Version:  2.0.40 2.2.26 2.4.37 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7

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

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us