Version:  2.6.34 2.6.35 2.6.36 2.6.37 2.6.38 2.6.39 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14

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  * Work out the current dirty-memory clamping and background writeout
160  * thresholds.
161  *
162  * The main aim here is to lower them aggressively if there is a lot of mapped
163  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
164  * pages.  It is better to clamp down on writers than to start swapping, and
165  * performing lots of scanning.
166  *
167  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
168  *
169  * We don't permit the clamping level to fall below 5% - that is getting rather
170  * excessive.
171  *
172  * We make sure that the background writeout level is below the adjusted
173  * clamping level.
174  */
175 
176 /*
177  * In a memory zone, there is a certain amount of pages we consider
178  * available for the page cache, which is essentially the number of
179  * free and reclaimable pages, minus some zone reserves to protect
180  * lowmem and the ability to uphold the zone's watermarks without
181  * requiring writeback.
182  *
183  * This number of dirtyable pages is the base value of which the
184  * user-configurable dirty ratio is the effictive number of pages that
185  * are allowed to be actually dirtied.  Per individual zone, or
186  * globally by using the sum of dirtyable pages over all zones.
187  *
188  * Because the user is allowed to specify the dirty limit globally as
189  * absolute number of bytes, calculating the per-zone dirty limit can
190  * require translating the configured limit into a percentage of
191  * global dirtyable memory first.
192  */
193 
194 /**
195  * zone_dirtyable_memory - number of dirtyable pages in a zone
196  * @zone: the zone
197  *
198  * Returns the zone's number of pages potentially available for dirty
199  * page cache.  This is the base value for the per-zone dirty limits.
200  */
201 static unsigned long zone_dirtyable_memory(struct zone *zone)
202 {
203         unsigned long nr_pages;
204 
205         nr_pages = zone_page_state(zone, NR_FREE_PAGES);
206         nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
207 
208         nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
209         nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
210 
211         return nr_pages;
212 }
213 
214 static unsigned long highmem_dirtyable_memory(unsigned long total)
215 {
216 #ifdef CONFIG_HIGHMEM
217         int node;
218         unsigned long x = 0;
219 
220         for_each_node_state(node, N_HIGH_MEMORY) {
221                 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
222 
223                 x += zone_dirtyable_memory(z);
224         }
225         /*
226          * Unreclaimable memory (kernel memory or anonymous memory
227          * without swap) can bring down the dirtyable pages below
228          * the zone's dirty balance reserve and the above calculation
229          * will underflow.  However we still want to add in nodes
230          * which are below threshold (negative values) to get a more
231          * accurate calculation but make sure that the total never
232          * underflows.
233          */
234         if ((long)x < 0)
235                 x = 0;
236 
237         /*
238          * Make sure that the number of highmem pages is never larger
239          * than the number of the total dirtyable memory. This can only
240          * occur in very strange VM situations but we want to make sure
241          * that this does not occur.
242          */
243         return min(x, total);
244 #else
245         return 0;
246 #endif
247 }
248 
249 /**
250  * global_dirtyable_memory - number of globally dirtyable pages
251  *
252  * Returns the global number of pages potentially available for dirty
253  * page cache.  This is the base value for the global dirty limits.
254  */
255 static unsigned long global_dirtyable_memory(void)
256 {
257         unsigned long x;
258 
259         x = global_page_state(NR_FREE_PAGES);
260         x -= min(x, dirty_balance_reserve);
261 
262         x += global_page_state(NR_INACTIVE_FILE);
263         x += global_page_state(NR_ACTIVE_FILE);
264 
265         if (!vm_highmem_is_dirtyable)
266                 x -= highmem_dirtyable_memory(x);
267 
268         return x + 1;   /* Ensure that we never return 0 */
269 }
270 
271 /*
272  * global_dirty_limits - background-writeback and dirty-throttling thresholds
273  *
274  * Calculate the dirty thresholds based on sysctl parameters
275  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
276  * - vm.dirty_ratio             or  vm.dirty_bytes
277  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
278  * real-time tasks.
279  */
280 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
281 {
282         unsigned long background;
283         unsigned long dirty;
284         unsigned long uninitialized_var(available_memory);
285         struct task_struct *tsk;
286 
287         if (!vm_dirty_bytes || !dirty_background_bytes)
288                 available_memory = global_dirtyable_memory();
289 
290         if (vm_dirty_bytes)
291                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
292         else
293                 dirty = (vm_dirty_ratio * available_memory) / 100;
294 
295         if (dirty_background_bytes)
296                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
297         else
298                 background = (dirty_background_ratio * available_memory) / 100;
299 
300         if (background >= dirty)
301                 background = dirty / 2;
302         tsk = current;
303         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
304                 background += background / 4;
305                 dirty += dirty / 4;
306         }
307         *pbackground = background;
308         *pdirty = dirty;
309         trace_global_dirty_state(background, dirty);
310 }
311 
312 /**
313  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
314  * @zone: the zone
315  *
316  * Returns the maximum number of dirty pages allowed in a zone, based
317  * on the zone's dirtyable memory.
318  */
319 static unsigned long zone_dirty_limit(struct zone *zone)
320 {
321         unsigned long zone_memory = zone_dirtyable_memory(zone);
322         struct task_struct *tsk = current;
323         unsigned long dirty;
324 
325         if (vm_dirty_bytes)
326                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
327                         zone_memory / global_dirtyable_memory();
328         else
329                 dirty = vm_dirty_ratio * zone_memory / 100;
330 
331         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
332                 dirty += dirty / 4;
333 
334         return dirty;
335 }
336 
337 /**
338  * zone_dirty_ok - tells whether a zone is within its dirty limits
339  * @zone: the zone to check
340  *
341  * Returns %true when the dirty pages in @zone are within the zone's
342  * dirty limit, %false if the limit is exceeded.
343  */
344 bool zone_dirty_ok(struct zone *zone)
345 {
346         unsigned long limit = zone_dirty_limit(zone);
347 
348         return zone_page_state(zone, NR_FILE_DIRTY) +
349                zone_page_state(zone, NR_UNSTABLE_NFS) +
350                zone_page_state(zone, NR_WRITEBACK) <= limit;
351 }
352 
353 int dirty_background_ratio_handler(struct ctl_table *table, int write,
354                 void __user *buffer, size_t *lenp,
355                 loff_t *ppos)
356 {
357         int ret;
358 
359         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
360         if (ret == 0 && write)
361                 dirty_background_bytes = 0;
362         return ret;
363 }
364 
365 int dirty_background_bytes_handler(struct ctl_table *table, int write,
366                 void __user *buffer, size_t *lenp,
367                 loff_t *ppos)
368 {
369         int ret;
370 
371         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
372         if (ret == 0 && write)
373                 dirty_background_ratio = 0;
374         return ret;
375 }
376 
377 int dirty_ratio_handler(struct ctl_table *table, int write,
378                 void __user *buffer, size_t *lenp,
379                 loff_t *ppos)
380 {
381         int old_ratio = vm_dirty_ratio;
382         int ret;
383 
384         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
385         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
386                 writeback_set_ratelimit();
387                 vm_dirty_bytes = 0;
388         }
389         return ret;
390 }
391 
392 int dirty_bytes_handler(struct ctl_table *table, int write,
393                 void __user *buffer, size_t *lenp,
394                 loff_t *ppos)
395 {
396         unsigned long old_bytes = vm_dirty_bytes;
397         int ret;
398 
399         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
400         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
401                 writeback_set_ratelimit();
402                 vm_dirty_ratio = 0;
403         }
404         return ret;
405 }
406 
407 static unsigned long wp_next_time(unsigned long cur_time)
408 {
409         cur_time += VM_COMPLETIONS_PERIOD_LEN;
410         /* 0 has a special meaning... */
411         if (!cur_time)
412                 return 1;
413         return cur_time;
414 }
415 
416 /*
417  * Increment the BDI's writeout completion count and the global writeout
418  * completion count. Called from test_clear_page_writeback().
419  */
420 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
421 {
422         __inc_bdi_stat(bdi, BDI_WRITTEN);
423         __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
424                                bdi->max_prop_frac);
425         /* First event after period switching was turned off? */
426         if (!unlikely(writeout_period_time)) {
427                 /*
428                  * We can race with other __bdi_writeout_inc calls here but
429                  * it does not cause any harm since the resulting time when
430                  * timer will fire and what is in writeout_period_time will be
431                  * roughly the same.
432                  */
433                 writeout_period_time = wp_next_time(jiffies);
434                 mod_timer(&writeout_period_timer, writeout_period_time);
435         }
436 }
437 
438 void bdi_writeout_inc(struct backing_dev_info *bdi)
439 {
440         unsigned long flags;
441 
442         local_irq_save(flags);
443         __bdi_writeout_inc(bdi);
444         local_irq_restore(flags);
445 }
446 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
447 
448 /*
449  * Obtain an accurate fraction of the BDI's portion.
450  */
451 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
452                 long *numerator, long *denominator)
453 {
454         fprop_fraction_percpu(&writeout_completions, &bdi->completions,
455                                 numerator, denominator);
456 }
457 
458 /*
459  * On idle system, we can be called long after we scheduled because we use
460  * deferred timers so count with missed periods.
461  */
462 static void writeout_period(unsigned long t)
463 {
464         int miss_periods = (jiffies - writeout_period_time) /
465                                                  VM_COMPLETIONS_PERIOD_LEN;
466 
467         if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
468                 writeout_period_time = wp_next_time(writeout_period_time +
469                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
470                 mod_timer(&writeout_period_timer, writeout_period_time);
471         } else {
472                 /*
473                  * Aging has zeroed all fractions. Stop wasting CPU on period
474                  * updates.
475                  */
476                 writeout_period_time = 0;
477         }
478 }
479 
480 /*
481  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
482  * registered backing devices, which, for obvious reasons, can not
483  * exceed 100%.
484  */
485 static unsigned int bdi_min_ratio;
486 
487 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
488 {
489         int ret = 0;
490 
491         spin_lock_bh(&bdi_lock);
492         if (min_ratio > bdi->max_ratio) {
493                 ret = -EINVAL;
494         } else {
495                 min_ratio -= bdi->min_ratio;
496                 if (bdi_min_ratio + min_ratio < 100) {
497                         bdi_min_ratio += min_ratio;
498                         bdi->min_ratio += min_ratio;
499                 } else {
500                         ret = -EINVAL;
501                 }
502         }
503         spin_unlock_bh(&bdi_lock);
504 
505         return ret;
506 }
507 
508 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
509 {
510         int ret = 0;
511 
512         if (max_ratio > 100)
513                 return -EINVAL;
514 
515         spin_lock_bh(&bdi_lock);
516         if (bdi->min_ratio > max_ratio) {
517                 ret = -EINVAL;
518         } else {
519                 bdi->max_ratio = max_ratio;
520                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
521         }
522         spin_unlock_bh(&bdi_lock);
523 
524         return ret;
525 }
526 EXPORT_SYMBOL(bdi_set_max_ratio);
527 
528 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
529                                            unsigned long bg_thresh)
530 {
531         return (thresh + bg_thresh) / 2;
532 }
533 
534 static unsigned long hard_dirty_limit(unsigned long thresh)
535 {
536         return max(thresh, global_dirty_limit);
537 }
538 
539 /**
540  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
541  * @bdi: the backing_dev_info to query
542  * @dirty: global dirty limit in pages
543  *
544  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
545  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
546  *
547  * Note that balance_dirty_pages() will only seriously take it as a hard limit
548  * when sleeping max_pause per page is not enough to keep the dirty pages under
549  * control. For example, when the device is completely stalled due to some error
550  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
551  * In the other normal situations, it acts more gently by throttling the tasks
552  * more (rather than completely block them) when the bdi dirty pages go high.
553  *
554  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
555  * - starving fast devices
556  * - piling up dirty pages (that will take long time to sync) on slow devices
557  *
558  * The bdi's share of dirty limit will be adapting to its throughput and
559  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
560  */
561 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
562 {
563         u64 bdi_dirty;
564         long numerator, denominator;
565 
566         /*
567          * Calculate this BDI's share of the dirty ratio.
568          */
569         bdi_writeout_fraction(bdi, &numerator, &denominator);
570 
571         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
572         bdi_dirty *= numerator;
573         do_div(bdi_dirty, denominator);
574 
575         bdi_dirty += (dirty * bdi->min_ratio) / 100;
576         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
577                 bdi_dirty = dirty * bdi->max_ratio / 100;
578 
579         return bdi_dirty;
580 }
581 
582 /*
583  *                           setpoint - dirty 3
584  *        f(dirty) := 1.0 + (----------------)
585  *                           limit - setpoint
586  *
587  * it's a 3rd order polynomial that subjects to
588  *
589  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
590  * (2) f(setpoint) = 1.0 => the balance point
591  * (3) f(limit)    = 0   => the hard limit
592  * (4) df/dx      <= 0   => negative feedback control
593  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
594  *     => fast response on large errors; small oscillation near setpoint
595  */
596 static inline long long pos_ratio_polynom(unsigned long setpoint,
597                                           unsigned long dirty,
598                                           unsigned long limit)
599 {
600         long long pos_ratio;
601         long x;
602 
603         x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
604                     limit - setpoint + 1);
605         pos_ratio = x;
606         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
607         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
608         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
609 
610         return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
611 }
612 
613 /*
614  * Dirty position control.
615  *
616  * (o) global/bdi setpoints
617  *
618  * We want the dirty pages be balanced around the global/bdi setpoints.
619  * When the number of dirty pages is higher/lower than the setpoint, the
620  * dirty position control ratio (and hence task dirty ratelimit) will be
621  * decreased/increased to bring the dirty pages back to the setpoint.
622  *
623  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
624  *
625  *     if (dirty < setpoint) scale up   pos_ratio
626  *     if (dirty > setpoint) scale down pos_ratio
627  *
628  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
629  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
630  *
631  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
632  *
633  * (o) global control line
634  *
635  *     ^ pos_ratio
636  *     |
637  *     |            |<===== global dirty control scope ======>|
638  * 2.0 .............*
639  *     |            .*
640  *     |            . *
641  *     |            .   *
642  *     |            .     *
643  *     |            .        *
644  *     |            .            *
645  * 1.0 ................................*
646  *     |            .                  .     *
647  *     |            .                  .          *
648  *     |            .                  .              *
649  *     |            .                  .                 *
650  *     |            .                  .                    *
651  *   0 +------------.------------------.----------------------*------------->
652  *           freerun^          setpoint^                 limit^   dirty pages
653  *
654  * (o) bdi control line
655  *
656  *     ^ pos_ratio
657  *     |
658  *     |            *
659  *     |              *
660  *     |                *
661  *     |                  *
662  *     |                    * |<=========== span ============>|
663  * 1.0 .......................*
664  *     |                      . *
665  *     |                      .   *
666  *     |                      .     *
667  *     |                      .       *
668  *     |                      .         *
669  *     |                      .           *
670  *     |                      .             *
671  *     |                      .               *
672  *     |                      .                 *
673  *     |                      .                   *
674  *     |                      .                     *
675  * 1/4 ...............................................* * * * * * * * * * * *
676  *     |                      .                         .
677  *     |                      .                           .
678  *     |                      .                             .
679  *   0 +----------------------.-------------------------------.------------->
680  *                bdi_setpoint^                    x_intercept^
681  *
682  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
683  * be smoothly throttled down to normal if it starts high in situations like
684  * - start writing to a slow SD card and a fast disk at the same time. The SD
685  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
686  * - the bdi dirty thresh drops quickly due to change of JBOD workload
687  */
688 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
689                                         unsigned long thresh,
690                                         unsigned long bg_thresh,
691                                         unsigned long dirty,
692                                         unsigned long bdi_thresh,
693                                         unsigned long bdi_dirty)
694 {
695         unsigned long write_bw = bdi->avg_write_bandwidth;
696         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
697         unsigned long limit = hard_dirty_limit(thresh);
698         unsigned long x_intercept;
699         unsigned long setpoint;         /* dirty pages' target balance point */
700         unsigned long bdi_setpoint;
701         unsigned long span;
702         long long pos_ratio;            /* for scaling up/down the rate limit */
703         long x;
704 
705         if (unlikely(dirty >= limit))
706                 return 0;
707 
708         /*
709          * global setpoint
710          *
711          * See comment for pos_ratio_polynom().
712          */
713         setpoint = (freerun + limit) / 2;
714         pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
715 
716         /*
717          * The strictlimit feature is a tool preventing mistrusted filesystems
718          * from growing a large number of dirty pages before throttling. For
719          * such filesystems balance_dirty_pages always checks bdi counters
720          * against bdi limits. Even if global "nr_dirty" is under "freerun".
721          * This is especially important for fuse which sets bdi->max_ratio to
722          * 1% by default. Without strictlimit feature, fuse writeback may
723          * consume arbitrary amount of RAM because it is accounted in
724          * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
725          *
726          * Here, in bdi_position_ratio(), we calculate pos_ratio based on
727          * two values: bdi_dirty and bdi_thresh. Let's consider an example:
728          * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
729          * limits are set by default to 10% and 20% (background and throttle).
730          * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
731          * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
732          * about ~6K pages (as the average of background and throttle bdi
733          * limits). The 3rd order polynomial will provide positive feedback if
734          * bdi_dirty is under bdi_setpoint and vice versa.
735          *
736          * Note, that we cannot use global counters in these calculations
737          * because we want to throttle process writing to a strictlimit BDI
738          * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
739          * in the example above).
740          */
741         if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
742                 long long bdi_pos_ratio;
743                 unsigned long bdi_bg_thresh;
744 
745                 if (bdi_dirty < 8)
746                         return min_t(long long, pos_ratio * 2,
747                                      2 << RATELIMIT_CALC_SHIFT);
748 
749                 if (bdi_dirty >= bdi_thresh)
750                         return 0;
751 
752                 bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
753                 bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
754                                                      bdi_bg_thresh);
755 
756                 if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
757                         return 0;
758 
759                 bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
760                                                   bdi_thresh);
761 
762                 /*
763                  * Typically, for strictlimit case, bdi_setpoint << setpoint
764                  * and pos_ratio >> bdi_pos_ratio. In the other words global
765                  * state ("dirty") is not limiting factor and we have to
766                  * make decision based on bdi counters. But there is an
767                  * important case when global pos_ratio should get precedence:
768                  * global limits are exceeded (e.g. due to activities on other
769                  * BDIs) while given strictlimit BDI is below limit.
770                  *
771                  * "pos_ratio * bdi_pos_ratio" would work for the case above,
772                  * but it would look too non-natural for the case of all
773                  * activity in the system coming from a single strictlimit BDI
774                  * with bdi->max_ratio == 100%.
775                  *
776                  * Note that min() below somewhat changes the dynamics of the
777                  * control system. Normally, pos_ratio value can be well over 3
778                  * (when globally we are at freerun and bdi is well below bdi
779                  * setpoint). Now the maximum pos_ratio in the same situation
780                  * is 2. We might want to tweak this if we observe the control
781                  * system is too slow to adapt.
782                  */
783                 return min(pos_ratio, bdi_pos_ratio);
784         }
785 
786         /*
787          * We have computed basic pos_ratio above based on global situation. If
788          * the bdi is over/under its share of dirty pages, we want to scale
789          * pos_ratio further down/up. That is done by the following mechanism.
790          */
791 
792         /*
793          * bdi setpoint
794          *
795          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
796          *
797          *                        x_intercept - bdi_dirty
798          *                     := --------------------------
799          *                        x_intercept - bdi_setpoint
800          *
801          * The main bdi control line is a linear function that subjects to
802          *
803          * (1) f(bdi_setpoint) = 1.0
804          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
805          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
806          *
807          * For single bdi case, the dirty pages are observed to fluctuate
808          * regularly within range
809          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
810          * for various filesystems, where (2) can yield in a reasonable 12.5%
811          * fluctuation range for pos_ratio.
812          *
813          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
814          * own size, so move the slope over accordingly and choose a slope that
815          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
816          */
817         if (unlikely(bdi_thresh > thresh))
818                 bdi_thresh = thresh;
819         /*
820          * It's very possible that bdi_thresh is close to 0 not because the
821          * device is slow, but that it has remained inactive for long time.
822          * Honour such devices a reasonable good (hopefully IO efficient)
823          * threshold, so that the occasional writes won't be blocked and active
824          * writes can rampup the threshold quickly.
825          */
826         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
827         /*
828          * scale global setpoint to bdi's:
829          *      bdi_setpoint = setpoint * bdi_thresh / thresh
830          */
831         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
832         bdi_setpoint = setpoint * (u64)x >> 16;
833         /*
834          * Use span=(8*write_bw) in single bdi case as indicated by
835          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
836          *
837          *        bdi_thresh                    thresh - bdi_thresh
838          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
839          *          thresh                            thresh
840          */
841         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
842         x_intercept = bdi_setpoint + span;
843 
844         if (bdi_dirty < x_intercept - span / 4) {
845                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
846                                     x_intercept - bdi_setpoint + 1);
847         } else
848                 pos_ratio /= 4;
849 
850         /*
851          * bdi reserve area, safeguard against dirty pool underrun and disk idle
852          * It may push the desired control point of global dirty pages higher
853          * than setpoint.
854          */
855         x_intercept = bdi_thresh / 2;
856         if (bdi_dirty < x_intercept) {
857                 if (bdi_dirty > x_intercept / 8)
858                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
859                 else
860                         pos_ratio *= 8;
861         }
862 
863         return pos_ratio;
864 }
865 
866 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
867                                        unsigned long elapsed,
868                                        unsigned long written)
869 {
870         const unsigned long period = roundup_pow_of_two(3 * HZ);
871         unsigned long avg = bdi->avg_write_bandwidth;
872         unsigned long old = bdi->write_bandwidth;
873         u64 bw;
874 
875         /*
876          * bw = written * HZ / elapsed
877          *
878          *                   bw * elapsed + write_bandwidth * (period - elapsed)
879          * write_bandwidth = ---------------------------------------------------
880          *                                          period
881          */
882         bw = written - bdi->written_stamp;
883         bw *= HZ;
884         if (unlikely(elapsed > period)) {
885                 do_div(bw, elapsed);
886                 avg = bw;
887                 goto out;
888         }
889         bw += (u64)bdi->write_bandwidth * (period - elapsed);
890         bw >>= ilog2(period);
891 
892         /*
893          * one more level of smoothing, for filtering out sudden spikes
894          */
895         if (avg > old && old >= (unsigned long)bw)
896                 avg -= (avg - old) >> 3;
897 
898         if (avg < old && old <= (unsigned long)bw)
899                 avg += (old - avg) >> 3;
900 
901 out:
902         bdi->write_bandwidth = bw;
903         bdi->avg_write_bandwidth = avg;
904 }
905 
906 /*
907  * The global dirtyable memory and dirty threshold could be suddenly knocked
908  * down by a large amount (eg. on the startup of KVM in a swapless system).
909  * This may throw the system into deep dirty exceeded state and throttle
910  * heavy/light dirtiers alike. To retain good responsiveness, maintain
911  * global_dirty_limit for tracking slowly down to the knocked down dirty
912  * threshold.
913  */
914 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
915 {
916         unsigned long limit = global_dirty_limit;
917 
918         /*
919          * Follow up in one step.
920          */
921         if (limit < thresh) {
922                 limit = thresh;
923                 goto update;
924         }
925 
926         /*
927          * Follow down slowly. Use the higher one as the target, because thresh
928          * may drop below dirty. This is exactly the reason to introduce
929          * global_dirty_limit which is guaranteed to lie above the dirty pages.
930          */
931         thresh = max(thresh, dirty);
932         if (limit > thresh) {
933                 limit -= (limit - thresh) >> 5;
934                 goto update;
935         }
936         return;
937 update:
938         global_dirty_limit = limit;
939 }
940 
941 static void global_update_bandwidth(unsigned long thresh,
942                                     unsigned long dirty,
943                                     unsigned long now)
944 {
945         static DEFINE_SPINLOCK(dirty_lock);
946         static unsigned long update_time;
947 
948         /*
949          * check locklessly first to optimize away locking for the most time
950          */
951         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
952                 return;
953 
954         spin_lock(&dirty_lock);
955         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
956                 update_dirty_limit(thresh, dirty);
957                 update_time = now;
958         }
959         spin_unlock(&dirty_lock);
960 }
961 
962 /*
963  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
964  *
965  * Normal bdi tasks will be curbed at or below it in long term.
966  * Obviously it should be around (write_bw / N) when there are N dd tasks.
967  */
968 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
969                                        unsigned long thresh,
970                                        unsigned long bg_thresh,
971                                        unsigned long dirty,
972                                        unsigned long bdi_thresh,
973                                        unsigned long bdi_dirty,
974                                        unsigned long dirtied,
975                                        unsigned long elapsed)
976 {
977         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
978         unsigned long limit = hard_dirty_limit(thresh);
979         unsigned long setpoint = (freerun + limit) / 2;
980         unsigned long write_bw = bdi->avg_write_bandwidth;
981         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
982         unsigned long dirty_rate;
983         unsigned long task_ratelimit;
984         unsigned long balanced_dirty_ratelimit;
985         unsigned long pos_ratio;
986         unsigned long step;
987         unsigned long x;
988 
989         /*
990          * The dirty rate will match the writeout rate in long term, except
991          * when dirty pages are truncated by userspace or re-dirtied by FS.
992          */
993         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
994 
995         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
996                                        bdi_thresh, bdi_dirty);
997         /*
998          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
999          */
1000         task_ratelimit = (u64)dirty_ratelimit *
1001                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
1002         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1003 
1004         /*
1005          * A linear estimation of the "balanced" throttle rate. The theory is,
1006          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
1007          * dirty_rate will be measured to be (N * task_ratelimit). So the below
1008          * formula will yield the balanced rate limit (write_bw / N).
1009          *
1010          * Note that the expanded form is not a pure rate feedback:
1011          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1012          * but also takes pos_ratio into account:
1013          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1014          *
1015          * (1) is not realistic because pos_ratio also takes part in balancing
1016          * the dirty rate.  Consider the state
1017          *      pos_ratio = 0.5                                              (3)
1018          *      rate = 2 * (write_bw / N)                                    (4)
1019          * If (1) is used, it will stuck in that state! Because each dd will
1020          * be throttled at
1021          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1022          * yielding
1023          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1024          * put (6) into (1) we get
1025          *      rate_(i+1) = rate_(i)                                        (7)
1026          *
1027          * So we end up using (2) to always keep
1028          *      rate_(i+1) ~= (write_bw / N)                                 (8)
1029          * regardless of the value of pos_ratio. As long as (8) is satisfied,
1030          * pos_ratio is able to drive itself to 1.0, which is not only where
1031          * the dirty count meet the setpoint, but also where the slope of
1032          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1033          */
1034         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1035                                            dirty_rate | 1);
1036         /*
1037          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1038          */
1039         if (unlikely(balanced_dirty_ratelimit > write_bw))
1040                 balanced_dirty_ratelimit = write_bw;
1041 
1042         /*
1043          * We could safely do this and return immediately:
1044          *
1045          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
1046          *
1047          * However to get a more stable dirty_ratelimit, the below elaborated
1048          * code makes use of task_ratelimit to filter out singular points and
1049          * limit the step size.
1050          *
1051          * The below code essentially only uses the relative value of
1052          *
1053          *      task_ratelimit - dirty_ratelimit
1054          *      = (pos_ratio - 1) * dirty_ratelimit
1055          *
1056          * which reflects the direction and size of dirty position error.
1057          */
1058 
1059         /*
1060          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1061          * task_ratelimit is on the same side of dirty_ratelimit, too.
1062          * For example, when
1063          * - dirty_ratelimit > balanced_dirty_ratelimit
1064          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1065          * lowering dirty_ratelimit will help meet both the position and rate
1066          * control targets. Otherwise, don't update dirty_ratelimit if it will
1067          * only help meet the rate target. After all, what the users ultimately
1068          * feel and care are stable dirty rate and small position error.
1069          *
1070          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1071          * and filter out the singular points of balanced_dirty_ratelimit. Which
1072          * keeps jumping around randomly and can even leap far away at times
1073          * due to the small 200ms estimation period of dirty_rate (we want to
1074          * keep that period small to reduce time lags).
1075          */
1076         step = 0;
1077 
1078         /*
1079          * For strictlimit case, calculations above were based on bdi counters
1080          * and limits (starting from pos_ratio = bdi_position_ratio() and up to
1081          * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1082          * Hence, to calculate "step" properly, we have to use bdi_dirty as
1083          * "dirty" and bdi_setpoint as "setpoint".
1084          *
1085          * We rampup dirty_ratelimit forcibly if bdi_dirty is low because
1086          * it's possible that bdi_thresh is close to zero due to inactivity
1087          * of backing device (see the implementation of bdi_dirty_limit()).
1088          */
1089         if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1090                 dirty = bdi_dirty;
1091                 if (bdi_dirty < 8)
1092                         setpoint = bdi_dirty + 1;
1093                 else
1094                         setpoint = (bdi_thresh +
1095                                     bdi_dirty_limit(bdi, bg_thresh)) / 2;
1096         }
1097 
1098         if (dirty < setpoint) {
1099                 x = min(bdi->balanced_dirty_ratelimit,
1100                          min(balanced_dirty_ratelimit, task_ratelimit));
1101                 if (dirty_ratelimit < x)
1102                         step = x - dirty_ratelimit;
1103         } else {
1104                 x = max(bdi->balanced_dirty_ratelimit,
1105                          max(balanced_dirty_ratelimit, task_ratelimit));
1106                 if (dirty_ratelimit > x)
1107                         step = dirty_ratelimit - x;
1108         }
1109 
1110         /*
1111          * Don't pursue 100% rate matching. It's impossible since the balanced
1112          * rate itself is constantly fluctuating. So decrease the track speed
1113          * when it gets close to the target. Helps eliminate pointless tremors.
1114          */
1115         step >>= dirty_ratelimit / (2 * step + 1);
1116         /*
1117          * Limit the tracking speed to avoid overshooting.
1118          */
1119         step = (step + 7) / 8;
1120 
1121         if (dirty_ratelimit < balanced_dirty_ratelimit)
1122                 dirty_ratelimit += step;
1123         else
1124                 dirty_ratelimit -= step;
1125 
1126         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1127         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1128 
1129         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1130 }
1131 
1132 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1133                             unsigned long thresh,
1134                             unsigned long bg_thresh,
1135                             unsigned long dirty,
1136                             unsigned long bdi_thresh,
1137                             unsigned long bdi_dirty,
1138                             unsigned long start_time)
1139 {
1140         unsigned long now = jiffies;
1141         unsigned long elapsed = now - bdi->bw_time_stamp;
1142         unsigned long dirtied;
1143         unsigned long written;
1144 
1145         /*
1146          * rate-limit, only update once every 200ms.
1147          */
1148         if (elapsed < BANDWIDTH_INTERVAL)
1149                 return;
1150 
1151         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1152         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1153 
1154         /*
1155          * Skip quiet periods when disk bandwidth is under-utilized.
1156          * (at least 1s idle time between two flusher runs)
1157          */
1158         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1159                 goto snapshot;
1160 
1161         if (thresh) {
1162                 global_update_bandwidth(thresh, dirty, now);
1163                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1164                                            bdi_thresh, bdi_dirty,
1165                                            dirtied, elapsed);
1166         }
1167         bdi_update_write_bandwidth(bdi, elapsed, written);
1168 
1169 snapshot:
1170         bdi->dirtied_stamp = dirtied;
1171         bdi->written_stamp = written;
1172         bdi->bw_time_stamp = now;
1173 }
1174 
1175 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1176                                  unsigned long thresh,
1177                                  unsigned long bg_thresh,
1178                                  unsigned long dirty,
1179                                  unsigned long bdi_thresh,
1180                                  unsigned long bdi_dirty,
1181                                  unsigned long start_time)
1182 {
1183         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1184                 return;
1185         spin_lock(&bdi->wb.list_lock);
1186         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1187                                bdi_thresh, bdi_dirty, start_time);
1188         spin_unlock(&bdi->wb.list_lock);
1189 }
1190 
1191 /*
1192  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1193  * will look to see if it needs to start dirty throttling.
1194  *
1195  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1196  * global_page_state() too often. So scale it near-sqrt to the safety margin
1197  * (the number of pages we may dirty without exceeding the dirty limits).
1198  */
1199 static unsigned long dirty_poll_interval(unsigned long dirty,
1200                                          unsigned long thresh)
1201 {
1202         if (thresh > dirty)
1203                 return 1UL << (ilog2(thresh - dirty) >> 1);
1204 
1205         return 1;
1206 }
1207 
1208 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1209                                    unsigned long bdi_dirty)
1210 {
1211         unsigned long bw = bdi->avg_write_bandwidth;
1212         unsigned long t;
1213 
1214         /*
1215          * Limit pause time for small memory systems. If sleeping for too long
1216          * time, a small pool of dirty/writeback pages may go empty and disk go
1217          * idle.
1218          *
1219          * 8 serves as the safety ratio.
1220          */
1221         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1222         t++;
1223 
1224         return min_t(unsigned long, t, MAX_PAUSE);
1225 }
1226 
1227 static long bdi_min_pause(struct backing_dev_info *bdi,
1228                           long max_pause,
1229                           unsigned long task_ratelimit,
1230                           unsigned long dirty_ratelimit,
1231                           int *nr_dirtied_pause)
1232 {
1233         long hi = ilog2(bdi->avg_write_bandwidth);
1234         long lo = ilog2(bdi->dirty_ratelimit);
1235         long t;         /* target pause */
1236         long pause;     /* estimated next pause */
1237         int pages;      /* target nr_dirtied_pause */
1238 
1239         /* target for 10ms pause on 1-dd case */
1240         t = max(1, HZ / 100);
1241 
1242         /*
1243          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1244          * overheads.
1245          *
1246          * (N * 10ms) on 2^N concurrent tasks.
1247          */
1248         if (hi > lo)
1249                 t += (hi - lo) * (10 * HZ) / 1024;
1250 
1251         /*
1252          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1253          * on the much more stable dirty_ratelimit. However the next pause time
1254          * will be computed based on task_ratelimit and the two rate limits may
1255          * depart considerably at some time. Especially if task_ratelimit goes
1256          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1257          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1258          * result task_ratelimit won't be executed faithfully, which could
1259          * eventually bring down dirty_ratelimit.
1260          *
1261          * We apply two rules to fix it up:
1262          * 1) try to estimate the next pause time and if necessary, use a lower
1263          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1264          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1265          * 2) limit the target pause time to max_pause/2, so that the normal
1266          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1267          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1268          */
1269         t = min(t, 1 + max_pause / 2);
1270         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1271 
1272         /*
1273          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1274          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1275          * When the 16 consecutive reads are often interrupted by some dirty
1276          * throttling pause during the async writes, cfq will go into idles
1277          * (deadline is fine). So push nr_dirtied_pause as high as possible
1278          * until reaches DIRTY_POLL_THRESH=32 pages.
1279          */
1280         if (pages < DIRTY_POLL_THRESH) {
1281                 t = max_pause;
1282                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1283                 if (pages > DIRTY_POLL_THRESH) {
1284                         pages = DIRTY_POLL_THRESH;
1285                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1286                 }
1287         }
1288 
1289         pause = HZ * pages / (task_ratelimit + 1);
1290         if (pause > max_pause) {
1291                 t = max_pause;
1292                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1293         }
1294 
1295         *nr_dirtied_pause = pages;
1296         /*
1297          * The minimal pause time will normally be half the target pause time.
1298          */
1299         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1300 }
1301 
1302 static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
1303                                     unsigned long dirty_thresh,
1304                                     unsigned long background_thresh,
1305                                     unsigned long *bdi_dirty,
1306                                     unsigned long *bdi_thresh,
1307                                     unsigned long *bdi_bg_thresh)
1308 {
1309         unsigned long bdi_reclaimable;
1310 
1311         /*
1312          * bdi_thresh is not treated as some limiting factor as
1313          * dirty_thresh, due to reasons
1314          * - in JBOD setup, bdi_thresh can fluctuate a lot
1315          * - in a system with HDD and USB key, the USB key may somehow
1316          *   go into state (bdi_dirty >> bdi_thresh) either because
1317          *   bdi_dirty starts high, or because bdi_thresh drops low.
1318          *   In this case we don't want to hard throttle the USB key
1319          *   dirtiers for 100 seconds until bdi_dirty drops under
1320          *   bdi_thresh. Instead the auxiliary bdi control line in
1321          *   bdi_position_ratio() will let the dirtier task progress
1322          *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1323          */
1324         *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1325 
1326         if (bdi_bg_thresh)
1327                 *bdi_bg_thresh = div_u64((u64)*bdi_thresh *
1328                                          background_thresh,
1329                                          dirty_thresh);
1330 
1331         /*
1332          * In order to avoid the stacked BDI deadlock we need
1333          * to ensure we accurately count the 'dirty' pages when
1334          * the threshold is low.
1335          *
1336          * Otherwise it would be possible to get thresh+n pages
1337          * reported dirty, even though there are thresh-m pages
1338          * actually dirty; with m+n sitting in the percpu
1339          * deltas.
1340          */
1341         if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
1342                 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1343                 *bdi_dirty = bdi_reclaimable +
1344                         bdi_stat_sum(bdi, BDI_WRITEBACK);
1345         } else {
1346                 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1347                 *bdi_dirty = bdi_reclaimable +
1348                         bdi_stat(bdi, BDI_WRITEBACK);
1349         }
1350 }
1351 
1352 /*
1353  * balance_dirty_pages() must be called by processes which are generating dirty
1354  * data.  It looks at the number of dirty pages in the machine and will force
1355  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1356  * If we're over `background_thresh' then the writeback threads are woken to
1357  * perform some writeout.
1358  */
1359 static void balance_dirty_pages(struct address_space *mapping,
1360                                 unsigned long pages_dirtied)
1361 {
1362         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1363         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1364         unsigned long background_thresh;
1365         unsigned long dirty_thresh;
1366         long period;
1367         long pause;
1368         long max_pause;
1369         long min_pause;
1370         int nr_dirtied_pause;
1371         bool dirty_exceeded = false;
1372         unsigned long task_ratelimit;
1373         unsigned long dirty_ratelimit;
1374         unsigned long pos_ratio;
1375         struct backing_dev_info *bdi = mapping->backing_dev_info;
1376         bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1377         unsigned long start_time = jiffies;
1378 
1379         for (;;) {
1380                 unsigned long now = jiffies;
1381                 unsigned long uninitialized_var(bdi_thresh);
1382                 unsigned long thresh;
1383                 unsigned long uninitialized_var(bdi_dirty);
1384                 unsigned long dirty;
1385                 unsigned long bg_thresh;
1386 
1387                 /*
1388                  * Unstable writes are a feature of certain networked
1389                  * filesystems (i.e. NFS) in which data may have been
1390                  * written to the server's write cache, but has not yet
1391                  * been flushed to permanent storage.
1392                  */
1393                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1394                                         global_page_state(NR_UNSTABLE_NFS);
1395                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1396 
1397                 global_dirty_limits(&background_thresh, &dirty_thresh);
1398 
1399                 if (unlikely(strictlimit)) {
1400                         bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1401                                          &bdi_dirty, &bdi_thresh, &bg_thresh);
1402 
1403                         dirty = bdi_dirty;
1404                         thresh = bdi_thresh;
1405                 } else {
1406                         dirty = nr_dirty;
1407                         thresh = dirty_thresh;
1408                         bg_thresh = background_thresh;
1409                 }
1410 
1411                 /*
1412                  * Throttle it only when the background writeback cannot
1413                  * catch-up. This avoids (excessively) small writeouts
1414                  * when the bdi limits are ramping up in case of !strictlimit.
1415                  *
1416                  * In strictlimit case make decision based on the bdi counters
1417                  * and limits. Small writeouts when the bdi limits are ramping
1418                  * up are the price we consciously pay for strictlimit-ing.
1419                  */
1420                 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1421                         current->dirty_paused_when = now;
1422                         current->nr_dirtied = 0;
1423                         current->nr_dirtied_pause =
1424                                 dirty_poll_interval(dirty, thresh);
1425                         break;
1426                 }
1427 
1428                 if (unlikely(!writeback_in_progress(bdi)))
1429                         bdi_start_background_writeback(bdi);
1430 
1431                 if (!strictlimit)
1432                         bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1433                                          &bdi_dirty, &bdi_thresh, NULL);
1434 
1435                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1436                                  ((nr_dirty > dirty_thresh) || strictlimit);
1437                 if (dirty_exceeded && !bdi->dirty_exceeded)
1438                         bdi->dirty_exceeded = 1;
1439 
1440                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1441                                      nr_dirty, bdi_thresh, bdi_dirty,
1442                                      start_time);
1443 
1444                 dirty_ratelimit = bdi->dirty_ratelimit;
1445                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1446                                                background_thresh, nr_dirty,
1447                                                bdi_thresh, bdi_dirty);
1448                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1449                                                         RATELIMIT_CALC_SHIFT;
1450                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1451                 min_pause = bdi_min_pause(bdi, max_pause,
1452                                           task_ratelimit, dirty_ratelimit,
1453                                           &nr_dirtied_pause);
1454 
1455                 if (unlikely(task_ratelimit == 0)) {
1456                         period = max_pause;
1457                         pause = max_pause;
1458                         goto pause;
1459                 }
1460                 period = HZ * pages_dirtied / task_ratelimit;
1461                 pause = period;
1462                 if (current->dirty_paused_when)
1463                         pause -= now - current->dirty_paused_when;
1464                 /*
1465                  * For less than 1s think time (ext3/4 may block the dirtier
1466                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1467                  * however at much less frequency), try to compensate it in
1468                  * future periods by updating the virtual time; otherwise just
1469                  * do a reset, as it may be a light dirtier.
1470                  */
1471                 if (pause < min_pause) {
1472                         trace_balance_dirty_pages(bdi,
1473                                                   dirty_thresh,
1474                                                   background_thresh,
1475                                                   nr_dirty,
1476                                                   bdi_thresh,
1477                                                   bdi_dirty,
1478                                                   dirty_ratelimit,
1479                                                   task_ratelimit,
1480                                                   pages_dirtied,
1481                                                   period,
1482                                                   min(pause, 0L),
1483                                                   start_time);
1484                         if (pause < -HZ) {
1485                                 current->dirty_paused_when = now;
1486                                 current->nr_dirtied = 0;
1487                         } else if (period) {
1488                                 current->dirty_paused_when += period;
1489                                 current->nr_dirtied = 0;
1490                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1491                                 current->nr_dirtied_pause += pages_dirtied;
1492                         break;
1493                 }
1494                 if (unlikely(pause > max_pause)) {
1495                         /* for occasional dropped task_ratelimit */
1496                         now += min(pause - max_pause, max_pause);
1497                         pause = max_pause;
1498                 }
1499 
1500 pause:
1501                 trace_balance_dirty_pages(bdi,
1502                                           dirty_thresh,
1503                                           background_thresh,
1504                                           nr_dirty,
1505                                           bdi_thresh,
1506                                           bdi_dirty,
1507                                           dirty_ratelimit,
1508                                           task_ratelimit,
1509                                           pages_dirtied,
1510                                           period,
1511                                           pause,
1512                                           start_time);
1513                 __set_current_state(TASK_KILLABLE);
1514                 io_schedule_timeout(pause);
1515 
1516                 current->dirty_paused_when = now + pause;
1517                 current->nr_dirtied = 0;
1518                 current->nr_dirtied_pause = nr_dirtied_pause;
1519 
1520                 /*
1521                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1522                  * also keep "1000+ dd on a slow USB stick" under control.
1523                  */
1524                 if (task_ratelimit)
1525                         break;
1526 
1527                 /*
1528                  * In the case of an unresponding NFS server and the NFS dirty
1529                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1530                  * to go through, so that tasks on them still remain responsive.
1531                  *
1532                  * In theory 1 page is enough to keep the comsumer-producer
1533                  * pipe going: the flusher cleans 1 page => the task dirties 1
1534                  * more page. However bdi_dirty has accounting errors.  So use
1535                  * the larger and more IO friendly bdi_stat_error.
1536                  */
1537                 if (bdi_dirty <= bdi_stat_error(bdi))
1538                         break;
1539 
1540                 if (fatal_signal_pending(current))
1541                         break;
1542         }
1543 
1544         if (!dirty_exceeded && bdi->dirty_exceeded)
1545                 bdi->dirty_exceeded = 0;
1546 
1547         if (writeback_in_progress(bdi))
1548                 return;
1549 
1550         /*
1551          * In laptop mode, we wait until hitting the higher threshold before
1552          * starting background writeout, and then write out all the way down
1553          * to the lower threshold.  So slow writers cause minimal disk activity.
1554          *
1555          * In normal mode, we start background writeout at the lower
1556          * background_thresh, to keep the amount of dirty memory low.
1557          */
1558         if (laptop_mode)
1559                 return;
1560 
1561         if (nr_reclaimable > background_thresh)
1562                 bdi_start_background_writeback(bdi);
1563 }
1564 
1565 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1566 {
1567         if (set_page_dirty(page) || page_mkwrite) {
1568                 struct address_space *mapping = page_mapping(page);
1569 
1570                 if (mapping)
1571                         balance_dirty_pages_ratelimited(mapping);
1572         }
1573 }
1574 
1575 static DEFINE_PER_CPU(int, bdp_ratelimits);
1576 
1577 /*
1578  * Normal tasks are throttled by
1579  *      loop {
1580  *              dirty tsk->nr_dirtied_pause pages;
1581  *              take a snap in balance_dirty_pages();
1582  *      }
1583  * However there is a worst case. If every task exit immediately when dirtied
1584  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1585  * called to throttle the page dirties. The solution is to save the not yet
1586  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1587  * randomly into the running tasks. This works well for the above worst case,
1588  * as the new task will pick up and accumulate the old task's leaked dirty
1589  * count and eventually get throttled.
1590  */
1591 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1592 
1593 /**
1594  * balance_dirty_pages_ratelimited - balance dirty memory state
1595  * @mapping: address_space which was dirtied
1596  *
1597  * Processes which are dirtying memory should call in here once for each page
1598  * which was newly dirtied.  The function will periodically check the system's
1599  * dirty state and will initiate writeback if needed.
1600  *
1601  * On really big machines, get_writeback_state is expensive, so try to avoid
1602  * calling it too often (ratelimiting).  But once we're over the dirty memory
1603  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1604  * from overshooting the limit by (ratelimit_pages) each.
1605  */
1606 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1607 {
1608         struct backing_dev_info *bdi = mapping->backing_dev_info;
1609         int ratelimit;
1610         int *p;
1611 
1612         if (!bdi_cap_account_dirty(bdi))
1613                 return;
1614 
1615         ratelimit = current->nr_dirtied_pause;
1616         if (bdi->dirty_exceeded)
1617                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1618 
1619         preempt_disable();
1620         /*
1621          * This prevents one CPU to accumulate too many dirtied pages without
1622          * calling into balance_dirty_pages(), which can happen when there are
1623          * 1000+ tasks, all of them start dirtying pages at exactly the same
1624          * time, hence all honoured too large initial task->nr_dirtied_pause.
1625          */
1626         p =  &__get_cpu_var(bdp_ratelimits);
1627         if (unlikely(current->nr_dirtied >= ratelimit))
1628                 *p = 0;
1629         else if (unlikely(*p >= ratelimit_pages)) {
1630                 *p = 0;
1631                 ratelimit = 0;
1632         }
1633         /*
1634          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1635          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1636          * the dirty throttling and livelock other long-run dirtiers.
1637          */
1638         p = &__get_cpu_var(dirty_throttle_leaks);
1639         if (*p > 0 && current->nr_dirtied < ratelimit) {
1640                 unsigned long nr_pages_dirtied;
1641                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1642                 *p -= nr_pages_dirtied;
1643                 current->nr_dirtied += nr_pages_dirtied;
1644         }
1645         preempt_enable();
1646 
1647         if (unlikely(current->nr_dirtied >= ratelimit))
1648                 balance_dirty_pages(mapping, current->nr_dirtied);
1649 }
1650 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1651 
1652 void throttle_vm_writeout(gfp_t gfp_mask)
1653 {
1654         unsigned long background_thresh;
1655         unsigned long dirty_thresh;
1656 
1657         for ( ; ; ) {
1658                 global_dirty_limits(&background_thresh, &dirty_thresh);
1659                 dirty_thresh = hard_dirty_limit(dirty_thresh);
1660 
1661                 /*
1662                  * Boost the allowable dirty threshold a bit for page
1663                  * allocators so they don't get DoS'ed by heavy writers
1664                  */
1665                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1666 
1667                 if (global_page_state(NR_UNSTABLE_NFS) +
1668                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1669                                 break;
1670                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1671 
1672                 /*
1673                  * The caller might hold locks which can prevent IO completion
1674                  * or progress in the filesystem.  So we cannot just sit here
1675                  * waiting for IO to complete.
1676                  */
1677                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1678                         break;
1679         }
1680 }
1681 
1682 /*
1683  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1684  */
1685 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1686         void __user *buffer, size_t *length, loff_t *ppos)
1687 {
1688         proc_dointvec(table, write, buffer, length, ppos);
1689         return 0;
1690 }
1691 
1692 #ifdef CONFIG_BLOCK
1693 void laptop_mode_timer_fn(unsigned long data)
1694 {
1695         struct request_queue *q = (struct request_queue *)data;
1696         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1697                 global_page_state(NR_UNSTABLE_NFS);
1698 
1699         /*
1700          * We want to write everything out, not just down to the dirty
1701          * threshold
1702          */
1703         if (bdi_has_dirty_io(&q->backing_dev_info))
1704                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1705                                         WB_REASON_LAPTOP_TIMER);
1706 }
1707 
1708 /*
1709  * We've spun up the disk and we're in laptop mode: schedule writeback
1710  * of all dirty data a few seconds from now.  If the flush is already scheduled
1711  * then push it back - the user is still using the disk.
1712  */
1713 void laptop_io_completion(struct backing_dev_info *info)
1714 {
1715         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1716 }
1717 
1718 /*
1719  * We're in laptop mode and we've just synced. The sync's writes will have
1720  * caused another writeback to be scheduled by laptop_io_completion.
1721  * Nothing needs to be written back anymore, so we unschedule the writeback.
1722  */
1723 void laptop_sync_completion(void)
1724 {
1725         struct backing_dev_info *bdi;
1726 
1727         rcu_read_lock();
1728 
1729         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1730                 del_timer(&bdi->laptop_mode_wb_timer);
1731 
1732         rcu_read_unlock();
1733 }
1734 #endif
1735 
1736 /*
1737  * If ratelimit_pages is too high then we can get into dirty-data overload
1738  * if a large number of processes all perform writes at the same time.
1739  * If it is too low then SMP machines will call the (expensive)
1740  * get_writeback_state too often.
1741  *
1742  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1743  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1744  * thresholds.
1745  */
1746 
1747 void writeback_set_ratelimit(void)
1748 {
1749         unsigned long background_thresh;
1750         unsigned long dirty_thresh;
1751         global_dirty_limits(&background_thresh, &dirty_thresh);
1752         global_dirty_limit = dirty_thresh;
1753         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1754         if (ratelimit_pages < 16)
1755                 ratelimit_pages = 16;
1756 }
1757 
1758 static int
1759 ratelimit_handler(struct notifier_block *self, unsigned long action,
1760                   void *hcpu)
1761 {
1762 
1763         switch (action & ~CPU_TASKS_FROZEN) {
1764         case CPU_ONLINE:
1765         case CPU_DEAD:
1766                 writeback_set_ratelimit();
1767                 return NOTIFY_OK;
1768         default:
1769                 return NOTIFY_DONE;
1770         }
1771 }
1772 
1773 static struct notifier_block ratelimit_nb = {
1774         .notifier_call  = ratelimit_handler,
1775         .next           = NULL,
1776 };
1777 
1778 /*
1779  * Called early on to tune the page writeback dirty limits.
1780  *
1781  * We used to scale dirty pages according to how total memory
1782  * related to pages that could be allocated for buffers (by
1783  * comparing nr_free_buffer_pages() to vm_total_pages.
1784  *
1785  * However, that was when we used "dirty_ratio" to scale with
1786  * all memory, and we don't do that any more. "dirty_ratio"
1787  * is now applied to total non-HIGHPAGE memory (by subtracting
1788  * totalhigh_pages from vm_total_pages), and as such we can't
1789  * get into the old insane situation any more where we had
1790  * large amounts of dirty pages compared to a small amount of
1791  * non-HIGHMEM memory.
1792  *
1793  * But we might still want to scale the dirty_ratio by how
1794  * much memory the box has..
1795  */
1796 void __init page_writeback_init(void)
1797 {
1798         writeback_set_ratelimit();
1799         register_cpu_notifier(&ratelimit_nb);
1800 
1801         fprop_global_init(&writeout_completions);
1802 }
1803 
1804 /**
1805  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1806  * @mapping: address space structure to write
1807  * @start: starting page index
1808  * @end: ending page index (inclusive)
1809  *
1810  * This function scans the page range from @start to @end (inclusive) and tags
1811  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1812  * that write_cache_pages (or whoever calls this function) will then use
1813  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1814  * used to avoid livelocking of writeback by a process steadily creating new
1815  * dirty pages in the file (thus it is important for this function to be quick
1816  * so that it can tag pages faster than a dirtying process can create them).
1817  */
1818 /*
1819  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1820  */
1821 void tag_pages_for_writeback(struct address_space *mapping,
1822                              pgoff_t start, pgoff_t end)
1823 {
1824 #define WRITEBACK_TAG_BATCH 4096
1825         unsigned long tagged;
1826 
1827         do {
1828                 spin_lock_irq(&mapping->tree_lock);
1829                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1830                                 &start, end, WRITEBACK_TAG_BATCH,
1831                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1832                 spin_unlock_irq(&mapping->tree_lock);
1833                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1834                 cond_resched();
1835                 /* We check 'start' to handle wrapping when end == ~0UL */
1836         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1837 }
1838 EXPORT_SYMBOL(tag_pages_for_writeback);
1839 
1840 /**
1841  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1842  * @mapping: address space structure to write
1843  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1844  * @writepage: function called for each page
1845  * @data: data passed to writepage function
1846  *
1847  * If a page is already under I/O, write_cache_pages() skips it, even
1848  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1849  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1850  * and msync() need to guarantee that all the data which was dirty at the time
1851  * the call was made get new I/O started against them.  If wbc->sync_mode is
1852  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1853  * existing IO to complete.
1854  *
1855  * To avoid livelocks (when other process dirties new pages), we first tag
1856  * pages which should be written back with TOWRITE tag and only then start
1857  * writing them. For data-integrity sync we have to be careful so that we do
1858  * not miss some pages (e.g., because some other process has cleared TOWRITE
1859  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1860  * by the process clearing the DIRTY tag (and submitting the page for IO).
1861  */
1862 int write_cache_pages(struct address_space *mapping,
1863                       struct writeback_control *wbc, writepage_t writepage,
1864                       void *data)
1865 {
1866         int ret = 0;
1867         int done = 0;
1868         struct pagevec pvec;
1869         int nr_pages;
1870         pgoff_t uninitialized_var(writeback_index);
1871         pgoff_t index;
1872         pgoff_t end;            /* Inclusive */
1873         pgoff_t done_index;
1874         int cycled;
1875         int range_whole = 0;
1876         int tag;
1877 
1878         pagevec_init(&pvec, 0);
1879         if (wbc->range_cyclic) {
1880                 writeback_index = mapping->writeback_index; /* prev offset */
1881                 index = writeback_index;
1882                 if (index == 0)
1883                         cycled = 1;
1884                 else
1885                         cycled = 0;
1886                 end = -1;
1887         } else {
1888                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1889                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1890                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1891                         range_whole = 1;
1892                 cycled = 1; /* ignore range_cyclic tests */
1893         }
1894         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1895                 tag = PAGECACHE_TAG_TOWRITE;
1896         else
1897                 tag = PAGECACHE_TAG_DIRTY;
1898 retry:
1899         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1900                 tag_pages_for_writeback(mapping, index, end);
1901         done_index = index;
1902         while (!done && (index <= end)) {
1903                 int i;
1904 
1905                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1906                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1907                 if (nr_pages == 0)
1908                         break;
1909 
1910                 for (i = 0; i < nr_pages; i++) {
1911                         struct page *page = pvec.pages[i];
1912 
1913                         /*
1914                          * At this point, the page may be truncated or
1915                          * invalidated (changing page->mapping to NULL), or
1916                          * even swizzled back from swapper_space to tmpfs file
1917                          * mapping. However, page->index will not change
1918                          * because we have a reference on the page.
1919                          */
1920                         if (page->index > end) {
1921                                 /*
1922                                  * can't be range_cyclic (1st pass) because
1923                                  * end == -1 in that case.
1924                                  */
1925                                 done = 1;
1926                                 break;
1927                         }
1928 
1929                         done_index = page->index;
1930 
1931                         lock_page(page);
1932 
1933                         /*
1934                          * Page truncated or invalidated. We can freely skip it
1935                          * then, even for data integrity operations: the page
1936                          * has disappeared concurrently, so there could be no
1937                          * real expectation of this data interity operation
1938                          * even if there is now a new, dirty page at the same
1939                          * pagecache address.
1940                          */
1941                         if (unlikely(page->mapping != mapping)) {
1942 continue_unlock:
1943                                 unlock_page(page);
1944                                 continue;
1945                         }
1946 
1947                         if (!PageDirty(page)) {
1948                                 /* someone wrote it for us */
1949                                 goto continue_unlock;
1950                         }
1951 
1952                         if (PageWriteback(page)) {
1953                                 if (wbc->sync_mode != WB_SYNC_NONE)
1954                                         wait_on_page_writeback(page);
1955                                 else
1956                                         goto continue_unlock;
1957                         }
1958 
1959                         BUG_ON(PageWriteback(page));
1960                         if (!clear_page_dirty_for_io(page))
1961                                 goto continue_unlock;
1962 
1963                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1964                         ret = (*writepage)(page, wbc, data);
1965                         if (unlikely(ret)) {
1966                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1967                                         unlock_page(page);
1968                                         ret = 0;
1969                                 } else {
1970                                         /*
1971                                          * done_index is set past this page,
1972                                          * so media errors will not choke
1973                                          * background writeout for the entire
1974                                          * file. This has consequences for
1975                                          * range_cyclic semantics (ie. it may
1976                                          * not be suitable for data integrity
1977                                          * writeout).
1978                                          */
1979                                         done_index = page->index + 1;
1980                                         done = 1;
1981                                         break;
1982                                 }
1983                         }
1984 
1985                         /*
1986                          * We stop writing back only if we are not doing
1987                          * integrity sync. In case of integrity sync we have to
1988                          * keep going until we have written all the pages
1989                          * we tagged for writeback prior to entering this loop.
1990                          */
1991                         if (--wbc->nr_to_write <= 0 &&
1992                             wbc->sync_mode == WB_SYNC_NONE) {
1993                                 done = 1;
1994                                 break;
1995                         }
1996                 }
1997                 pagevec_release(&pvec);
1998                 cond_resched();
1999         }
2000         if (!cycled && !done) {
2001                 /*
2002                  * range_cyclic:
2003                  * We hit the last page and there is more work to be done: wrap
2004                  * back to the start of the file
2005                  */
2006                 cycled = 1;
2007                 index = 0;
2008                 end = writeback_index - 1;
2009                 goto retry;
2010         }
2011         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2012                 mapping->writeback_index = done_index;
2013 
2014         return ret;
2015 }
2016 EXPORT_SYMBOL(write_cache_pages);
2017 
2018 /*
2019  * Function used by generic_writepages to call the real writepage
2020  * function and set the mapping flags on error
2021  */
2022 static int __writepage(struct page *page, struct writeback_control *wbc,
2023                        void *data)
2024 {
2025         struct address_space *mapping = data;
2026         int ret = mapping->a_ops->writepage(page, wbc);
2027         mapping_set_error(mapping, ret);
2028         return ret;
2029 }
2030 
2031 /**
2032  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2033  * @mapping: address space structure to write
2034  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2035  *
2036  * This is a library function, which implements the writepages()
2037  * address_space_operation.
2038  */
2039 int generic_writepages(struct address_space *mapping,
2040                        struct writeback_control *wbc)
2041 {
2042         struct blk_plug plug;
2043         int ret;
2044 
2045         /* deal with chardevs and other special file */
2046         if (!mapping->a_ops->writepage)
2047                 return 0;
2048 
2049         blk_start_plug(&plug);
2050         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2051         blk_finish_plug(&plug);
2052         return ret;
2053 }
2054 
2055 EXPORT_SYMBOL(generic_writepages);
2056 
2057 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2058 {
2059         int ret;
2060 
2061         if (wbc->nr_to_write <= 0)
2062                 return 0;
2063         if (mapping->a_ops->writepages)
2064                 ret = mapping->a_ops->writepages(mapping, wbc);
2065         else
2066                 ret = generic_writepages(mapping, wbc);
2067         return ret;
2068 }
2069 
2070 /**
2071  * write_one_page - write out a single page and optionally wait on I/O
2072  * @page: the page to write
2073  * @wait: if true, wait on writeout
2074  *
2075  * The page must be locked by the caller and will be unlocked upon return.
2076  *
2077  * write_one_page() returns a negative error code if I/O failed.
2078  */
2079 int write_one_page(struct page *page, int wait)
2080 {
2081         struct address_space *mapping = page->mapping;
2082         int ret = 0;
2083         struct writeback_control wbc = {
2084                 .sync_mode = WB_SYNC_ALL,
2085                 .nr_to_write = 1,
2086         };
2087 
2088         BUG_ON(!PageLocked(page));
2089 
2090         if (wait)
2091                 wait_on_page_writeback(page);
2092 
2093         if (clear_page_dirty_for_io(page)) {
2094                 page_cache_get(page);
2095                 ret = mapping->a_ops->writepage(page, &wbc);
2096                 if (ret == 0 && wait) {
2097                         wait_on_page_writeback(page);
2098                         if (PageError(page))
2099                                 ret = -EIO;
2100                 }
2101                 page_cache_release(page);
2102         } else {
2103                 unlock_page(page);
2104         }
2105         return ret;
2106 }
2107 EXPORT_SYMBOL(write_one_page);
2108 
2109 /*
2110  * For address_spaces which do not use buffers nor write back.
2111  */
2112 int __set_page_dirty_no_writeback(struct page *page)
2113 {
2114         if (!PageDirty(page))
2115                 return !TestSetPageDirty(page);
2116         return 0;
2117 }
2118 
2119 /*
2120  * Helper function for set_page_dirty family.
2121  * NOTE: This relies on being atomic wrt interrupts.
2122  */
2123 void account_page_dirtied(struct page *page, struct address_space *mapping)
2124 {
2125         trace_writeback_dirty_page(page, mapping);
2126 
2127         if (mapping_cap_account_dirty(mapping)) {
2128                 __inc_zone_page_state(page, NR_FILE_DIRTY);
2129                 __inc_zone_page_state(page, NR_DIRTIED);
2130                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
2131                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2132                 task_io_account_write(PAGE_CACHE_SIZE);
2133                 current->nr_dirtied++;
2134                 this_cpu_inc(bdp_ratelimits);
2135         }
2136 }
2137 EXPORT_SYMBOL(account_page_dirtied);
2138 
2139 /*
2140  * Helper function for set_page_writeback family.
2141  *
2142  * The caller must hold mem_cgroup_begin/end_update_page_stat() lock
2143  * while calling this function.
2144  * See test_set_page_writeback for example.
2145  *
2146  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2147  * wrt interrupts.
2148  */
2149 void account_page_writeback(struct page *page)
2150 {
2151         mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2152         inc_zone_page_state(page, NR_WRITEBACK);
2153 }
2154 EXPORT_SYMBOL(account_page_writeback);
2155 
2156 /*
2157  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2158  * its radix tree.
2159  *
2160  * This is also used when a single buffer is being dirtied: we want to set the
2161  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2162  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2163  *
2164  * Most callers have locked the page, which pins the address_space in memory.
2165  * But zap_pte_range() does not lock the page, however in that case the
2166  * mapping is pinned by the vma's ->vm_file reference.
2167  *
2168  * We take care to handle the case where the page was truncated from the
2169  * mapping by re-checking page_mapping() inside tree_lock.
2170  */
2171 int __set_page_dirty_nobuffers(struct page *page)
2172 {
2173         if (!TestSetPageDirty(page)) {
2174                 struct address_space *mapping = page_mapping(page);
2175                 struct address_space *mapping2;
2176                 unsigned long flags;
2177 
2178                 if (!mapping)
2179                         return 1;
2180 
2181                 spin_lock_irqsave(&mapping->tree_lock, flags);
2182                 mapping2 = page_mapping(page);
2183                 if (mapping2) { /* Race with truncate? */
2184                         BUG_ON(mapping2 != mapping);
2185                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2186                         account_page_dirtied(page, mapping);
2187                         radix_tree_tag_set(&mapping->page_tree,
2188                                 page_index(page), PAGECACHE_TAG_DIRTY);
2189                 }
2190                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2191                 if (mapping->host) {
2192                         /* !PageAnon && !swapper_space */
2193                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2194                 }
2195                 return 1;
2196         }
2197         return 0;
2198 }
2199 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2200 
2201 /*
2202  * Call this whenever redirtying a page, to de-account the dirty counters
2203  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2204  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2205  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2206  * control.
2207  */
2208 void account_page_redirty(struct page *page)
2209 {
2210         struct address_space *mapping = page->mapping;
2211         if (mapping && mapping_cap_account_dirty(mapping)) {
2212                 current->nr_dirtied--;
2213                 dec_zone_page_state(page, NR_DIRTIED);
2214                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2215         }
2216 }
2217 EXPORT_SYMBOL(account_page_redirty);
2218 
2219 /*
2220  * When a writepage implementation decides that it doesn't want to write this
2221  * page for some reason, it should redirty the locked page via
2222  * redirty_page_for_writepage() and it should then unlock the page and return 0
2223  */
2224 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2225 {
2226         wbc->pages_skipped++;
2227         account_page_redirty(page);
2228         return __set_page_dirty_nobuffers(page);
2229 }
2230 EXPORT_SYMBOL(redirty_page_for_writepage);
2231 
2232 /*
2233  * Dirty a page.
2234  *
2235  * For pages with a mapping this should be done under the page lock
2236  * for the benefit of asynchronous memory errors who prefer a consistent
2237  * dirty state. This rule can be broken in some special cases,
2238  * but should be better not to.
2239  *
2240  * If the mapping doesn't provide a set_page_dirty a_op, then
2241  * just fall through and assume that it wants buffer_heads.
2242  */
2243 int set_page_dirty(struct page *page)
2244 {
2245         struct address_space *mapping = page_mapping(page);
2246 
2247         if (likely(mapping)) {
2248                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2249                 /*
2250                  * readahead/lru_deactivate_page could remain
2251                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2252                  * About readahead, if the page is written, the flags would be
2253                  * reset. So no problem.
2254                  * About lru_deactivate_page, if the page is redirty, the flag
2255                  * will be reset. So no problem. but if the page is used by readahead
2256                  * it will confuse readahead and make it restart the size rampup
2257                  * process. But it's a trivial problem.
2258                  */
2259                 ClearPageReclaim(page);
2260 #ifdef CONFIG_BLOCK
2261                 if (!spd)
2262                         spd = __set_page_dirty_buffers;
2263 #endif
2264                 return (*spd)(page);
2265         }
2266         if (!PageDirty(page)) {
2267                 if (!TestSetPageDirty(page))
2268                         return 1;
2269         }
2270         return 0;
2271 }
2272 EXPORT_SYMBOL(set_page_dirty);
2273 
2274 /*
2275  * set_page_dirty() is racy if the caller has no reference against
2276  * page->mapping->host, and if the page is unlocked.  This is because another
2277  * CPU could truncate the page off the mapping and then free the mapping.
2278  *
2279  * Usually, the page _is_ locked, or the caller is a user-space process which
2280  * holds a reference on the inode by having an open file.
2281  *
2282  * In other cases, the page should be locked before running set_page_dirty().
2283  */
2284 int set_page_dirty_lock(struct page *page)
2285 {
2286         int ret;
2287 
2288         lock_page(page);
2289         ret = set_page_dirty(page);
2290         unlock_page(page);
2291         return ret;
2292 }
2293 EXPORT_SYMBOL(set_page_dirty_lock);
2294 
2295 /*
2296  * Clear a page's dirty flag, while caring for dirty memory accounting.
2297  * Returns true if the page was previously dirty.
2298  *
2299  * This is for preparing to put the page under writeout.  We leave the page
2300  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2301  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2302  * implementation will run either set_page_writeback() or set_page_dirty(),
2303  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2304  * back into sync.
2305  *
2306  * This incoherency between the page's dirty flag and radix-tree tag is
2307  * unfortunate, but it only exists while the page is locked.
2308  */
2309 int clear_page_dirty_for_io(struct page *page)
2310 {
2311         struct address_space *mapping = page_mapping(page);
2312 
2313         BUG_ON(!PageLocked(page));
2314 
2315         if (mapping && mapping_cap_account_dirty(mapping)) {
2316                 /*
2317                  * Yes, Virginia, this is indeed insane.
2318                  *
2319                  * We use this sequence to make sure that
2320                  *  (a) we account for dirty stats properly
2321                  *  (b) we tell the low-level filesystem to
2322                  *      mark the whole page dirty if it was
2323                  *      dirty in a pagetable. Only to then
2324                  *  (c) clean the page again and return 1 to
2325                  *      cause the writeback.
2326                  *
2327                  * This way we avoid all nasty races with the
2328                  * dirty bit in multiple places and clearing
2329                  * them concurrently from different threads.
2330                  *
2331                  * Note! Normally the "set_page_dirty(page)"
2332                  * has no effect on the actual dirty bit - since
2333                  * that will already usually be set. But we
2334                  * need the side effects, and it can help us
2335                  * avoid races.
2336                  *
2337                  * We basically use the page "master dirty bit"
2338                  * as a serialization point for all the different
2339                  * threads doing their things.
2340                  */
2341                 if (page_mkclean(page))
2342                         set_page_dirty(page);
2343                 /*
2344                  * We carefully synchronise fault handlers against
2345                  * installing a dirty pte and marking the page dirty
2346                  * at this point. We do this by having them hold the
2347                  * page lock at some point after installing their
2348                  * pte, but before marking the page dirty.
2349                  * Pages are always locked coming in here, so we get
2350                  * the desired exclusion. See mm/memory.c:do_wp_page()
2351                  * for more comments.
2352                  */
2353                 if (TestClearPageDirty(page)) {
2354                         dec_zone_page_state(page, NR_FILE_DIRTY);
2355                         dec_bdi_stat(mapping->backing_dev_info,
2356                                         BDI_RECLAIMABLE);
2357                         return 1;
2358                 }
2359                 return 0;
2360         }
2361         return TestClearPageDirty(page);
2362 }
2363 EXPORT_SYMBOL(clear_page_dirty_for_io);
2364 
2365 int test_clear_page_writeback(struct page *page)
2366 {
2367         struct address_space *mapping = page_mapping(page);
2368         int ret;
2369         bool locked;
2370         unsigned long memcg_flags;
2371 
2372         mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2373         if (mapping) {
2374                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2375                 unsigned long flags;
2376 
2377                 spin_lock_irqsave(&mapping->tree_lock, flags);
2378                 ret = TestClearPageWriteback(page);
2379                 if (ret) {
2380                         radix_tree_tag_clear(&mapping->page_tree,
2381                                                 page_index(page),
2382                                                 PAGECACHE_TAG_WRITEBACK);
2383                         if (bdi_cap_account_writeback(bdi)) {
2384                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2385                                 __bdi_writeout_inc(bdi);
2386                         }
2387                 }
2388                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2389         } else {
2390                 ret = TestClearPageWriteback(page);
2391         }
2392         if (ret) {
2393                 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2394                 dec_zone_page_state(page, NR_WRITEBACK);
2395                 inc_zone_page_state(page, NR_WRITTEN);
2396         }
2397         mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2398         return ret;
2399 }
2400 
2401 int test_set_page_writeback(struct page *page)
2402 {
2403         struct address_space *mapping = page_mapping(page);
2404         int ret;
2405         bool locked;
2406         unsigned long memcg_flags;
2407 
2408         mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2409         if (mapping) {
2410                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2411                 unsigned long flags;
2412 
2413                 spin_lock_irqsave(&mapping->tree_lock, flags);
2414                 ret = TestSetPageWriteback(page);
2415                 if (!ret) {
2416                         radix_tree_tag_set(&mapping->page_tree,
2417                                                 page_index(page),
2418                                                 PAGECACHE_TAG_WRITEBACK);
2419                         if (bdi_cap_account_writeback(bdi))
2420                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2421                 }
2422                 if (!PageDirty(page))
2423                         radix_tree_tag_clear(&mapping->page_tree,
2424                                                 page_index(page),
2425                                                 PAGECACHE_TAG_DIRTY);
2426                 radix_tree_tag_clear(&mapping->page_tree,
2427                                      page_index(page),
2428                                      PAGECACHE_TAG_TOWRITE);
2429                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2430         } else {
2431                 ret = TestSetPageWriteback(page);
2432         }
2433         if (!ret)
2434                 account_page_writeback(page);
2435         mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2436         return ret;
2437 
2438 }
2439 EXPORT_SYMBOL(test_set_page_writeback);
2440 
2441 /*
2442  * Return true if any of the pages in the mapping are marked with the
2443  * passed tag.
2444  */
2445 int mapping_tagged(struct address_space *mapping, int tag)
2446 {
2447         return radix_tree_tagged(&mapping->page_tree, tag);
2448 }
2449 EXPORT_SYMBOL(mapping_tagged);
2450 
2451 /**
2452  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2453  * @page:       The page to wait on.
2454  *
2455  * This function determines if the given page is related to a backing device
2456  * that requires page contents to be held stable during writeback.  If so, then
2457  * it will wait for any pending writeback to complete.
2458  */
2459 void wait_for_stable_page(struct page *page)
2460 {
2461         struct address_space *mapping = page_mapping(page);
2462         struct backing_dev_info *bdi = mapping->backing_dev_info;
2463 
2464         if (!bdi_cap_stable_pages_required(bdi))
2465                 return;
2466 
2467         wait_on_page_writeback(page);
2468 }
2469 EXPORT_SYMBOL_GPL(wait_for_stable_page);
2470 

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