~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/page-writeback.c

Version: ~ [ linux-5.3-rc5 ] ~ [ linux-5.2.9 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.67 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.139 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.72 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp