_timer(&dom->period_timer, dom->period_time);
593         }
594 }
595
596 /*
597  * Increment @wb's writeout completion count and the global writeout
598  * completion count. Called from test_clear_page_writeback().
599  */
600 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
601 {
602         struct wb_domain *cgdom;
603
604         inc_wb_stat(wb, WB_WRITTEN);
605         wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
606                                wb->bdi->max_prop_frac);
607
608         cgdom = mem_cgroup_wb_domain(wb);
609         if (cgdom)
610                 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
611                                        wb->bdi->max_prop_frac);
612 }
613
614 void wb_writeout_inc(struct bdi_writeback *wb)
615 {
616         unsigned long flags;
617
618         local_irq_save(flags);
619         __wb_writeout_inc(wb);
620         local_irq_restore(flags);
621 }
622 EXPORT_SYMBOL_GPL(wb_writeout_inc);
623
624 /*
625  * On idle system, we can be called long after we scheduled because we use
626  * deferred timers so count with missed periods.
627  */
628 static void writeout_period(struct timer_list *t)
629 {
630         struct wb_domain *dom = from_timer(dom, t, period_timer);
631         int miss_periods = (jiffies - dom->period_time) /
632                                                  VM_COMPLETIONS_PERIOD_LEN;
633
634         if (fprop_new_period(&dom->completions, miss_periods + 1)) {
635                 dom->period_time = wp_next_time(dom->period_time +
636                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
637                 mod_timer(&dom->period_timer, dom->period_time);
638         } else {
639                 /*
640                  * Aging has zeroed all fractions. Stop wasting CPU on period
641                  * updates.
642                  */
643                 dom->period_time = 0;
644         }
645 }
646
647 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
648 {
649         memset(dom, 0, sizeof(*dom));
650
651         spin_lock_init(&dom->lock);
652
653         timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
654
655         dom->dirty_limit_tstamp = jiffies;
656
657         return fprop_global_init(&dom->completions, gfp);
658 }
659
660 #ifdef CONFIG_CGROUP_WRITEBACK
661 void wb_domain_exit(struct wb_domain *dom)
662 {
663         del_timer_sync(&dom->period_timer);
664         fprop_global_destroy(&dom->completions);
665 }
666 #endif
667
668 /*
669  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
670  * registered backing devices, which, for obvious reasons, can not
671  * exceed 100%.
672  */
673 static unsigned int bdi_min_ratio;
674
675 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
676 {
677         int ret = 0;
678
679         spin_lock_bh(&bdi_lock);
680         if (min_ratio > bdi->max_ratio) {
681                 ret = -EINVAL;
682         } else {
683                 min_ratio -= bdi->min_ratio;
684                 if (bdi_min_ratio + min_ratio < 100) {
685                         bdi_min_ratio += min_ratio;
686                         bdi->min_ratio += min_ratio;
687                 } else {
688                         ret = -EINVAL;
689                 }
690         }
691         spin_unlock_bh(&bdi_lock);
692
693         return ret;
694 }
695
696 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
697 {
698         int ret = 0;
699
700         if (max_ratio > 100)
701                 return -EINVAL;
702
703         spin_lock_bh(&bdi_lock);
704         if (bdi->min_ratio > max_ratio) {
705                 ret = -EINVAL;
706         } else {
707                 bdi->max_ratio = max_ratio;
708                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
709         }
710         spin_unlock_bh(&bdi_lock);
711
712         return ret;
713 }
714 EXPORT_SYMBOL(bdi_set_max_ratio);
715
716 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
717                                            unsigned long bg_thresh)
718 {
719         return (thresh + bg_thresh) / 2;
720 }
721
722 static unsigned long hard_dirty_limit(struct wb_domain *dom,
723                                       unsigned long thresh)
724 {
725         return max(thresh, dom->dirty_limit);
726 }
727
728 /*
729  * Memory which can be further allocated to a memcg domain is capped by
730  * system-wide clean memory excluding the amount being used in the domain.
731  */
732 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
733                             unsigned long filepages, unsigned long headroom)
734 {
735         struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
736         unsigned long clean = filepages - min(filepages, mdtc->dirty);
737         unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
738         unsigned long other_clean = global_clean - min(global_clean, clean);
739
740         mdtc->avail = filepages + min(headroom, other_clean);
741 }
742
743 /**
744  * __wb_calc_thresh - @wb's share of dirty throttling threshold
745  * @dtc: dirty_throttle_context of interest
746  *
747  * Returns @wb's dirty limit in pages. The term "dirty" in the context of
748  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
749  *
750  * Note that balance_dirty_pages() will only seriously take it as a hard limit
751  * when sleeping max_pause per page is not enough to keep the dirty pages under
752  * control. For example, when the device is completely stalled due to some error
753  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
754  * In the other normal situations, it acts more gently by throttling the tasks
755  * more (rather than completely block them) when the wb dirty pages go high.
756  *
757  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
758  * - starving fast devices
759  * - piling up dirty pages (that will take long time to sync) on slow devices
760  *
761  * The wb's share of dirty limit will be adapting to its throughput and
762  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
763  */
764 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
765 {
766         struct wb_domain *dom = dtc_dom(dtc);
767         unsigned long thresh = dtc->thresh;
768         u64 wb_thresh;
769         long numerator, denominator;
770         unsigned long wb_min_ratio, wb_max_ratio;
771
772         /*
773          * Calculate this BDI's share of the thresh ratio.
774          */
775         fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
776                               &numerator, &denominator);
777
778         wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
779         wb_thresh *= numerator;
780         do_div(wb_thresh, denominator);
781
782         wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
783
784         wb_thresh += (thresh * wb_min_ratio) / 100;
785         if (wb_thresh > (thresh * wb_max_ratio) / 100)
786                 wb_thresh = thresh * wb_max_ratio / 100;
787
788         return wb_thresh;
789 }
790
791 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
792 {
793         struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
794                                                .thresh = thresh };
795         return __wb_calc_thresh(&gdtc);
796 }
797
798 /*
799  *                           setpoint - dirty 3
800  *        f(dirty) := 1.0 + (----------------)
801  *                           limit - setpoint
802  *
803  * it's a 3rd order polynomial that subjects to
804  *
805  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
806  * (2) f(setpoint) = 1.0 => the balance point
807  * (3) f(limit)    = 0   => the hard limit
808  * (4) df/dx      <= 0   => negative feedback control
809  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
810  *     => fast response on large errors; small oscillation near setpoint
811  */
812 static long long pos_ratio_polynom(unsigned long setpoint,
813                                           unsigned long dirty,
814                                           unsigned long limit)
815 {
816         long long pos_ratio;
817         long x;
818
819         x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
820                       (limit - setpoint) | 1);
821         pos_ratio = x;
822         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
823         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
824         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
825
826         return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
827 }
828
829 /*
830  * Dirty position control.
831  *
832  * (o) global/bdi setpoints
833  *
834  * We want the dirty pages be balanced around the global/wb setpoints.
835  * When the number of dirty pages is higher/lower than the setpoint, the
836  * dirty position control ratio (and hence task dirty ratelimit) will be
837  * decreased/increased to bring the dirty pages back to the setpoint.
838  *
839  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
840  *
841  *     if (dirty < setpoint) scale up   pos_ratio
842  *     if (dirty > setpoint) scale down pos_ratio
843  *
844  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
845  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
846  *
847  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
848  *
849  * (o) global control line
850  *
851  *     ^ pos_ratio
852  *     |
853  *     |            |<===== global dirty control scope ======>|
854  * 2.0 .............*
855  *     |            .*
856  *     |            . *
857  *     |            .   *
858  *     |            .     *
859  *     |            .        *
860  *     |            .            *
861  * 1.0 ................................*
862  *     |            .                  .     *
863  *     |            .                  .          *
864  *     |            .                  .              *
865  *     |            .                  .                 *
866  *     |            .                  .                    *
867  *   0 +------------.------------------.----------------------*------------->
868  *           freerun^          setpoint^                 limit^   dirty pages
869  *
870  * (o) wb control line
871  *
872  *     ^ pos_ratio
873  *     |
874  *     |            *
875  *     |              *
876  *     |                *
877  *     |                  *
878  *     |                    * |<=========== span ============>|
879  * 1.0 .......................*
880  *     |                      . *
881  *     |                      .   *
882  *     |                      .     *
883  *     |                      .       *
884  *     |                      .         *
885  *     |                      .           *
886  *     |                      .             *
887  *     |                      .               *
888  *     |                      .                 *
889  *     |                      .                   *
890  *     |                      .                     *
891  * 1/4 ...............................................* * * * * * * * * * * *
892  *     |                      .                         .
893  *     |                      .                           .
894  *     |                      .                             .
895  *   0 +----------------------.-------------------------------.------------->
896  *                wb_setpoint^                    x_intercept^
897  *
898  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
899  * be smoothly throttled down to normal if it starts high in situations like
900  * - start writing to a slow SD card and a fast disk at the same time. The SD
901  *   card's wb_dirty may rush to many times higher than wb_setpoint.
902  * - the wb dirty thresh drops quickly due to change of JBOD workload
903  */
904 static void wb_position_ratio(struct dirty_throttle_control *dtc)
905 {
906         struct bdi_writeback *wb = dtc->wb;
907         unsigned long write_bw = wb->avg_write_bandwidth;
908         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
909         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
910         unsigned long wb_thresh = dtc->wb_thresh;
911         unsigned long x_intercept;
912         unsigned long setpoint;         /* dirty pages' target balance point */
913         unsigned long wb_setpoint;
914         unsigned long span;
915         long long pos_ratio;            /* for scaling up/down the rate limit */
916         long x;
917
918         dtc->pos_ratio = 0;
919
920         if (unlikely(dtc->dirty >= limit))
921                 return;
922
923         /*
924          * global setpoint
925          *
926          * See comment for pos_ratio_polynom().
927          */
928         setpoint = (freerun + limit) / 2;
929         pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
930
931         /*
932          * The strictlimit feature is a tool preventing mistrusted filesystems
933          * from growing a large number of dirty pages before throttling. For
934          * such filesystems balance_dirty_pages always checks wb counters
935          * against wb limits. Even if global "nr_dirty" is under "freerun".
936          * This is especially important for fuse which sets bdi->max_ratio to
937          * 1% by default. Without strictlimit feature, fuse writeback may
938          * consume arbitrary amount of RAM because it is accounted in
939          * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
940          *
941          * Here, in wb_position_ratio(), we calculate pos_ratio based on
942          * two values: wb_dirty and wb_thresh. Let's consider an example:
943          * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
944          * limits are set by default to 10% and 20% (background and throttle).
945          * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
946          * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
947          * about ~6K pages (as the average of background and throttle wb
948          * limits). The 3rd order polynomial will provide positive feedback if
949          * wb_dirty is under wb_setpoint and vice versa.
950          *
951          * Note, that we cannot use global counters in these calculations
952          * because we want to throttle process writing to a strictlimit wb
953          * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
954          * in the example above).
955          */
956         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
957                 long long wb_pos_ratio;
958
959                 if (dtc->wb_dirty < 8) {
960                         dtc->pos_ratio = min_t(long long, pos_ratio * 2,
961                                            2 << RATELIMIT_CALC_SHIFT);
962                         return;
963                 }
964
965                 if (dtc->wb_dirty >= wb_thresh)
966                         return;
967
968                 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
969                                                     dtc->wb_bg_thresh);
970
971                 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
972                         return;
973
974                 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
975                                                  wb_thresh);
976
977                 /*
978                  * Typically, for strictlimit case, wb_setpoint << setpoint
979                  * and pos_ratio >> wb_pos_ratio. In the other words global
980                  * state ("dirty") is not limiting factor and we have to
981                  * make decision based on wb counters. But there is an
982                  * important case when global pos_ratio should get precedence:
983                  * global limits are exceeded (e.g. due to activities on other
984                  * wb's) while given strictlimit wb is below limit.
985                  *
986                  * "pos_ratio * wb_pos_ratio" would work for the case above,
987                  * but it would look too non-natural for the case of all
988                  * activity in the system coming from a single strictlimit wb
989                  * with bdi->max_ratio == 100%.
990                  *
991                  * Note that min() below somewhat changes the dynamics of the
992                  * control system. Normally, pos_ratio value can be well over 3
993                  * (when globally we are at freerun and wb is well below wb
994                  * setpoint). Now the maximum pos_ratio in the same situation
995                  * is 2. We might want to tweak this if we observe the control
996                  * system is too slow to adapt.
997                  */
998                 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
999                 return;
1000         }
1001
1002         /*
1003          * We have computed basic pos_ratio above based on global situation. If
1004          * the wb is over/under its share of dirty pages, we want to scale
1005          * pos_ratio further down/up. That is done by the following mechanism.
1006          */
1007
1008         /*
1009          * wb setpoint
1010          *
1011          *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1012          *
1013          *                        x_intercept - wb_dirty
1014          *                     := --------------------------
1015          *                        x_intercept - wb_setpoint
1016          *
1017          * The main wb control line is a linear function that subjects to
1018          *
1019          * (1) f(wb_setpoint) = 1.0
1020          * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1021          *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1022          *
1023          * For single wb case, the dirty pages are observed to fluctuate
1024          * regularly within range
1025          *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1026          * for various filesystems, where (2) can yield in a reasonable 12.5%
1027          * fluctuation range for pos_ratio.
1028          *
1029          * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1030          * own size, so move the slope over accordingly and choose a slope that
1031          * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1032          */
1033         if (unlikely(wb_thresh > dtc->thresh))
1034                 wb_thresh = dtc->thresh;
1035         /*
1036          * It's very possible that wb_thresh is close to 0 not because the
1037          * device is slow, but that it has remained inactive for long time.
1038          * Honour such devices a reasonable good (hopefully IO efficient)
1039          * threshold, so that the occasional writes won't be blocked and active
1040          * writes can rampup the threshold quickly.
1041          */
1042         wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1043         /*
1044          * scale global setpoint to wb's:
1045          *      wb_setpoint = setpoint * wb_thresh / thresh
1046          */
1047         x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1048         wb_setpoint = setpoint * (u64)x >> 16;
1049         /*
1050          * Use span=(8*write_bw) in single wb case as indicated by
1051          * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1052          *
1053          *        wb_thresh                    thresh - wb_thresh
1054          * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1055          *         thresh                           thresh
1056          */
1057         span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1058         x_intercept = wb_setpoint + span;
1059
1060         if (dtc->wb_dirty < x_intercept - span / 4) {
1061                 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1062                                       (x_intercept - wb_setpoint) | 1);
1063         } else
1064                 pos_ratio /= 4;
1065
1066         /*
1067          * wb reserve area, safeguard against dirty pool underrun and disk idle
1068          * It may push the desired control point of global dirty pages higher
1069          * than setpoint.
1070          */
1071         x_intercept = wb_thresh / 2;
1072         if (dtc->wb_dirty < x_intercept) {
1073                 if (dtc->wb_dirty > x_intercept / 8)
1074                         pos_ratio = div_u64(pos_ratio * x_intercept,
1075                                             dtc->wb_dirty);
1076                 else
1077                         pos_ratio *= 8;
1078         }
1079
1080         dtc->pos_ratio = pos_ratio;
1081 }
1082
1083 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1084                                       unsigned long elapsed,
1085                                       unsigned long written)
1086 {
1087         const unsigned long period = roundup_pow_of_two(3 * HZ);
1088         unsigned long avg = wb->avg_write_bandwidth;
1089         unsigned long old = wb->write_bandwidth;
1090         u64 bw;
1091
1092         /*
1093          * bw = written * HZ / elapsed
1094          *
1095          *                   bw * elapsed + write_bandwidth * (period - elapsed)
1096          * write_bandwidth = ---------------------------------------------------
1097          *                                          period
1098          *
1099          * @written may have decreased due to account_page_redirty().
1100          * Avoid underflowing @bw calculation.
1101          */
1102         bw = written - min(written, wb->written_stamp);
1103         bw *= HZ;
1104         if (unlikely(elapsed > period)) {
1105                 do_div(bw, elapsed);
1106                 avg = bw;
1107                 goto out;
1108         }
1109         bw += (u64)wb->write_bandwidth * (period - elapsed);
1110         bw >>= ilog2(period);
1111
1112         /*
1113          * one more level of smoothing, for filtering out sudden spikes
1114          */
1115         if (avg > old && old >= (unsigned long)bw)
1116                 avg -= (avg - old) >> 3;
1117
1118         if (avg < old && old <= (unsigned long)bw)
1119                 avg += (old - avg) >> 3;
1120
1121 out:
1122         /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1123         avg = max(avg, 1LU);
1124         if (wb_has_dirty_io(wb)) {
1125                 long delta = avg - wb->avg_write_bandwidth;
1126                 WARN_ON_ONCE(atomic_long_add_return(delta,
1127                                         &wb->bdi->tot_write_bandwidth) <= 0);
1128         }
1129         wb->write_bandwidth = bw;
1130         wb->avg_write_bandwidth = avg;
1131 }
1132
1133 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1134 {
1135         struct wb_domain *dom = dtc_dom(dtc);
1136         unsigned long thresh = dtc->thresh;
1137         unsigned long limit = dom->dirty_limit;
1138
1139         /*
1140          * Follow up in one step.
1141          */
1142         if (limit < thresh) {
1143                 limit = thresh;
1144                 goto update;
1145         }
1146
1147         /*
1148          * Follow down slowly. Use the higher one as the target, because thresh
1149          * may drop below dirty. This is exactly the reason to introduce
1150          * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1151          */
1152         thresh = max(thresh, dtc->dirty);
1153         if (limit > thresh) {
1154                 limit -= (limit - thresh) >> 5;
1155                 goto update;
1156         }
1157         return;
1158 update:
1159         dom->dirty_limit = limit;
1160 }
1161
1162 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1163                                     unsigned long now)
1164 {
1165         struct wb_domain *dom = dtc_dom(dtc);
1166
1167         /*
1168          * check locklessly first to optimize away locking for the most time
1169          */
1170         if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1171                 return;
1172
1173         spin_lock(&dom->lock);
1174         if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1175                 update_dirty_limit(dtc);
1176                 dom->dirty_limit_tstamp = now;
1177         }
1178         spin_unlock(&dom->lock);
1179 }
1180
1181 /*
1182  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1183  *
1184  * Normal wb tasks will be curbed at or below it in long term.
1185  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1186  */
1187 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1188                                       unsigned long dirtied,
1189                                       unsigned long elapsed)
1190 {
1191         struct bdi_writeback *wb = dtc->wb;
1192         unsigned long dirty = dtc->dirty;
1193         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1194         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1195         unsigned long setpoint = (freerun + limit) / 2;
1196         unsigned long write_bw = wb->avg_write_bandwidth;
1197         unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1198         unsigned long dirty_rate;
1199         unsigned long task_ratelimit;
1200         unsigned long balanced_dirty_ratelimit;
1201         unsigned long step;
1202         unsigned long x;
1203         unsigned long shift;
1204
1205         /*
1206          * The dirty rate will match the writeout rate in long term, except
1207          * when dirty pages are truncated by userspace or re-dirtied by FS.
1208          */
1209         dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1210
1211         /*
1212          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1213          */
1214         task_ratelimit = (u64)dirty_ratelimit *
1215                                         dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1216         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1217
1218         /*
1219          * A linear estimation of the "balanced" throttle rate. The theory is,
1220          * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1221          * dirty_rate will be measured to be (N * task_ratelimit). So the below
1222          * formula will yield the balanced rate limit (write_bw / N).
1223          *
1224          * Note that the expanded form is not a pure rate feedback:
1225          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1226          * but also takes pos_ratio into account:
1227          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1228          *
1229          * (1) is not realistic because pos_ratio also takes part in balancing
1230          * the dirty rate.  Consider the state
1231          *      pos_ratio = 0.5                                              (3)
1232          *      rate = 2 * (write_bw / N)                                    (4)
1233          * If (1) is used, it will stuck in that state! Because each dd will
1234          * be throttled at
1235          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1236          * yielding
1237          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1238          * put (6) into (1) we get
1239          *      rate_(i+1) = rate_(i)                                        (7)
1240          *
1241          * So we end up using (2) to always keep
1242          *      rate_(i+1) ~= (write_bw / N)                                 (8)
1243          * regardless of the value of pos_ratio. As long as (8) is satisfied,
1244          * pos_ratio is able to drive itself to 1.0, which is not only where
1245          * the dirty count meet the setpoint, but also where the slope of
1246          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1247          */
1248         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1249                                            dirty_rate | 1);
1250         /*
1251          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1252          */
1253         if (unlikely(balanced_dirty_ratelimit > write_bw))
1254                 balanced_dirty_ratelimit = write_bw;
1255
1256         /*
1257          * We could safely do this and return immediately:
1258          *
1259          *      wb->dirty_ratelimit = balanced_dirty_ratelimit;
1260          *
1261          * However to get a more stable dirty_ratelimit, the below elaborated
1262          * code makes use of task_ratelimit to filter out singular points and
1263          * limit the step size.
1264          *
1265          * The below code essentially only uses the relative value of
1266          *
1267          *      task_ratelimit - dirty_ratelimit
1268          *      = (pos_ratio - 1) * dirty_ratelimit
1269          *
1270          * which reflects the direction and size of dirty position error.
1271          */
1272
1273         /*
1274          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1275          * task_ratelimit is on the same side of dirty_ratelimit, too.
1276          * For example, when
1277          * - dirty_ratelimit > balanced_dirty_ratelimit
1278          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1279          * lowering dirty_ratelimit will help meet both the position and rate
1280          * control targets. Otherwise, don't update dirty_ratelimit if it will
1281          * only help meet the rate target. After all, what the users ultimately
1282          * feel and care are stable dirty rate and small position error.
1283          *
1284          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1285          * and filter out the singular points of balanced_dirty_ratelimit. Which
1286          * keeps jumping around randomly and can even leap far away at times
1287          * due to the small 200ms estimation period of dirty_rate (we want to
1288          * keep that period small to reduce time lags).
1289          */
1290         step = 0;
1291
1292         /*
1293          * For strictlimit case, calculations above were based on wb counters
1294          * and limits (starting from pos_ratio = wb_position_ratio() and up to
1295          * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1296          * Hence, to calculate "step" properly, we have to use wb_dirty as
1297          * "dirty" and wb_setpoint as "setpoint".
1298          *
1299          * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1300          * it's possible that wb_thresh is close to zero due to inactivity
1301          * of backing device.
1302          */
1303         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1304                 dirty = dtc->wb_dirty;
1305                 if (dtc->wb_dirty < 8)
1306                         setpoint = dtc->wb_dirty + 1;
1307                 else
1308                         setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1309         }
1310
1311         if (dirty < setpoint) {
1312                 x = min3(wb->balanced_dirty_ratelimit,
1313                          balanced_dirty_ratelimit, task_ratelimit);
1314                 if (dirty_ratelimit < x)
1315                         step = x - dirty_ratelimit;
1316         } else {
1317                 x = max3(wb->balanced_dirty_ratelimit,
1318                          balanced_dirty_ratelimit, task_ratelimit);
1319                 if (dirty_ratelimit > x)
1320                         step = dirty_ratelimit - x;
1321         }
1322
1323         /*
1324          * Don't pursue 100% rate matching. It's impossible since the balanced
1325          * rate itself is constantly fluctuating. So decrease the track speed
1326          * when it gets close to the target. Helps eliminate pointless tremors.
1327          */
1328         shift = dirty_ratelimit / (2 * step + 1);
1329         if (shift < BITS_PER_LONG)
1330                 step = DIV_ROUND_UP(step >> shift, 8);
1331         else
1332                 step = 0;
1333
1334         if (dirty_ratelimit < balanced_dirty_ratelimit)
1335                 dirty_ratelimit += step;
1336         else
1337                 dirty_ratelimit -= step;
1338
1339         wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1340         wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1341
1342         trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1343 }
1344
1345 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1346                                   struct dirty_throttle_control *mdtc,
1347                                   unsigned long start_time,
1348                                   bool update_ratelimit)
1349 {
1350         struct bdi_writeback *wb = gdtc->wb;
1351         unsigned long now = jiffies;
1352         unsigned long elapsed = now - wb->bw_time_stamp;
1353         unsigned long dirtied;
1354         unsigned long written;
1355
1356         lockdep_assert_held(&wb->list_lock);
1357
1358         /*
1359          * rate-limit, only update once every 200ms.
1360          */
1361         if (elapsed < BANDWIDTH_INTERVAL)
1362                 return;
1363
1364         dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1365         written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1366
1367         /*
1368          * Skip quiet periods when disk bandwidth is under-utilized.
1369          * (at least 1s idle time between two flusher runs)
1370          */
1371         if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1372                 goto snapshot;
1373
1374         if (update_ratelimit) {
1375                 domain_update_bandwidth(gdtc, now);
1376                 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1377
1378                 /*
1379                  * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1380                  * compiler has no way to figure that out.  Help it.
1381                  */
1382                 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1383                         domain_update_bandwidth(mdtc, now);
1384                         wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1385                 }
1386         }
1387         wb_update_write_bandwidth(wb, elapsed, written);
1388
1389 snapshot:
1390         wb->dirtied_stamp = dirtied;
1391         wb->written_stamp = written;
1392         wb->bw_time_stamp = now;
1393 }
1394
1395 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1396 {
1397         struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1398
1399         __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1400 }
1401
1402 /*
1403  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1404  * will look to see if it needs to start dirty throttling.
1405  *
1406  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1407  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1408  * (the number of pages we may dirty without exceeding the dirty limits).
1409  */
1410 static unsigned long dirty_poll_interval(unsigned long dirty,
1411                                          unsigned long thresh)
1412 {
1413         if (thresh > dirty)
1414                 return 1UL << (ilog2(thresh - dirty) >> 1);
1415
1416         return 1;
1417 }
1418
1419 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1420                                   unsigned long wb_dirty)
1421 {
1422         unsigned long bw = wb->avg_write_bandwidth;
1423         unsigned long t;
1424
1425         /*
1426          * Limit pause time for small memory systems. If sleeping for too long
1427          * time, a small pool of dirty/writeback pages may go empty and disk go
1428          * idle.
1429          *
1430          * 8 serves as the safety ratio.
1431          */
1432         t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1433         t++;
1434
1435         return min_t(unsigned long, t, MAX_PAUSE);
1436 }
1437
1438 static long wb_min_pause(struct bdi_writeback *wb,
1439                          long max_pause,
1440                          unsigned long task_ratelimit,
1441                          unsigned long dirty_ratelimit,
1442                          int *nr_dirtied_pause)
1443 {
1444         long hi = ilog2(wb->avg_write_bandwidth);
1445         long lo = ilog2(wb->dirty_ratelimit);
1446         long t;         /* target pause */
1447         long pause;     /* estimated next pause */
1448         int pages;      /* target nr_dirtied_pause */
1449
1450         /* target for 10ms pause on 1-dd case */
1451         t = max(1, HZ / 100);
1452
1453         /*
1454          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1455          * overheads.
1456          *
1457          * (N * 10ms) on 2^N concurrent tasks.
1458          */
1459         if (hi > lo)
1460                 t += (hi - lo) * (10 * HZ) / 1024;
1461
1462         /*
1463          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1464          * on the much more stable dirty_ratelimit. However the next pause time
1465          * will be computed based on task_ratelimit and the two rate limits may
1466          * depart considerably at some time. Especially if task_ratelimit goes
1467          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1468          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1469          * result task_ratelimit won't be executed faithfully, which could
1470          * eventually bring down dirty_ratelimit.
1471          *
1472          * We apply two rules to fix it up:
1473          * 1) try to estimate the next pause time and if necessary, use a lower
1474          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1475          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1476          * 2) limit the target pause time to max_pause/2, so that the normal
1477          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1478          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1479          */
1480         t = min(t, 1 + max_pause / 2);
1481         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482
1483         /*
1484          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1485          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1486          * When the 16 consecutive reads are often interrupted by some dirty
1487          * throttling pause during the async writes, cfq will go into idles
1488          * (deadline is fine). So push nr_dirtied_pause as high as possible
1489          * until reaches DIRTY_POLL_THRESH=32 pages.
1490          */
1491         if (pages < DIRTY_POLL_THRESH) {
1492                 t = max_pause;
1493                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1494                 if (pages > DIRTY_POLL_THRESH) {
1495                         pages = DIRTY_POLL_THRESH;
1496                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1497                 }
1498         }
1499
1500         pause = HZ * pages / (task_ratelimit + 1);
1501         if (pause > max_pause) {
1502                 t = max_pause;
1503                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1504         }
1505
1506         *nr_dirtied_pause = pages;
1507         /*
1508          * The minimal pause time will normally be half the target pause time.
1509          */
1510         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1511 }
1512
1513 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1514 {
1515         struct bdi_writeback *wb = dtc->wb;
1516         unsigned long wb_reclaimable;
1517
1518         /*
1519          * wb_thresh is not treated as some limiting factor as
1520          * dirty_thresh, due to reasons
1521          * - in JBOD setup, wb_thresh can fluctuate a lot
1522          * - in a system with HDD and USB key, the USB key may somehow
1523          *   go into state (wb_dirty >> wb_thresh) either because
1524          *   wb_dirty starts high, or because wb_thresh drops low.
1525          *   In this case we don't want to hard throttle the USB key
1526          *   dirtiers for 100 seconds until wb_dirty drops under
1527          *   wb_thresh. Instead the auxiliary wb control line in
1528          *   wb_position_ratio() will let the dirtier task progress
1529          *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1530          */
1531         dtc->wb_thresh = __wb_calc_thresh(dtc);
1532         dtc->wb_bg_thresh = dtc->thresh ?
1533                 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1534
1535         /*
1536          * In order to avoid the stacked BDI deadlock we need
1537          * to ensure we accurately count the 'dirty' pages when
1538          * the threshold is low.
1539          *
1540          * Otherwise it would be possible to get thresh+n pages
1541          * reported dirty, even though there are thresh-m pages
1542          * actually dirty; with m+n sitting in the percpu
1543          * deltas.
1544          */
1545         if (dtc->wb_thresh < 2 * wb_stat_error()) {
1546                 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1547                 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1548         } else {
1549                 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1550                 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1551         }
1552 }
1553
1554 /*
1555  * balance_dirty_pages() must be called by processes which are generating dirty
1556  * data.  It looks at the number of dirty pages in the machine and will force
1557  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1558  * If we're over `background_thresh' then the writeback threads are woken to
1559  * perform some writeout.
1560  */
1561 static void balance_dirty_pages(struct bdi_writeback *wb,
1562                                 unsigned long pages_dirtied)
1563 {
1564         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1565         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1566         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1567         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1568                                                      &mdtc_stor : NULL;
1569         struct dirty_throttle_control *sdtc;
1570         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1571         long period;
1572         long pause;
1573         long max_pause;
1574         long min_pause;
1575         int nr_dirtied_pause;
1576         bool dirty_exceeded = false;
1577         unsigned long task_ratelimit;
1578         unsigned long dirty_ratelimit;
1579         struct backing_dev_info *bdi = wb->bdi;
1580         bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1581         unsigned long start_time = jiffies;
1582
1583         for (;;) {
1584                 unsigned long now = jiffies;
1585                 unsigned long dirty, thresh, bg_thresh;
1586                 unsigned long m_dirty = 0;      /* stop bogus uninit warnings */
1587                 unsigned long m_thresh = 0;
1588                 unsigned long m_bg_thresh = 0;
1589
1590                 /*
1591                  * Unstable writes are a feature of certain networked
1592                  * filesystems (i.e. NFS) in which data may have been
1593                  * written to the server's write cache, but has not yet
1594                  * been flushed to permanent storage.
1595                  */
1596                 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1597                                         global_node_page_state(NR_UNSTABLE_NFS);
1598                 gdtc->avail = global_dirtyable_memory();
1599                 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1600
1601                 domain_dirty_limits(gdtc);
1602
1603                 if (unlikely(strictlimit)) {
1604                         wb_dirty_limits(gdtc);
1605
1606                         dirty = gdtc->wb_dirty;
1607                         thresh = gdtc->wb_thresh;
1608                         bg_thresh = gdtc->wb_bg_thresh;
1609                 } else {
1610                         dirty = gdtc->dirty;
1611                         thresh = gdtc->thresh;
1612                         bg_thresh = gdtc->bg_thresh;
1613                 }
1614
1615                 if (mdtc) {
1616                         unsigned long filepages, headroom, writeback;
1617
1618                         /*
1619                          * If @wb belongs to !root memcg, repeat the same
1620                          * basic calculations for the memcg domain.
1621                          */
1622                         mem_cgroup_wb_stats(wb, &filepages, &headroom,
1623                                             &mdtc->dirty, &writeback);
1624                         mdtc->dirty += writeback;
1625                         mdtc_calc_avail(mdtc, filepages, headroom);
1626
1627                         domain_dirty_limits(mdtc);
1628
1629                         if (unlikely(strictlimit)) {
1630                                 wb_dirty_limits(mdtc);
1631                                 m_dirty = mdtc->wb_dirty;
1632                                 m_thresh = mdtc->wb_thresh;
1633                                 m_bg_thresh = mdtc->wb_bg_thresh;
1634                         } else {
1635                                 m_dirty = mdtc->dirty;
1636                                 m_thresh = mdtc->thresh;
1637                                 m_bg_thresh = mdtc->bg_thresh;
1638                         }
1639                 }
1640
1641                 /*
1642                  * Throttle it only when the background writeback cannot
1643                  * catch-up. This avoids (excessively) small writeouts
1644                  * when the wb limits are ramping up in case of !strictlimit.
1645                  *
1646                  * In strictlimit case make decision based on the wb counters
1647                  * and limits. Small writeouts when the wb limits are ramping
1648                  * up are the price we consciously pay for strictlimit-ing.
1649                  *
1650                  * If memcg domain is in effect, @dirty should be under
1651                  * both global and memcg freerun ceilings.
1652                  */
1653                 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1654                     (!mdtc ||
1655                      m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1656                         unsigned long intv = dirty_poll_interval(dirty, thresh);
1657                         unsigned long m_intv = ULONG_MAX;
1658
1659                         current->dirty_paused_when = now;
1660                         current->nr_dirtied = 0;
1661                         if (mdtc)
1662                                 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1663                         current->nr_dirtied_pause = min(intv, m_intv);
1664                         break;
1665                 }
1666
1667                 if (unlikely(!writeback_in_progress(wb)))
1668                         wb_start_background_writeback(wb);
1669
1670                 /*
1671                  * Calculate global domain's pos_ratio and select the
1672                  * global dtc by default.
1673                  */
1674                 if (!strictlimit)
1675                         wb_dirty_limits(gdtc);
1676
1677                 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1678                         ((gdtc->dirty > gdtc->thresh) || strictlimit);
1679
1680                 wb_position_ratio(gdtc);
1681                 sdtc = gdtc;
1682
1683                 if (mdtc) {
1684                         /*
1685                          * If memcg domain is in effect, calculate its
1686                          * pos_ratio.  @wb should satisfy constraints from
1687                          * both global and memcg domains.  Choose the one
1688                          * w/ lower pos_ratio.
1689                          */
1690                         if (!strictlimit)
1691                                 wb_dirty_limits(mdtc);
1692
1693                         dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1694                                 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1695
1696                         wb_position_ratio(mdtc);
1697                         if (mdtc->pos_ratio < gdtc->pos_ratio)
1698                                 sdtc = mdtc;
1699                 }
1700
1701                 if (dirty_exceeded && !wb->dirty_exceeded)
1702                         wb->dirty_exceeded = 1;
1703
1704                 if (time_is_before_jiffies(wb->bw_time_stamp +
1705                                            BANDWIDTH_INTERVAL)) {
1706                         spin_lock(&wb->list_lock);
1707                         __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1708                         spin_unlock(&wb->list_lock);
1709                 }
1710
1711                 /* throttle according to the chosen dtc */
1712                 dirty_ratelimit = wb->dirty_ratelimit;
1713                 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1714                                                         RATELIMIT_CALC_SHIFT;
1715                 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1716                 min_pause = wb_min_pause(wb, max_pause,
1717                                          task_ratelimit, dirty_ratelimit,
1718                                          &nr_dirtied_pause);
1719
1720                 if (unlikely(task_ratelimit == 0)) {
1721                         period = max_pause;
1722                         pause = max_pause;
1723                         goto pause;
1724                 }
1725                 period = HZ * pages_dirtied / task_ratelimit;
1726                 pause = period;
1727                 if (current->dirty_paused_when)
1728                         pause -= now - current->dirty_paused_when;
1729                 /*
1730                  * For less than 1s think time (ext3/4 may block the dirtier
1731                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1732                  * however at much less frequency), try to compensate it in
1733                  * future periods by updating the virtual time; otherwise just
1734                  * do a reset, as it may be a light dirtier.
1735                  */
1736                 if (pause < min_pause) {
1737                         trace_balance_dirty_pages(wb,
1738                                                   sdtc->thresh,
1739                                                   sdtc->bg_thresh,
1740                                                   sdtc->dirty,
1741                                                   sdtc->wb_thresh,
1742                                                   sdtc->wb_dirty,
1743                                                   dirty_ratelimit,
1744                                                   task_ratelimit,
1745                                                   pages_dirtied,
1746                                                   period,
1747                                                   min(pause, 0L),
1748                                                   start_time);
1749                         if (pause < -HZ) {
1750                                 current->dirty_paused_when = now;
1751                                 current->nr_dirtied = 0;
1752                         } else if (period) {
1753                                 current->dirty_paused_when += period;
1754                                 current->nr_dirtied = 0;
1755                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1756                                 current->nr_dirtied_pause += pages_dirtied;
1757                         break;
1758                 }
1759                 if (unlikely(pause > max_pause)) {
1760                         /* for occasional dropped task_ratelimit */
1761                         now += min(pause - max_pause, max_pause);
1762                         pause = max_pause;
1763                 }
1764
1765 pause:
1766                 trace_balance_dirty_pages(wb,
1767                                           sdtc->thresh,
1768                                           sdtc->bg_thresh,
1769                                           sdtc->dirty,
1770                                           sdtc->wb_thresh,
1771                                           sdtc->wb_dirty,
1772                                           dirty_ratelimit,
1773                                           task_ratelimit,
1774                                           pages_dirtied,
1775                                           period,
1776                                           pause,
1777                                           start_time);
1778                 __set_current_state(TASK_KILLABLE);
1779                 wb->dirty_sleep = now;
1780                 io_schedule_timeout(pause);
1781
1782                 current->dirty_paused_when = now + pause;
1783                 current->nr_dirtied = 0;
1784                 current->nr_dirtied_pause = nr_dirtied_pause;
1785
1786                 /*
1787                  * This is typically equal to (dirty < thresh) and can also
1788                  * keep "1000+ dd on a slow USB stick" under control.
1789                  */
1790                 if (task_ratelimit)
1791                         break;
1792
1793                 /*
1794                  * In the case of an unresponding NFS server and the NFS dirty
1795                  * pages exceeds dirty_thresh, give the other good wb's a pipe
1796                  * to go through, so that tasks on them still remain responsive.
1797                  *
1798                  * In theory 1 page is enough to keep the consumer-producer
1799                  * pipe going: the flusher cleans 1 page => the task dirties 1
1800                  * more page. However wb_dirty has accounting errors.  So use
1801                  * the larger and more IO friendly wb_stat_error.
1802                  */
1803                 if (sdtc->wb_dirty <= wb_stat_error())
1804                         break;
1805
1806                 if (fatal_signal_pending(current))
1807                         break;
1808         }
1809
1810         if (!dirty_exceeded && wb->dirty_exceeded)
1811                 wb->dirty_exceeded = 0;
1812
1813         if (writeback_in_progress(wb))
1814                 return;
1815
1816         /*
1817          * In laptop mode, we wait until hitting the higher threshold before
1818          * starting background writeout, and then write out all the way down
1819          * to the lower threshold.  So slow writers cause minimal disk activity.
1820          *
1821          * In normal mode, we start background writeout at the lower
1822          * background_thresh, to keep the amount of dirty memory low.
1823          */
1824         if (laptop_mode)
1825                 return;
1826
1827         if (nr_reclaimable > gdtc->bg_thresh)
1828                 wb_start_background_writeback(wb);
1829 }
1830
1831 static DEFINE_PER_CPU(int, bdp_ratelimits);
1832
1833 /*
1834  * Normal tasks are throttled by
1835  *      loop {
1836  *              dirty tsk->nr_dirtied_pause pages;
1837  *              take a snap in balance_dirty_pages();
1838  *      }
1839  * However there is a worst case. If every task exit immediately when dirtied
1840  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1841  * called to throttle the page dirties. The solution is to save the not yet
1842  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1843  * randomly into the running tasks. This works well for the above worst case,
1844  * as the new task will pick up and accumulate the old task's leaked dirty
1845  * count and eventually get throttled.
1846  */
1847 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1848
1849 /**
1850  * balance_dirty_pages_ratelimited - balance dirty memory state
1851  * @mapping: address_space which was dirtied
1852  *
1853  * Processes which are dirtying memory should call in here once for each page
1854  * which was newly dirtied.  The function will periodically check the system's
1855  * dirty state and will initiate writeback if needed.
1856  *
1857  * On really big machines, get_writeback_state is expensive, so try to avoid
1858  * calling it too often (ratelimiting).  But once we're over the dirty memory
1859  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1860  * from overshooting the limit by (ratelimit_pages) each.
1861  */
1862 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1863 {
1864         struct inode *inode = mapping->host;
1865         struct backing_dev_info *bdi = inode_to_bdi(inode);
1866         struct bdi_writeback *wb = NULL;
1867         int ratelimit;
1868         int *p;
1869
1870         if (!bdi_cap_account_dirty(bdi))
1871                 return;
1872
1873         if (inode_cgwb_enabled(inode))
1874                 wb = wb_get_create_current(bdi, GFP_KERNEL);
1875         if (!wb)
1876                 wb = &bdi->wb;
1877
1878         ratelimit = current->nr_dirtied_pause;
1879         if (wb->dirty_exceeded)
1880                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1881
1882         preempt_disable();
1883         /*
1884          * This prevents one CPU to accumulate too many dirtied pages without
1885          * calling into balance_dirty_pages(), which can happen when there are
1886          * 1000+ tasks, all of them start dirtying pages at exactly the same
1887          * time, hence all honoured too large initial task->nr_dirtied_pause.
1888          */
1889         p =  this_cpu_ptr(&bdp_ratelimits);
1890         if (unlikely(current->nr_dirtied >= ratelimit))
1891                 *p = 0;
1892         else if (unlikely(*p >= ratelimit_pages)) {
1893                 *p = 0;
1894                 ratelimit = 0;
1895         }
1896         /*
1897          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1898          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1899          * the dirty throttling and livelock other long-run dirtiers.
1900          */
1901         p = this_cpu_ptr(&dirty_throttle_leaks);
1902         if (*p > 0 && current->nr_dirtied < ratelimit) {
1903                 unsigned long nr_pages_dirtied;
1904                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1905                 *p -= nr_pages_dirtied;
1906                 current->nr_dirtied += nr_pages_dirtied;
1907         }
1908         preempt_enable();
1909
1910         if (unlikely(current->nr_dirtied >= ratelimit))
1911                 balance_dirty_pages(wb, current->nr_dirtied);
1912
1913         wb_put(wb);
1914 }
1915 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1916
1917 /**
1918  * wb_over_bg_thresh - does @wb need to be written back?
1919  * @wb: bdi_writeback of interest
1920  *
1921  * Determines whether background writeback should keep writing @wb or it's
1922  * clean enough.  Returns %true if writeback should continue.
1923  */
1924 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1925 {
1926         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1927         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1928         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1929         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1930                                                      &mdtc_stor : NULL;
1931
1932         /*
1933          * Similar to balance_dirty_pages() but ignores pages being written
1934          * as we're trying to decide whether to put more under writeback.
1935          */
1936         gdtc->avail = global_dirtyable_memory();
1937         gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1938                       global_node_page_state(NR_UNSTABLE_NFS);
1939         domain_dirty_limits(gdtc);
1940
1941         if (gdtc->dirty > gdtc->bg_thresh)
1942                 return true;
1943
1944         if (wb_stat(wb, WB_RECLAIMABLE) >
1945             wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1946                 return true;
1947
1948         if (mdtc) {
1949                 unsigned long filepages, headroom, writeback;
1950
1951                 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1952                                     &writeback);
1953                 mdtc_calc_avail(mdtc, filepages, headroom);
1954                 domain_dirty_limits(mdtc);      /* ditto, ignore writeback */
1955
1956                 if (mdtc->dirty > mdtc->bg_thresh)
1957                         return true;
1958
1959                 if (wb_stat(wb, WB_RECLAIMABLE) >
1960                     wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1961                         return true;
1962         }
1963
1964         return false;
1965 }
1966
1967 /*
1968  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1969  */
1970 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1971         void __user *buffer, size_t *length, loff_t *ppos)
1972 {
1973         unsigned int old_interval = dirty_writeback_interval;
1974         int ret;
1975
1976         ret = proc_dointvec(table, write, buffer, length, ppos);
1977
1978         /*
1979          * Writing 0 to dirty_writeback_interval will disable periodic writeback
1980          * and a different non-zero value will wakeup the writeback threads.
1981          * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1982          * iterate over all bdis and wbs.
1983          * The reason we do this is to make the change take effect immediately.
1984          */
1985         if (!ret && write && dirty_writeback_interval &&
1986                 dirty_writeback_interval != old_interval)
1987                 wakeup_flusher_threads(WB_REASON_PERIODIC);
1988
1989         return ret;
1990 }
1991
1992 #ifdef CONFIG_BLOCK
1993 void laptop_mode_timer_fn(struct timer_list *t)
1994 {
1995         struct backing_dev_info *backing_dev_info =
1996                 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
1997
1998         wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
1999 }
2000
2001 /*
2002  * We've spun up the disk and we're in laptop mode: schedule writeback
2003  * of all dirty data a few seconds from now.  If the flush is already scheduled
2004  * then push it back - the user is still using the disk.
2005  */
2006 void laptop_io_completion(struct backing_dev_info *info)
2007 {
2008         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2009 }
2010
2011 /*
2012  * We're in laptop mode and we've just synced. The sync's writes will have
2013  * caused another writeback to be scheduled by laptop_io_completion.
2014  * Nothing needs to be written back anymore, so we unschedule the writeback.
2015  */
2016 void laptop_sync_completion(void)
2017 {
2018         struct backing_dev_info *bdi;
2019
2020         rcu_read_lock();
2021
2022         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2023                 del_timer(&bdi->laptop_mode_wb_timer);
2024
2025         rcu_read_unlock();
2026 }
2027 #endif
2028
2029 /*
2030  * If ratelimit_pages is too high then we can get into dirty-data overload
2031  * if a large number of processes all perform writes at the same time.
2032  * If it is too low then SMP machines will call the (expensive)
2033  * get_writeback_state too often.
2034  *
2035  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2036  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2037  * thresholds.
2038  */
2039
2040 void writeback_set_ratelimit(void)
2041 {
2042         struct wb_domain *dom = &global_wb_domain;
2043         unsigned long background_thresh;
2044         unsigned long dirty_thresh;
2045
2046         global_dirty_limits(&background_thresh, &dirty_thresh);
2047         dom->dirty_limit = dirty_thresh;
2048         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2049         if (ratelimit_pages < 16)
2050                 ratelimit_pages = 16;
2051 }
2052
2053 static int page_writeback_cpu_online(unsigned int cpu)
2054 {
2055         writeback_set_ratelimit();
2056         return 0;
2057 }
2058
2059 /*
2060  * Called early on to tune the page writeback dirty limits.
2061  *
2062  * We used to scale dirty pages according to how total memory
2063  * related to pages that could be allocated for buffers (by
2064  * comparing nr_free_buffer_pages() to vm_total_pages.
2065  *
2066  * However, that was when we used "dirty_ratio" to scale with
2067  * all memory, and we don't do that any more. "dirty_ratio"
2068  * is now applied to total non-HIGHPAGE memory (by subtracting
2069  * totalhigh_pages from vm_total_pages), and as such we can't
2070  * get into the old insane situation any more where we had
2071  * large amounts of dirty pages compared to a small amount of
2072  * non-HIGHMEM memory.
2073  *
2074  * But we might still want to scale the dirty_ratio by how
2075  * much memory the box has..
2076  */
2077 void __init page_writeback_init(void)
2078 {
2079         BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2080
2081         cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2082                           page_writeback_cpu_online, NULL);
2083         cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2084                           page_writeback_cpu_online);
2085 }
2086
2087 /**
2088  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2089  * @mapping: address space structure to write
2090  * @start: starting page index
2091  * @end: ending page index (inclusive)
2092  *
2093  * This function scans the page range from @start to @end (inclusive) and tags
2094  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2095  * that write_cache_pages (or whoever calls this function) will then use
2096  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2097  * used to avoid livelocking of writeback by a process steadily creating new
2098  * dirty pages in the file (thus it is important for this function to be quick
2099  * so that it can tag pages faster than a dirtying process can create them).
2100  */
2101 /*
2102  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce the i_pages lock
2103  * latency.
2104  */
2105 void tag_pages_for_writeback(struct address_space *mapping,
2106                              pgoff_t start, pgoff_t end)
2107 {
2108 #define WRITEBACK_TAG_BATCH 4096
2109         unsigned long tagged = 0;
2110         struct radix_tree_iter iter;
2111         void **slot;
2112
2113         xa_lock_irq(&mapping->i_pages);
2114         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start,
2115                                                         PAGECACHE_TAG_DIRTY) {
2116                 if (iter.index > end)
2117                         break;
2118                 radix_tree_iter_tag_set(&mapping->i_pages, &iter,
2119                                                         PAGECACHE_TAG_TOWRITE);
2120                 tagged++;
2121                 if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2122                         continue;
2123                 slot = radix_tree_iter_resume(slot, &iter);
2124                 xa_unlock_irq(&mapping->i_pages);
2125                 cond_resched();
2126                 xa_lock_irq(&mapping->i_pages);
2127         }
2128         xa_unlock_irq(&mapping->i_pages);
2129 }
2130 EXPORT_SYMBOL(tag_pages_for_writeback);
2131
2132 /**
2133  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2134  * @mapping: address space structure to write
2135  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2136  * @writepage: function called for each page
2137  * @data: data passed to writepage function
2138  *
2139  * If a page is already under I/O, write_cache_pages() skips it, even
2140  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2141  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2142  * and msync() need to guarantee that all the data which was dirty at the time
2143  * the call was made get new I/O started against them.  If wbc->sync_mode is
2144  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2145  * existing IO to complete.
2146  *
2147  * To avoid livelocks (when other process dirties new pages), we first tag
2148  * pages which should be written back with TOWRITE tag and only then start
2149  * writing them. For data-integrity sync we have to be careful so that we do
2150  * not miss some pages (e.g., because some other process has cleared TOWRITE
2151  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2152  * by the process clearing the DIRTY tag (and submitting the page for IO).
2153  */
2154 int write_cache_pages(struct address_space *mapping,
2155                       struct writeback_control *wbc, writepage_t writepage,
2156                       void *data)
2157 {
2158         int ret = 0;
2159         int done = 0;
2160         struct pagevec pvec;
2161         int nr_pages;
2162         pgoff_t uninitialized_var(writeback_index);
2163         pgoff_t index;
2164         pgoff_t end;            /* Inclusive */
2165         pgoff_t done_index;
2166         int cycled;
2167         int range_whole = 0;
2168         int tag;
2169
2170         pagevec_init(&pvec);
2171         if (wbc->range_cyclic) {
2172                 writeback_index = mapping->writeback_index; /* prev offset */
2173                 index = writeback_index;
2174                 if (index == 0)
2175                         cycled = 1;
2176                 else
2177                         cycled = 0;
2178                 end = -1;
2179         } else {
2180                 index = wbc->range_start >> PAGE_SHIFT;
2181                 end = wbc->range_end >> PAGE_SHIFT;
2182                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2183                         range_whole = 1;
2184                 cycled = 1; /* ignore range_cyclic tests */
2185         }
2186         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2187                 tag = PAGECACHE_TAG_TOWRITE;
2188         else
2189                 tag = PAGECACHE_TAG_DIRTY;
2190 retry:
2191         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2192                 tag_pages_for_writeback(mapping, index, end);
2193         done_index = index;
2194         while (!done && (index <= end)) {
2195                 int i;
2196
2197                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2198                                 tag);
2199                 if (nr_pages == 0)
2200                         break;
2201
2202                 for (i = 0; i < nr_pages; i++) {
2203                         struct page *page = pvec.pages[i];
2204
2205                         done_index = page->index;
2206
2207                         lock_page(page);
2208
2209                         /*
2210                          * Page truncated or invalidated. We can freely skip it
2211                          * then, even for data integrity operations: the page
2212                          * has disappeared concurrently, so there could be no
2213                          * real expectation of this data interity operation
2214                          * even if there is now a new, dirty page at the same
2215                          * pagecache address.
2216                          */
2217                         if (unlikely(page->mapping != mapping)) {
2218 continue_unlock:
2219                                 unlock_page(page);
2220                                 continue;
2221                         }
2222
2223                         if (!PageDirty(page)) {
2224                                 /* someone wrote it for us */
2225                                 goto continue_unlock;
2226                         }
2227
2228                         if (PageWriteback(page)) {
2229                                 if (wbc->sync_mode != WB_SYNC_NONE)
2230                                         wait_on_page_writeback(page);
2231                                 else
2232                                         goto continue_unlock;
2233                         }
2234
2235                         BUG_ON(PageWriteback(page));
2236                         if (!clear_page_dirty_for_io(page))
2237                                 goto continue_unlock;
2238
2239                         trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2240                         ret = (*writepage)(page, wbc, data);
2241                         if (unlikely(ret)) {
2242                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2243                                         unlock_page(page);
2244                                         ret = 0;
2245                                 } else {
2246                                         /*
2247                                          * done_index is set past this page,
2248                                          * so media errors will not choke
2249                                          * background writeout for the entire
2250                                          * file. This has consequences for
2251                                          * range_cyclic semantics (ie. it may
2252                                          * not be suitable for data integrity
2253                                          * writeout).
2254                                          */
2255                                         done_index = page->index + 1;
2256                                         done = 1;
2257                                         break;
2258                                 }
2259                         }
2260
2261                         /*
2262                          * We stop writing back only if we are not doing
2263                          * integrity sync. In case of integrity sync we have to
2264                          * keep going until we have written all the pages
2265                          * we tagged for writeback prior to entering this loop.
2266                          */
2267                         if (--wbc->nr_to_write <= 0 &&
2268                             wbc->sync_mode == WB_SYNC_NONE) {
2269                                 done = 1;
2270                                 break;
2271                         }
2272                 }
2273                 pagevec_release(&pvec);
2274                 cond_resched();
2275         }
2276         if (!cycled && !done) {
2277                 /*
2278                  * range_cyclic:
2279                  * We hit the last page and there is more work to be done: wrap
2280                  * back to the start of the file
2281                  */
2282                 cycled = 1;
2283                 index = 0;
2284                 end = writeback_index - 1;
2285                 goto retry;
2286         }
2287         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2288                 mapping->writeback_index = done_index;
2289
2290         return ret;
2291 }
2292 EXPORT_SYMBOL(write_cache_pages);
2293
2294 /*
2295  * Function used by generic_writepages to call the real writepage
2296  * function and set the mapping flags on error
2297  */
2298 static int __writepage(struct page *page, struct writeback_control *wbc,
2299                        void *data)
2300 {
2301         struct address_space *mapping = data;
2302         int ret = mapping->a_ops->writepage(page, wbc);
2303         mapping_set_error(mapping, ret);
2304         return ret;
2305 }
2306
2307 /**
2308  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2309  * @mapping: address space structure to write
2310  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2311  *
2312  * This is a library function, which implements the writepages()
2313  * address_space_operation.
2314  */
2315 int generic_writepages(struct address_space *mapping,
2316                        struct writeback_control *wbc)
2317 {
2318         struct blk_plug plug;
2319         int ret;
2320
2321         /* deal with chardevs and other special file */
2322         if (!mapping->a_ops->writepage)
2323                 return 0;
2324
2325         blk_start_plug(&plug);
2326         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2327         blk_finish_plug(&plug);
2328         return ret;
2329 }
2330
2331 EXPORT_SYMBOL(generic_writepages);
2332
2333 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2334 {
2335         int ret;
2336
2337         if (wbc->nr_to_write <= 0)
2338                 return 0;
2339         while (1) {
2340                 if (mapping->a_ops->writepages)
2341                         ret = mapping->a_ops->writepages(mapping, wbc);
2342                 else
2343                         ret = generic_writepages(mapping, wbc);
2344                 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2345                         break;
2346                 cond_resched();
2347                 congestion_wait(BLK_RW_ASYNC, HZ/50);
2348         }
2349         return ret;
2350 }
2351
2352 /**
2353  * write_one_page - write out a single page and wait on I/O
2354  * @page: the page to write
2355  *
2356  * The page must be locked by the caller and will be unlocked upon return.
2357  *
2358  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2359  * function returns.
2360  */
2361 int write_one_page(struct page *page)
2362 {
2363         struct address_space *mapping = page->mapping;
2364         int ret = 0;
2365         struct writeback_control wbc = {
2366                 .sync_mode = WB_SYNC_ALL,
2367                 .nr_to_write = 1,
2368         };
2369
2370         BUG_ON(!PageLocked(page));
2371
2372         wait_on_page_writeback(page);
2373
2374         if (clear_page_dirty_for_io(page)) {
2375                 get_page(page);
2376                 ret = mapping->a_ops->writepage(page, &wbc);
2377                 if (ret == 0)
2378                         wait_on_page_writeback(page);
2379                 put_page(page);
2380         } else {
2381                 unlock_page(page);
2382         }
2383
2384         if (!ret)
2385                 ret = filemap_check_errors(mapping);
2386         return ret;
2387 }
2388 EXPORT_SYMBOL(write_one_page);
2389
2390 /*
2391  * For address_spaces which do not use buffers nor write back.
2392  */
2393 int __set_page_dirty_no_writeback(struct page *page)
2394 {
2395         if (!PageDirty(page))
2396                 return !TestSetPageDirty(page);
2397         return 0;
2398 }
2399
2400 /*
2401  * Helper function for set_page_dirty family.
2402  *
2403  * Caller must hold lock_page_memcg().
2404  *
2405  * NOTE: This relies on being atomic wrt interrupts.
2406  */
2407 void account_page_dirtied(struct page *page, struct address_space *mapping)
2408 {
2409         struct inode *inode = mapping->host;
2410
2411         trace_writeback_dirty_page(page, mapping);
2412
2413         if (mapping_cap_account_dirty(mapping)) {
2414                 struct bdi_writeback *wb;
2415
2416                 inode_attach_wb(inode, page);
2417                 wb = inode_to_wb(inode);
2418
2419                 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2420                 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2421                 __inc_node_page_state(page, NR_DIRTIED);
2422                 inc_wb_stat(wb, WB_RECLAIMABLE);
2423                 inc_wb_stat(wb, WB_DIRTIED);
2424                 task_io_account_write(PAGE_SIZE);
2425                 current->nr_dirtied++;
2426                 this_cpu_inc(bdp_ratelimits);
2427         }
2428 }
2429 EXPORT_SYMBOL(account_page_dirtied);
2430
2431 /*
2432  * Helper function for deaccounting dirty page without writeback.
2433  *
2434  * Caller must hold lock_page_memcg().
2435  */
2436 void account_page_cleaned(struct page *page, struct address_space *mapping,
2437                           struct bdi_writeback *wb)
2438 {
2439         if (mapping_cap_account_dirty(mapping)) {
2440                 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2441                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2442                 dec_wb_stat(wb, WB_RECLAIMABLE);
2443                 task_io_account_cancelled_write(PAGE_SIZE);
2444         }
2445 }
2446
2447 /*
2448  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2449  * its radix tree.
2450  *
2451  * This is also used when a single buffer is being dirtied: we want to set the
2452  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2453  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454  *
2455  * The caller must ensure this doesn't race with truncation.  Most will simply
2456  * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457  * the pte lock held, which also locks out truncation.
2458  */
2459 int __set_page_dirty_nobuffers(struct page *page)
2460 {
2461         lock_page_memcg(page);
2462         if (!TestSetPageDirty(page)) {
2463                 struct address_space *mapping = page_mapping(page);
2464                 unsigned long flags;
2465
2466                 if (!mapping) {
2467                         unlock_page_memcg(page);
2468                         return 1;
2469                 }
2470
2471                 xa_lock_irqsave(&mapping->i_pages, flags);
2472                 BUG_ON(page_mapping(page) != mapping);
2473                 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2474                 account_page_dirtied(page, mapping);
2475                 radix_tree_tag_set(&mapping->i_pages, page_index(page),
2476                                    PAGECACHE_TAG_DIRTY);
2477                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2478                 unlock_page_memcg(page);
2479
2480                 if (mapping->host) {
2481                         /* !PageAnon && !swapper_space */
2482                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2483                 }
2484                 return 1;
2485         }
2486         unlock_page_memcg(page);
2487         return 0;
2488 }
2489 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2490
2491 /*
2492  * Call this whenever redirtying a page, to de-account the dirty counters
2493  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2494  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2495  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2496  * control.
2497  */
2498 void account_page_redirty(struct page *page)
2499 {
2500         struct address_space *mapping = page->mapping;
2501
2502         if (mapping && mapping_cap_account_dirty(mapping)) {
2503                 struct inode *inode = mapping->host;
2504                 struct bdi_writeback *wb;
2505                 struct wb_lock_cookie cookie = {};
2506
2507                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2508                 current->nr_dirtied--;
2509                 dec_node_page_state(page, NR_DIRTIED);
2510                 dec_wb_stat(wb, WB_DIRTIED);
2511                 unlocked_inode_to_wb_end(inode, &cookie);
2512         }
2513 }
2514 EXPORT_SYMBOL(account_page_redirty);
2515
2516 /*
2517  * When a writepage implementation decides that it doesn't want to write this
2518  * page for some reason, it should redirty the locked page via
2519  * redirty_page_for_writepage() and it should then unlock the page and return 0
2520  */
2521 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2522 {
2523         int ret;
2524
2525         wbc->pages_skipped++;
2526         ret = __set_page_dirty_nobuffers(page);
2527         account_page_redirty(page);
2528         return ret;
2529 }
2530 EXPORT_SYMBOL(redirty_page_for_writepage);
2531
2532 /*
2533  * Dirty a page.
2534  *
2535  * For pages with a mapping this should be done under the page lock
2536  * for the benefit of asynchronous memory errors who prefer a consistent
2537  * dirty state. This rule can be broken in some special cases,
2538  * but should be better not to.
2539  *
2540  * If the mapping doesn't provide a set_page_dirty a_op, then
2541  * just fall through and assume that it wants buffer_heads.
2542  */
2543 int set_page_dirty(struct page *page)
2544 {
2545         struct address_space *mapping = page_mapping(page);
2546
2547         page = compound_head(page);
2548         if (likely(mapping)) {
2549                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2550                 /*
2551                  * readahead/lru_deactivate_page could remain
2552                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2553                  * About readahead, if the page is written, the flags would be
2554                  * reset. So no problem.
2555                  * About lru_deactivate_page, if the page is redirty, the flag
2556                  * will be reset. So no problem. but if the page is used by readahead
2557                  * it will confuse readahead and make it restart the size rampup
2558                  * process. But it's a trivial problem.
2559                  */
2560                 if (PageReclaim(page))
2561                         ClearPageReclaim(page);
2562 #ifdef CONFIG_BLOCK
2563                 if (!spd)
2564                         spd = __set_page_dirty_buffers;
2565 #endif
2566                 return (*spd)(page);
2567         }
2568         if (!PageDirty(page)) {
2569                 if (!TestSetPageDirty(page))
2570                         return 1;
2571         }
2572         return 0;
2573 }
2574 EXPORT_SYMBOL(set_page_dirty);
2575
2576 /*
2577  * set_page_dirty() is racy if the caller has no reference against
2578  * page->mapping->host, and if the page is unlocked.  This is because another
2579  * CPU could truncate the page off the mapping and then free the mapping.
2580  *
2581  * Usually, the page _is_ locked, or the caller is a user-space process which
2582  * holds a reference on the inode by having an open file.
2583  *
2584  * In other cases, the page should be locked before running set_page_dirty().
2585  */
2586 int set_page_dirty_lock(struct page *page)
2587 {
2588         int ret;
2589
2590         lock_page(page);
2591         ret = set_page_dirty(page);
2592         unlock_page(page);
2593         return ret;
2594 }
2595 EXPORT_SYMBOL(set_page_dirty_lock);
2596
2597 /*
2598  * This cancels just the dirty bit on the kernel page itself, it does NOT
2599  * actually remove dirty bits on any mmap's that may be around. It also
2600  * leaves the page tagged dirty, so any sync activity will still find it on
2601  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2602  * look at the dirty bits in the VM.
2603  *
2604  * Doing this should *normally* only ever be done when a page is truncated,
2605  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2606  * this when it notices that somebody has cleaned out all the buffers on a
2607  * page without actually doing it through the VM. Can you say "ext3 is
2608  * horribly ugly"? Thought you could.
2609  */
2610 void __cancel_dirty_page(struct page *page)
2611 {
2612         struct address_space *mapping = page_mapping(page);
2613
2614         if (mapping_cap_account_dirty(mapping)) {
2615                 struct inode *inode = mapping->host;
2616                 struct bdi_writeback *wb;
2617                 struct wb_lock_cookie cookie = {};
2618
2619                 lock_page_memcg(page);
2620                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2621
2622                 if (TestClearPageDirty(page))
2623                         account_page_cleaned(page, mapping, wb);
2624
2625                 unlocked_inode_to_wb_end(inode, &cookie);
2626                 unlock_page_memcg(page);
2627         } else {
2628                 ClearPageDirty(page);
2629         }
2630 }
2631 EXPORT_SYMBOL(__cancel_dirty_page);
2632
2633 /*
2634  * Clear a page's dirty flag, while caring for dirty memory accounting.
2635  * Returns true if the page was previously dirty.
2636  *
2637  * This is for preparing to put the page under writeout.  We leave the page
2638  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2639  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2640  * implementation will run either set_page_writeback() or set_page_dirty(),
2641  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2642  * back into sync.
2643  *
2644  * This incoherency between the page's dirty flag and radix-tree tag is
2645  * unfortunate, but it only exists while the page is locked.
2646  */
2647 int clear_page_dirty_for_io(struct page *page)
2648 {
2649         struct address_space *mapping = page_mapping(page);
2650         int ret = 0;
2651
2652         BUG_ON(!PageLocked(page));
2653
2654         if (mapping && mapping_cap_account_dirty(mapping)) {
2655                 struct inode *inode = mapping->host;
2656                 struct bdi_writeback *wb;
2657                 struct wb_lock_cookie cookie = {};
2658
2659                 /*
2660                  * Yes, Virginia, this is indeed insane.
2661                  *
2662                  * We use this sequence to make sure that
2663                  *  (a) we account for dirty stats properly
2664                  *  (b) we tell the low-level filesystem to
2665                  *      mark the whole page dirty if it was
2666                  *      dirty in a pagetable. Only to then
2667                  *  (c) clean the page again and return 1 to
2668                  *      cause the writeback.
2669                  *
2670                  * This way we avoid all nasty races with the
2671                  * dirty bit in multiple places and clearing
2672                  * them concurrently from different threads.
2673                  *
2674                  * Note! Normally the "set_page_dirty(page)"
2675                  * has no effect on the actual dirty bit - since
2676                  * that will already usually be set. But we
2677                  * need the side effects, and it can help us
2678                  * avoid races.
2679                  *
2680                  * We basically use the page "master dirty bit"
2681                  * as a serialization point for all the different
2682                  * threads doing their things.
2683                  */
2684                 if (page_mkclean(page))
2685                         set_page_dirty(page);
2686                 /*
2687                  * We carefully synchronise fault handlers against
2688                  * installing a dirty pte and marking the page dirty
2689                  * at this point.  We do this by having them hold the
2690                  * page lock while dirtying the page, and pages are
2691                  * always locked coming in here, so we get the desired
2692                  * exclusion.
2693                  */
2694                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2695                 if (TestClearPageDirty(page)) {
2696                         dec_lruvec_page_state(page, NR_FILE_DIRTY);
2697                         dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2698                         dec_wb_stat(wb, WB_RECLAIMABLE);
2699                         ret = 1;
2700                 }
2701                 unlocked_inode_to_wb_end(inode, &cookie);
2702                 return ret;
2703         }
2704         return TestClearPageDirty(page);
2705 }
2706 EXPORT_SYMBOL(clear_page_dirty_for_io);
2707
2708 int test_clear_page_writeback(struct page *page)
2709 {
2710         struct address_space *mapping = page_mapping(page);
2711         struct mem_cgroup *memcg;
2712         struct lruvec *lruvec;
2713         int ret;
2714
2715         memcg = lock_page_memcg(page);
2716         lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2717         if (mapping && mapping_use_writeback_tags(mapping)) {
2718                 struct inode *inode = mapping->host;
2719                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2720                 unsigned long flags;
2721
2722                 xa_lock_irqsave(&mapping->i_pages, flags);
2723                 ret = TestClearPageWriteback(page);
2724                 if (ret) {
2725                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2726                                                 PAGECACHE_TAG_WRITEBACK);
2727                         if (bdi_cap_account_writeback(bdi)) {
2728                                 struct bdi_writeback *wb = inode_to_wb(inode);
2729
2730                                 dec_wb_stat(wb, WB_WRITEBACK);
2731                                 __wb_writeout_inc(wb);
2732                         }
2733                 }
2734
2735                 if (mapping->host && !mapping_tagged(mapping,
2736                                                      PAGECACHE_TAG_WRITEBACK))
2737                         sb_clear_inode_writeback(mapping->host);
2738
2739                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2740         } else {
2741                 ret = TestClearPageWriteback(page);
2742         }
2743         /*
2744          * NOTE: Page might be free now! Writeback doesn't hold a page
2745          * reference on its own, it relies on truncation to wait for
2746          * the clearing of PG_writeback. The below can only access
2747          * page state that is static across allocation cycles.
2748          */
2749         if (ret) {
2750                 dec_lruvec_state(lruvec, NR_WRITEBACK);
2751                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2752                 inc_node_page_state(page, NR_WRITTEN);
2753         }
2754         __unlock_page_memcg(memcg);
2755         return ret;
2756 }
2757
2758 int __test_set_page_writeback(struct page *page, bool keep_write)
2759 {
2760         struct address_space *mapping = page_mapping(page);
2761         int ret;
2762
2763         lock_page_memcg(page);
2764         if (mapping && mapping_use_writeback_tags(mapping)) {
2765                 struct inode *inode = mapping->host;
2766                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2767                 unsigned long flags;
2768
2769                 xa_lock_irqsave(&mapping->i_pages, flags);
2770                 ret = TestSetPageWriteback(page);
2771                 if (!ret) {
2772                         bool on_wblist;
2773
2774                         on_wblist = mapping_tagged(mapping,
2775                                                    PAGECACHE_TAG_WRITEBACK);
2776
2777                         radix_tree_tag_set(&mapping->i_pages, page_index(page),
2778                                                 PAGECACHE_TAG_WRITEBACK);
2779                         if (bdi_cap_account_writeback(bdi))
2780                                 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2781
2782                         /*
2783                          * We can come through here when swapping anonymous
2784                          * pages, so we don't necessarily have an inode to track
2785                          * for sync.
2786                          */
2787                         if (mapping->host && !on_wblist)
2788                                 sb_mark_inode_writeback(mapping->host);
2789                 }
2790                 if (!PageDirty(page))
2791                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2792                                                 PAGECACHE_TAG_DIRTY);
2793                 if (!keep_write)
2794                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2795                                                 PAGECACHE_TAG_TOWRITE);
2796                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2797         } else {
2798                 ret = TestSetPageWriteback(page);
2799         }
2800         if (!ret) {
2801                 inc_lruvec_page_state(page, NR_WRITEBACK);
2802                 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2803         }
2804         unlock_page_memcg(page);
2805         return ret;
2806
2807 }
2808 EXPORT_SYMBOL(__test_set_page_writeback);
2809
2810 /*
2811  * Return true if any of the pages in the mapping are marked with the
2812  * passed tag.
2813  */
2814 int mapping_tagged(struct address_space *mapping, int tag)
2815 {
2816         return radix_tree_tagged(&mapping->i_pages, tag);
2817 }
2818 EXPORT_SYMBOL(mapping_tagged);
2819
2820 /**
2821  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2822  * @page:       The page to wait on.
2823  *
2824  * This function determines if the given page is related to a backing device
2825  * that requires page contents to be held stable during writeback.  If so, then
2826  * it will wait for any pending writeback to complete.
2827  */
2828 void wait_for_stable_page(struct page *page)
2829 {
2830         if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2831                 wait_on_page_writeback(page);
2832 }
2833 EXPORT_SYMBOL_GPL(wait_for_stable_page);