db30ce0b7d80c18a06a7994a89725994485e55c3
[muen/linux.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_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(unsigned long t)
629 {
630         struct wb_domain *dom = (void *)t;
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         setup_deferrable_timer(&dom->period_timer, writeout_period,
654                                (unsigned long)dom);
655
656         dom->dirty_limit_tstamp = jiffies;
657
658         return fprop_global_init(&dom->completions, gfp);
659 }
660
661 #ifdef CONFIG_CGROUP_WRITEBACK
662 void wb_domain_exit(struct wb_domain *dom)
663 {
664         del_timer_sync(&dom->period_timer);
665         fprop_global_destroy(&dom->completions);
666 }
667 #endif
668
669 /*
670  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
671  * registered backing devices, which, for obvious reasons, can not
672  * exceed 100%.
673  */
674 static unsigned int bdi_min_ratio;
675
676 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
677 {
678         int ret = 0;
679
680         spin_lock_bh(&bdi_lock);
681         if (min_ratio > bdi->max_ratio) {
682                 ret = -EINVAL;
683         } else {
684                 min_ratio -= bdi->min_ratio;
685                 if (bdi_min_ratio + min_ratio < 100) {
686                         bdi_min_ratio += min_ratio;
687                         bdi->min_ratio += min_ratio;
688                 } else {
689                         ret = -EINVAL;
690                 }
691         }
692         spin_unlock_bh(&bdi_lock);
693
694         return ret;
695 }
696
697 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
698 {
699         int ret = 0;
700
701         if (max_ratio > 100)
702                 return -EINVAL;
703
704         spin_lock_bh(&bdi_lock);
705         if (bdi->min_ratio > max_ratio) {
706                 ret = -EINVAL;
707         } else {
708                 bdi->max_ratio = max_ratio;
709                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
710         }
711         spin_unlock_bh(&bdi_lock);
712
713         return ret;
714 }
715 EXPORT_SYMBOL(bdi_set_max_ratio);
716
717 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
718                                            unsigned long bg_thresh)
719 {
720         return (thresh + bg_thresh) / 2;
721 }
722
723 static unsigned long hard_dirty_limit(struct wb_domain *dom,
724                                       unsigned long thresh)
725 {
726         return max(thresh, dom->dirty_limit);
727 }
728
729 /*
730  * Memory which can be further allocated to a memcg domain is capped by
731  * system-wide clean memory excluding the amount being used in the domain.
732  */
733 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
734                             unsigned long filepages, unsigned long headroom)
735 {
736         struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
737         unsigned long clean = filepages - min(filepages, mdtc->dirty);
738         unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
739         unsigned long other_clean = global_clean - min(global_clean, clean);
740
741         mdtc->avail = filepages + min(headroom, other_clean);
742 }
743
744 /**
745  * __wb_calc_thresh - @wb's share of dirty throttling threshold
746  * @dtc: dirty_throttle_context of interest
747  *
748  * Returns @wb's dirty limit in pages. The term "dirty" in the context of
749  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
750  *
751  * Note that balance_dirty_pages() will only seriously take it as a hard limit
752  * when sleeping max_pause per page is not enough to keep the dirty pages under
753  * control. For example, when the device is completely stalled due to some error
754  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
755  * In the other normal situations, it acts more gently by throttling the tasks
756  * more (rather than completely block them) when the wb dirty pages go high.
757  *
758  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
759  * - starving fast devices
760  * - piling up dirty pages (that will take long time to sync) on slow devices
761  *
762  * The wb's share of dirty limit will be adapting to its throughput and
763  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
764  */
765 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
766 {
767         struct wb_domain *dom = dtc_dom(dtc);
768         unsigned long thresh = dtc->thresh;
769         u64 wb_thresh;
770         long numerator, denominator;
771         unsigned long wb_min_ratio, wb_max_ratio;
772
773         /*
774          * Calculate this BDI's share of the thresh ratio.
775          */
776         fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
777                               &numerator, &denominator);
778
779         wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
780         wb_thresh *= numerator;
781         do_div(wb_thresh, denominator);
782
783         wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
784
785         wb_thresh += (thresh * wb_min_ratio) / 100;
786         if (wb_thresh > (thresh * wb_max_ratio) / 100)
787                 wb_thresh = thresh * wb_max_ratio / 100;
788
789         return wb_thresh;
790 }
791
792 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
793 {
794         struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
795                                                .thresh = thresh };
796         return __wb_calc_thresh(&gdtc);
797 }
798
799 /*
800  *                           setpoint - dirty 3
801  *        f(dirty) := 1.0 + (----------------)
802  *                           limit - setpoint
803  *
804  * it's a 3rd order polynomial that subjects to
805  *
806  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
807  * (2) f(setpoint) = 1.0 => the balance point
808  * (3) f(limit)    = 0   => the hard limit
809  * (4) df/dx      <= 0   => negative feedback control
810  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
811  *     => fast response on large errors; small oscillation near setpoint
812  */
813 static long long pos_ratio_polynom(unsigned long setpoint,
814                                           unsigned long dirty,
815                                           unsigned long limit)
816 {
817         long long pos_ratio;
818         long x;
819
820         x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
821                       (limit - setpoint) | 1);
822         pos_ratio = x;
823         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
824         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
825         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
826
827         return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
828 }
829
830 /*
831  * Dirty position control.
832  *
833  * (o) global/bdi setpoints
834  *
835  * We want the dirty pages be balanced around the global/wb setpoints.
836  * When the number of dirty pages is higher/lower than the setpoint, the
837  * dirty position control ratio (and hence task dirty ratelimit) will be
838  * decreased/increased to bring the dirty pages back to the setpoint.
839  *
840  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
841  *
842  *     if (dirty < setpoint) scale up   pos_ratio
843  *     if (dirty > setpoint) scale down pos_ratio
844  *
845  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
846  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
847  *
848  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
849  *
850  * (o) global control line
851  *
852  *     ^ pos_ratio
853  *     |
854  *     |            |<===== global dirty control scope ======>|
855  * 2.0 .............*
856  *     |            .*
857  *     |            . *
858  *     |            .   *
859  *     |            .     *
860  *     |            .        *
861  *     |            .            *
862  * 1.0 ................................*
863  *     |            .                  .     *
864  *     |            .                  .          *
865  *     |            .                  .              *
866  *     |            .                  .                 *
867  *     |            .                  .                    *
868  *   0 +------------.------------------.----------------------*------------->
869  *           freerun^          setpoint^                 limit^   dirty pages
870  *
871  * (o) wb control line
872  *
873  *     ^ pos_ratio
874  *     |
875  *     |            *
876  *     |              *
877  *     |                *
878  *     |                  *
879  *     |                    * |<=========== span ============>|
880  * 1.0 .......................*
881  *     |                      . *
882  *     |                      .   *
883  *     |                      .     *
884  *     |                      .       *
885  *     |                      .         *
886  *     |                      .           *
887  *     |                      .             *
888  *     |                      .               *
889  *     |                      .                 *
890  *     |                      .                   *
891  *     |                      .                     *
892  * 1/4 ...............................................* * * * * * * * * * * *
893  *     |                      .                         .
894  *     |                      .                           .
895  *     |                      .                             .
896  *   0 +----------------------.-------------------------------.------------->
897  *                wb_setpoint^                    x_intercept^
898  *
899  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
900  * be smoothly throttled down to normal if it starts high in situations like
901  * - start writing to a slow SD card and a fast disk at the same time. The SD
902  *   card's wb_dirty may rush to many times higher than wb_setpoint.
903  * - the wb dirty thresh drops quickly due to change of JBOD workload
904  */
905 static void wb_position_ratio(struct dirty_throttle_control *dtc)
906 {
907         struct bdi_writeback *wb = dtc->wb;
908         unsigned long write_bw = wb->avg_write_bandwidth;
909         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
910         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
911         unsigned long wb_thresh = dtc->wb_thresh;
912         unsigned long x_intercept;
913         unsigned long setpoint;         /* dirty pages' target balance point */
914         unsigned long wb_setpoint;
915         unsigned long span;
916         long long pos_ratio;            /* for scaling up/down the rate limit */
917         long x;
918
919         dtc->pos_ratio = 0;
920
921         if (unlikely(dtc->dirty >= limit))
922                 return;
923
924         /*
925          * global setpoint
926          *
927          * See comment for pos_ratio_polynom().
928          */
929         setpoint = (freerun + limit) / 2;
930         pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
931
932         /*
933          * The strictlimit feature is a tool preventing mistrusted filesystems
934          * from growing a large number of dirty pages before throttling. For
935          * such filesystems balance_dirty_pages always checks wb counters
936          * against wb limits. Even if global "nr_dirty" is under "freerun".
937          * This is especially important for fuse which sets bdi->max_ratio to
938          * 1% by default. Without strictlimit feature, fuse writeback may
939          * consume arbitrary amount of RAM because it is accounted in
940          * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
941          *
942          * Here, in wb_position_ratio(), we calculate pos_ratio based on
943          * two values: wb_dirty and wb_thresh. Let's consider an example:
944          * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
945          * limits are set by default to 10% and 20% (background and throttle).
946          * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
947          * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
948          * about ~6K pages (as the average of background and throttle wb
949          * limits). The 3rd order polynomial will provide positive feedback if
950          * wb_dirty is under wb_setpoint and vice versa.
951          *
952          * Note, that we cannot use global counters in these calculations
953          * because we want to throttle process writing to a strictlimit wb
954          * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
955          * in the example above).
956          */
957         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
958                 long long wb_pos_ratio;
959
960                 if (dtc->wb_dirty < 8) {
961                         dtc->pos_ratio = min_t(long long, pos_ratio * 2,
962                                            2 << RATELIMIT_CALC_SHIFT);
963                         return;
964                 }
965
966                 if (dtc->wb_dirty >= wb_thresh)
967                         return;
968
969                 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
970                                                     dtc->wb_bg_thresh);
971
972                 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
973                         return;
974
975                 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
976                                                  wb_thresh);
977
978                 /*
979                  * Typically, for strictlimit case, wb_setpoint << setpoint
980                  * and pos_ratio >> wb_pos_ratio. In the other words global
981                  * state ("dirty") is not limiting factor and we have to
982                  * make decision based on wb counters. But there is an
983                  * important case when global pos_ratio should get precedence:
984                  * global limits are exceeded (e.g. due to activities on other
985                  * wb's) while given strictlimit wb is below limit.
986                  *
987                  * "pos_ratio * wb_pos_ratio" would work for the case above,
988                  * but it would look too non-natural for the case of all
989                  * activity in the system coming from a single strictlimit wb
990                  * with bdi->max_ratio == 100%.
991                  *
992                  * Note that min() below somewhat changes the dynamics of the
993                  * control system. Normally, pos_ratio value can be well over 3
994                  * (when globally we are at freerun and wb is well below wb
995                  * setpoint). Now the maximum pos_ratio in the same situation
996                  * is 2. We might want to tweak this if we observe the control
997                  * system is too slow to adapt.
998                  */
999                 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1000                 return;
1001         }
1002
1003         /*
1004          * We have computed basic pos_ratio above based on global situation. If
1005          * the wb is over/under its share of dirty pages, we want to scale
1006          * pos_ratio further down/up. That is done by the following mechanism.
1007          */
1008
1009         /*
1010          * wb setpoint
1011          *
1012          *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1013          *
1014          *                        x_intercept - wb_dirty
1015          *                     := --------------------------
1016          *                        x_intercept - wb_setpoint
1017          *
1018          * The main wb control line is a linear function that subjects to
1019          *
1020          * (1) f(wb_setpoint) = 1.0
1021          * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1022          *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1023          *
1024          * For single wb case, the dirty pages are observed to fluctuate
1025          * regularly within range
1026          *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1027          * for various filesystems, where (2) can yield in a reasonable 12.5%
1028          * fluctuation range for pos_ratio.
1029          *
1030          * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1031          * own size, so move the slope over accordingly and choose a slope that
1032          * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1033          */
1034         if (unlikely(wb_thresh > dtc->thresh))
1035                 wb_thresh = dtc->thresh;
1036         /*
1037          * It's very possible that wb_thresh is close to 0 not because the
1038          * device is slow, but that it has remained inactive for long time.
1039          * Honour such devices a reasonable good (hopefully IO efficient)
1040          * threshold, so that the occasional writes won't be blocked and active
1041          * writes can rampup the threshold quickly.
1042          */
1043         wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1044         /*
1045          * scale global setpoint to wb's:
1046          *      wb_setpoint = setpoint * wb_thresh / thresh
1047          */
1048         x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1049         wb_setpoint = setpoint * (u64)x >> 16;
1050         /*
1051          * Use span=(8*write_bw) in single wb case as indicated by
1052          * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1053          *
1054          *        wb_thresh                    thresh - wb_thresh
1055          * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1056          *         thresh                           thresh
1057          */
1058         span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1059         x_intercept = wb_setpoint + span;
1060
1061         if (dtc->wb_dirty < x_intercept - span / 4) {
1062                 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1063                                       (x_intercept - wb_setpoint) | 1);
1064         } else
1065                 pos_ratio /= 4;
1066
1067         /*
1068          * wb reserve area, safeguard against dirty pool underrun and disk idle
1069          * It may push the desired control point of global dirty pages higher
1070          * than setpoint.
1071          */
1072         x_intercept = wb_thresh / 2;
1073         if (dtc->wb_dirty < x_intercept) {
1074                 if (dtc->wb_dirty > x_intercept / 8)
1075                         pos_ratio = div_u64(pos_ratio * x_intercept,
1076                                             dtc->wb_dirty);
1077                 else
1078                         pos_ratio *= 8;
1079         }
1080
1081         dtc->pos_ratio = pos_ratio;
1082 }
1083
1084 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1085                                       unsigned long elapsed,
1086                                       unsigned long written)
1087 {
1088         const unsigned long period = roundup_pow_of_two(3 * HZ);
1089         unsigned long avg = wb->avg_write_bandwidth;
1090         unsigned long old = wb->write_bandwidth;
1091         u64 bw;
1092
1093         /*
1094          * bw = written * HZ / elapsed
1095          *
1096          *                   bw * elapsed + write_bandwidth * (period - elapsed)
1097          * write_bandwidth = ---------------------------------------------------
1098          *                                          period
1099          *
1100          * @written may have decreased due to account_page_redirty().
1101          * Avoid underflowing @bw calculation.
1102          */
1103         bw = written - min(written, wb->written_stamp);
1104         bw *= HZ;
1105         if (unlikely(elapsed > period)) {
1106                 do_div(bw, elapsed);
1107                 avg = bw;
1108                 goto out;
1109         }
1110         bw += (u64)wb->write_bandwidth * (period - elapsed);
1111         bw >>= ilog2(period);
1112
1113         /*
1114          * one more level of smoothing, for filtering out sudden spikes
1115          */
1116         if (avg > old && old >= (unsigned long)bw)
1117                 avg -= (avg - old) >> 3;
1118
1119         if (avg < old && old <= (unsigned long)bw)
1120                 avg += (old - avg) >> 3;
1121
1122 out:
1123         /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1124         avg = max(avg, 1LU);
1125         if (wb_has_dirty_io(wb)) {
1126                 long delta = avg - wb->avg_write_bandwidth;
1127                 WARN_ON_ONCE(atomic_long_add_return(delta,
1128                                         &wb->bdi->tot_write_bandwidth) <= 0);
1129         }
1130         wb->write_bandwidth = bw;
1131         wb->avg_write_bandwidth = avg;
1132 }
1133
1134 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1135 {
1136         struct wb_domain *dom = dtc_dom(dtc);
1137         unsigned long thresh = dtc->thresh;
1138         unsigned long limit = dom->dirty_limit;
1139
1140         /*
1141          * Follow up in one step.
1142          */
1143         if (limit < thresh) {
1144                 limit = thresh;
1145                 goto update;
1146         }
1147
1148         /*
1149          * Follow down slowly. Use the higher one as the target, because thresh
1150          * may drop below dirty. This is exactly the reason to introduce
1151          * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1152          */
1153         thresh = max(thresh, dtc->dirty);
1154         if (limit > thresh) {
1155                 limit -= (limit - thresh) >> 5;
1156                 goto update;
1157         }
1158         return;
1159 update:
1160         dom->dirty_limit = limit;
1161 }
1162
1163 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1164                                     unsigned long now)
1165 {
1166         struct wb_domain *dom = dtc_dom(dtc);
1167
1168         /*
1169          * check locklessly first to optimize away locking for the most time
1170          */
1171         if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1172                 return;
1173
1174         spin_lock(&dom->lock);
1175         if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1176                 update_dirty_limit(dtc);
1177                 dom->dirty_limit_tstamp = now;
1178         }
1179         spin_unlock(&dom->lock);
1180 }
1181
1182 /*
1183  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1184  *
1185  * Normal wb tasks will be curbed at or below it in long term.
1186  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1187  */
1188 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1189                                       unsigned long dirtied,
1190                                       unsigned long elapsed)
1191 {
1192         struct bdi_writeback *wb = dtc->wb;
1193         unsigned long dirty = dtc->dirty;
1194         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1195         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1196         unsigned long setpoint = (freerun + limit) / 2;
1197         unsigned long write_bw = wb->avg_write_bandwidth;
1198         unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1199         unsigned long dirty_rate;
1200         unsigned long task_ratelimit;
1201         unsigned long balanced_dirty_ratelimit;
1202         unsigned long step;
1203         unsigned long x;
1204         unsigned long shift;
1205
1206         /*
1207          * The dirty rate will match the writeout rate in long term, except
1208          * when dirty pages are truncated by userspace or re-dirtied by FS.
1209          */
1210         dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1211
1212         /*
1213          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1214          */
1215         task_ratelimit = (u64)dirty_ratelimit *
1216                                         dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1217         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1218
1219         /*
1220          * A linear estimation of the "balanced" throttle rate. The theory is,
1221          * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1222          * dirty_rate will be measured to be (N * task_ratelimit). So the below
1223          * formula will yield the balanced rate limit (write_bw / N).
1224          *
1225          * Note that the expanded form is not a pure rate feedback:
1226          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1227          * but also takes pos_ratio into account:
1228          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1229          *
1230          * (1) is not realistic because pos_ratio also takes part in balancing
1231          * the dirty rate.  Consider the state
1232          *      pos_ratio = 0.5                                              (3)
1233          *      rate = 2 * (write_bw / N)                                    (4)
1234          * If (1) is used, it will stuck in that state! Because each dd will
1235          * be throttled at
1236          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1237          * yielding
1238          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1239          * put (6) into (1) we get
1240          *      rate_(i+1) = rate_(i)                                        (7)
1241          *
1242          * So we end up using (2) to always keep
1243          *      rate_(i+1) ~= (write_bw / N)                                 (8)
1244          * regardless of the value of pos_ratio. As long as (8) is satisfied,
1245          * pos_ratio is able to drive itself to 1.0, which is not only where
1246          * the dirty count meet the setpoint, but also where the slope of
1247          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1248          */
1249         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1250                                            dirty_rate | 1);
1251         /*
1252          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1253          */
1254         if (unlikely(balanced_dirty_ratelimit > write_bw))
1255                 balanced_dirty_ratelimit = write_bw;
1256
1257         /*
1258          * We could safely do this and return immediately:
1259          *
1260          *      wb->dirty_ratelimit = balanced_dirty_ratelimit;
1261          *
1262          * However to get a more stable dirty_ratelimit, the below elaborated
1263          * code makes use of task_ratelimit to filter out singular points and
1264          * limit the step size.
1265          *
1266          * The below code essentially only uses the relative value of
1267          *
1268          *      task_ratelimit - dirty_ratelimit
1269          *      = (pos_ratio - 1) * dirty_ratelimit
1270          *
1271          * which reflects the direction and size of dirty position error.
1272          */
1273
1274         /*
1275          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1276          * task_ratelimit is on the same side of dirty_ratelimit, too.
1277          * For example, when
1278          * - dirty_ratelimit > balanced_dirty_ratelimit
1279          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1280          * lowering dirty_ratelimit will help meet both the position and rate
1281          * control targets. Otherwise, don't update dirty_ratelimit if it will
1282          * only help meet the rate target. After all, what the users ultimately
1283          * feel and care are stable dirty rate and small position error.
1284          *
1285          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1286          * and filter out the singular points of balanced_dirty_ratelimit. Which
1287          * keeps jumping around randomly and can even leap far away at times
1288          * due to the small 200ms estimation period of dirty_rate (we want to
1289          * keep that period small to reduce time lags).
1290          */
1291         step = 0;
1292
1293         /*
1294          * For strictlimit case, calculations above were based on wb counters
1295          * and limits (starting from pos_ratio = wb_position_ratio() and up to
1296          * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1297          * Hence, to calculate "step" properly, we have to use wb_dirty as
1298          * "dirty" and wb_setpoint as "setpoint".
1299          *
1300          * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1301          * it's possible that wb_thresh is close to zero due to inactivity
1302          * of backing device.
1303          */
1304         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1305                 dirty = dtc->wb_dirty;
1306                 if (dtc->wb_dirty < 8)
1307                         setpoint = dtc->wb_dirty + 1;
1308                 else
1309                         setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1310         }
1311
1312         if (dirty < setpoint) {
1313                 x = min3(wb->balanced_dirty_ratelimit,
1314                          balanced_dirty_ratelimit, task_ratelimit);
1315                 if (dirty_ratelimit < x)
1316                         step = x - dirty_ratelimit;
1317         } else {
1318                 x = max3(wb->balanced_dirty_ratelimit,
1319                          balanced_dirty_ratelimit, task_ratelimit);
1320                 if (dirty_ratelimit > x)
1321                         step = dirty_ratelimit - x;
1322         }
1323
1324         /*
1325          * Don't pursue 100% rate matching. It's impossible since the balanced
1326          * rate itself is constantly fluctuating. So decrease the track speed
1327          * when it gets close to the target. Helps eliminate pointless tremors.
1328          */
1329         shift = dirty_ratelimit / (2 * step + 1);
1330         if (shift < BITS_PER_LONG)
1331                 step = DIV_ROUND_UP(step >> shift, 8);
1332         else
1333                 step = 0;
1334
1335         if (dirty_ratelimit < balanced_dirty_ratelimit)
1336                 dirty_ratelimit += step;
1337         else
1338                 dirty_ratelimit -= step;
1339
1340         wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1341         wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1342
1343         trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1344 }
1345
1346 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1347                                   struct dirty_throttle_control *mdtc,
1348                                   unsigned long start_time,
1349                                   bool update_ratelimit)
1350 {
1351         struct bdi_writeback *wb = gdtc->wb;
1352         unsigned long now = jiffies;
1353         unsigned long elapsed = now - wb->bw_time_stamp;
1354         unsigned long dirtied;
1355         unsigned long written;
1356
1357         lockdep_assert_held(&wb->list_lock);
1358
1359         /*
1360          * rate-limit, only update once every 200ms.
1361          */
1362         if (elapsed < BANDWIDTH_INTERVAL)
1363                 return;
1364
1365         dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1366         written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1367
1368         /*
1369          * Skip quiet periods when disk bandwidth is under-utilized.
1370          * (at least 1s idle time between two flusher runs)
1371          */
1372         if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1373                 goto snapshot;
1374
1375         if (update_ratelimit) {
1376                 domain_update_bandwidth(gdtc, now);
1377                 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1378
1379                 /*
1380                  * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1381                  * compiler has no way to figure that out.  Help it.
1382                  */
1383                 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1384                         domain_update_bandwidth(mdtc, now);
1385                         wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1386                 }
1387         }
1388         wb_update_write_bandwidth(wb, elapsed, written);
1389
1390 snapshot:
1391         wb->dirtied_stamp = dirtied;
1392         wb->written_stamp = written;
1393         wb->bw_time_stamp = now;
1394 }
1395
1396 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1397 {
1398         struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1399
1400         __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1401 }
1402
1403 /*
1404  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1405  * will look to see if it needs to start dirty throttling.
1406  *
1407  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1408  * global_page_state() too often. So scale it near-sqrt to the safety margin
1409  * (the number of pages we may dirty without exceeding the dirty limits).
1410  */
1411 static unsigned long dirty_poll_interval(unsigned long dirty,
1412                                          unsigned long thresh)
1413 {
1414         if (thresh > dirty)
1415                 return 1UL << (ilog2(thresh - dirty) >> 1);
1416
1417         return 1;
1418 }
1419
1420 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1421                                   unsigned long wb_dirty)
1422 {
1423         unsigned long bw = wb->avg_write_bandwidth;
1424         unsigned long t;
1425
1426         /*
1427          * Limit pause time for small memory systems. If sleeping for too long
1428          * time, a small pool of dirty/writeback pages may go empty and disk go
1429          * idle.
1430          *
1431          * 8 serves as the safety ratio.
1432          */
1433         t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1434         t++;
1435
1436         return min_t(unsigned long, t, MAX_PAUSE);
1437 }
1438
1439 static long wb_min_pause(struct bdi_writeback *wb,
1440                          long max_pause,
1441                          unsigned long task_ratelimit,
1442                          unsigned long dirty_ratelimit,
1443                          int *nr_dirtied_pause)
1444 {
1445         long hi = ilog2(wb->avg_write_bandwidth);
1446         long lo = ilog2(wb->dirty_ratelimit);
1447         long t;         /* target pause */
1448         long pause;     /* estimated next pause */
1449         int pages;      /* target nr_dirtied_pause */
1450
1451         /* target for 10ms pause on 1-dd case */
1452         t = max(1, HZ / 100);
1453
1454         /*
1455          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1456          * overheads.
1457          *
1458          * (N * 10ms) on 2^N concurrent tasks.
1459          */
1460         if (hi > lo)
1461                 t += (hi - lo) * (10 * HZ) / 1024;
1462
1463         /*
1464          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1465          * on the much more stable dirty_ratelimit. However the next pause time
1466          * will be computed based on task_ratelimit and the two rate limits may
1467          * depart considerably at some time. Especially if task_ratelimit goes
1468          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1469          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1470          * result task_ratelimit won't be executed faithfully, which could
1471          * eventually bring down dirty_ratelimit.
1472          *
1473          * We apply two rules to fix it up:
1474          * 1) try to estimate the next pause time and if necessary, use a lower
1475          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1476          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1477          * 2) limit the target pause time to max_pause/2, so that the normal
1478          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1479          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1480          */
1481         t = min(t, 1 + max_pause / 2);
1482         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1483
1484         /*
1485          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1486          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1487          * When the 16 consecutive reads are often interrupted by some dirty
1488          * throttling pause during the async writes, cfq will go into idles
1489          * (deadline is fine). So push nr_dirtied_pause as high as possible
1490          * until reaches DIRTY_POLL_THRESH=32 pages.
1491          */
1492         if (pages < DIRTY_POLL_THRESH) {
1493                 t = max_pause;
1494                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1495                 if (pages > DIRTY_POLL_THRESH) {
1496                         pages = DIRTY_POLL_THRESH;
1497                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1498                 }
1499         }
1500
1501         pause = HZ * pages / (task_ratelimit + 1);
1502         if (pause > max_pause) {
1503                 t = max_pause;
1504                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1505         }
1506
1507         *nr_dirtied_pause = pages;
1508         /*
1509          * The minimal pause time will normally be half the target pause time.
1510          */
1511         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1512 }
1513
1514 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1515 {
1516         struct bdi_writeback *wb = dtc->wb;
1517         unsigned long wb_reclaimable;
1518
1519         /*
1520          * wb_thresh is not treated as some limiting factor as
1521          * dirty_thresh, due to reasons
1522          * - in JBOD setup, wb_thresh can fluctuate a lot
1523          * - in a system with HDD and USB key, the USB key may somehow
1524          *   go into state (wb_dirty >> wb_thresh) either because
1525          *   wb_dirty starts high, or because wb_thresh drops low.
1526          *   In this case we don't want to hard throttle the USB key
1527          *   dirtiers for 100 seconds until wb_dirty drops under
1528          *   wb_thresh. Instead the auxiliary wb control line in
1529          *   wb_position_ratio() will let the dirtier task progress
1530          *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1531          */
1532         dtc->wb_thresh = __wb_calc_thresh(dtc);
1533         dtc->wb_bg_thresh = dtc->thresh ?
1534                 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1535
1536         /*
1537          * In order to avoid the stacked BDI deadlock we need
1538          * to ensure we accurately count the 'dirty' pages when
1539          * the threshold is low.
1540          *
1541          * Otherwise it would be possible to get thresh+n pages
1542          * reported dirty, even though there are thresh-m pages
1543          * actually dirty; with m+n sitting in the percpu
1544          * deltas.
1545          */
1546         if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1547                 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1548                 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1549         } else {
1550                 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1551                 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1552         }
1553 }
1554
1555 /*
1556  * balance_dirty_pages() must be called by processes which are generating dirty
1557  * data.  It looks at the number of dirty pages in the machine and will force
1558  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1559  * If we're over `background_thresh' then the writeback threads are woken to
1560  * perform some writeout.
1561  */
1562 static void balance_dirty_pages(struct address_space *mapping,
1563                                 struct bdi_writeback *wb,
1564                                 unsigned long pages_dirtied)
1565 {
1566         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1567         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1568         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1569         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1570                                                      &mdtc_stor : NULL;
1571         struct dirty_throttle_control *sdtc;
1572         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1573         long period;
1574         long pause;
1575         long max_pause;
1576         long min_pause;
1577         int nr_dirtied_pause;
1578         bool dirty_exceeded = false;
1579         unsigned long task_ratelimit;
1580         unsigned long dirty_ratelimit;
1581         struct backing_dev_info *bdi = wb->bdi;
1582         bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1583         unsigned long start_time = jiffies;
1584
1585         for (;;) {
1586                 unsigned long now = jiffies;
1587                 unsigned long dirty, thresh, bg_thresh;
1588                 unsigned long m_dirty = 0;      /* stop bogus uninit warnings */
1589                 unsigned long m_thresh = 0;
1590                 unsigned long m_bg_thresh = 0;
1591
1592                 /*
1593                  * Unstable writes are a feature of certain networked
1594                  * filesystems (i.e. NFS) in which data may have been
1595                  * written to the server's write cache, but has not yet
1596                  * been flushed to permanent storage.
1597                  */
1598                 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1599                                         global_node_page_state(NR_UNSTABLE_NFS);
1600                 gdtc->avail = global_dirtyable_memory();
1601                 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1602
1603                 domain_dirty_limits(gdtc);
1604
1605                 if (unlikely(strictlimit)) {
1606                         wb_dirty_limits(gdtc);
1607
1608                         dirty = gdtc->wb_dirty;
1609                         thresh = gdtc->wb_thresh;
1610                         bg_thresh = gdtc->wb_bg_thresh;
1611                 } else {
1612                         dirty = gdtc->dirty;
1613                         thresh = gdtc->thresh;
1614                         bg_thresh = gdtc->bg_thresh;
1615                 }
1616
1617                 if (mdtc) {
1618                         unsigned long filepages, headroom, writeback;
1619
1620                         /*
1621                          * If @wb belongs to !root memcg, repeat the same
1622                          * basic calculations for the memcg domain.
1623                          */
1624                         mem_cgroup_wb_stats(wb, &filepages, &headroom,
1625                                             &mdtc->dirty, &writeback);
1626                         mdtc->dirty += writeback;
1627                         mdtc_calc_avail(mdtc, filepages, headroom);
1628
1629                         domain_dirty_limits(mdtc);
1630
1631                         if (unlikely(strictlimit)) {
1632                                 wb_dirty_limits(mdtc);
1633                                 m_dirty = mdtc->wb_dirty;
1634                                 m_thresh = mdtc->wb_thresh;
1635                                 m_bg_thresh = mdtc->wb_bg_thresh;
1636                         } else {
1637                                 m_dirty = mdtc->dirty;
1638                                 m_thresh = mdtc->thresh;
1639                                 m_bg_thresh = mdtc->bg_thresh;
1640                         }
1641                 }
1642
1643                 /*
1644                  * Throttle it only when the background writeback cannot
1645                  * catch-up. This avoids (excessively) small writeouts
1646                  * when the wb limits are ramping up in case of !strictlimit.
1647                  *
1648                  * In strictlimit case make decision based on the wb counters
1649                  * and limits. Small writeouts when the wb limits are ramping
1650                  * up are the price we consciously pay for strictlimit-ing.
1651                  *
1652                  * If memcg domain is in effect, @dirty should be under
1653                  * both global and memcg freerun ceilings.
1654                  */
1655                 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1656                     (!mdtc ||
1657                      m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1658                         unsigned long intv = dirty_poll_interval(dirty, thresh);
1659                         unsigned long m_intv = ULONG_MAX;
1660
1661                         current->dirty_paused_when = now;
1662                         current->nr_dirtied = 0;
1663                         if (mdtc)
1664                                 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1665                         current->nr_dirtied_pause = min(intv, m_intv);
1666                         break;
1667                 }
1668
1669                 if (unlikely(!writeback_in_progress(wb)))
1670                         wb_start_background_writeback(wb);
1671
1672                 /*
1673                  * Calculate global domain's pos_ratio and select the
1674                  * global dtc by default.
1675                  */
1676                 if (!strictlimit)
1677                         wb_dirty_limits(gdtc);
1678
1679                 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1680                         ((gdtc->dirty > gdtc->thresh) || strictlimit);
1681
1682                 wb_position_ratio(gdtc);
1683                 sdtc = gdtc;
1684
1685                 if (mdtc) {
1686                         /*
1687                          * If memcg domain is in effect, calculate its
1688                          * pos_ratio.  @wb should satisfy constraints from
1689                          * both global and memcg domains.  Choose the one
1690                          * w/ lower pos_ratio.
1691                          */
1692                         if (!strictlimit)
1693                                 wb_dirty_limits(mdtc);
1694
1695                         dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1696                                 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1697
1698                         wb_position_ratio(mdtc);
1699                         if (mdtc->pos_ratio < gdtc->pos_ratio)
1700                                 sdtc = mdtc;
1701                 }
1702
1703                 if (dirty_exceeded && !wb->dirty_exceeded)
1704                         wb->dirty_exceeded = 1;
1705
1706                 if (time_is_before_jiffies(wb->bw_time_stamp +
1707                                            BANDWIDTH_INTERVAL)) {
1708                         spin_lock(&wb->list_lock);
1709                         __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1710                         spin_unlock(&wb->list_lock);
1711                 }
1712
1713                 /* throttle according to the chosen dtc */
1714                 dirty_ratelimit = wb->dirty_ratelimit;
1715                 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1716                                                         RATELIMIT_CALC_SHIFT;
1717                 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1718                 min_pause = wb_min_pause(wb, max_pause,
1719                                          task_ratelimit, dirty_ratelimit,
1720                                          &nr_dirtied_pause);
1721
1722                 if (unlikely(task_ratelimit == 0)) {
1723                         period = max_pause;
1724                         pause = max_pause;
1725                         goto pause;
1726                 }
1727                 period = HZ * pages_dirtied / task_ratelimit;
1728                 pause = period;
1729                 if (current->dirty_paused_when)
1730                         pause -= now - current->dirty_paused_when;
1731                 /*
1732                  * For less than 1s think time (ext3/4 may block the dirtier
1733                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1734                  * however at much less frequency), try to compensate it in
1735                  * future periods by updating the virtual time; otherwise just
1736                  * do a reset, as it may be a light dirtier.
1737                  */
1738                 if (pause < min_pause) {
1739                         trace_balance_dirty_pages(wb,
1740                                                   sdtc->thresh,
1741                                                   sdtc->bg_thresh,
1742                                                   sdtc->dirty,
1743                                                   sdtc->wb_thresh,
1744                                                   sdtc->wb_dirty,
1745                                                   dirty_ratelimit,
1746                                                   task_ratelimit,
1747                                                   pages_dirtied,
1748                                                   period,
1749                                                   min(pause, 0L),
1750                                                   start_time);
1751                         if (pause < -HZ) {
1752                                 current->dirty_paused_when = now;
1753                                 current->nr_dirtied = 0;
1754                         } else if (period) {
1755                                 current->dirty_paused_when += period;
1756                                 current->nr_dirtied = 0;
1757                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1758                                 current->nr_dirtied_pause += pages_dirtied;
1759                         break;
1760                 }
1761                 if (unlikely(pause > max_pause)) {
1762                         /* for occasional dropped task_ratelimit */
1763                         now += min(pause - max_pause, max_pause);
1764                         pause = max_pause;
1765                 }
1766
1767 pause:
1768                 trace_balance_dirty_pages(wb,
1769                                           sdtc->thresh,
1770                                           sdtc->bg_thresh,
1771                                           sdtc->dirty,
1772                                           sdtc->wb_thresh,
1773                                           sdtc->wb_dirty,
1774                                           dirty_ratelimit,
1775                                           task_ratelimit,
1776                                           pages_dirtied,
1777                                           period,
1778                                           pause,
1779                                           start_time);
1780                 __set_current_state(TASK_KILLABLE);
1781                 wb->dirty_sleep = now;
1782                 io_schedule_timeout(pause);
1783
1784                 current->dirty_paused_when = now + pause;
1785                 current->nr_dirtied = 0;
1786                 current->nr_dirtied_pause = nr_dirtied_pause;
1787
1788                 /*
1789                  * This is typically equal to (dirty < thresh) and can also
1790                  * keep "1000+ dd on a slow USB stick" under control.
1791                  */
1792                 if (task_ratelimit)
1793                         break;
1794
1795                 /*
1796                  * In the case of an unresponding NFS server and the NFS dirty
1797                  * pages exceeds dirty_thresh, give the other good wb's a pipe
1798                  * to go through, so that tasks on them still remain responsive.
1799                  *
1800                  * In theory 1 page is enough to keep the consumer-producer
1801                  * pipe going: the flusher cleans 1 page => the task dirties 1
1802                  * more page. However wb_dirty has accounting errors.  So use
1803                  * the larger and more IO friendly wb_stat_error.
1804                  */
1805                 if (sdtc->wb_dirty <= wb_stat_error(wb))
1806                         break;
1807
1808                 if (fatal_signal_pending(current))
1809                         break;
1810         }
1811
1812         if (!dirty_exceeded && wb->dirty_exceeded)
1813                 wb->dirty_exceeded = 0;
1814
1815         if (writeback_in_progress(wb))
1816                 return;
1817
1818         /*
1819          * In laptop mode, we wait until hitting the higher threshold before
1820          * starting background writeout, and then write out all the way down
1821          * to the lower threshold.  So slow writers cause minimal disk activity.
1822          *
1823          * In normal mode, we start background writeout at the lower
1824          * background_thresh, to keep the amount of dirty memory low.
1825          */
1826         if (laptop_mode)
1827                 return;
1828
1829         if (nr_reclaimable > gdtc->bg_thresh)
1830                 wb_start_background_writeback(wb);
1831 }
1832
1833 static DEFINE_PER_CPU(int, bdp_ratelimits);
1834
1835 /*
1836  * Normal tasks are throttled by
1837  *      loop {
1838  *              dirty tsk->nr_dirtied_pause pages;
1839  *              take a snap in balance_dirty_pages();
1840  *      }
1841  * However there is a worst case. If every task exit immediately when dirtied
1842  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1843  * called to throttle the page dirties. The solution is to save the not yet
1844  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1845  * randomly into the running tasks. This works well for the above worst case,
1846  * as the new task will pick up and accumulate the old task's leaked dirty
1847  * count and eventually get throttled.
1848  */
1849 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1850
1851 /**
1852  * balance_dirty_pages_ratelimited - balance dirty memory state
1853  * @mapping: address_space which was dirtied
1854  *
1855  * Processes which are dirtying memory should call in here once for each page
1856  * which was newly dirtied.  The function will periodically check the system's
1857  * dirty state and will initiate writeback if needed.
1858  *
1859  * On really big machines, get_writeback_state is expensive, so try to avoid
1860  * calling it too often (ratelimiting).  But once we're over the dirty memory
1861  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1862  * from overshooting the limit by (ratelimit_pages) each.
1863  */
1864 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1865 {
1866         struct inode *inode = mapping->host;
1867         struct backing_dev_info *bdi = inode_to_bdi(inode);
1868         struct bdi_writeback *wb = NULL;
1869         int ratelimit;
1870         int *p;
1871
1872         if (!bdi_cap_account_dirty(bdi))
1873                 return;
1874
1875         if (inode_cgwb_enabled(inode))
1876                 wb = wb_get_create_current(bdi, GFP_KERNEL);
1877         if (!wb)
1878                 wb = &bdi->wb;
1879
1880         ratelimit = current->nr_dirtied_pause;
1881         if (wb->dirty_exceeded)
1882                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1883
1884         preempt_disable();
1885         /*
1886          * This prevents one CPU to accumulate too many dirtied pages without
1887          * calling into balance_dirty_pages(), which can happen when there are
1888          * 1000+ tasks, all of them start dirtying pages at exactly the same
1889          * time, hence all honoured too large initial task->nr_dirtied_pause.
1890          */
1891         p =  this_cpu_ptr(&bdp_ratelimits);
1892         if (unlikely(current->nr_dirtied >= ratelimit))
1893                 *p = 0;
1894         else if (unlikely(*p >= ratelimit_pages)) {
1895                 *p = 0;
1896                 ratelimit = 0;
1897         }
1898         /*
1899          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1900          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1901          * the dirty throttling and livelock other long-run dirtiers.
1902          */
1903         p = this_cpu_ptr(&dirty_throttle_leaks);
1904         if (*p > 0 && current->nr_dirtied < ratelimit) {
1905                 unsigned long nr_pages_dirtied;
1906                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1907                 *p -= nr_pages_dirtied;
1908                 current->nr_dirtied += nr_pages_dirtied;
1909         }
1910         preempt_enable();
1911
1912         if (unlikely(current->nr_dirtied >= ratelimit))
1913                 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1914
1915         wb_put(wb);
1916 }
1917 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1918
1919 /**
1920  * wb_over_bg_thresh - does @wb need to be written back?
1921  * @wb: bdi_writeback of interest
1922  *
1923  * Determines whether background writeback should keep writing @wb or it's
1924  * clean enough.  Returns %true if writeback should continue.
1925  */
1926 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1927 {
1928         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1929         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1930         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1931         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1932                                                      &mdtc_stor : NULL;
1933
1934         /*
1935          * Similar to balance_dirty_pages() but ignores pages being written
1936          * as we're trying to decide whether to put more under writeback.
1937          */
1938         gdtc->avail = global_dirtyable_memory();
1939         gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1940                       global_node_page_state(NR_UNSTABLE_NFS);
1941         domain_dirty_limits(gdtc);
1942
1943         if (gdtc->dirty > gdtc->bg_thresh)
1944                 return true;
1945
1946         if (wb_stat(wb, WB_RECLAIMABLE) >
1947             wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1948                 return true;
1949
1950         if (mdtc) {
1951                 unsigned long filepages, headroom, writeback;
1952
1953                 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1954                                     &writeback);
1955                 mdtc_calc_avail(mdtc, filepages, headroom);
1956                 domain_dirty_limits(mdtc);      /* ditto, ignore writeback */
1957
1958                 if (mdtc->dirty > mdtc->bg_thresh)
1959                         return true;
1960
1961                 if (wb_stat(wb, WB_RECLAIMABLE) >
1962                     wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1963                         return true;
1964         }
1965
1966         return false;
1967 }
1968
1969 /*
1970  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1971  */
1972 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1973         void __user *buffer, size_t *length, loff_t *ppos)
1974 {
1975         proc_dointvec(table, write, buffer, length, ppos);
1976         return 0;
1977 }
1978
1979 #ifdef CONFIG_BLOCK
1980 void laptop_mode_timer_fn(unsigned long data)
1981 {
1982         struct request_queue *q = (struct request_queue *)data;
1983         int nr_pages = global_node_page_state(NR_FILE_DIRTY) +
1984                 global_node_page_state(NR_UNSTABLE_NFS);
1985         struct bdi_writeback *wb;
1986
1987         /*
1988          * We want to write everything out, not just down to the dirty
1989          * threshold
1990          */
1991         if (!bdi_has_dirty_io(q->backing_dev_info))
1992                 return;
1993
1994         rcu_read_lock();
1995         list_for_each_entry_rcu(wb, &q->backing_dev_info->wb_list, bdi_node)
1996                 if (wb_has_dirty_io(wb))
1997                         wb_start_writeback(wb, nr_pages, true,
1998                                            WB_REASON_LAPTOP_TIMER);
1999         rcu_read_unlock();
2000 }
2001
2002 /*
2003  * We've spun up the disk and we're in laptop mode: schedule writeback
2004  * of all dirty data a few seconds from now.  If the flush is already scheduled
2005  * then push it back - the user is still using the disk.
2006  */
2007 void laptop_io_completion(struct backing_dev_info *info)
2008 {
2009         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2010 }
2011
2012 /*
2013  * We're in laptop mode and we've just synced. The sync's writes will have
2014  * caused another writeback to be scheduled by laptop_io_completion.
2015  * Nothing needs to be written back anymore, so we unschedule the writeback.
2016  */
2017 void laptop_sync_completion(void)
2018 {
2019         struct backing_dev_info *bdi;
2020
2021         rcu_read_lock();
2022
2023         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2024                 del_timer(&bdi->laptop_mode_wb_timer);
2025
2026         rcu_read_unlock();
2027 }
2028 #endif
2029
2030 /*
2031  * If ratelimit_pages is too high then we can get into dirty-data overload
2032  * if a large number of processes all perform writes at the same time.
2033  * If it is too low then SMP machines will call the (expensive)
2034  * get_writeback_state too often.
2035  *
2036  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2037  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2038  * thresholds.
2039  */
2040
2041 void writeback_set_ratelimit(void)
2042 {
2043         struct wb_domain *dom = &global_wb_domain;
2044         unsigned long background_thresh;
2045         unsigned long dirty_thresh;
2046
2047         global_dirty_limits(&background_thresh, &dirty_thresh);
2048         dom->dirty_limit = dirty_thresh;
2049         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2050         if (ratelimit_pages < 16)
2051                 ratelimit_pages = 16;
2052 }
2053
2054 static int page_writeback_cpu_online(unsigned int cpu)
2055 {
2056         writeback_set_ratelimit();
2057         return 0;
2058 }
2059
2060 /*
2061  * Called early on to tune the page writeback dirty limits.
2062  *
2063  * We used to scale dirty pages according to how total memory
2064  * related to pages that could be allocated for buffers (by
2065  * comparing nr_free_buffer_pages() to vm_total_pages.
2066  *
2067  * However, that was when we used "dirty_ratio" to scale with
2068  * all memory, and we don't do that any more. "dirty_ratio"
2069  * is now applied to total non-HIGHPAGE memory (by subtracting
2070  * totalhigh_pages from vm_total_pages), and as such we can't
2071  * get into the old insane situation any more where we had
2072  * large amounts of dirty pages compared to a small amount of
2073  * non-HIGHMEM memory.
2074  *
2075  * But we might still want to scale the dirty_ratio by how
2076  * much memory the box has..
2077  */
2078 void __init page_writeback_init(void)
2079 {
2080         BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2081
2082         cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2083                           page_writeback_cpu_online, NULL);
2084         cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2085                           page_writeback_cpu_online);
2086 }
2087
2088 /**
2089  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2090  * @mapping: address space structure to write
2091  * @start: starting page index
2092  * @end: ending page index (inclusive)
2093  *
2094  * This function scans the page range from @start to @end (inclusive) and tags
2095  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2096  * that write_cache_pages (or whoever calls this function) will then use
2097  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2098  * used to avoid livelocking of writeback by a process steadily creating new
2099  * dirty pages in the file (thus it is important for this function to be quick
2100  * so that it can tag pages faster than a dirtying process can create them).
2101  */
2102 /*
2103  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock 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         spin_lock_irq(&mapping->tree_lock);
2114         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, start,
2115                                                         PAGECACHE_TAG_DIRTY) {
2116                 if (iter.index > end)
2117                         break;
2118                 radix_tree_iter_tag_set(&mapping->page_tree, &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                 spin_unlock_irq(&mapping->tree_lock);
2125                 cond_resched();
2126                 spin_lock_irq(&mapping->tree_lock);
2127         }
2128         spin_unlock_irq(&mapping->tree_lock);
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, 0);
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_tag(&pvec, mapping, &index, tag,
2198                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
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                         /*
2206                          * At this point, the page may be truncated or
2207                          * invalidated (changing page->mapping to NULL), or
2208                          * even swizzled back from swapper_space to tmpfs file
2209                          * mapping. However, page->index will not change
2210                          * because we have a reference on the page.
2211                          */
2212                         if (page->index > end) {
2213                                 /*
2214                                  * can't be range_cyclic (1st pass) because
2215                                  * end == -1 in that case.
2216                                  */
2217                                 done = 1;
2218                                 break;
2219                         }
2220
2221                         done_index = page->index;
2222
2223                         lock_page(page);
2224
2225                         /*
2226                          * Page truncated or invalidated. We can freely skip it
2227                          * then, even for data integrity operations: the page
2228                          * has disappeared concurrently, so there could be no
2229                          * real expectation of this data interity operation
2230                          * even if there is now a new, dirty page at the same
2231                          * pagecache address.
2232                          */
2233                         if (unlikely(page->mapping != mapping)) {
2234 continue_unlock:
2235                                 unlock_page(page);
2236                                 continue;
2237                         }
2238
2239                         if (!PageDirty(page)) {
2240                                 /* someone wrote it for us */
2241                                 goto continue_unlock;
2242                         }
2243
2244                         if (PageWriteback(page)) {
2245                                 if (wbc->sync_mode != WB_SYNC_NONE)
2246                                         wait_on_page_writeback(page);
2247                                 else
2248                                         goto continue_unlock;
2249                         }
2250
2251                         BUG_ON(PageWriteback(page));
2252                         if (!clear_page_dirty_for_io(page))
2253                                 goto continue_unlock;
2254
2255                         trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2256                         ret = (*writepage)(page, wbc, data);
2257                         if (unlikely(ret)) {
2258                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2259                                         unlock_page(page);
2260                                         ret = 0;
2261                                 } else {
2262                                         /*
2263                                          * done_index is set past this page,
2264                                          * so media errors will not choke
2265                                          * background writeout for the entire
2266                                          * file. This has consequences for
2267                                          * range_cyclic semantics (ie. it may
2268                                          * not be suitable for data integrity
2269                                          * writeout).
2270                                          */
2271                                         done_index = page->index + 1;
2272                                         done = 1;
2273                                         break;
2274                                 }
2275                         }
2276
2277                         /*
2278                          * We stop writing back only if we are not doing
2279                          * integrity sync. In case of integrity sync we have to
2280                          * keep going until we have written all the pages
2281                          * we tagged for writeback prior to entering this loop.
2282                          */
2283                         if (--wbc->nr_to_write <= 0 &&
2284                             wbc->sync_mode == WB_SYNC_NONE) {
2285                                 done = 1;
2286                                 break;
2287                         }
2288                 }
2289                 pagevec_release(&pvec);
2290                 cond_resched();
2291         }
2292         if (!cycled && !done) {
2293                 /*
2294                  * range_cyclic:
2295                  * We hit the last page and there is more work to be done: wrap
2296                  * back to the start of the file
2297                  */
2298                 cycled = 1;
2299                 index = 0;
2300                 end = writeback_index - 1;
2301                 goto retry;
2302         }
2303         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2304                 mapping->writeback_index = done_index;
2305
2306         return ret;
2307 }
2308 EXPORT_SYMBOL(write_cache_pages);
2309
2310 /*
2311  * Function used by generic_writepages to call the real writepage
2312  * function and set the mapping flags on error
2313  */
2314 static int __writepage(struct page *page, struct writeback_control *wbc,
2315                        void *data)
2316 {
2317         struct address_space *mapping = data;
2318         int ret = mapping->a_ops->writepage(page, wbc);
2319         mapping_set_error(mapping, ret);
2320         return ret;
2321 }
2322
2323 /**
2324  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2325  * @mapping: address space structure to write
2326  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2327  *
2328  * This is a library function, which implements the writepages()
2329  * address_space_operation.
2330  */
2331 int generic_writepages(struct address_space *mapping,
2332                        struct writeback_control *wbc)
2333 {
2334         struct blk_plug plug;
2335         int ret;
2336
2337         /* deal with chardevs and other special file */
2338         if (!mapping->a_ops->writepage)
2339                 return 0;
2340
2341         blk_start_plug(&plug);
2342         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2343         blk_finish_plug(&plug);
2344         return ret;
2345 }
2346
2347 EXPORT_SYMBOL(generic_writepages);
2348
2349 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2350 {
2351         int ret;
2352
2353         if (wbc->nr_to_write <= 0)
2354                 return 0;
2355         while (1) {
2356                 if (mapping->a_ops->writepages)
2357                         ret = mapping->a_ops->writepages(mapping, wbc);
2358                 else
2359                         ret = generic_writepages(mapping, wbc);
2360                 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2361                         break;
2362                 cond_resched();
2363                 congestion_wait(BLK_RW_ASYNC, HZ/50);
2364         }
2365         return ret;
2366 }
2367
2368 /**
2369  * write_one_page - write out a single page and wait on I/O
2370  * @page: the page to write
2371  *
2372  * The page must be locked by the caller and will be unlocked upon return.
2373  *
2374  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2375  * function returns.
2376  */
2377 int write_one_page(struct page *page)
2378 {
2379         struct address_space *mapping = page->mapping;
2380         int ret = 0;
2381         struct writeback_control wbc = {
2382                 .sync_mode = WB_SYNC_ALL,
2383                 .nr_to_write = 1,
2384         };
2385
2386         BUG_ON(!PageLocked(page));
2387
2388         wait_on_page_writeback(page);
2389
2390         if (clear_page_dirty_for_io(page)) {
2391                 get_page(page);
2392                 ret = mapping->a_ops->writepage(page, &wbc);
2393                 if (ret == 0)
2394                         wait_on_page_writeback(page);
2395                 put_page(page);
2396         } else {
2397                 unlock_page(page);
2398         }
2399
2400         if (!ret)
2401                 ret = filemap_check_errors(mapping);
2402         return ret;
2403 }
2404 EXPORT_SYMBOL(write_one_page);
2405
2406 /*
2407  * For address_spaces which do not use buffers nor write back.
2408  */
2409 int __set_page_dirty_no_writeback(struct page *page)
2410 {
2411         if (!PageDirty(page))
2412                 return !TestSetPageDirty(page);
2413         return 0;
2414 }
2415
2416 /*
2417  * Helper function for set_page_dirty family.
2418  *
2419  * Caller must hold lock_page_memcg().
2420  *
2421  * NOTE: This relies on being atomic wrt interrupts.
2422  */
2423 void account_page_dirtied(struct page *page, struct address_space *mapping)
2424 {
2425         struct inode *inode = mapping->host;
2426
2427         trace_writeback_dirty_page(page, mapping);
2428
2429         if (mapping_cap_account_dirty(mapping)) {
2430                 struct bdi_writeback *wb;
2431
2432                 inode_attach_wb(inode, page);
2433                 wb = inode_to_wb(inode);
2434
2435                 inc_memcg_page_state(page, NR_FILE_DIRTY);
2436                 __inc_node_page_state(page, NR_FILE_DIRTY);
2437                 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2438                 __inc_node_page_state(page, NR_DIRTIED);
2439                 __inc_wb_stat(wb, WB_RECLAIMABLE);
2440                 __inc_wb_stat(wb, WB_DIRTIED);
2441                 task_io_account_write(PAGE_SIZE);
2442                 current->nr_dirtied++;
2443                 this_cpu_inc(bdp_ratelimits);
2444         }
2445 }
2446 EXPORT_SYMBOL(account_page_dirtied);
2447
2448 /*
2449  * Helper function for deaccounting dirty page without writeback.
2450  *
2451  * Caller must hold lock_page_memcg().
2452  */
2453 void account_page_cleaned(struct page *page, struct address_space *mapping,
2454                           struct bdi_writeback *wb)
2455 {
2456         if (mapping_cap_account_dirty(mapping)) {
2457                 dec_memcg_page_state(page, NR_FILE_DIRTY);
2458                 dec_node_page_state(page, NR_FILE_DIRTY);
2459                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2460                 dec_wb_stat(wb, WB_RECLAIMABLE);
2461                 task_io_account_cancelled_write(PAGE_SIZE);
2462         }
2463 }
2464
2465 /*
2466  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2467  * its radix tree.
2468  *
2469  * This is also used when a single buffer is being dirtied: we want to set the
2470  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2471  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2472  *
2473  * The caller must ensure this doesn't race with truncation.  Most will simply
2474  * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2475  * the pte lock held, which also locks out truncation.
2476  */
2477 int __set_page_dirty_nobuffers(struct page *page)
2478 {
2479         lock_page_memcg(page);
2480         if (!TestSetPageDirty(page)) {
2481                 struct address_space *mapping = page_mapping(page);
2482                 unsigned long flags;
2483
2484                 if (!mapping) {
2485                         unlock_page_memcg(page);
2486                         return 1;
2487                 }
2488
2489                 spin_lock_irqsave(&mapping->tree_lock, flags);
2490                 BUG_ON(page_mapping(page) != mapping);
2491                 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2492                 account_page_dirtied(page, mapping);
2493                 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2494                                    PAGECACHE_TAG_DIRTY);
2495                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2496                 unlock_page_memcg(page);
2497
2498                 if (mapping->host) {
2499                         /* !PageAnon && !swapper_space */
2500                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2501                 }
2502                 return 1;
2503         }
2504         unlock_page_memcg(page);
2505         return 0;
2506 }
2507 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2508
2509 /*
2510  * Call this whenever redirtying a page, to de-account the dirty counters
2511  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2512  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2513  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2514  * control.
2515  */
2516 void account_page_redirty(struct page *page)
2517 {
2518         struct address_space *mapping = page->mapping;
2519
2520         if (mapping && mapping_cap_account_dirty(mapping)) {
2521                 struct inode *inode = mapping->host;
2522                 struct bdi_writeback *wb;
2523                 bool locked;
2524
2525                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2526                 current->nr_dirtied--;
2527                 dec_node_page_state(page, NR_DIRTIED);
2528                 dec_wb_stat(wb, WB_DIRTIED);
2529                 unlocked_inode_to_wb_end(inode, locked);
2530         }
2531 }
2532 EXPORT_SYMBOL(account_page_redirty);
2533
2534 /*
2535  * When a writepage implementation decides that it doesn't want to write this
2536  * page for some reason, it should redirty the locked page via
2537  * redirty_page_for_writepage() and it should then unlock the page and return 0
2538  */
2539 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2540 {
2541         int ret;
2542
2543         wbc->pages_skipped++;
2544         ret = __set_page_dirty_nobuffers(page);
2545         account_page_redirty(page);
2546         return ret;
2547 }
2548 EXPORT_SYMBOL(redirty_page_for_writepage);
2549
2550 /*
2551  * Dirty a page.
2552  *
2553  * For pages with a mapping this should be done under the page lock
2554  * for the benefit of asynchronous memory errors who prefer a consistent
2555  * dirty state. This rule can be broken in some special cases,
2556  * but should be better not to.
2557  *
2558  * If the mapping doesn't provide a set_page_dirty a_op, then
2559  * just fall through and assume that it wants buffer_heads.
2560  */
2561 int set_page_dirty(struct page *page)
2562 {
2563         struct address_space *mapping = page_mapping(page);
2564
2565         page = compound_head(page);
2566         if (likely(mapping)) {
2567                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2568                 /*
2569                  * readahead/lru_deactivate_page could remain
2570                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2571                  * About readahead, if the page is written, the flags would be
2572                  * reset. So no problem.
2573                  * About lru_deactivate_page, if the page is redirty, the flag
2574                  * will be reset. So no problem. but if the page is used by readahead
2575                  * it will confuse readahead and make it restart the size rampup
2576                  * process. But it's a trivial problem.
2577                  */
2578                 if (PageReclaim(page))
2579                         ClearPageReclaim(page);
2580 #ifdef CONFIG_BLOCK
2581                 if (!spd)
2582                         spd = __set_page_dirty_buffers;
2583 #endif
2584                 return (*spd)(page);
2585         }
2586         if (!PageDirty(page)) {
2587                 if (!TestSetPageDirty(page))
2588                         return 1;
2589         }
2590         return 0;
2591 }
2592 EXPORT_SYMBOL(set_page_dirty);
2593
2594 /*
2595  * set_page_dirty() is racy if the caller has no reference against
2596  * page->mapping->host, and if the page is unlocked.  This is because another
2597  * CPU could truncate the page off the mapping and then free the mapping.
2598  *
2599  * Usually, the page _is_ locked, or the caller is a user-space process which
2600  * holds a reference on the inode by having an open file.
2601  *
2602  * In other cases, the page should be locked before running set_page_dirty().
2603  */
2604 int set_page_dirty_lock(struct page *page)
2605 {
2606         int ret;
2607
2608         lock_page(page);
2609         ret = set_page_dirty(page);
2610         unlock_page(page);
2611         return ret;
2612 }
2613 EXPORT_SYMBOL(set_page_dirty_lock);
2614
2615 /*
2616  * This cancels just the dirty bit on the kernel page itself, it does NOT
2617  * actually remove dirty bits on any mmap's that may be around. It also
2618  * leaves the page tagged dirty, so any sync activity will still find it on
2619  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2620  * look at the dirty bits in the VM.
2621  *
2622  * Doing this should *normally* only ever be done when a page is truncated,
2623  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2624  * this when it notices that somebody has cleaned out all the buffers on a
2625  * page without actually doing it through the VM. Can you say "ext3 is
2626  * horribly ugly"? Thought you could.
2627  */
2628 void cancel_dirty_page(struct page *page)
2629 {
2630         struct address_space *mapping = page_mapping(page);
2631
2632         if (mapping_cap_account_dirty(mapping)) {
2633                 struct inode *inode = mapping->host;
2634                 struct bdi_writeback *wb;
2635                 bool locked;
2636
2637                 lock_page_memcg(page);
2638                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2639
2640                 if (TestClearPageDirty(page))
2641                         account_page_cleaned(page, mapping, wb);
2642
2643                 unlocked_inode_to_wb_end(inode, locked);
2644                 unlock_page_memcg(page);
2645         } else {
2646                 ClearPageDirty(page);
2647         }
2648 }
2649 EXPORT_SYMBOL(cancel_dirty_page);
2650
2651 /*
2652  * Clear a page's dirty flag, while caring for dirty memory accounting.
2653  * Returns true if the page was previously dirty.
2654  *
2655  * This is for preparing to put the page under writeout.  We leave the page
2656  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2657  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2658  * implementation will run either set_page_writeback() or set_page_dirty(),
2659  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2660  * back into sync.
2661  *
2662  * This incoherency between the page's dirty flag and radix-tree tag is
2663  * unfortunate, but it only exists while the page is locked.
2664  */
2665 int clear_page_dirty_for_io(struct page *page)
2666 {
2667         struct address_space *mapping = page_mapping(page);
2668         int ret = 0;
2669
2670         BUG_ON(!PageLocked(page));
2671
2672         if (mapping && mapping_cap_account_dirty(mapping)) {
2673                 struct inode *inode = mapping->host;
2674                 struct bdi_writeback *wb;
2675                 bool locked;
2676
2677                 /*
2678                  * Yes, Virginia, this is indeed insane.
2679                  *
2680                  * We use this sequence to make sure that
2681                  *  (a) we account for dirty stats properly
2682                  *  (b) we tell the low-level filesystem to
2683                  *      mark the whole page dirty if it was
2684                  *      dirty in a pagetable. Only to then
2685                  *  (c) clean the page again and return 1 to
2686                  *      cause the writeback.
2687                  *
2688                  * This way we avoid all nasty races with the
2689                  * dirty bit in multiple places and clearing
2690                  * them concurrently from different threads.
2691                  *
2692                  * Note! Normally the "set_page_dirty(page)"
2693                  * has no effect on the actual dirty bit - since
2694                  * that will already usually be set. But we
2695                  * need the side effects, and it can help us
2696                  * avoid races.
2697                  *
2698                  * We basically use the page "master dirty bit"
2699                  * as a serialization point for all the different
2700                  * threads doing their things.
2701                  */
2702                 if (page_mkclean(page))
2703                         set_page_dirty(page);
2704                 /*
2705                  * We carefully synchronise fault handlers against
2706                  * installing a dirty pte and marking the page dirty
2707                  * at this point.  We do this by having them hold the
2708                  * page lock while dirtying the page, and pages are
2709                  * always locked coming in here, so we get the desired
2710                  * exclusion.
2711                  */
2712                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2713                 if (TestClearPageDirty(page)) {
2714                         dec_memcg_page_state(page, NR_FILE_DIRTY);
2715                         dec_node_page_state(page, NR_FILE_DIRTY);
2716                         dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2717                         dec_wb_stat(wb, WB_RECLAIMABLE);
2718                         ret = 1;
2719                 }
2720                 unlocked_inode_to_wb_end(inode, locked);
2721                 return ret;
2722         }
2723         return TestClearPageDirty(page);
2724 }
2725 EXPORT_SYMBOL(clear_page_dirty_for_io);
2726
2727 int test_clear_page_writeback(struct page *page)
2728 {
2729         struct address_space *mapping = page_mapping(page);
2730         int ret;
2731
2732         lock_page_memcg(page);
2733         if (mapping && mapping_use_writeback_tags(mapping)) {
2734                 struct inode *inode = mapping->host;
2735                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2736                 unsigned long flags;
2737
2738                 spin_lock_irqsave(&mapping->tree_lock, flags);
2739                 ret = TestClearPageWriteback(page);
2740                 if (ret) {
2741                         radix_tree_tag_clear(&mapping->page_tree,
2742                                                 page_index(page),
2743                                                 PAGECACHE_TAG_WRITEBACK);
2744                         if (bdi_cap_account_writeback(bdi)) {
2745                                 struct bdi_writeback *wb = inode_to_wb(inode);
2746
2747                                 __dec_wb_stat(wb, WB_WRITEBACK);
2748                                 __wb_writeout_inc(wb);
2749                         }
2750                 }
2751
2752                 if (mapping->host && !mapping_tagged(mapping,
2753                                                      PAGECACHE_TAG_WRITEBACK))
2754                         sb_clear_inode_writeback(mapping->host);
2755
2756                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2757         } else {
2758                 ret = TestClearPageWriteback(page);
2759         }
2760         if (ret) {
2761                 dec_memcg_page_state(page, NR_WRITEBACK);
2762                 dec_node_page_state(page, NR_WRITEBACK);
2763                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2764                 inc_node_page_state(page, NR_WRITTEN);
2765         }
2766         unlock_page_memcg(page);
2767         return ret;
2768 }
2769
2770 int __test_set_page_writeback(struct page *page, bool keep_write)
2771 {
2772         struct address_space *mapping = page_mapping(page);
2773         int ret;
2774
2775         lock_page_memcg(page);
2776         if (mapping && mapping_use_writeback_tags(mapping)) {
2777                 struct inode *inode = mapping->host;
2778                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2779                 unsigned long flags;
2780
2781                 spin_lock_irqsave(&mapping->tree_lock, flags);
2782                 ret = TestSetPageWriteback(page);
2783                 if (!ret) {
2784                         bool on_wblist;
2785
2786                         on_wblist = mapping_tagged(mapping,
2787                                                    PAGECACHE_TAG_WRITEBACK);
2788
2789                         radix_tree_tag_set(&mapping->page_tree,
2790                                                 page_index(page),
2791                                                 PAGECACHE_TAG_WRITEBACK);
2792                         if (bdi_cap_account_writeback(bdi))
2793                                 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2794
2795                         /*
2796                          * We can come through here when swapping anonymous
2797                          * pages, so we don't necessarily have an inode to track
2798                          * for sync.
2799                          */
2800                         if (mapping->host && !on_wblist)
2801                                 sb_mark_inode_writeback(mapping->host);
2802                 }
2803                 if (!PageDirty(page))
2804                         radix_tree_tag_clear(&mapping->page_tree,
2805                                                 page_index(page),
2806                                                 PAGECACHE_TAG_DIRTY);
2807                 if (!keep_write)
2808                         radix_tree_tag_clear(&mapping->page_tree,
2809                                                 page_index(page),
2810                                                 PAGECACHE_TAG_TOWRITE);
2811                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2812         } else {
2813                 ret = TestSetPageWriteback(page);
2814         }
2815         if (!ret) {
2816                 inc_memcg_page_state(page, NR_WRITEBACK);
2817                 inc_node_page_state(page, NR_WRITEBACK);
2818                 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2819         }
2820         unlock_page_memcg(page);
2821         return ret;
2822
2823 }
2824 EXPORT_SYMBOL(__test_set_page_writeback);
2825
2826 /*
2827  * Return true if any of the pages in the mapping are marked with the
2828  * passed tag.
2829  */
2830 int mapping_tagged(struct address_space *mapping, int tag)
2831 {
2832         return radix_tree_tagged(&mapping->page_tree, tag);
2833 }
2834 EXPORT_SYMBOL(mapping_tagged);
2835
2836 /**
2837  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2838  * @page:       The page to wait on.
2839  *
2840  * This function determines if the given page is related to a backing device
2841  * that requires page contents to be held stable during writeback.  If so, then
2842  * it will wait for any pending writeback to complete.
2843  */
2844 void wait_for_stable_page(struct page *page)
2845 {
2846         if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2847                 wait_on_page_writeback(page);
2848 }
2849 EXPORT_SYMBOL_GPL(wait_for_stable_page);