Version:  2.0.40 2.2.26 2.4.37 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1

Linux/mm/page-writeback.c

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

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