mm: use pagevec_lookup_range_tag() in write_cache_pages()
[sfrench/cifs-2.6.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    Andrew Morton
11  *              Initial version
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/sched/signal.h>
40 #include <linux/mm_inline.h>
41 #include <trace/events/writeback.h>
42
43 #include "internal.h"
44
45 /*
46  * Sleep at most 200ms at a time in balance_dirty_pages().
47  */
48 #define MAX_PAUSE               max(HZ/5, 1)
49
50 /*
51  * Try to keep balance_dirty_pages() call intervals higher than this many pages
52  * by raising pause time to max_pause when falls below it.
53  */
54 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
55
56 /*
57  * Estimate write bandwidth at 200ms intervals.
58  */
59 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
60
61 #define RATELIMIT_CALC_SHIFT    10
62
63 /*
64  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65  * will look to see if it needs to force writeback or throttling.
66  */
67 static long ratelimit_pages = 32;
68
69 /* The following parameters are exported via /proc/sys/vm */
70
71 /*
72  * Start background writeback (via writeback threads) at this percentage
73  */
74 int dirty_background_ratio = 10;
75
76 /*
77  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78  * dirty_background_ratio * the amount of dirtyable memory
79  */
80 unsigned long dirty_background_bytes;
81
82 /*
83  * free highmem will not be subtracted from the total free memory
84  * for calculating free ratios if vm_highmem_is_dirtyable is true
85  */
86 int vm_highmem_is_dirtyable;
87
88 /*
89  * The generator of dirty data starts writeback at this percentage
90  */
91 int vm_dirty_ratio = 20;
92
93 /*
94  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95  * vm_dirty_ratio * the amount of dirtyable memory
96  */
97 unsigned long vm_dirty_bytes;
98
99 /*
100  * The interval between `kupdate'-style writebacks
101  */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106 /*
107  * The longest time for which data is allowed to remain dirty
108  */
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111 /*
112  * Flag that makes the machine dump writes/reads and block dirtyings.
113  */
114 int block_dump;
115
116 /*
117  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118  * a full sync is triggered after this time elapses without any disk activity.
119  */
120 int laptop_mode;
121
122 EXPORT_SYMBOL(laptop_mode);
123
124 /* End of sysctl-exported parameters */
125
126 struct wb_domain global_wb_domain;
127
128 /* consolidated parameters for balance_dirty_pages() and its subroutines */
129 struct dirty_throttle_control {
130 #ifdef CONFIG_CGROUP_WRITEBACK
131         struct wb_domain        *dom;
132         struct dirty_throttle_control *gdtc;    /* only set in memcg dtc's */
133 #endif
134         struct bdi_writeback    *wb;
135         struct fprop_local_percpu *wb_completions;
136
137         unsigned long           avail;          /* dirtyable */
138         unsigned long           dirty;          /* file_dirty + write + nfs */
139         unsigned long           thresh;         /* dirty threshold */
140         unsigned long           bg_thresh;      /* dirty background threshold */
141
142         unsigned long           wb_dirty;       /* per-wb counterparts */
143         unsigned long           wb_thresh;
144         unsigned long           wb_bg_thresh;
145
146         unsigned long           pos_ratio;
147 };
148
149 /*
150  * Length of period for aging writeout fractions of bdis. This is an
151  * arbitrarily chosen number. The longer the period, the slower fractions will
152  * reflect changes in current writeout rate.
153  */
154 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155
156 #ifdef CONFIG_CGROUP_WRITEBACK
157
158 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
159                                 .dom = &global_wb_domain,               \
160                                 .wb_completions = &(__wb)->completions
161
162 #define GDTC_INIT_NO_WB         .dom = &global_wb_domain
163
164 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb),                           \
165                                 .dom = mem_cgroup_wb_domain(__wb),      \
166                                 .wb_completions = &(__wb)->memcg_completions, \
167                                 .gdtc = __gdtc
168
169 static bool mdtc_valid(struct dirty_throttle_control *dtc)
170 {
171         return dtc->dom;
172 }
173
174 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175 {
176         return dtc->dom;
177 }
178
179 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180 {
181         return mdtc->gdtc;
182 }
183
184 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185 {
186         return &wb->memcg_completions;
187 }
188
189 static void wb_min_max_ratio(struct bdi_writeback *wb,
190                              unsigned long *minp, unsigned long *maxp)
191 {
192         unsigned long this_bw = wb->avg_write_bandwidth;
193         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194         unsigned long long min = wb->bdi->min_ratio;
195         unsigned long long max = wb->bdi->max_ratio;
196
197         /*
198          * @wb may already be clean by the time control reaches here and
199          * the total may not include its bw.
200          */
201         if (this_bw < tot_bw) {
202                 if (min) {
203                         min *= this_bw;
204                         do_div(min, tot_bw);
205                 }
206                 if (max < 100) {
207                         max *= this_bw;
208                         do_div(max, tot_bw);
209                 }
210         }
211
212         *minp = min;
213         *maxp = max;
214 }
215
216 #else   /* CONFIG_CGROUP_WRITEBACK */
217
218 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
219                                 .wb_completions = &(__wb)->completions
220 #define GDTC_INIT_NO_WB
221 #define MDTC_INIT(__wb, __gdtc)
222
223 static bool mdtc_valid(struct dirty_throttle_control *dtc)
224 {
225         return false;
226 }
227
228 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229 {
230         return &global_wb_domain;
231 }
232
233 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234 {
235         return NULL;
236 }
237
238 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239 {
240         return NULL;
241 }
242
243 static void wb_min_max_ratio(struct bdi_writeback *wb,
244                              unsigned long *minp, unsigned long *maxp)
245 {
246         *minp = wb->bdi->min_ratio;
247         *maxp = wb->bdi->max_ratio;
248 }
249
250 #endif  /* CONFIG_CGROUP_WRITEBACK */
251
252 /*
253  * In a memory zone, there is a certain amount of pages we consider
254  * available for the page cache, which is essentially the number of
255  * free and reclaimable pages, minus some zone reserves to protect
256  * lowmem and the ability to uphold the zone's watermarks without
257  * requiring writeback.
258  *
259  * This number of dirtyable pages is the base value of which the
260  * user-configurable dirty ratio is the effictive number of pages that
261  * are allowed to be actually dirtied.  Per individual zone, or
262  * globally by using the sum of dirtyable pages over all zones.
263  *
264  * Because the user is allowed to specify the dirty limit globally as
265  * absolute number of bytes, calculating the per-zone dirty limit can
266  * require translating the configured limit into a percentage of
267  * global dirtyable memory first.
268  */
269
270 /**
271  * node_dirtyable_memory - number of dirtyable pages in a node
272  * @pgdat: the node
273  *
274  * Returns the node's number of pages potentially available for dirty
275  * page cache.  This is the base value for the per-node dirty limits.
276  */
277 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278 {
279         unsigned long nr_pages = 0;
280         int z;
281
282         for (z = 0; z < MAX_NR_ZONES; z++) {
283                 struct zone *zone = pgdat->node_zones + z;
284
285                 if (!populated_zone(zone))
286                         continue;
287
288                 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289         }
290
291         /*
292          * Pages reserved for the kernel should not be considered
293          * dirtyable, to prevent a situation where reclaim has to
294          * clean pages in order to balance the zones.
295          */
296         nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297
298         nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299         nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
300
301         return nr_pages;
302 }
303
304 static unsigned long highmem_dirtyable_memory(unsigned long total)
305 {
306 #ifdef CONFIG_HIGHMEM
307         int node;
308         unsigned long x = 0;
309         int i;
310
311         for_each_node_state(node, N_HIGH_MEMORY) {
312                 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313                         struct zone *z;
314                         unsigned long nr_pages;
315
316                         if (!is_highmem_idx(i))
317                                 continue;
318
319                         z = &NODE_DATA(node)->node_zones[i];
320                         if (!populated_zone(z))
321                                 continue;
322
323                         nr_pages = zone_page_state(z, NR_FREE_PAGES);
324                         /* watch for underflows */
325                         nr_pages -= min(nr_pages, high_wmark_pages(z));
326                         nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327                         nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328                         x += nr_pages;
329                 }
330         }
331
332         /*
333          * Unreclaimable memory (kernel memory or anonymous memory
334          * without swap) can bring down the dirtyable pages below
335          * the zone's dirty balance reserve and the above calculation
336          * will underflow.  However we still want to add in nodes
337          * which are below threshold (negative values) to get a more
338          * accurate calculation but make sure that the total never
339          * underflows.
340          */
341         if ((long)x < 0)
342                 x = 0;
343
344         /*
345          * Make sure that the number of highmem pages is never larger
346          * than the number of the total dirtyable memory. This can only
347          * occur in very strange VM situations but we want to make sure
348          * that this does not occur.
349          */
350         return min(x, total);
351 #else
352         return 0;
353 #endif
354 }
355
356 /**
357  * global_dirtyable_memory - number of globally dirtyable pages
358  *
359  * Returns the global number of pages potentially available for dirty
360  * page cache.  This is the base value for the global dirty limits.
361  */
362 static unsigned long global_dirtyable_memory(void)
363 {
364         unsigned long x;
365
366         x = global_zone_page_state(NR_FREE_PAGES);
367         /*
368          * Pages reserved for the kernel should not be considered
369          * dirtyable, to prevent a situation where reclaim has to
370          * clean pages in order to balance the zones.
371          */
372         x -= min(x, totalreserve_pages);
373
374         x += global_node_page_state(NR_INACTIVE_FILE);
375         x += global_node_page_state(NR_ACTIVE_FILE);
376
377         if (!vm_highmem_is_dirtyable)
378                 x -= highmem_dirtyable_memory(x);
379
380         return x + 1;   /* Ensure that we never return 0 */
381 }
382
383 /**
384  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385  * @dtc: dirty_throttle_control of interest
386  *
387  * Calculate @dtc->thresh and ->bg_thresh considering
388  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
389  * must ensure that @dtc->avail is set before calling this function.  The
390  * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
391  * real-time tasks.
392  */
393 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
394 {
395         const unsigned long available_memory = dtc->avail;
396         struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
397         unsigned long bytes = vm_dirty_bytes;
398         unsigned long bg_bytes = dirty_background_bytes;
399         /* convert ratios to per-PAGE_SIZE for higher precision */
400         unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
401         unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
402         unsigned long thresh;
403         unsigned long bg_thresh;
404         struct task_struct *tsk;
405
406         /* gdtc is !NULL iff @dtc is for memcg domain */
407         if (gdtc) {
408                 unsigned long global_avail = gdtc->avail;
409
410                 /*
411                  * The byte settings can't be applied directly to memcg
412                  * domains.  Convert them to ratios by scaling against
413                  * globally available memory.  As the ratios are in
414                  * per-PAGE_SIZE, they can be obtained by dividing bytes by
415                  * number of pages.
416                  */
417                 if (bytes)
418                         ratio = min(DIV_ROUND_UP(bytes, global_avail),
419                                     PAGE_SIZE);
420                 if (bg_bytes)
421                         bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
422                                        PAGE_SIZE);
423                 bytes = bg_bytes = 0;
424         }
425
426         if (bytes)
427                 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
428         else
429                 thresh = (ratio * available_memory) / PAGE_SIZE;
430
431         if (bg_bytes)
432                 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
433         else
434                 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
435
436         if (unlikely(bg_thresh >= thresh)) {
437                 pr_warn("vm direct limit must be set greater than background limit.\n");
438                 bg_thresh = thresh / 2;
439         }
440
441         tsk = current;
442         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
443                 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
444                 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
445         }
446         dtc->thresh = thresh;
447         dtc->bg_thresh = bg_thresh;
448
449         /* we should eventually report the domain in the TP */
450         if (!gdtc)
451                 trace_global_dirty_state(bg_thresh, thresh);
452 }
453
454 /**
455  * global_dirty_limits - background-writeback and dirty-throttling thresholds
456  * @pbackground: out parameter for bg_thresh
457  * @pdirty: out parameter for thresh
458  *
459  * Calculate bg_thresh and thresh for global_wb_domain.  See
460  * domain_dirty_limits() for details.
461  */
462 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
463 {
464         struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
465
466         gdtc.avail = global_dirtyable_memory();
467         domain_dirty_limits(&gdtc);
468
469         *pbackground = gdtc.bg_thresh;
470         *pdirty = gdtc.thresh;
471 }
472
473 /**
474  * node_dirty_limit - maximum number of dirty pages allowed in a node
475  * @pgdat: the node
476  *
477  * Returns the maximum number of dirty pages allowed in a node, based
478  * on the node's dirtyable memory.
479  */
480 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
481 {
482         unsigned long node_memory = node_dirtyable_memory(pgdat);
483         struct task_struct *tsk = current;
484         unsigned long dirty;
485
486         if (vm_dirty_bytes)
487                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
488                         node_memory / global_dirtyable_memory();
489         else
490                 dirty = vm_dirty_ratio * node_memory / 100;
491
492         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
493                 dirty += dirty / 4;
494
495         return dirty;
496 }
497
498 /**
499  * node_dirty_ok - tells whether a node is within its dirty limits
500  * @pgdat: the node to check
501  *
502  * Returns %true when the dirty pages in @pgdat are within the node's
503  * dirty limit, %false if the limit is exceeded.
504  */
505 bool node_dirty_ok(struct pglist_data *pgdat)
506 {
507         unsigned long limit = node_dirty_limit(pgdat);
508         unsigned long nr_pages = 0;
509
510         nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
511         nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
512         nr_pages += node_page_state(pgdat, NR_WRITEBACK);
513
514         return nr_pages <= limit;
515 }
516
517 int dirty_background_ratio_handler(struct ctl_table *table, int write,
518                 void __user *buffer, size_t *lenp,
519                 loff_t *ppos)
520 {
521         int ret;
522
523         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524         if (ret == 0 && write)
525                 dirty_background_bytes = 0;
526         return ret;
527 }
528
529 int dirty_background_bytes_handler(struct ctl_table *table, int write,
530                 void __user *buffer, size_t *lenp,
531                 loff_t *ppos)
532 {
533         int ret;
534
535         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
536         if (ret == 0 && write)
537                 dirty_background_ratio = 0;
538         return ret;
539 }
540
541 int dirty_ratio_handler(struct ctl_table *table, int write,
542                 void __user *buffer, size_t *lenp,
543                 loff_t *ppos)
544 {
545         int old_ratio = vm_dirty_ratio;
546         int ret;
547
548         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
549         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
550                 writeback_set_ratelimit();
551                 vm_dirty_bytes = 0;
552         }
553         return ret;
554 }
555
556 int dirty_bytes_handler(struct ctl_table *table, int write,
557                 void __user *buffer, size_t *lenp,
558                 loff_t *ppos)
559 {
560         unsigned long old_bytes = vm_dirty_bytes;
561         int ret;
562
563         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
564         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
565                 writeback_set_ratelimit();
566                 vm_dirty_ratio = 0;
567         }
568         return ret;
569 }
570
571 static unsigned long wp_next_time(unsigned long cur_time)
572 {
573         cur_time += VM_COMPLETIONS_PERIOD_LEN;
574         /* 0 has a special meaning... */
575         if (!cur_time)
576                 return 1;
577         return cur_time;
578 }
579
580 static void wb_domain_writeout_inc(struct wb_domain *dom,
581                                    struct fprop_local_percpu *completions,
582                                    unsigned int max_prop_frac)
583 {
584         __fprop_inc_percpu_max(&dom->completions, completions,
585                                max_prop_frac);
586         /* First event after period switching was turned off? */
587         if (unlikely(!dom->period_time)) {
588                 /*
589                  * We can race with other __bdi_writeout_inc calls here but
590                  * it does not cause any harm since the resulting time when
591                  * timer will fire and what is in writeout_period_time will be
592                  * roughly the same.
593                  */
594                 dom->period_time = wp_next_time(jiffies);
595                 mod_timer(&dom->period_timer, dom->period_time);
596         }
597 }
598
599 /*
600  * Increment @wb's writeout completion count and the global writeout
601  * completion count. Called from test_clear_page_writeback().
602  */
603 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
604 {
605         struct wb_domain *cgdom;
606
607         inc_wb_stat(wb, WB_WRITTEN);
608         wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
609                                wb->bdi->max_prop_frac);
610
611         cgdom = mem_cgroup_wb_domain(wb);
612         if (cgdom)
613                 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
614                                        wb->bdi->max_prop_frac);
615 }
616
617 void wb_writeout_inc(struct bdi_writeback *wb)
618 {
619         unsigned long flags;
620
621         local_irq_save(flags);
622         __wb_writeout_inc(wb);
623         local_irq_restore(flags);
624 }
625 EXPORT_SYMBOL_GPL(wb_writeout_inc);
626
627 /*
628  * On idle system, we can be called long after we scheduled because we use
629  * deferred timers so count with missed periods.
630  */
631 static void writeout_period(unsigned long t)
632 {
633         struct wb_domain *dom = (void *)t;
634         int miss_periods = (jiffies - dom->period_time) /
635                                                  VM_COMPLETIONS_PERIOD_LEN;
636
637         if (fprop_new_period(&dom->completions, miss_periods + 1)) {
638                 dom->period_time = wp_next_time(dom->period_time +
639                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
640                 mod_timer(&dom->period_timer, dom->period_time);
641         } else {
642                 /*
643                  * Aging has zeroed all fractions. Stop wasting CPU on period
644                  * updates.
645                  */
646                 dom->period_time = 0;
647         }
648 }
649
650 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
651 {
652         memset(dom, 0, sizeof(*dom));
653
654         spin_lock_init(&dom->lock);
655
656         setup_deferrable_timer(&dom->period_timer, writeout_period,
657                                (unsigned long)dom);
658
659         dom->dirty_limit_tstamp = jiffies;
660
661         return fprop_global_init(&dom->completions, gfp);
662 }
663
664 #ifdef CONFIG_CGROUP_WRITEBACK
665 void wb_domain_exit(struct wb_domain *dom)
666 {
667         del_timer_sync(&dom->period_timer);
668         fprop_global_destroy(&dom->completions);
669 }
670 #endif
671
672 /*
673  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
674  * registered backing devices, which, for obvious reasons, can not
675  * exceed 100%.
676  */
677 static unsigned int bdi_min_ratio;
678
679 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
680 {
681         int ret = 0;
682
683         spin_lock_bh(&bdi_lock);
684         if (min_ratio > bdi->max_ratio) {
685                 ret = -EINVAL;
686         } else {
687                 min_ratio -= bdi->min_ratio;
688                 if (bdi_min_ratio + min_ratio < 100) {
689                         bdi_min_ratio += min_ratio;
690                         bdi->min_ratio += min_ratio;
691                 } else {
692                         ret = -EINVAL;
693                 }
694         }
695         spin_unlock_bh(&bdi_lock);
696
697         return ret;
698 }
699
700 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
701 {
702         int ret = 0;
703
704         if (max_ratio > 100)
705                 return -EINVAL;
706
707         spin_lock_bh(&bdi_lock);
708         if (bdi->min_ratio > max_ratio) {
709                 ret = -EINVAL;
710         } else {
711                 bdi->max_ratio = max_ratio;
712                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
713         }
714         spin_unlock_bh(&bdi_lock);
715
716         return ret;
717 }
718 EXPORT_SYMBOL(bdi_set_max_ratio);
719
720 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
721                                            unsigned long bg_thresh)
722 {
723         return (thresh + bg_thresh) / 2;
724 }
725
726 static unsigned long hard_dirty_limit(struct wb_domain *dom,
727                                       unsigned long thresh)
728 {
729         return max(thresh, dom->dirty_limit);
730 }
731
732 /*
733  * Memory which can be further allocated to a memcg domain is capped by
734  * system-wide clean memory excluding the amount being used in the domain.
735  */
736 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
737                             unsigned long filepages, unsigned long headroom)
738 {
739         struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
740         unsigned long clean = filepages - min(filepages, mdtc->dirty);
741         unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
742         unsigned long other_clean = global_clean - min(global_clean, clean);
743
744         mdtc->avail = filepages + min(headroom, other_clean);
745 }
746
747 /**
748  * __wb_calc_thresh - @wb's share of dirty throttling threshold
749  * @dtc: dirty_throttle_context of interest
750  *
751  * Returns @wb's dirty limit in pages. The term "dirty" in the context of
752  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
753  *
754  * Note that balance_dirty_pages() will only seriously take it as a hard limit
755  * when sleeping max_pause per page is not enough to keep the dirty pages under
756  * control. For example, when the device is completely stalled due to some error
757  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
758  * In the other normal situations, it acts more gently by throttling the tasks
759  * more (rather than completely block them) when the wb dirty pages go high.
760  *
761  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
762  * - starving fast devices
763  * - piling up dirty pages (that will take long time to sync) on slow devices
764  *
765  * The wb's share of dirty limit will be adapting to its throughput and
766  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
767  */
768 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
769 {
770         struct wb_domain *dom = dtc_dom(dtc);
771         unsigned long thresh = dtc->thresh;
772         u64 wb_thresh;
773         long numerator, denominator;
774         unsigned long wb_min_ratio, wb_max_ratio;
775
776         /*
777          * Calculate this BDI's share of the thresh ratio.
778          */
779         fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
780                               &numerator, &denominator);
781
782         wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
783         wb_thresh *= numerator;
784         do_div(wb_thresh, denominator);
785
786         wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
787
788         wb_thresh += (thresh * wb_min_ratio) / 100;
789         if (wb_thresh > (thresh * wb_max_ratio) / 100)
790                 wb_thresh = thresh * wb_max_ratio / 100;
791
792         return wb_thresh;
793 }
794
795 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
796 {
797         struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
798                                                .thresh = thresh };
799         return __wb_calc_thresh(&gdtc);
800 }
801
802 /*
803  *                           setpoint - dirty 3
804  *        f(dirty) := 1.0 + (----------------)
805  *                           limit - setpoint
806  *
807  * it's a 3rd order polynomial that subjects to
808  *
809  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
810  * (2) f(setpoint) = 1.0 => the balance point
811  * (3) f(limit)    = 0   => the hard limit
812  * (4) df/dx      <= 0   => negative feedback control
813  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
814  *     => fast response on large errors; small oscillation near setpoint
815  */
816 static long long pos_ratio_polynom(unsigned long setpoint,
817                                           unsigned long dirty,
818                                           unsigned long limit)
819 {
820         long long pos_ratio;
821         long x;
822
823         x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
824                       (limit - setpoint) | 1);
825         pos_ratio = x;
826         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
827         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
828         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
829
830         return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
831 }
832
833 /*
834  * Dirty position control.
835  *
836  * (o) global/bdi setpoints
837  *
838  * We want the dirty pages be balanced around the global/wb setpoints.
839  * When the number of dirty pages is higher/lower than the setpoint, the
840  * dirty position control ratio (and hence task dirty ratelimit) will be
841  * decreased/increased to bring the dirty pages back to the setpoint.
842  *
843  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
844  *
845  *     if (dirty < setpoint) scale up   pos_ratio
846  *     if (dirty > setpoint) scale down pos_ratio
847  *
848  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
849  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
850  *
851  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
852  *
853  * (o) global control line
854  *
855  *     ^ pos_ratio
856  *     |
857  *     |            |<===== global dirty control scope ======>|
858  * 2.0 .............*
859  *     |            .*
860  *     |            . *
861  *     |            .   *
862  *     |            .     *
863  *     |            .        *
864  *     |            .            *
865  * 1.0 ................................*
866  *     |            .                  .     *
867  *     |            .                  .          *
868  *     |            .                  .              *
869  *     |            .                  .                 *
870  *     |            .                  .                    *
871  *   0 +------------.------------------.----------------------*------------->
872  *           freerun^          setpoint^                 limit^   dirty pages
873  *
874  * (o) wb control line
875  *
876  *     ^ pos_ratio
877  *     |
878  *     |            *
879  *     |              *
880  *     |                *
881  *     |                  *
882  *     |                    * |<=========== span ============>|
883  * 1.0 .......................*
884  *     |                      . *
885  *     |                      .   *
886  *     |                      .     *
887  *     |                      .       *
888  *     |                      .         *
889  *     |                      .           *
890  *     |                      .             *
891  *     |                      .               *
892  *     |                      .                 *
893  *     |                      .                   *
894  *     |                      .                     *
895  * 1/4 ...............................................* * * * * * * * * * * *
896  *     |                      .                         .
897  *     |                      .                           .
898  *     |                      .                             .
899  *   0 +----------------------.-------------------------------.------------->
900  *                wb_setpoint^                    x_intercept^
901  *
902  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
903  * be smoothly throttled down to normal if it starts high in situations like
904  * - start writing to a slow SD card and a fast disk at the same time. The SD
905  *   card's wb_dirty may rush to many times higher than wb_setpoint.
906  * - the wb dirty thresh drops quickly due to change of JBOD workload
907  */
908 static void wb_position_ratio(struct dirty_throttle_control *dtc)
909 {
910         struct bdi_writeback *wb = dtc->wb;
911         unsigned long write_bw = wb->avg_write_bandwidth;
912         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
913         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
914         unsigned long wb_thresh = dtc->wb_thresh;
915         unsigned long x_intercept;
916         unsigned long setpoint;         /* dirty pages' target balance point */
917         unsigned long wb_setpoint;
918         unsigned long span;
919         long long pos_ratio;            /* for scaling up/down the rate limit */
920         long x;
921
922         dtc->pos_ratio = 0;
923
924         if (unlikely(dtc->dirty >= limit))
925                 return;
926
927         /*
928          * global setpoint
929          *
930          * See comment for pos_ratio_polynom().
931          */
932         setpoint = (freerun + limit) / 2;
933         pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
934
935         /*
936          * The strictlimit feature is a tool preventing mistrusted filesystems
937          * from growing a large number of dirty pages before throttling. For
938          * such filesystems balance_dirty_pages always checks wb counters
939          * against wb limits. Even if global "nr_dirty" is under "freerun".
940          * This is especially important for fuse which sets bdi->max_ratio to
941          * 1% by default. Without strictlimit feature, fuse writeback may
942          * consume arbitrary amount of RAM because it is accounted in
943          * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
944          *
945          * Here, in wb_position_ratio(), we calculate pos_ratio based on
946          * two values: wb_dirty and wb_thresh. Let's consider an example:
947          * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
948          * limits are set by default to 10% and 20% (background and throttle).
949          * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
950          * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
951          * about ~6K pages (as the average of background and throttle wb
952          * limits). The 3rd order polynomial will provide positive feedback if
953          * wb_dirty is under wb_setpoint and vice versa.
954          *
955          * Note, that we cannot use global counters in these calculations
956          * because we want to throttle process writing to a strictlimit wb
957          * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
958          * in the example above).
959          */
960         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
961                 long long wb_pos_ratio;
962
963                 if (dtc->wb_dirty < 8) {
964                         dtc->pos_ratio = min_t(long long, pos_ratio * 2,
965                                            2 << RATELIMIT_CALC_SHIFT);
966                         return;
967                 }
968
969                 if (dtc->wb_dirty >= wb_thresh)
970                         return;
971
972                 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
973                                                     dtc->wb_bg_thresh);
974
975                 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
976                         return;
977
978                 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
979                                                  wb_thresh);
980
981                 /*
982                  * Typically, for strictlimit case, wb_setpoint << setpoint
983                  * and pos_ratio >> wb_pos_ratio. In the other words global
984                  * state ("dirty") is not limiting factor and we have to
985                  * make decision based on wb counters. But there is an
986                  * important case when global pos_ratio should get precedence:
987                  * global limits are exceeded (e.g. due to activities on other
988                  * wb's) while given strictlimit wb is below limit.
989                  *
990                  * "pos_ratio * wb_pos_ratio" would work for the case above,
991                  * but it would look too non-natural for the case of all
992                  * activity in the system coming from a single strictlimit wb
993                  * with bdi->max_ratio == 100%.
994                  *
995                  * Note that min() below somewhat changes the dynamics of the
996                  * control system. Normally, pos_ratio value can be well over 3
997                  * (when globally we are at freerun and wb is well below wb
998                  * setpoint). Now the maximum pos_ratio in the same situation
999                  * is 2. We might want to tweak this if we observe the control
1000                  * system is too slow to adapt.
1001                  */
1002                 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1003                 return;
1004         }
1005
1006         /*
1007          * We have computed basic pos_ratio above based on global situation. If
1008          * the wb is over/under its share of dirty pages, we want to scale
1009          * pos_ratio further down/up. That is done by the following mechanism.
1010          */
1011
1012         /*
1013          * wb setpoint
1014          *
1015          *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1016          *
1017          *                        x_intercept - wb_dirty
1018          *                     := --------------------------
1019          *                        x_intercept - wb_setpoint
1020          *
1021          * The main wb control line is a linear function that subjects to
1022          *
1023          * (1) f(wb_setpoint) = 1.0
1024          * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1025          *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1026          *
1027          * For single wb case, the dirty pages are observed to fluctuate
1028          * regularly within range
1029          *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1030          * for various filesystems, where (2) can yield in a reasonable 12.5%
1031          * fluctuation range for pos_ratio.
1032          *
1033          * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1034          * own size, so move the slope over accordingly and choose a slope that
1035          * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1036          */
1037         if (unlikely(wb_thresh > dtc->thresh))
1038                 wb_thresh = dtc->thresh;
1039         /*
1040          * It's very possible that wb_thresh is close to 0 not because the
1041          * device is slow, but that it has remained inactive for long time.
1042          * Honour such devices a reasonable good (hopefully IO efficient)
1043          * threshold, so that the occasional writes won't be blocked and active
1044          * writes can rampup the threshold quickly.
1045          */
1046         wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1047         /*
1048          * scale global setpoint to wb's:
1049          *      wb_setpoint = setpoint * wb_thresh / thresh
1050          */
1051         x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1052         wb_setpoint = setpoint * (u64)x >> 16;
1053         /*
1054          * Use span=(8*write_bw) in single wb case as indicated by
1055          * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1056          *
1057          *        wb_thresh                    thresh - wb_thresh
1058          * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1059          *         thresh                           thresh
1060          */
1061         span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1062         x_intercept = wb_setpoint + span;
1063
1064         if (dtc->wb_dirty < x_intercept - span / 4) {
1065                 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1066                                       (x_intercept - wb_setpoint) | 1);
1067         } else
1068                 pos_ratio /= 4;
1069
1070         /*
1071          * wb reserve area, safeguard against dirty pool underrun and disk idle
1072          * It may push the desired control point of global dirty pages higher
1073          * than setpoint.
1074          */
1075         x_intercept = wb_thresh / 2;
1076         if (dtc->wb_dirty < x_intercept) {
1077                 if (dtc->wb_dirty > x_intercept / 8)
1078                         pos_ratio = div_u64(pos_ratio * x_intercept,
1079                                             dtc->wb_dirty);
1080                 else
1081                         pos_ratio *= 8;
1082         }
1083
1084         dtc->pos_ratio = pos_ratio;
1085 }
1086
1087 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1088                                       unsigned long elapsed,
1089                                       unsigned long written)
1090 {
1091         const unsigned long period = roundup_pow_of_two(3 * HZ);
1092         unsigned long avg = wb->avg_write_bandwidth;
1093         unsigned long old = wb->write_bandwidth;
1094         u64 bw;
1095
1096         /*
1097          * bw = written * HZ / elapsed
1098          *
1099          *                   bw * elapsed + write_bandwidth * (period - elapsed)
1100          * write_bandwidth = ---------------------------------------------------
1101          *                                          period
1102          *
1103          * @written may have decreased due to account_page_redirty().
1104          * Avoid underflowing @bw calculation.
1105          */
1106         bw = written - min(written, wb->written_stamp);
1107         bw *= HZ;
1108         if (unlikely(elapsed > period)) {
1109                 do_div(bw, elapsed);
1110                 avg = bw;
1111                 goto out;
1112         }
1113         bw += (u64)wb->write_bandwidth * (period - elapsed);
1114         bw >>= ilog2(period);
1115
1116         /*
1117          * one more level of smoothing, for filtering out sudden spikes
1118          */
1119         if (avg > old && old >= (unsigned long)bw)
1120                 avg -= (avg - old) >> 3;
1121
1122         if (avg < old && old <= (unsigned long)bw)
1123                 avg += (old - avg) >> 3;
1124
1125 out:
1126         /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1127         avg = max(avg, 1LU);
1128         if (wb_has_dirty_io(wb)) {
1129                 long delta = avg - wb->avg_write_bandwidth;
1130                 WARN_ON_ONCE(atomic_long_add_return(delta,
1131                                         &wb->bdi->tot_write_bandwidth) <= 0);
1132         }
1133         wb->write_bandwidth = bw;
1134         wb->avg_write_bandwidth = avg;
1135 }
1136
1137 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1138 {
1139         struct wb_domain *dom = dtc_dom(dtc);
1140         unsigned long thresh = dtc->thresh;
1141         unsigned long limit = dom->dirty_limit;
1142
1143         /*
1144          * Follow up in one step.
1145          */
1146         if (limit < thresh) {
1147                 limit = thresh;
1148                 goto update;
1149         }
1150
1151         /*
1152          * Follow down slowly. Use the higher one as the target, because thresh
1153          * may drop below dirty. This is exactly the reason to introduce
1154          * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1155          */
1156         thresh = max(thresh, dtc->dirty);
1157         if (limit > thresh) {
1158                 limit -= (limit - thresh) >> 5;
1159                 goto update;
1160         }
1161         return;
1162 update:
1163         dom->dirty_limit = limit;
1164 }
1165
1166 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1167                                     unsigned long now)
1168 {
1169         struct wb_domain *dom = dtc_dom(dtc);
1170
1171         /*
1172          * check locklessly first to optimize away locking for the most time
1173          */
1174         if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1175                 return;
1176
1177         spin_lock(&dom->lock);
1178         if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1179                 update_dirty_limit(dtc);
1180                 dom->dirty_limit_tstamp = now;
1181         }
1182         spin_unlock(&dom->lock);
1183 }
1184
1185 /*
1186  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1187  *
1188  * Normal wb tasks will be curbed at or below it in long term.
1189  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1190  */
1191 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1192                                       unsigned long dirtied,
1193                                       unsigned long elapsed)
1194 {
1195         struct bdi_writeback *wb = dtc->wb;
1196         unsigned long dirty = dtc->dirty;
1197         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1198         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1199         unsigned long setpoint = (freerun + limit) / 2;
1200         unsigned long write_bw = wb->avg_write_bandwidth;
1201         unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1202         unsigned long dirty_rate;
1203         unsigned long task_ratelimit;
1204         unsigned long balanced_dirty_ratelimit;
1205         unsigned long step;
1206         unsigned long x;
1207         unsigned long shift;
1208
1209         /*
1210          * The dirty rate will match the writeout rate in long term, except
1211          * when dirty pages are truncated by userspace or re-dirtied by FS.
1212          */
1213         dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1214
1215         /*
1216          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1217          */
1218         task_ratelimit = (u64)dirty_ratelimit *
1219                                         dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1220         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1221
1222         /*
1223          * A linear estimation of the "balanced" throttle rate. The theory is,
1224          * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1225          * dirty_rate will be measured to be (N * task_ratelimit). So the below
1226          * formula will yield the balanced rate limit (write_bw / N).
1227          *
1228          * Note that the expanded form is not a pure rate feedback:
1229          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1230          * but also takes pos_ratio into account:
1231          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1232          *
1233          * (1) is not realistic because pos_ratio also takes part in balancing
1234          * the dirty rate.  Consider the state
1235          *      pos_ratio = 0.5                                              (3)
1236          *      rate = 2 * (write_bw / N)                                    (4)
1237          * If (1) is used, it will stuck in that state! Because each dd will
1238          * be throttled at
1239          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1240          * yielding
1241          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1242          * put (6) into (1) we get
1243          *      rate_(i+1) = rate_(i)                                        (7)
1244          *
1245          * So we end up using (2) to always keep
1246          *      rate_(i+1) ~= (write_bw / N)                                 (8)
1247          * regardless of the value of pos_ratio. As long as (8) is satisfied,
1248          * pos_ratio is able to drive itself to 1.0, which is not only where
1249          * the dirty count meet the setpoint, but also where the slope of
1250          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1251          */
1252         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1253                                            dirty_rate | 1);
1254         /*
1255          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1256          */
1257         if (unlikely(balanced_dirty_ratelimit > write_bw))
1258                 balanced_dirty_ratelimit = write_bw;
1259
1260         /*
1261          * We could safely do this and return immediately:
1262          *
1263          *      wb->dirty_ratelimit = balanced_dirty_ratelimit;
1264          *
1265          * However to get a more stable dirty_ratelimit, the below elaborated
1266          * code makes use of task_ratelimit to filter out singular points and
1267          * limit the step size.
1268          *
1269          * The below code essentially only uses the relative value of
1270          *
1271          *      task_ratelimit - dirty_ratelimit
1272          *      = (pos_ratio - 1) * dirty_ratelimit
1273          *
1274          * which reflects the direction and size of dirty position error.
1275          */
1276
1277         /*
1278          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1279          * task_ratelimit is on the same side of dirty_ratelimit, too.
1280          * For example, when
1281          * - dirty_ratelimit > balanced_dirty_ratelimit
1282          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1283          * lowering dirty_ratelimit will help meet both the position and rate
1284          * control targets. Otherwise, don't update dirty_ratelimit if it will
1285          * only help meet the rate target. After all, what the users ultimately
1286          * feel and care are stable dirty rate and small position error.
1287          *
1288          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1289          * and filter out the singular points of balanced_dirty_ratelimit. Which
1290          * keeps jumping around randomly and can even leap far away at times
1291          * due to the small 200ms estimation period of dirty_rate (we want to
1292          * keep that period small to reduce time lags).
1293          */
1294         step = 0;
1295
1296         /*
1297          * For strictlimit case, calculations above were based on wb counters
1298          * and limits (starting from pos_ratio = wb_position_ratio() and up to
1299          * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1300          * Hence, to calculate "step" properly, we have to use wb_dirty as
1301          * "dirty" and wb_setpoint as "setpoint".
1302          *
1303          * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1304          * it's possible that wb_thresh is close to zero due to inactivity
1305          * of backing device.
1306          */
1307         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1308                 dirty = dtc->wb_dirty;
1309                 if (dtc->wb_dirty < 8)
1310                         setpoint = dtc->wb_dirty + 1;
1311                 else
1312                         setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1313         }
1314
1315         if (dirty < setpoint) {
1316                 x = min3(wb->balanced_dirty_ratelimit,
1317                          balanced_dirty_ratelimit, task_ratelimit);
1318                 if (dirty_ratelimit < x)
1319                         step = x - dirty_ratelimit;
1320         } else {
1321                 x = max3(wb->balanced_dirty_ratelimit,
1322                          balanced_dirty_ratelimit, task_ratelimit);
1323                 if (dirty_ratelimit > x)
1324                         step = dirty_ratelimit - x;
1325         }
1326
1327         /*
1328          * Don't pursue 100% rate matching. It's impossible since the balanced
1329          * rate itself is constantly fluctuating. So decrease the track speed
1330          * when it gets close to the target. Helps eliminate pointless tremors.
1331          */
1332         shift = dirty_ratelimit / (2 * step + 1);
1333         if (shift < BITS_PER_LONG)
1334                 step = DIV_ROUND_UP(step >> shift, 8);
1335         else
1336                 step = 0;
1337
1338         if (dirty_ratelimit < balanced_dirty_ratelimit)
1339                 dirty_ratelimit += step;
1340         else
1341                 dirty_ratelimit -= step;
1342
1343         wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1344         wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1345
1346         trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1347 }
1348
1349 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1350                                   struct dirty_throttle_control *mdtc,
1351                                   unsigned long start_time,
1352                                   bool update_ratelimit)
1353 {
1354         struct bdi_writeback *wb = gdtc->wb;
1355         unsigned long now = jiffies;
1356         unsigned long elapsed = now - wb->bw_time_stamp;
1357         unsigned long dirtied;
1358         unsigned long written;
1359
1360         lockdep_assert_held(&wb->list_lock);
1361
1362         /*
1363          * rate-limit, only update once every 200ms.
1364          */
1365         if (elapsed < BANDWIDTH_INTERVAL)
1366                 return;
1367
1368         dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1369         written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1370
1371         /*
1372          * Skip quiet periods when disk bandwidth is under-utilized.
1373          * (at least 1s idle time between two flusher runs)
1374          */
1375         if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1376                 goto snapshot;
1377
1378         if (update_ratelimit) {
1379                 domain_update_bandwidth(gdtc, now);
1380                 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1381
1382                 /*
1383                  * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1384                  * compiler has no way to figure that out.  Help it.
1385                  */
1386                 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1387                         domain_update_bandwidth(mdtc, now);
1388                         wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1389                 }
1390         }
1391         wb_update_write_bandwidth(wb, elapsed, written);
1392
1393 snapshot:
1394         wb->dirtied_stamp = dirtied;
1395         wb->written_stamp = written;
1396         wb->bw_time_stamp = now;
1397 }
1398
1399 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1400 {
1401         struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1402
1403         __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1404 }
1405
1406 /*
1407  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1408  * will look to see if it needs to start dirty throttling.
1409  *
1410  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1411  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1412  * (the number of pages we may dirty without exceeding the dirty limits).
1413  */
1414 static unsigned long dirty_poll_interval(unsigned long dirty,
1415                                          unsigned long thresh)
1416 {
1417         if (thresh > dirty)
1418                 return 1UL << (ilog2(thresh - dirty) >> 1);
1419
1420         return 1;
1421 }
1422
1423 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1424                                   unsigned long wb_dirty)
1425 {
1426         unsigned long bw = wb->avg_write_bandwidth;
1427         unsigned long t;
1428
1429         /*
1430          * Limit pause time for small memory systems. If sleeping for too long
1431          * time, a small pool of dirty/writeback pages may go empty and disk go
1432          * idle.
1433          *
1434          * 8 serves as the safety ratio.
1435          */
1436         t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1437         t++;
1438
1439         return min_t(unsigned long, t, MAX_PAUSE);
1440 }
1441
1442 static long wb_min_pause(struct bdi_writeback *wb,
1443                          long max_pause,
1444                          unsigned long task_ratelimit,
1445                          unsigned long dirty_ratelimit,
1446                          int *nr_dirtied_pause)
1447 {
1448         long hi = ilog2(wb->avg_write_bandwidth);
1449         long lo = ilog2(wb->dirty_ratelimit);
1450         long t;         /* target pause */
1451         long pause;     /* estimated next pause */
1452         int pages;      /* target nr_dirtied_pause */
1453
1454         /* target for 10ms pause on 1-dd case */
1455         t = max(1, HZ / 100);
1456
1457         /*
1458          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1459          * overheads.
1460          *
1461          * (N * 10ms) on 2^N concurrent tasks.
1462          */
1463         if (hi > lo)
1464                 t += (hi - lo) * (10 * HZ) / 1024;
1465
1466         /*
1467          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1468          * on the much more stable dirty_ratelimit. However the next pause time
1469          * will be computed based on task_ratelimit and the two rate limits may
1470          * depart considerably at some time. Especially if task_ratelimit goes
1471          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1472          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1473          * result task_ratelimit won't be executed faithfully, which could
1474          * eventually bring down dirty_ratelimit.
1475          *
1476          * We apply two rules to fix it up:
1477          * 1) try to estimate the next pause time and if necessary, use a lower
1478          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1479          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1480          * 2) limit the target pause time to max_pause/2, so that the normal
1481          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1482          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1483          */
1484         t = min(t, 1 + max_pause / 2);
1485         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1486
1487         /*
1488          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1489          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1490          * When the 16 consecutive reads are often interrupted by some dirty
1491          * throttling pause during the async writes, cfq will go into idles
1492          * (deadline is fine). So push nr_dirtied_pause as high as possible
1493          * until reaches DIRTY_POLL_THRESH=32 pages.
1494          */
1495         if (pages < DIRTY_POLL_THRESH) {
1496                 t = max_pause;
1497                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1498                 if (pages > DIRTY_POLL_THRESH) {
1499                         pages = DIRTY_POLL_THRESH;
1500                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1501                 }
1502         }
1503
1504         pause = HZ * pages / (task_ratelimit + 1);
1505         if (pause > max_pause) {
1506                 t = max_pause;
1507                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1508         }
1509
1510         *nr_dirtied_pause = pages;
1511         /*
1512          * The minimal pause time will normally be half the target pause time.
1513          */
1514         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1515 }
1516
1517 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1518 {
1519         struct bdi_writeback *wb = dtc->wb;
1520         unsigned long wb_reclaimable;
1521
1522         /*
1523          * wb_thresh is not treated as some limiting factor as
1524          * dirty_thresh, due to reasons
1525          * - in JBOD setup, wb_thresh can fluctuate a lot
1526          * - in a system with HDD and USB key, the USB key may somehow
1527          *   go into state (wb_dirty >> wb_thresh) either because
1528          *   wb_dirty starts high, or because wb_thresh drops low.
1529          *   In this case we don't want to hard throttle the USB key
1530          *   dirtiers for 100 seconds until wb_dirty drops under
1531          *   wb_thresh. Instead the auxiliary wb control line in
1532          *   wb_position_ratio() will let the dirtier task progress
1533          *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1534          */
1535         dtc->wb_thresh = __wb_calc_thresh(dtc);
1536         dtc->wb_bg_thresh = dtc->thresh ?
1537                 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1538
1539         /*
1540          * In order to avoid the stacked BDI deadlock we need
1541          * to ensure we accurately count the 'dirty' pages when
1542          * the threshold is low.
1543          *
1544          * Otherwise it would be possible to get thresh+n pages
1545          * reported dirty, even though there are thresh-m pages
1546          * actually dirty; with m+n sitting in the percpu
1547          * deltas.
1548          */
1549         if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1550                 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1551                 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1552         } else {
1553                 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1554                 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1555         }
1556 }
1557
1558 /*
1559  * balance_dirty_pages() must be called by processes which are generating dirty
1560  * data.  It looks at the number of dirty pages in the machine and will force
1561  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1562  * If we're over `background_thresh' then the writeback threads are woken to
1563  * perform some writeout.
1564  */
1565 static void balance_dirty_pages(struct bdi_writeback *wb,
1566                                 unsigned long pages_dirtied)
1567 {
1568         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1569         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1570         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1571         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1572                                                      &mdtc_stor : NULL;
1573         struct dirty_throttle_control *sdtc;
1574         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1575         long period;
1576         long pause;
1577         long max_pause;
1578         long min_pause;
1579         int nr_dirtied_pause;
1580         bool dirty_exceeded = false;
1581         unsigned long task_ratelimit;
1582         unsigned long dirty_ratelimit;
1583         struct backing_dev_info *bdi = wb->bdi;
1584         bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1585         unsigned long start_time = jiffies;
1586
1587         for (;;) {
1588                 unsigned long now = jiffies;
1589                 unsigned long dirty, thresh, bg_thresh;
1590                 unsigned long m_dirty = 0;      /* stop bogus uninit warnings */
1591                 unsigned long m_thresh = 0;
1592                 unsigned long m_bg_thresh = 0;
1593
1594                 /*
1595                  * Unstable writes are a feature of certain networked
1596                  * filesystems (i.e. NFS) in which data may have been
1597                  * written to the server's write cache, but has not yet
1598                  * been flushed to permanent storage.
1599                  */
1600                 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1601                                         global_node_page_state(NR_UNSTABLE_NFS);
1602                 gdtc->avail = global_dirtyable_memory();
1603                 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1604
1605                 domain_dirty_limits(gdtc);
1606
1607                 if (unlikely(strictlimit)) {
1608                         wb_dirty_limits(gdtc);
1609
1610                         dirty = gdtc->wb_dirty;
1611                         thresh = gdtc->wb_thresh;
1612                         bg_thresh = gdtc->wb_bg_thresh;
1613                 } else {
1614                         dirty = gdtc->dirty;
1615                         thresh = gdtc->thresh;
1616                         bg_thresh = gdtc->bg_thresh;
1617                 }
1618
1619                 if (mdtc) {
1620                         unsigned long filepages, headroom, writeback;
1621
1622                         /*
1623                          * If @wb belongs to !root memcg, repeat the same
1624                          * basic calculations for the memcg domain.
1625                          */
1626                         mem_cgroup_wb_stats(wb, &filepages, &headroom,
1627                                             &mdtc->dirty, &writeback);
1628                         mdtc->dirty += writeback;
1629                         mdtc_calc_avail(mdtc, filepages, headroom);
1630
1631                         domain_dirty_limits(mdtc);
1632
1633                         if (unlikely(strictlimit)) {
1634                                 wb_dirty_limits(mdtc);
1635                                 m_dirty = mdtc->wb_dirty;
1636                                 m_thresh = mdtc->wb_thresh;
1637                                 m_bg_thresh = mdtc->wb_bg_thresh;
1638                         } else {
1639                                 m_dirty = mdtc->dirty;
1640                                 m_thresh = mdtc->thresh;
1641                                 m_bg_thresh = mdtc->bg_thresh;
1642                         }
1643                 }
1644
1645                 /*
1646                  * Throttle it only when the background writeback cannot
1647                  * catch-up. This avoids (excessively) small writeouts
1648                  * when the wb limits are ramping up in case of !strictlimit.
1649                  *
1650                  * In strictlimit case make decision based on the wb counters
1651                  * and limits. Small writeouts when the wb limits are ramping
1652                  * up are the price we consciously pay for strictlimit-ing.
1653                  *
1654                  * If memcg domain is in effect, @dirty should be under
1655                  * both global and memcg freerun ceilings.
1656                  */
1657                 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1658                     (!mdtc ||
1659                      m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1660                         unsigned long intv = dirty_poll_interval(dirty, thresh);
1661                         unsigned long m_intv = ULONG_MAX;
1662
1663                         current->dirty_paused_when = now;
1664                         current->nr_dirtied = 0;
1665                         if (mdtc)
1666                                 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1667                         current->nr_dirtied_pause = min(intv, m_intv);
1668                         break;
1669                 }
1670
1671                 if (unlikely(!writeback_in_progress(wb)))
1672                         wb_start_background_writeback(wb);
1673
1674                 /*
1675                  * Calculate global domain's pos_ratio and select the
1676                  * global dtc by default.
1677                  */
1678                 if (!strictlimit)
1679                         wb_dirty_limits(gdtc);
1680
1681                 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1682                         ((gdtc->dirty > gdtc->thresh) || strictlimit);
1683
1684                 wb_position_ratio(gdtc);
1685                 sdtc = gdtc;
1686
1687                 if (mdtc) {
1688                         /*
1689                          * If memcg domain is in effect, calculate its
1690                          * pos_ratio.  @wb should satisfy constraints from
1691                          * both global and memcg domains.  Choose the one
1692                          * w/ lower pos_ratio.
1693                          */
1694                         if (!strictlimit)
1695                                 wb_dirty_limits(mdtc);
1696
1697                         dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1698                                 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1699
1700                         wb_position_ratio(mdtc);
1701                         if (mdtc->pos_ratio < gdtc->pos_ratio)
1702                                 sdtc = mdtc;
1703                 }
1704
1705                 if (dirty_exceeded && !wb->dirty_exceeded)
1706                         wb->dirty_exceeded = 1;
1707
1708                 if (time_is_before_jiffies(wb->bw_time_stamp +
1709                                            BANDWIDTH_INTERVAL)) {
1710                         spin_lock(&wb->list_lock);
1711                         __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1712                         spin_unlock(&wb->list_lock);
1713                 }
1714
1715                 /* throttle according to the chosen dtc */
1716                 dirty_ratelimit = wb->dirty_ratelimit;
1717                 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1718                                                         RATELIMIT_CALC_SHIFT;
1719                 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1720                 min_pause = wb_min_pause(wb, max_pause,
1721                                          task_ratelimit, dirty_ratelimit,
1722                                          &nr_dirtied_pause);
1723
1724                 if (unlikely(task_ratelimit == 0)) {
1725                         period = max_pause;
1726                         pause = max_pause;
1727                         goto pause;
1728                 }
1729                 period = HZ * pages_dirtied / task_ratelimit;
1730                 pause = period;
1731                 if (current->dirty_paused_when)
1732                         pause -= now - current->dirty_paused_when;
1733                 /*
1734                  * For less than 1s think time (ext3/4 may block the dirtier
1735                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1736                  * however at much less frequency), try to compensate it in
1737                  * future periods by updating the virtual time; otherwise just
1738                  * do a reset, as it may be a light dirtier.
1739                  */
1740                 if (pause < min_pause) {
1741                         trace_balance_dirty_pages(wb,
1742                                                   sdtc->thresh,
1743                                                   sdtc->bg_thresh,
1744                                                   sdtc->dirty,
1745                                                   sdtc->wb_thresh,
1746                                                   sdtc->wb_dirty,
1747                                                   dirty_ratelimit,
1748                                                   task_ratelimit,
1749                                                   pages_dirtied,
1750                                                   period,
1751                                                   min(pause, 0L),
1752                                                   start_time);
1753                         if (pause < -HZ) {
1754                                 current->dirty_paused_when = now;
1755                                 current->nr_dirtied = 0;
1756                         } else if (period) {
1757                                 current->dirty_paused_when += period;
1758                                 current->nr_dirtied = 0;
1759                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1760                                 current->nr_dirtied_pause += pages_dirtied;
1761                         break;
1762                 }
1763                 if (unlikely(pause > max_pause)) {
1764                         /* for occasional dropped task_ratelimit */
1765                         now += min(pause - max_pause, max_pause);
1766                         pause = max_pause;
1767                 }
1768
1769 pause:
1770                 trace_balance_dirty_pages(wb,
1771                                           sdtc->thresh,
1772                                           sdtc->bg_thresh,
1773                                           sdtc->dirty,
1774                                           sdtc->wb_thresh,
1775                                           sdtc->wb_dirty,
1776                                           dirty_ratelimit,
1777                                           task_ratelimit,
1778                                           pages_dirtied,
1779                                           period,
1780                                           pause,
1781                                           start_time);
1782                 __set_current_state(TASK_KILLABLE);
1783                 wb->dirty_sleep = now;
1784                 io_schedule_timeout(pause);
1785
1786                 current->dirty_paused_when = now + pause;
1787                 current->nr_dirtied = 0;
1788                 current->nr_dirtied_pause = nr_dirtied_pause;
1789
1790                 /*
1791                  * This is typically equal to (dirty < thresh) and can also
1792                  * keep "1000+ dd on a slow USB stick" under control.
1793                  */
1794                 if (task_ratelimit)
1795                         break;
1796
1797                 /*
1798                  * In the case of an unresponding NFS server and the NFS dirty
1799                  * pages exceeds dirty_thresh, give the other good wb's a pipe
1800                  * to go through, so that tasks on them still remain responsive.
1801                  *
1802                  * In theory 1 page is enough to keep the consumer-producer
1803                  * pipe going: the flusher cleans 1 page => the task dirties 1
1804                  * more page. However wb_dirty has accounting errors.  So use
1805                  * the larger and more IO friendly wb_stat_error.
1806                  */
1807                 if (sdtc->wb_dirty <= wb_stat_error(wb))
1808                         break;
1809
1810                 if (fatal_signal_pending(current))
1811                         break;
1812         }
1813
1814         if (!dirty_exceeded && wb->dirty_exceeded)
1815                 wb->dirty_exceeded = 0;
1816
1817         if (writeback_in_progress(wb))
1818                 return;
1819
1820         /*
1821          * In laptop mode, we wait until hitting the higher threshold before
1822          * starting background writeout, and then write out all the way down
1823          * to the lower threshold.  So slow writers cause minimal disk activity.
1824          *
1825          * In normal mode, we start background writeout at the lower
1826          * background_thresh, to keep the amount of dirty memory low.
1827          */
1828         if (laptop_mode)
1829                 return;
1830
1831         if (nr_reclaimable > gdtc->bg_thresh)
1832                 wb_start_background_writeback(wb);
1833 }
1834
1835 static DEFINE_PER_CPU(int, bdp_ratelimits);
1836
1837 /*
1838  * Normal tasks are throttled by
1839  *      loop {
1840  *              dirty tsk->nr_dirtied_pause pages;
1841  *              take a snap in balance_dirty_pages();
1842  *      }
1843  * However there is a worst case. If every task exit immediately when dirtied
1844  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1845  * called to throttle the page dirties. The solution is to save the not yet
1846  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1847  * randomly into the running tasks. This works well for the above worst case,
1848  * as the new task will pick up and accumulate the old task's leaked dirty
1849  * count and eventually get throttled.
1850  */
1851 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1852
1853 /**
1854  * balance_dirty_pages_ratelimited - balance dirty memory state
1855  * @mapping: address_space which was dirtied
1856  *
1857  * Processes which are dirtying memory should call in here once for each page
1858  * which was newly dirtied.  The function will periodically check the system's
1859  * dirty state and will initiate writeback if needed.
1860  *
1861  * On really big machines, get_writeback_state is expensive, so try to avoid
1862  * calling it too often (ratelimiting).  But once we're over the dirty memory
1863  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1864  * from overshooting the limit by (ratelimit_pages) each.
1865  */
1866 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1867 {
1868         struct inode *inode = mapping->host;
1869         struct backing_dev_info *bdi = inode_to_bdi(inode);
1870         struct bdi_writeback *wb = NULL;
1871         int ratelimit;
1872         int *p;
1873
1874         if (!bdi_cap_account_dirty(bdi))
1875                 return;
1876
1877         if (inode_cgwb_enabled(inode))
1878                 wb = wb_get_create_current(bdi, GFP_KERNEL);
1879         if (!wb)
1880                 wb = &bdi->wb;
1881
1882         ratelimit = current->nr_dirtied_pause;
1883         if (wb->dirty_exceeded)
1884                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1885
1886         preempt_disable();
1887         /*
1888          * This prevents one CPU to accumulate too many dirtied pages without
1889          * calling into balance_dirty_pages(), which can happen when there are
1890          * 1000+ tasks, all of them start dirtying pages at exactly the same
1891          * time, hence all honoured too large initial task->nr_dirtied_pause.
1892          */
1893         p =  this_cpu_ptr(&bdp_ratelimits);
1894         if (unlikely(current->nr_dirtied >= ratelimit))
1895                 *p = 0;
1896         else if (unlikely(*p >= ratelimit_pages)) {
1897                 *p = 0;
1898                 ratelimit = 0;
1899         }
1900         /*
1901          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1902          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1903          * the dirty throttling and livelock other long-run dirtiers.
1904          */
1905         p = this_cpu_ptr(&dirty_throttle_leaks);
1906         if (*p > 0 && current->nr_dirtied < ratelimit) {
1907                 unsigned long nr_pages_dirtied;
1908                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1909                 *p -= nr_pages_dirtied;
1910                 current->nr_dirtied += nr_pages_dirtied;
1911         }
1912         preempt_enable();
1913
1914         if (unlikely(current->nr_dirtied >= ratelimit))
1915                 balance_dirty_pages(wb, current->nr_dirtied);
1916
1917         wb_put(wb);
1918 }
1919 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1920
1921 /**
1922  * wb_over_bg_thresh - does @wb need to be written back?
1923  * @wb: bdi_writeback of interest
1924  *
1925  * Determines whether background writeback should keep writing @wb or it's
1926  * clean enough.  Returns %true if writeback should continue.
1927  */
1928 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1929 {
1930         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1931         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1932         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1933         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1934                                                      &mdtc_stor : NULL;
1935
1936         /*
1937          * Similar to balance_dirty_pages() but ignores pages being written
1938          * as we're trying to decide whether to put more under writeback.
1939          */
1940         gdtc->avail = global_dirtyable_memory();
1941         gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1942                       global_node_page_state(NR_UNSTABLE_NFS);
1943         domain_dirty_limits(gdtc);
1944
1945         if (gdtc->dirty > gdtc->bg_thresh)
1946                 return true;
1947
1948         if (wb_stat(wb, WB_RECLAIMABLE) >
1949             wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1950                 return true;
1951
1952         if (mdtc) {
1953                 unsigned long filepages, headroom, writeback;
1954
1955                 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1956                                     &writeback);
1957                 mdtc_calc_avail(mdtc, filepages, headroom);
1958                 domain_dirty_limits(mdtc);      /* ditto, ignore writeback */
1959
1960                 if (mdtc->dirty > mdtc->bg_thresh)
1961                         return true;
1962
1963                 if (wb_stat(wb, WB_RECLAIMABLE) >
1964                     wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1965                         return true;
1966         }
1967
1968         return false;
1969 }
1970
1971 /*
1972  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1973  */
1974 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1975         void __user *buffer, size_t *length, loff_t *ppos)
1976 {
1977         unsigned int old_interval = dirty_writeback_interval;
1978         int ret;
1979
1980         ret = proc_dointvec(table, write, buffer, length, ppos);
1981
1982         /*
1983          * Writing 0 to dirty_writeback_interval will disable periodic writeback
1984          * and a different non-zero value will wakeup the writeback threads.
1985          * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1986          * iterate over all bdis and wbs.
1987          * The reason we do this is to make the change take effect immediately.
1988          */
1989         if (!ret && write && dirty_writeback_interval &&
1990                 dirty_writeback_interval != old_interval)
1991                 wakeup_flusher_threads(WB_REASON_PERIODIC);
1992
1993         return ret;
1994 }
1995
1996 #ifdef CONFIG_BLOCK
1997 void laptop_mode_timer_fn(unsigned long data)
1998 {
1999         struct request_queue *q = (struct request_queue *)data;
2000
2001         wakeup_flusher_threads_bdi(q->backing_dev_info, WB_REASON_LAPTOP_TIMER);
2002 }
2003
2004 /*
2005  * We've spun up the disk and we're in laptop mode: schedule writeback
2006  * of all dirty data a few seconds from now.  If the flush is already scheduled
2007  * then push it back - the user is still using the disk.
2008  */
2009 void laptop_io_completion(struct backing_dev_info *info)
2010 {
2011         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2012 }
2013
2014 /*
2015  * We're in laptop mode and we've just synced. The sync's writes will have
2016  * caused another writeback to be scheduled by laptop_io_completion.
2017  * Nothing needs to be written back anymore, so we unschedule the writeback.
2018  */
2019 void laptop_sync_completion(void)
2020 {
2021         struct backing_dev_info *bdi;
2022
2023         rcu_read_lock();
2024
2025         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2026                 del_timer(&bdi->laptop_mode_wb_timer);
2027
2028         rcu_read_unlock();
2029 }
2030 #endif
2031
2032 /*
2033  * If ratelimit_pages is too high then we can get into dirty-data overload
2034  * if a large number of processes all perform writes at the same time.
2035  * If it is too low then SMP machines will call the (expensive)
2036  * get_writeback_state too often.
2037  *
2038  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2039  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2040  * thresholds.
2041  */
2042
2043 void writeback_set_ratelimit(void)
2044 {
2045         struct wb_domain *dom = &global_wb_domain;
2046         unsigned long background_thresh;
2047         unsigned long dirty_thresh;
2048
2049         global_dirty_limits(&background_thresh, &dirty_thresh);
2050         dom->dirty_limit = dirty_thresh;
2051         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2052         if (ratelimit_pages < 16)
2053                 ratelimit_pages = 16;
2054 }
2055
2056 static int page_writeback_cpu_online(unsigned int cpu)
2057 {
2058         writeback_set_ratelimit();
2059         return 0;
2060 }
2061
2062 /*
2063  * Called early on to tune the page writeback dirty limits.
2064  *
2065  * We used to scale dirty pages according to how total memory
2066  * related to pages that could be allocated for buffers (by
2067  * comparing nr_free_buffer_pages() to vm_total_pages.
2068  *
2069  * However, that was when we used "dirty_ratio" to scale with
2070  * all memory, and we don't do that any more. "dirty_ratio"
2071  * is now applied to total non-HIGHPAGE memory (by subtracting
2072  * totalhigh_pages from vm_total_pages), and as such we can't
2073  * get into the old insane situation any more where we had
2074  * large amounts of dirty pages compared to a small amount of
2075  * non-HIGHMEM memory.
2076  *
2077  * But we might still want to scale the dirty_ratio by how
2078  * much memory the box has..
2079  */
2080 void __init page_writeback_init(void)
2081 {
2082         BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2083
2084         cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2085                           page_writeback_cpu_online, NULL);
2086         cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2087                           page_writeback_cpu_online);
2088 }
2089
2090 /**
2091  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2092  * @mapping: address space structure to write
2093  * @start: starting page index
2094  * @end: ending page index (inclusive)
2095  *
2096  * This function scans the page range from @start to @end (inclusive) and tags
2097  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2098  * that write_cache_pages (or whoever calls this function) will then use
2099  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2100  * used to avoid livelocking of writeback by a process steadily creating new
2101  * dirty pages in the file (thus it is important for this function to be quick
2102  * so that it can tag pages faster than a dirtying process can create them).
2103  */
2104 /*
2105  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2106  */
2107 void tag_pages_for_writeback(struct address_space *mapping,
2108                              pgoff_t start, pgoff_t end)
2109 {
2110 #define WRITEBACK_TAG_BATCH 4096
2111         unsigned long tagged = 0;
2112         struct radix_tree_iter iter;
2113         void **slot;
2114
2115         spin_lock_irq(&mapping->tree_lock);
2116         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, start,
2117                                                         PAGECACHE_TAG_DIRTY) {
2118                 if (iter.index > end)
2119                         break;
2120                 radix_tree_iter_tag_set(&mapping->page_tree, &iter,
2121                                                         PAGECACHE_TAG_TOWRITE);
2122                 tagged++;
2123                 if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2124                         continue;
2125                 slot = radix_tree_iter_resume(slot, &iter);
2126                 spin_unlock_irq(&mapping->tree_lock);
2127                 cond_resched();
2128                 spin_lock_irq(&mapping->tree_lock);
2129         }
2130         spin_unlock_irq(&mapping->tree_lock);
2131 }
2132 EXPORT_SYMBOL(tag_pages_for_writeback);
2133
2134 /**
2135  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2136  * @mapping: address space structure to write
2137  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2138  * @writepage: function called for each page
2139  * @data: data passed to writepage function
2140  *
2141  * If a page is already under I/O, write_cache_pages() skips it, even
2142  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2143  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2144  * and msync() need to guarantee that all the data which was dirty at the time
2145  * the call was made get new I/O started against them.  If wbc->sync_mode is
2146  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2147  * existing IO to complete.
2148  *
2149  * To avoid livelocks (when other process dirties new pages), we first tag
2150  * pages which should be written back with TOWRITE tag and only then start
2151  * writing them. For data-integrity sync we have to be careful so that we do
2152  * not miss some pages (e.g., because some other process has cleared TOWRITE
2153  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2154  * by the process clearing the DIRTY tag (and submitting the page for IO).
2155  */
2156 int write_cache_pages(struct address_space *mapping,
2157                       struct writeback_control *wbc, writepage_t writepage,
2158                       void *data)
2159 {
2160         int ret = 0;
2161         int done = 0;
2162         struct pagevec pvec;
2163         int nr_pages;
2164         pgoff_t uninitialized_var(writeback_index);
2165         pgoff_t index;
2166         pgoff_t end;            /* Inclusive */
2167         pgoff_t done_index;
2168         int cycled;
2169         int range_whole = 0;
2170         int tag;
2171
2172         pagevec_init(&pvec, 0);
2173         if (wbc->range_cyclic) {
2174                 writeback_index = mapping->writeback_index; /* prev offset */
2175                 index = writeback_index;
2176                 if (index == 0)
2177                         cycled = 1;
2178                 else
2179                         cycled = 0;
2180                 end = -1;
2181         } else {
2182                 index = wbc->range_start >> PAGE_SHIFT;
2183                 end = wbc->range_end >> PAGE_SHIFT;
2184                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2185                         range_whole = 1;
2186                 cycled = 1; /* ignore range_cyclic tests */
2187         }
2188         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2189                 tag = PAGECACHE_TAG_TOWRITE;
2190         else
2191                 tag = PAGECACHE_TAG_DIRTY;
2192 retry:
2193         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2194                 tag_pages_for_writeback(mapping, index, end);
2195         done_index = index;
2196         while (!done && (index <= end)) {
2197                 int i;
2198
2199                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2200                                 tag, PAGEVEC_SIZE);
2201                 if (nr_pages == 0)
2202                         break;
2203
2204                 for (i = 0; i < nr_pages; i++) {
2205                         struct page *page = pvec.pages[i];
2206
2207                         done_index = page->index;
2208
2209                         lock_page(page);
2210
2211                         /*
2212                          * Page truncated or invalidated. We can freely skip it
2213                          * then, even for data integrity operations: the page
2214                          * has disappeared concurrently, so there could be no
2215                          * real expectation of this data interity operation
2216                          * even if there is now a new, dirty page at the same
2217                          * pagecache address.
2218                          */
2219                         if (unlikely(page->mapping != mapping)) {
2220 continue_unlock:
2221                                 unlock_page(page);
2222                                 continue;
2223                         }
2224
2225                         if (!PageDirty(page)) {
2226                                 /* someone wrote it for us */
2227                                 goto continue_unlock;
2228                         }
2229
2230                         if (PageWriteback(page)) {
2231                                 if (wbc->sync_mode != WB_SYNC_NONE)
2232                                         wait_on_page_writeback(page);
2233                                 else
2234                                         goto continue_unlock;
2235                         }
2236
2237                         BUG_ON(PageWriteback(page));
2238                         if (!clear_page_dirty_for_io(page))
2239                                 goto continue_unlock;
2240
2241                         trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2242                         ret = (*writepage)(page, wbc, data);
2243                         if (unlikely(ret)) {
2244                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2245                                         unlock_page(page);
2246                                         ret = 0;
2247                                 } else {
2248                                         /*
2249                                          * done_index is set past this page,
2250                                          * so media errors will not choke
2251                                          * background writeout for the entire
2252                                          * file. This has consequences for
2253                                          * range_cyclic semantics (ie. it may
2254                                          * not be suitable for data integrity
2255                                          * writeout).
2256                                          */
2257                                         done_index = page->index + 1;
2258                                         done = 1;
2259                                         break;
2260                                 }
2261                         }
2262
2263                         /*
2264                          * We stop writing back only if we are not doing
2265                          * integrity sync. In case of integrity sync we have to
2266                          * keep going until we have written all the pages
2267                          * we tagged for writeback prior to entering this loop.
2268                          */
2269                         if (--wbc->nr_to_write <= 0 &&
2270                             wbc->sync_mode == WB_SYNC_NONE) {
2271                                 done = 1;
2272                                 break;
2273                         }
2274                 }
2275                 pagevec_release(&pvec);
2276                 cond_resched();
2277         }
2278         if (!cycled && !done) {
2279                 /*
2280                  * range_cyclic:
2281                  * We hit the last page and there is more work to be done: wrap
2282                  * back to the start of the file
2283                  */
2284                 cycled = 1;
2285                 index = 0;
2286                 end = writeback_index - 1;
2287                 goto retry;
2288         }
2289         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2290                 mapping->writeback_index = done_index;
2291
2292         return ret;
2293 }
2294 EXPORT_SYMBOL(write_cache_pages);
2295
2296 /*
2297  * Function used by generic_writepages to call the real writepage
2298  * function and set the mapping flags on error
2299  */
2300 static int __writepage(struct page *page, struct writeback_control *wbc,
2301                        void *data)
2302 {
2303         struct address_space *mapping = data;
2304         int ret = mapping->a_ops->writepage(page, wbc);
2305         mapping_set_error(mapping, ret);
2306         return ret;
2307 }
2308
2309 /**
2310  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2311  * @mapping: address space structure to write
2312  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2313  *
2314  * This is a library function, which implements the writepages()
2315  * address_space_operation.
2316  */
2317 int generic_writepages(struct address_space *mapping,
2318                        struct writeback_control *wbc)
2319 {
2320         struct blk_plug plug;
2321         int ret;
2322
2323         /* deal with chardevs and other special file */
2324         if (!mapping->a_ops->writepage)
2325                 return 0;
2326
2327         blk_start_plug(&plug);
2328         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2329         blk_finish_plug(&plug);
2330         return ret;
2331 }
2332
2333 EXPORT_SYMBOL(generic_writepages);
2334
2335 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2336 {
2337         int ret;
2338
2339         if (wbc->nr_to_write <= 0)
2340                 return 0;
2341         while (1) {
2342                 if (mapping->a_ops->writepages)
2343                         ret = mapping->a_ops->writepages(mapping, wbc);
2344                 else
2345                         ret = generic_writepages(mapping, wbc);
2346                 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2347                         break;
2348                 cond_resched();
2349                 congestion_wait(BLK_RW_ASYNC, HZ/50);
2350         }
2351         return ret;
2352 }
2353
2354 /**
2355  * write_one_page - write out a single page and wait on I/O
2356  * @page: the page to write
2357  *
2358  * The page must be locked by the caller and will be unlocked upon return.
2359  *
2360  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2361  * function returns.
2362  */
2363 int write_one_page(struct page *page)
2364 {
2365         struct address_space *mapping = page->mapping;
2366         int ret = 0;
2367         struct writeback_control wbc = {
2368                 .sync_mode = WB_SYNC_ALL,
2369                 .nr_to_write = 1,
2370         };
2371
2372         BUG_ON(!PageLocked(page));
2373
2374         wait_on_page_writeback(page);
2375
2376         if (clear_page_dirty_for_io(page)) {
2377                 get_page(page);
2378                 ret = mapping->a_ops->writepage(page, &wbc);
2379                 if (ret == 0)
2380                         wait_on_page_writeback(page);
2381                 put_page(page);
2382         } else {
2383                 unlock_page(page);
2384         }
2385
2386         if (!ret)
2387                 ret = filemap_check_errors(mapping);
2388         return ret;
2389 }
2390 EXPORT_SYMBOL(write_one_page);
2391
2392 /*
2393  * For address_spaces which do not use buffers nor write back.
2394  */
2395 int __set_page_dirty_no_writeback(struct page *page)
2396 {
2397         if (!PageDirty(page))
2398                 return !TestSetPageDirty(page);
2399         return 0;
2400 }
2401
2402 /*
2403  * Helper function for set_page_dirty family.
2404  *
2405  * Caller must hold lock_page_memcg().
2406  *
2407  * NOTE: This relies on being atomic wrt interrupts.
2408  */
2409 void account_page_dirtied(struct page *page, struct address_space *mapping)
2410 {
2411         struct inode *inode = mapping->host;
2412
2413         trace_writeback_dirty_page(page, mapping);
2414
2415         if (mapping_cap_account_dirty(mapping)) {
2416                 struct bdi_writeback *wb;
2417
2418                 inode_attach_wb(inode, page);
2419                 wb = inode_to_wb(inode);
2420
2421                 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2422                 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2423                 __inc_node_page_state(page, NR_DIRTIED);
2424                 inc_wb_stat(wb, WB_RECLAIMABLE);
2425                 inc_wb_stat(wb, WB_DIRTIED);
2426                 task_io_account_write(PAGE_SIZE);
2427                 current->nr_dirtied++;
2428                 this_cpu_inc(bdp_ratelimits);
2429         }
2430 }
2431 EXPORT_SYMBOL(account_page_dirtied);
2432
2433 /*
2434  * Helper function for deaccounting dirty page without writeback.
2435  *
2436  * Caller must hold lock_page_memcg().
2437  */
2438 void account_page_cleaned(struct page *page, struct address_space *mapping,
2439                           struct bdi_writeback *wb)
2440 {
2441         if (mapping_cap_account_dirty(mapping)) {
2442                 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2443                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2444                 dec_wb_stat(wb, WB_RECLAIMABLE);
2445                 task_io_account_cancelled_write(PAGE_SIZE);
2446         }
2447 }
2448
2449 /*
2450  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2451  * its radix tree.
2452  *
2453  * This is also used when a single buffer is being dirtied: we want to set the
2454  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2455  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2456  *
2457  * The caller must ensure this doesn't race with truncation.  Most will simply
2458  * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2459  * the pte lock held, which also locks out truncation.
2460  */
2461 int __set_page_dirty_nobuffers(struct page *page)
2462 {
2463         lock_page_memcg(page);
2464         if (!TestSetPageDirty(page)) {
2465                 struct address_space *mapping = page_mapping(page);
2466                 unsigned long flags;
2467
2468                 if (!mapping) {
2469                         unlock_page_memcg(page);
2470                         return 1;
2471                 }
2472
2473                 spin_lock_irqsave(&mapping->tree_lock, flags);
2474                 BUG_ON(page_mapping(page) != mapping);
2475                 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2476                 account_page_dirtied(page, mapping);
2477                 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2478                                    PAGECACHE_TAG_DIRTY);
2479                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2480                 unlock_page_memcg(page);
2481
2482                 if (mapping->host) {
2483                         /* !PageAnon && !swapper_space */
2484                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2485                 }
2486                 return 1;
2487         }
2488         unlock_page_memcg(page);
2489         return 0;
2490 }
2491 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2492
2493 /*
2494  * Call this whenever redirtying a page, to de-account the dirty counters
2495  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2496  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2497  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2498  * control.
2499  */
2500 void account_page_redirty(struct page *page)
2501 {
2502         struct address_space *mapping = page->mapping;
2503
2504         if (mapping && mapping_cap_account_dirty(mapping)) {
2505                 struct inode *inode = mapping->host;
2506                 struct bdi_writeback *wb;
2507                 bool locked;
2508
2509                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2510                 current->nr_dirtied--;
2511                 dec_node_page_state(page, NR_DIRTIED);
2512                 dec_wb_stat(wb, WB_DIRTIED);
2513                 unlocked_inode_to_wb_end(inode, locked);
2514         }
2515 }
2516 EXPORT_SYMBOL(account_page_redirty);
2517
2518 /*
2519  * When a writepage implementation decides that it doesn't want to write this
2520  * page for some reason, it should redirty the locked page via
2521  * redirty_page_for_writepage() and it should then unlock the page and return 0
2522  */
2523 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2524 {
2525         int ret;
2526
2527         wbc->pages_skipped++;
2528         ret = __set_page_dirty_nobuffers(page);
2529         account_page_redirty(page);
2530         return ret;
2531 }
2532 EXPORT_SYMBOL(redirty_page_for_writepage);
2533
2534 /*
2535  * Dirty a page.
2536  *
2537  * For pages with a mapping this should be done under the page lock
2538  * for the benefit of asynchronous memory errors who prefer a consistent
2539  * dirty state. This rule can be broken in some special cases,
2540  * but should be better not to.
2541  *
2542  * If the mapping doesn't provide a set_page_dirty a_op, then
2543  * just fall through and assume that it wants buffer_heads.
2544  */
2545 int set_page_dirty(struct page *page)
2546 {
2547         struct address_space *mapping = page_mapping(page);
2548
2549         page = compound_head(page);
2550         if (likely(mapping)) {
2551                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2552                 /*
2553                  * readahead/lru_deactivate_page could remain
2554                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2555                  * About readahead, if the page is written, the flags would be
2556                  * reset. So no problem.
2557                  * About lru_deactivate_page, if the page is redirty, the flag
2558                  * will be reset. So no problem. but if the page is used by readahead
2559                  * it will confuse readahead and make it restart the size rampup
2560                  * process. But it's a trivial problem.
2561                  */
2562                 if (PageReclaim(page))
2563                         ClearPageReclaim(page);
2564 #ifdef CONFIG_BLOCK
2565                 if (!spd)
2566                         spd = __set_page_dirty_buffers;
2567 #endif
2568                 return (*spd)(page);
2569         }
2570         if (!PageDirty(page)) {
2571                 if (!TestSetPageDirty(page))
2572                         return 1;
2573         }
2574         return 0;
2575 }
2576 EXPORT_SYMBOL(set_page_dirty);
2577
2578 /*
2579  * set_page_dirty() is racy if the caller has no reference against
2580  * page->mapping->host, and if the page is unlocked.  This is because another
2581  * CPU could truncate the page off the mapping and then free the mapping.
2582  *
2583  * Usually, the page _is_ locked, or the caller is a user-space process which
2584  * holds a reference on the inode by having an open file.
2585  *
2586  * In other cases, the page should be locked before running set_page_dirty().
2587  */
2588 int set_page_dirty_lock(struct page *page)
2589 {
2590         int ret;
2591
2592         lock_page(page);
2593         ret = set_page_dirty(page);
2594         unlock_page(page);
2595         return ret;
2596 }
2597 EXPORT_SYMBOL(set_page_dirty_lock);
2598
2599 /*
2600  * This cancels just the dirty bit on the kernel page itself, it does NOT
2601  * actually remove dirty bits on any mmap's that may be around. It also
2602  * leaves the page tagged dirty, so any sync activity will still find it on
2603  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2604  * look at the dirty bits in the VM.
2605  *
2606  * Doing this should *normally* only ever be done when a page is truncated,
2607  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2608  * this when it notices that somebody has cleaned out all the buffers on a
2609  * page without actually doing it through the VM. Can you say "ext3 is
2610  * horribly ugly"? Thought you could.
2611  */
2612 void cancel_dirty_page(struct page *page)
2613 {
2614         struct address_space *mapping = page_mapping(page);
2615
2616         if (mapping_cap_account_dirty(mapping)) {
2617                 struct inode *inode = mapping->host;
2618                 struct bdi_writeback *wb;
2619                 bool locked;
2620
2621                 lock_page_memcg(page);
2622                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2623
2624                 if (TestClearPageDirty(page))
2625                         account_page_cleaned(page, mapping, wb);
2626
2627                 unlocked_inode_to_wb_end(inode, locked);
2628                 unlock_page_memcg(page);
2629         } else {
2630                 ClearPageDirty(page);
2631         }
2632 }
2633 EXPORT_SYMBOL(cancel_dirty_page);
2634
2635 /*
2636  * Clear a page's dirty flag, while caring for dirty memory accounting.
2637  * Returns true if the page was previously dirty.
2638  *
2639  * This is for preparing to put the page under writeout.  We leave the page
2640  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2641  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2642  * implementation will run either set_page_writeback() or set_page_dirty(),
2643  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2644  * back into sync.
2645  *
2646  * This incoherency between the page's dirty flag and radix-tree tag is
2647  * unfortunate, but it only exists while the page is locked.
2648  */
2649 int clear_page_dirty_for_io(struct page *page)
2650 {
2651         struct address_space *mapping = page_mapping(page);
2652         int ret = 0;
2653
2654         BUG_ON(!PageLocked(page));
2655
2656         if (mapping && mapping_cap_account_dirty(mapping)) {
2657                 struct inode *inode = mapping->host;
2658                 struct bdi_writeback *wb;
2659                 bool locked;
2660
2661                 /*
2662                  * Yes, Virginia, this is indeed insane.
2663                  *
2664                  * We use this sequence to make sure that
2665                  *  (a) we account for dirty stats properly
2666                  *  (b) we tell the low-level filesystem to
2667                  *      mark the whole page dirty if it was
2668                  *      dirty in a pagetable. Only to then
2669                  *  (c) clean the page again and return 1 to
2670                  *      cause the writeback.
2671                  *
2672                  * This way we avoid all nasty races with the
2673                  * dirty bit in multiple places and clearing
2674                  * them concurrently from different threads.
2675                  *
2676                  * Note! Normally the "set_page_dirty(page)"
2677                  * has no effect on the actual dirty bit - since
2678                  * that will already usually be set. But we
2679                  * need the side effects, and it can help us
2680                  * avoid races.
2681                  *
2682                  * We basically use the page "master dirty bit"
2683                  * as a serialization point for all the different
2684                  * threads doing their things.
2685                  */
2686                 if (page_mkclean(page))
2687                         set_page_dirty(page);
2688                 /*
2689                  * We carefully synchronise fault handlers against
2690                  * installing a dirty pte and marking the page dirty
2691                  * at this point.  We do this by having them hold the
2692                  * page lock while dirtying the page, and pages are
2693                  * always locked coming in here, so we get the desired
2694                  * exclusion.
2695                  */
2696                 wb = unlocked_inode_to_wb_begin(inode, &locked);
2697                 if (TestClearPageDirty(page)) {
2698                         dec_lruvec_page_state(page, NR_FILE_DIRTY);
2699                         dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2700                         dec_wb_stat(wb, WB_RECLAIMABLE);
2701                         ret = 1;
2702                 }
2703                 unlocked_inode_to_wb_end(inode, locked);
2704                 return ret;
2705         }
2706         return TestClearPageDirty(page);
2707 }
2708 EXPORT_SYMBOL(clear_page_dirty_for_io);
2709
2710 int test_clear_page_writeback(struct page *page)
2711 {
2712         struct address_space *mapping = page_mapping(page);
2713         struct mem_cgroup *memcg;
2714         struct lruvec *lruvec;
2715         int ret;
2716
2717         memcg = lock_page_memcg(page);
2718         lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2719         if (mapping && mapping_use_writeback_tags(mapping)) {
2720                 struct inode *inode = mapping->host;
2721                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2722                 unsigned long flags;
2723
2724                 spin_lock_irqsave(&mapping->tree_lock, flags);
2725                 ret = TestClearPageWriteback(page);
2726                 if (ret) {
2727                         radix_tree_tag_clear(&mapping->page_tree,
2728                                                 page_index(page),
2729                                                 PAGECACHE_TAG_WRITEBACK);
2730                         if (bdi_cap_account_writeback(bdi)) {
2731                                 struct bdi_writeback *wb = inode_to_wb(inode);
2732
2733                                 dec_wb_stat(wb, WB_WRITEBACK);
2734                                 __wb_writeout_inc(wb);
2735                         }
2736                 }
2737
2738                 if (mapping->host && !mapping_tagged(mapping,
2739                                                      PAGECACHE_TAG_WRITEBACK))
2740                         sb_clear_inode_writeback(mapping->host);
2741
2742                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2743         } else {
2744                 ret = TestClearPageWriteback(page);
2745         }
2746         /*
2747          * NOTE: Page might be free now! Writeback doesn't hold a page
2748          * reference on its own, it relies on truncation to wait for
2749          * the clearing of PG_writeback. The below can only access
2750          * page state that is static across allocation cycles.
2751          */
2752         if (ret) {
2753                 dec_lruvec_state(lruvec, NR_WRITEBACK);
2754                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2755                 inc_node_page_state(page, NR_WRITTEN);
2756         }
2757         __unlock_page_memcg(memcg);
2758         return ret;
2759 }
2760
2761 int __test_set_page_writeback(struct page *page, bool keep_write)
2762 {
2763         struct address_space *mapping = page_mapping(page);
2764         int ret;
2765
2766         lock_page_memcg(page);
2767         if (mapping && mapping_use_writeback_tags(mapping)) {
2768                 struct inode *inode = mapping->host;
2769                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2770                 unsigned long flags;
2771
2772                 spin_lock_irqsave(&mapping->tree_lock, flags);
2773                 ret = TestSetPageWriteback(page);
2774                 if (!ret) {
2775                         bool on_wblist;
2776
2777                         on_wblist = mapping_tagged(mapping,
2778                                                    PAGECACHE_TAG_WRITEBACK);
2779
2780                         radix_tree_tag_set(&mapping->page_tree,
2781                                                 page_index(page),
2782                                                 PAGECACHE_TAG_WRITEBACK);
2783                         if (bdi_cap_account_writeback(bdi))
2784                                 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2785
2786                         /*
2787                          * We can come through here when swapping anonymous
2788                          * pages, so we don't necessarily have an inode to track
2789                          * for sync.
2790                          */
2791                         if (mapping->host && !on_wblist)
2792                                 sb_mark_inode_writeback(mapping->host);
2793                 }
2794                 if (!PageDirty(page))
2795                         radix_tree_tag_clear(&mapping->page_tree,
2796                                                 page_index(page),
2797                                                 PAGECACHE_TAG_DIRTY);
2798                 if (!keep_write)
2799                         radix_tree_tag_clear(&mapping->page_tree,
2800                                                 page_index(page),
2801                                                 PAGECACHE_TAG_TOWRITE);
2802                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2803         } else {
2804                 ret = TestSetPageWriteback(page);
2805         }
2806         if (!ret) {
2807                 inc_lruvec_page_state(page, NR_WRITEBACK);
2808                 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2809         }
2810         unlock_page_memcg(page);
2811         return ret;
2812
2813 }
2814 EXPORT_SYMBOL(__test_set_page_writeback);
2815
2816 /*
2817  * Return true if any of the pages in the mapping are marked with the
2818  * passed tag.
2819  */
2820 int mapping_tagged(struct address_space *mapping, int tag)
2821 {
2822         return radix_tree_tagged(&mapping->page_tree, tag);
2823 }
2824 EXPORT_SYMBOL(mapping_tagged);
2825
2826 /**
2827  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2828  * @page:       The page to wait on.
2829  *
2830  * This function determines if the given page is related to a backing device
2831  * that requires page contents to be held stable during writeback.  If so, then
2832  * it will wait for any pending writeback to complete.
2833  */
2834 void wait_for_stable_page(struct page *page)
2835 {
2836         if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2837                 wait_on_page_writeback(page);
2838 }
2839 EXPORT_SYMBOL_GPL(wait_for_stable_page);