Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris...
[linux-3.10.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <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 <trace/events/writeback.h>
39
40 /*
41  * Sleep at most 200ms at a time in balance_dirty_pages().
42  */
43 #define MAX_PAUSE               max(HZ/5, 1)
44
45 /*
46  * Try to keep balance_dirty_pages() call intervals higher than this many pages
47  * by raising pause time to max_pause when falls below it.
48  */
49 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
50
51 /*
52  * Estimate write bandwidth at 200ms intervals.
53  */
54 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
55
56 #define RATELIMIT_CALC_SHIFT    10
57
58 /*
59  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
60  * will look to see if it needs to force writeback or throttling.
61  */
62 static long ratelimit_pages = 32;
63
64 /* The following parameters are exported via /proc/sys/vm */
65
66 /*
67  * Start background writeback (via writeback threads) at this percentage
68  */
69 int dirty_background_ratio = 10;
70
71 /*
72  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
73  * dirty_background_ratio * the amount of dirtyable memory
74  */
75 unsigned long dirty_background_bytes;
76
77 /*
78  * free highmem will not be subtracted from the total free memory
79  * for calculating free ratios if vm_highmem_is_dirtyable is true
80  */
81 int vm_highmem_is_dirtyable;
82
83 /*
84  * The generator of dirty data starts writeback at this percentage
85  */
86 int vm_dirty_ratio = 20;
87
88 /*
89  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
90  * vm_dirty_ratio * the amount of dirtyable memory
91  */
92 unsigned long vm_dirty_bytes;
93
94 /*
95  * The interval between `kupdate'-style writebacks
96  */
97 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
98
99 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
100
101 /*
102  * The longest time for which data is allowed to remain dirty
103  */
104 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
105
106 /*
107  * Flag that makes the machine dump writes/reads and block dirtyings.
108  */
109 int block_dump;
110
111 /*
112  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113  * a full sync is triggered after this time elapses without any disk activity.
114  */
115 int laptop_mode;
116
117 EXPORT_SYMBOL(laptop_mode);
118
119 /* End of sysctl-exported parameters */
120
121 unsigned long global_dirty_limit;
122
123 /*
124  * Scale the writeback cache size proportional to the relative writeout speeds.
125  *
126  * We do this by keeping a floating proportion between BDIs, based on page
127  * writeback completions [end_page_writeback()]. Those devices that write out
128  * pages fastest will get the larger share, while the slower will get a smaller
129  * share.
130  *
131  * We use page writeout completions because we are interested in getting rid of
132  * dirty pages. Having them written out is the primary goal.
133  *
134  * We introduce a concept of time, a period over which we measure these events,
135  * because demand can/will vary over time. The length of this period itself is
136  * measured in page writeback completions.
137  *
138  */
139 static struct fprop_global writeout_completions;
140
141 static void writeout_period(unsigned long t);
142 /* Timer for aging of writeout_completions */
143 static struct timer_list writeout_period_timer =
144                 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
145 static unsigned long writeout_period_time = 0;
146
147 /*
148  * Length of period for aging writeout fractions of bdis. This is an
149  * arbitrarily chosen number. The longer the period, the slower fractions will
150  * reflect changes in current writeout rate.
151  */
152 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
153
154 /*
155  * Work out the current dirty-memory clamping and background writeout
156  * thresholds.
157  *
158  * The main aim here is to lower them aggressively if there is a lot of mapped
159  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
160  * pages.  It is better to clamp down on writers than to start swapping, and
161  * performing lots of scanning.
162  *
163  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
164  *
165  * We don't permit the clamping level to fall below 5% - that is getting rather
166  * excessive.
167  *
168  * We make sure that the background writeout level is below the adjusted
169  * clamping level.
170  */
171
172 /*
173  * In a memory zone, there is a certain amount of pages we consider
174  * available for the page cache, which is essentially the number of
175  * free and reclaimable pages, minus some zone reserves to protect
176  * lowmem and the ability to uphold the zone's watermarks without
177  * requiring writeback.
178  *
179  * This number of dirtyable pages is the base value of which the
180  * user-configurable dirty ratio is the effictive number of pages that
181  * are allowed to be actually dirtied.  Per individual zone, or
182  * globally by using the sum of dirtyable pages over all zones.
183  *
184  * Because the user is allowed to specify the dirty limit globally as
185  * absolute number of bytes, calculating the per-zone dirty limit can
186  * require translating the configured limit into a percentage of
187  * global dirtyable memory first.
188  */
189
190 static unsigned long highmem_dirtyable_memory(unsigned long total)
191 {
192 #ifdef CONFIG_HIGHMEM
193         int node;
194         unsigned long x = 0;
195
196         for_each_node_state(node, N_HIGH_MEMORY) {
197                 struct zone *z =
198                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
199
200                 x += zone_page_state(z, NR_FREE_PAGES) +
201                      zone_reclaimable_pages(z) - z->dirty_balance_reserve;
202         }
203         /*
204          * Unreclaimable memory (kernel memory or anonymous memory
205          * without swap) can bring down the dirtyable pages below
206          * the zone's dirty balance reserve and the above calculation
207          * will underflow.  However we still want to add in nodes
208          * which are below threshold (negative values) to get a more
209          * accurate calculation but make sure that the total never
210          * underflows.
211          */
212         if ((long)x < 0)
213                 x = 0;
214
215         /*
216          * Make sure that the number of highmem pages is never larger
217          * than the number of the total dirtyable memory. This can only
218          * occur in very strange VM situations but we want to make sure
219          * that this does not occur.
220          */
221         return min(x, total);
222 #else
223         return 0;
224 #endif
225 }
226
227 /**
228  * global_dirtyable_memory - number of globally dirtyable pages
229  *
230  * Returns the global number of pages potentially available for dirty
231  * page cache.  This is the base value for the global dirty limits.
232  */
233 static unsigned long global_dirtyable_memory(void)
234 {
235         unsigned long x;
236
237         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
238         x -= min(x, dirty_balance_reserve);
239
240         if (!vm_highmem_is_dirtyable)
241                 x -= highmem_dirtyable_memory(x);
242
243         return x + 1;   /* Ensure that we never return 0 */
244 }
245
246 /*
247  * global_dirty_limits - background-writeback and dirty-throttling thresholds
248  *
249  * Calculate the dirty thresholds based on sysctl parameters
250  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
251  * - vm.dirty_ratio             or  vm.dirty_bytes
252  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
253  * real-time tasks.
254  */
255 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
256 {
257         unsigned long background;
258         unsigned long dirty;
259         unsigned long uninitialized_var(available_memory);
260         struct task_struct *tsk;
261
262         if (!vm_dirty_bytes || !dirty_background_bytes)
263                 available_memory = global_dirtyable_memory();
264
265         if (vm_dirty_bytes)
266                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
267         else
268                 dirty = (vm_dirty_ratio * available_memory) / 100;
269
270         if (dirty_background_bytes)
271                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
272         else
273                 background = (dirty_background_ratio * available_memory) / 100;
274
275         if (background >= dirty)
276                 background = dirty / 2;
277         tsk = current;
278         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
279                 background += background / 4;
280                 dirty += dirty / 4;
281         }
282         *pbackground = background;
283         *pdirty = dirty;
284         trace_global_dirty_state(background, dirty);
285 }
286
287 /**
288  * zone_dirtyable_memory - number of dirtyable pages in a zone
289  * @zone: the zone
290  *
291  * Returns the zone's number of pages potentially available for dirty
292  * page cache.  This is the base value for the per-zone dirty limits.
293  */
294 static unsigned long zone_dirtyable_memory(struct zone *zone)
295 {
296         /*
297          * The effective global number of dirtyable pages may exclude
298          * highmem as a big-picture measure to keep the ratio between
299          * dirty memory and lowmem reasonable.
300          *
301          * But this function is purely about the individual zone and a
302          * highmem zone can hold its share of dirty pages, so we don't
303          * care about vm_highmem_is_dirtyable here.
304          */
305         unsigned long nr_pages = zone_page_state(zone, NR_FREE_PAGES) +
306                 zone_reclaimable_pages(zone);
307
308         /* don't allow this to underflow */
309         nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
310         return nr_pages;
311 }
312
313 /**
314  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
315  * @zone: the zone
316  *
317  * Returns the maximum number of dirty pages allowed in a zone, based
318  * on the zone's dirtyable memory.
319  */
320 static unsigned long zone_dirty_limit(struct zone *zone)
321 {
322         unsigned long zone_memory = zone_dirtyable_memory(zone);
323         struct task_struct *tsk = current;
324         unsigned long dirty;
325
326         if (vm_dirty_bytes)
327                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
328                         zone_memory / global_dirtyable_memory();
329         else
330                 dirty = vm_dirty_ratio * zone_memory / 100;
331
332         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
333                 dirty += dirty / 4;
334
335         return dirty;
336 }
337
338 /**
339  * zone_dirty_ok - tells whether a zone is within its dirty limits
340  * @zone: the zone to check
341  *
342  * Returns %true when the dirty pages in @zone are within the zone's
343  * dirty limit, %false if the limit is exceeded.
344  */
345 bool zone_dirty_ok(struct zone *zone)
346 {
347         unsigned long limit = zone_dirty_limit(zone);
348
349         return zone_page_state(zone, NR_FILE_DIRTY) +
350                zone_page_state(zone, NR_UNSTABLE_NFS) +
351                zone_page_state(zone, NR_WRITEBACK) <= limit;
352 }
353
354 int dirty_background_ratio_handler(struct ctl_table *table, int write,
355                 void __user *buffer, size_t *lenp,
356                 loff_t *ppos)
357 {
358         int ret;
359
360         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
361         if (ret == 0 && write)
362                 dirty_background_bytes = 0;
363         return ret;
364 }
365
366 int dirty_background_bytes_handler(struct ctl_table *table, int write,
367                 void __user *buffer, size_t *lenp,
368                 loff_t *ppos)
369 {
370         int ret;
371
372         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
373         if (ret == 0 && write)
374                 dirty_background_ratio = 0;
375         return ret;
376 }
377
378 int dirty_ratio_handler(struct ctl_table *table, int write,
379                 void __user *buffer, size_t *lenp,
380                 loff_t *ppos)
381 {
382         int old_ratio = vm_dirty_ratio;
383         int ret;
384
385         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
386         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
387                 writeback_set_ratelimit();
388                 vm_dirty_bytes = 0;
389         }
390         return ret;
391 }
392
393 int dirty_bytes_handler(struct ctl_table *table, int write,
394                 void __user *buffer, size_t *lenp,
395                 loff_t *ppos)
396 {
397         unsigned long old_bytes = vm_dirty_bytes;
398         int ret;
399
400         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
401         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
402                 writeback_set_ratelimit();
403                 vm_dirty_ratio = 0;
404         }
405         return ret;
406 }
407
408 static unsigned long wp_next_time(unsigned long cur_time)
409 {
410         cur_time += VM_COMPLETIONS_PERIOD_LEN;
411         /* 0 has a special meaning... */
412         if (!cur_time)
413                 return 1;
414         return cur_time;
415 }
416
417 /*
418  * Increment the BDI's writeout completion count and the global writeout
419  * completion count. Called from test_clear_page_writeback().
420  */
421 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
422 {
423         __inc_bdi_stat(bdi, BDI_WRITTEN);
424         __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
425                                bdi->max_prop_frac);
426         /* First event after period switching was turned off? */
427         if (!unlikely(writeout_period_time)) {
428                 /*
429                  * We can race with other __bdi_writeout_inc calls here but
430                  * it does not cause any harm since the resulting time when
431                  * timer will fire and what is in writeout_period_time will be
432                  * roughly the same.
433                  */
434                 writeout_period_time = wp_next_time(jiffies);
435                 mod_timer(&writeout_period_timer, writeout_period_time);
436         }
437 }
438
439 void bdi_writeout_inc(struct backing_dev_info *bdi)
440 {
441         unsigned long flags;
442
443         local_irq_save(flags);
444         __bdi_writeout_inc(bdi);
445         local_irq_restore(flags);
446 }
447 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
448
449 /*
450  * Obtain an accurate fraction of the BDI's portion.
451  */
452 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
453                 long *numerator, long *denominator)
454 {
455         fprop_fraction_percpu(&writeout_completions, &bdi->completions,
456                                 numerator, denominator);
457 }
458
459 /*
460  * On idle system, we can be called long after we scheduled because we use
461  * deferred timers so count with missed periods.
462  */
463 static void writeout_period(unsigned long t)
464 {
465         int miss_periods = (jiffies - writeout_period_time) /
466                                                  VM_COMPLETIONS_PERIOD_LEN;
467
468         if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
469                 writeout_period_time = wp_next_time(writeout_period_time +
470                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
471                 mod_timer(&writeout_period_timer, writeout_period_time);
472         } else {
473                 /*
474                  * Aging has zeroed all fractions. Stop wasting CPU on period
475                  * updates.
476                  */
477                 writeout_period_time = 0;
478         }
479 }
480
481 /*
482  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
483  * registered backing devices, which, for obvious reasons, can not
484  * exceed 100%.
485  */
486 static unsigned int bdi_min_ratio;
487
488 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
489 {
490         int ret = 0;
491
492         spin_lock_bh(&bdi_lock);
493         if (min_ratio > bdi->max_ratio) {
494                 ret = -EINVAL;
495         } else {
496                 min_ratio -= bdi->min_ratio;
497                 if (bdi_min_ratio + min_ratio < 100) {
498                         bdi_min_ratio += min_ratio;
499                         bdi->min_ratio += min_ratio;
500                 } else {
501                         ret = -EINVAL;
502                 }
503         }
504         spin_unlock_bh(&bdi_lock);
505
506         return ret;
507 }
508
509 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
510 {
511         int ret = 0;
512
513         if (max_ratio > 100)
514                 return -EINVAL;
515
516         spin_lock_bh(&bdi_lock);
517         if (bdi->min_ratio > max_ratio) {
518                 ret = -EINVAL;
519         } else {
520                 bdi->max_ratio = max_ratio;
521                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
522         }
523         spin_unlock_bh(&bdi_lock);
524
525         return ret;
526 }
527 EXPORT_SYMBOL(bdi_set_max_ratio);
528
529 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
530                                            unsigned long bg_thresh)
531 {
532         return (thresh + bg_thresh) / 2;
533 }
534
535 static unsigned long hard_dirty_limit(unsigned long thresh)
536 {
537         return max(thresh, global_dirty_limit);
538 }
539
540 /**
541  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
542  * @bdi: the backing_dev_info to query
543  * @dirty: global dirty limit in pages
544  *
545  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
546  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
547  *
548  * Note that balance_dirty_pages() will only seriously take it as a hard limit
549  * when sleeping max_pause per page is not enough to keep the dirty pages under
550  * control. For example, when the device is completely stalled due to some error
551  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
552  * In the other normal situations, it acts more gently by throttling the tasks
553  * more (rather than completely block them) when the bdi dirty pages go high.
554  *
555  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
556  * - starving fast devices
557  * - piling up dirty pages (that will take long time to sync) on slow devices
558  *
559  * The bdi's share of dirty limit will be adapting to its throughput and
560  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
561  */
562 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
563 {
564         u64 bdi_dirty;
565         long numerator, denominator;
566
567         /*
568          * Calculate this BDI's share of the dirty ratio.
569          */
570         bdi_writeout_fraction(bdi, &numerator, &denominator);
571
572         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
573         bdi_dirty *= numerator;
574         do_div(bdi_dirty, denominator);
575
576         bdi_dirty += (dirty * bdi->min_ratio) / 100;
577         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
578                 bdi_dirty = dirty * bdi->max_ratio / 100;
579
580         return bdi_dirty;
581 }
582
583 /*
584  * Dirty position control.
585  *
586  * (o) global/bdi setpoints
587  *
588  * We want the dirty pages be balanced around the global/bdi setpoints.
589  * When the number of dirty pages is higher/lower than the setpoint, the
590  * dirty position control ratio (and hence task dirty ratelimit) will be
591  * decreased/increased to bring the dirty pages back to the setpoint.
592  *
593  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
594  *
595  *     if (dirty < setpoint) scale up   pos_ratio
596  *     if (dirty > setpoint) scale down pos_ratio
597  *
598  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
599  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
600  *
601  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
602  *
603  * (o) global control line
604  *
605  *     ^ pos_ratio
606  *     |
607  *     |            |<===== global dirty control scope ======>|
608  * 2.0 .............*
609  *     |            .*
610  *     |            . *
611  *     |            .   *
612  *     |            .     *
613  *     |            .        *
614  *     |            .            *
615  * 1.0 ................................*
616  *     |            .                  .     *
617  *     |            .                  .          *
618  *     |            .                  .              *
619  *     |            .                  .                 *
620  *     |            .                  .                    *
621  *   0 +------------.------------------.----------------------*------------->
622  *           freerun^          setpoint^                 limit^   dirty pages
623  *
624  * (o) bdi control line
625  *
626  *     ^ pos_ratio
627  *     |
628  *     |            *
629  *     |              *
630  *     |                *
631  *     |                  *
632  *     |                    * |<=========== span ============>|
633  * 1.0 .......................*
634  *     |                      . *
635  *     |                      .   *
636  *     |                      .     *
637  *     |                      .       *
638  *     |                      .         *
639  *     |                      .           *
640  *     |                      .             *
641  *     |                      .               *
642  *     |                      .                 *
643  *     |                      .                   *
644  *     |                      .                     *
645  * 1/4 ...............................................* * * * * * * * * * * *
646  *     |                      .                         .
647  *     |                      .                           .
648  *     |                      .                             .
649  *   0 +----------------------.-------------------------------.------------->
650  *                bdi_setpoint^                    x_intercept^
651  *
652  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
653  * be smoothly throttled down to normal if it starts high in situations like
654  * - start writing to a slow SD card and a fast disk at the same time. The SD
655  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
656  * - the bdi dirty thresh drops quickly due to change of JBOD workload
657  */
658 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
659                                         unsigned long thresh,
660                                         unsigned long bg_thresh,
661                                         unsigned long dirty,
662                                         unsigned long bdi_thresh,
663                                         unsigned long bdi_dirty)
664 {
665         unsigned long write_bw = bdi->avg_write_bandwidth;
666         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
667         unsigned long limit = hard_dirty_limit(thresh);
668         unsigned long x_intercept;
669         unsigned long setpoint;         /* dirty pages' target balance point */
670         unsigned long bdi_setpoint;
671         unsigned long span;
672         long long pos_ratio;            /* for scaling up/down the rate limit */
673         long x;
674
675         if (unlikely(dirty >= limit))
676                 return 0;
677
678         /*
679          * global setpoint
680          *
681          *                           setpoint - dirty 3
682          *        f(dirty) := 1.0 + (----------------)
683          *                           limit - setpoint
684          *
685          * it's a 3rd order polynomial that subjects to
686          *
687          * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
688          * (2) f(setpoint) = 1.0 => the balance point
689          * (3) f(limit)    = 0   => the hard limit
690          * (4) df/dx      <= 0   => negative feedback control
691          * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
692          *     => fast response on large errors; small oscillation near setpoint
693          */
694         setpoint = (freerun + limit) / 2;
695         x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
696                     limit - setpoint + 1);
697         pos_ratio = x;
698         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
699         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
700         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
701
702         /*
703          * We have computed basic pos_ratio above based on global situation. If
704          * the bdi is over/under its share of dirty pages, we want to scale
705          * pos_ratio further down/up. That is done by the following mechanism.
706          */
707
708         /*
709          * bdi setpoint
710          *
711          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
712          *
713          *                        x_intercept - bdi_dirty
714          *                     := --------------------------
715          *                        x_intercept - bdi_setpoint
716          *
717          * The main bdi control line is a linear function that subjects to
718          *
719          * (1) f(bdi_setpoint) = 1.0
720          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
721          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
722          *
723          * For single bdi case, the dirty pages are observed to fluctuate
724          * regularly within range
725          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
726          * for various filesystems, where (2) can yield in a reasonable 12.5%
727          * fluctuation range for pos_ratio.
728          *
729          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
730          * own size, so move the slope over accordingly and choose a slope that
731          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
732          */
733         if (unlikely(bdi_thresh > thresh))
734                 bdi_thresh = thresh;
735         /*
736          * It's very possible that bdi_thresh is close to 0 not because the
737          * device is slow, but that it has remained inactive for long time.
738          * Honour such devices a reasonable good (hopefully IO efficient)
739          * threshold, so that the occasional writes won't be blocked and active
740          * writes can rampup the threshold quickly.
741          */
742         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
743         /*
744          * scale global setpoint to bdi's:
745          *      bdi_setpoint = setpoint * bdi_thresh / thresh
746          */
747         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
748         bdi_setpoint = setpoint * (u64)x >> 16;
749         /*
750          * Use span=(8*write_bw) in single bdi case as indicated by
751          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
752          *
753          *        bdi_thresh                    thresh - bdi_thresh
754          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
755          *          thresh                            thresh
756          */
757         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
758         x_intercept = bdi_setpoint + span;
759
760         if (bdi_dirty < x_intercept - span / 4) {
761                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
762                                     x_intercept - bdi_setpoint + 1);
763         } else
764                 pos_ratio /= 4;
765
766         /*
767          * bdi reserve area, safeguard against dirty pool underrun and disk idle
768          * It may push the desired control point of global dirty pages higher
769          * than setpoint.
770          */
771         x_intercept = bdi_thresh / 2;
772         if (bdi_dirty < x_intercept) {
773                 if (bdi_dirty > x_intercept / 8)
774                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
775                 else
776                         pos_ratio *= 8;
777         }
778
779         return pos_ratio;
780 }
781
782 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
783                                        unsigned long elapsed,
784                                        unsigned long written)
785 {
786         const unsigned long period = roundup_pow_of_two(3 * HZ);
787         unsigned long avg = bdi->avg_write_bandwidth;
788         unsigned long old = bdi->write_bandwidth;
789         u64 bw;
790
791         /*
792          * bw = written * HZ / elapsed
793          *
794          *                   bw * elapsed + write_bandwidth * (period - elapsed)
795          * write_bandwidth = ---------------------------------------------------
796          *                                          period
797          */
798         bw = written - bdi->written_stamp;
799         bw *= HZ;
800         if (unlikely(elapsed > period)) {
801                 do_div(bw, elapsed);
802                 avg = bw;
803                 goto out;
804         }
805         bw += (u64)bdi->write_bandwidth * (period - elapsed);
806         bw >>= ilog2(period);
807
808         /*
809          * one more level of smoothing, for filtering out sudden spikes
810          */
811         if (avg > old && old >= (unsigned long)bw)
812                 avg -= (avg - old) >> 3;
813
814         if (avg < old && old <= (unsigned long)bw)
815                 avg += (old - avg) >> 3;
816
817 out:
818         bdi->write_bandwidth = bw;
819         bdi->avg_write_bandwidth = avg;
820 }
821
822 /*
823  * The global dirtyable memory and dirty threshold could be suddenly knocked
824  * down by a large amount (eg. on the startup of KVM in a swapless system).
825  * This may throw the system into deep dirty exceeded state and throttle
826  * heavy/light dirtiers alike. To retain good responsiveness, maintain
827  * global_dirty_limit for tracking slowly down to the knocked down dirty
828  * threshold.
829  */
830 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
831 {
832         unsigned long limit = global_dirty_limit;
833
834         /*
835          * Follow up in one step.
836          */
837         if (limit < thresh) {
838                 limit = thresh;
839                 goto update;
840         }
841
842         /*
843          * Follow down slowly. Use the higher one as the target, because thresh
844          * may drop below dirty. This is exactly the reason to introduce
845          * global_dirty_limit which is guaranteed to lie above the dirty pages.
846          */
847         thresh = max(thresh, dirty);
848         if (limit > thresh) {
849                 limit -= (limit - thresh) >> 5;
850                 goto update;
851         }
852         return;
853 update:
854         global_dirty_limit = limit;
855 }
856
857 static void global_update_bandwidth(unsigned long thresh,
858                                     unsigned long dirty,
859                                     unsigned long now)
860 {
861         static DEFINE_SPINLOCK(dirty_lock);
862         static unsigned long update_time;
863
864         /*
865          * check locklessly first to optimize away locking for the most time
866          */
867         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
868                 return;
869
870         spin_lock(&dirty_lock);
871         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
872                 update_dirty_limit(thresh, dirty);
873                 update_time = now;
874         }
875         spin_unlock(&dirty_lock);
876 }
877
878 /*
879  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
880  *
881  * Normal bdi tasks will be curbed at or below it in long term.
882  * Obviously it should be around (write_bw / N) when there are N dd tasks.
883  */
884 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
885                                        unsigned long thresh,
886                                        unsigned long bg_thresh,
887                                        unsigned long dirty,
888                                        unsigned long bdi_thresh,
889                                        unsigned long bdi_dirty,
890                                        unsigned long dirtied,
891                                        unsigned long elapsed)
892 {
893         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
894         unsigned long limit = hard_dirty_limit(thresh);
895         unsigned long setpoint = (freerun + limit) / 2;
896         unsigned long write_bw = bdi->avg_write_bandwidth;
897         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
898         unsigned long dirty_rate;
899         unsigned long task_ratelimit;
900         unsigned long balanced_dirty_ratelimit;
901         unsigned long pos_ratio;
902         unsigned long step;
903         unsigned long x;
904
905         /*
906          * The dirty rate will match the writeout rate in long term, except
907          * when dirty pages are truncated by userspace or re-dirtied by FS.
908          */
909         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
910
911         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
912                                        bdi_thresh, bdi_dirty);
913         /*
914          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
915          */
916         task_ratelimit = (u64)dirty_ratelimit *
917                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
918         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
919
920         /*
921          * A linear estimation of the "balanced" throttle rate. The theory is,
922          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
923          * dirty_rate will be measured to be (N * task_ratelimit). So the below
924          * formula will yield the balanced rate limit (write_bw / N).
925          *
926          * Note that the expanded form is not a pure rate feedback:
927          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
928          * but also takes pos_ratio into account:
929          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
930          *
931          * (1) is not realistic because pos_ratio also takes part in balancing
932          * the dirty rate.  Consider the state
933          *      pos_ratio = 0.5                                              (3)
934          *      rate = 2 * (write_bw / N)                                    (4)
935          * If (1) is used, it will stuck in that state! Because each dd will
936          * be throttled at
937          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
938          * yielding
939          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
940          * put (6) into (1) we get
941          *      rate_(i+1) = rate_(i)                                        (7)
942          *
943          * So we end up using (2) to always keep
944          *      rate_(i+1) ~= (write_bw / N)                                 (8)
945          * regardless of the value of pos_ratio. As long as (8) is satisfied,
946          * pos_ratio is able to drive itself to 1.0, which is not only where
947          * the dirty count meet the setpoint, but also where the slope of
948          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
949          */
950         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
951                                            dirty_rate | 1);
952         /*
953          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
954          */
955         if (unlikely(balanced_dirty_ratelimit > write_bw))
956                 balanced_dirty_ratelimit = write_bw;
957
958         /*
959          * We could safely do this and return immediately:
960          *
961          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
962          *
963          * However to get a more stable dirty_ratelimit, the below elaborated
964          * code makes use of task_ratelimit to filter out singular points and
965          * limit the step size.
966          *
967          * The below code essentially only uses the relative value of
968          *
969          *      task_ratelimit - dirty_ratelimit
970          *      = (pos_ratio - 1) * dirty_ratelimit
971          *
972          * which reflects the direction and size of dirty position error.
973          */
974
975         /*
976          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
977          * task_ratelimit is on the same side of dirty_ratelimit, too.
978          * For example, when
979          * - dirty_ratelimit > balanced_dirty_ratelimit
980          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
981          * lowering dirty_ratelimit will help meet both the position and rate
982          * control targets. Otherwise, don't update dirty_ratelimit if it will
983          * only help meet the rate target. After all, what the users ultimately
984          * feel and care are stable dirty rate and small position error.
985          *
986          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
987          * and filter out the singular points of balanced_dirty_ratelimit. Which
988          * keeps jumping around randomly and can even leap far away at times
989          * due to the small 200ms estimation period of dirty_rate (we want to
990          * keep that period small to reduce time lags).
991          */
992         step = 0;
993         if (dirty < setpoint) {
994                 x = min(bdi->balanced_dirty_ratelimit,
995                          min(balanced_dirty_ratelimit, task_ratelimit));
996                 if (dirty_ratelimit < x)
997                         step = x - dirty_ratelimit;
998         } else {
999                 x = max(bdi->balanced_dirty_ratelimit,
1000                          max(balanced_dirty_ratelimit, task_ratelimit));
1001                 if (dirty_ratelimit > x)
1002                         step = dirty_ratelimit - x;
1003         }
1004
1005         /*
1006          * Don't pursue 100% rate matching. It's impossible since the balanced
1007          * rate itself is constantly fluctuating. So decrease the track speed
1008          * when it gets close to the target. Helps eliminate pointless tremors.
1009          */
1010         step >>= dirty_ratelimit / (2 * step + 1);
1011         /*
1012          * Limit the tracking speed to avoid overshooting.
1013          */
1014         step = (step + 7) / 8;
1015
1016         if (dirty_ratelimit < balanced_dirty_ratelimit)
1017                 dirty_ratelimit += step;
1018         else
1019                 dirty_ratelimit -= step;
1020
1021         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1022         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1023
1024         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1025 }
1026
1027 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1028                             unsigned long thresh,
1029                             unsigned long bg_thresh,
1030                             unsigned long dirty,
1031                             unsigned long bdi_thresh,
1032                             unsigned long bdi_dirty,
1033                             unsigned long start_time)
1034 {
1035         unsigned long now = jiffies;
1036         unsigned long elapsed = now - bdi->bw_time_stamp;
1037         unsigned long dirtied;
1038         unsigned long written;
1039
1040         /*
1041          * rate-limit, only update once every 200ms.
1042          */
1043         if (elapsed < BANDWIDTH_INTERVAL)
1044                 return;
1045
1046         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1047         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1048
1049         /*
1050          * Skip quiet periods when disk bandwidth is under-utilized.
1051          * (at least 1s idle time between two flusher runs)
1052          */
1053         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1054                 goto snapshot;
1055
1056         if (thresh) {
1057                 global_update_bandwidth(thresh, dirty, now);
1058                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1059                                            bdi_thresh, bdi_dirty,
1060                                            dirtied, elapsed);
1061         }
1062         bdi_update_write_bandwidth(bdi, elapsed, written);
1063
1064 snapshot:
1065         bdi->dirtied_stamp = dirtied;
1066         bdi->written_stamp = written;
1067         bdi->bw_time_stamp = now;
1068 }
1069
1070 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1071                                  unsigned long thresh,
1072                                  unsigned long bg_thresh,
1073                                  unsigned long dirty,
1074                                  unsigned long bdi_thresh,
1075                                  unsigned long bdi_dirty,
1076                                  unsigned long start_time)
1077 {
1078         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1079                 return;
1080         spin_lock(&bdi->wb.list_lock);
1081         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1082                                bdi_thresh, bdi_dirty, start_time);
1083         spin_unlock(&bdi->wb.list_lock);
1084 }
1085
1086 /*
1087  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1088  * will look to see if it needs to start dirty throttling.
1089  *
1090  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1091  * global_page_state() too often. So scale it near-sqrt to the safety margin
1092  * (the number of pages we may dirty without exceeding the dirty limits).
1093  */
1094 static unsigned long dirty_poll_interval(unsigned long dirty,
1095                                          unsigned long thresh)
1096 {
1097         if (thresh > dirty)
1098                 return 1UL << (ilog2(thresh - dirty) >> 1);
1099
1100         return 1;
1101 }
1102
1103 static long bdi_max_pause(struct backing_dev_info *bdi,
1104                           unsigned long bdi_dirty)
1105 {
1106         long bw = bdi->avg_write_bandwidth;
1107         long t;
1108
1109         /*
1110          * Limit pause time for small memory systems. If sleeping for too long
1111          * time, a small pool of dirty/writeback pages may go empty and disk go
1112          * idle.
1113          *
1114          * 8 serves as the safety ratio.
1115          */
1116         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1117         t++;
1118
1119         return min_t(long, t, MAX_PAUSE);
1120 }
1121
1122 static long bdi_min_pause(struct backing_dev_info *bdi,
1123                           long max_pause,
1124                           unsigned long task_ratelimit,
1125                           unsigned long dirty_ratelimit,
1126                           int *nr_dirtied_pause)
1127 {
1128         long hi = ilog2(bdi->avg_write_bandwidth);
1129         long lo = ilog2(bdi->dirty_ratelimit);
1130         long t;         /* target pause */
1131         long pause;     /* estimated next pause */
1132         int pages;      /* target nr_dirtied_pause */
1133
1134         /* target for 10ms pause on 1-dd case */
1135         t = max(1, HZ / 100);
1136
1137         /*
1138          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1139          * overheads.
1140          *
1141          * (N * 10ms) on 2^N concurrent tasks.
1142          */
1143         if (hi > lo)
1144                 t += (hi - lo) * (10 * HZ) / 1024;
1145
1146         /*
1147          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1148          * on the much more stable dirty_ratelimit. However the next pause time
1149          * will be computed based on task_ratelimit and the two rate limits may
1150          * depart considerably at some time. Especially if task_ratelimit goes
1151          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1152          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1153          * result task_ratelimit won't be executed faithfully, which could
1154          * eventually bring down dirty_ratelimit.
1155          *
1156          * We apply two rules to fix it up:
1157          * 1) try to estimate the next pause time and if necessary, use a lower
1158          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1159          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1160          * 2) limit the target pause time to max_pause/2, so that the normal
1161          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1162          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1163          */
1164         t = min(t, 1 + max_pause / 2);
1165         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1166
1167         /*
1168          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1169          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1170          * When the 16 consecutive reads are often interrupted by some dirty
1171          * throttling pause during the async writes, cfq will go into idles
1172          * (deadline is fine). So push nr_dirtied_pause as high as possible
1173          * until reaches DIRTY_POLL_THRESH=32 pages.
1174          */
1175         if (pages < DIRTY_POLL_THRESH) {
1176                 t = max_pause;
1177                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1178                 if (pages > DIRTY_POLL_THRESH) {
1179                         pages = DIRTY_POLL_THRESH;
1180                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1181                 }
1182         }
1183
1184         pause = HZ * pages / (task_ratelimit + 1);
1185         if (pause > max_pause) {
1186                 t = max_pause;
1187                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1188         }
1189
1190         *nr_dirtied_pause = pages;
1191         /*
1192          * The minimal pause time will normally be half the target pause time.
1193          */
1194         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1195 }
1196
1197 /*
1198  * balance_dirty_pages() must be called by processes which are generating dirty
1199  * data.  It looks at the number of dirty pages in the machine and will force
1200  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1201  * If we're over `background_thresh' then the writeback threads are woken to
1202  * perform some writeout.
1203  */
1204 static void balance_dirty_pages(struct address_space *mapping,
1205                                 unsigned long pages_dirtied)
1206 {
1207         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1208         unsigned long bdi_reclaimable;
1209         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1210         unsigned long bdi_dirty;
1211         unsigned long freerun;
1212         unsigned long background_thresh;
1213         unsigned long dirty_thresh;
1214         unsigned long bdi_thresh;
1215         long period;
1216         long pause;
1217         long max_pause;
1218         long min_pause;
1219         int nr_dirtied_pause;
1220         bool dirty_exceeded = false;
1221         unsigned long task_ratelimit;
1222         unsigned long dirty_ratelimit;
1223         unsigned long pos_ratio;
1224         struct backing_dev_info *bdi = mapping->backing_dev_info;
1225         unsigned long start_time = jiffies;
1226
1227         for (;;) {
1228                 unsigned long now = jiffies;
1229
1230                 /*
1231                  * Unstable writes are a feature of certain networked
1232                  * filesystems (i.e. NFS) in which data may have been
1233                  * written to the server's write cache, but has not yet
1234                  * been flushed to permanent storage.
1235                  */
1236                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1237                                         global_page_state(NR_UNSTABLE_NFS);
1238                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1239
1240                 global_dirty_limits(&background_thresh, &dirty_thresh);
1241
1242                 /*
1243                  * Throttle it only when the background writeback cannot
1244                  * catch-up. This avoids (excessively) small writeouts
1245                  * when the bdi limits are ramping up.
1246                  */
1247                 freerun = dirty_freerun_ceiling(dirty_thresh,
1248                                                 background_thresh);
1249                 if (nr_dirty <= freerun) {
1250                         current->dirty_paused_when = now;
1251                         current->nr_dirtied = 0;
1252                         current->nr_dirtied_pause =
1253                                 dirty_poll_interval(nr_dirty, dirty_thresh);
1254                         break;
1255                 }
1256
1257                 if (unlikely(!writeback_in_progress(bdi)))
1258                         bdi_start_background_writeback(bdi);
1259
1260                 /*
1261                  * bdi_thresh is not treated as some limiting factor as
1262                  * dirty_thresh, due to reasons
1263                  * - in JBOD setup, bdi_thresh can fluctuate a lot
1264                  * - in a system with HDD and USB key, the USB key may somehow
1265                  *   go into state (bdi_dirty >> bdi_thresh) either because
1266                  *   bdi_dirty starts high, or because bdi_thresh drops low.
1267                  *   In this case we don't want to hard throttle the USB key
1268                  *   dirtiers for 100 seconds until bdi_dirty drops under
1269                  *   bdi_thresh. Instead the auxiliary bdi control line in
1270                  *   bdi_position_ratio() will let the dirtier task progress
1271                  *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1272                  */
1273                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1274
1275                 /*
1276                  * In order to avoid the stacked BDI deadlock we need
1277                  * to ensure we accurately count the 'dirty' pages when
1278                  * the threshold is low.
1279                  *
1280                  * Otherwise it would be possible to get thresh+n pages
1281                  * reported dirty, even though there are thresh-m pages
1282                  * actually dirty; with m+n sitting in the percpu
1283                  * deltas.
1284                  */
1285                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1286                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1287                         bdi_dirty = bdi_reclaimable +
1288                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
1289                 } else {
1290                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1291                         bdi_dirty = bdi_reclaimable +
1292                                     bdi_stat(bdi, BDI_WRITEBACK);
1293                 }
1294
1295                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1296                                   (nr_dirty > dirty_thresh);
1297                 if (dirty_exceeded && !bdi->dirty_exceeded)
1298                         bdi->dirty_exceeded = 1;
1299
1300                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1301                                      nr_dirty, bdi_thresh, bdi_dirty,
1302                                      start_time);
1303
1304                 dirty_ratelimit = bdi->dirty_ratelimit;
1305                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1306                                                background_thresh, nr_dirty,
1307                                                bdi_thresh, bdi_dirty);
1308                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1309                                                         RATELIMIT_CALC_SHIFT;
1310                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1311                 min_pause = bdi_min_pause(bdi, max_pause,
1312                                           task_ratelimit, dirty_ratelimit,
1313                                           &nr_dirtied_pause);
1314
1315                 if (unlikely(task_ratelimit == 0)) {
1316                         period = max_pause;
1317                         pause = max_pause;
1318                         goto pause;
1319                 }
1320                 period = HZ * pages_dirtied / task_ratelimit;
1321                 pause = period;
1322                 if (current->dirty_paused_when)
1323                         pause -= now - current->dirty_paused_when;
1324                 /*
1325                  * For less than 1s think time (ext3/4 may block the dirtier
1326                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1327                  * however at much less frequency), try to compensate it in
1328                  * future periods by updating the virtual time; otherwise just
1329                  * do a reset, as it may be a light dirtier.
1330                  */
1331                 if (pause < min_pause) {
1332                         trace_balance_dirty_pages(bdi,
1333                                                   dirty_thresh,
1334                                                   background_thresh,
1335                                                   nr_dirty,
1336                                                   bdi_thresh,
1337                                                   bdi_dirty,
1338                                                   dirty_ratelimit,
1339                                                   task_ratelimit,
1340                                                   pages_dirtied,
1341                                                   period,
1342                                                   min(pause, 0L),
1343                                                   start_time);
1344                         if (pause < -HZ) {
1345                                 current->dirty_paused_when = now;
1346                                 current->nr_dirtied = 0;
1347                         } else if (period) {
1348                                 current->dirty_paused_when += period;
1349                                 current->nr_dirtied = 0;
1350                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1351                                 current->nr_dirtied_pause += pages_dirtied;
1352                         break;
1353                 }
1354                 if (unlikely(pause > max_pause)) {
1355                         /* for occasional dropped task_ratelimit */
1356                         now += min(pause - max_pause, max_pause);
1357                         pause = max_pause;
1358                 }
1359
1360 pause:
1361                 trace_balance_dirty_pages(bdi,
1362                                           dirty_thresh,
1363                                           background_thresh,
1364                                           nr_dirty,
1365                                           bdi_thresh,
1366                                           bdi_dirty,
1367                                           dirty_ratelimit,
1368                                           task_ratelimit,
1369                                           pages_dirtied,
1370                                           period,
1371                                           pause,
1372                                           start_time);
1373                 __set_current_state(TASK_KILLABLE);
1374                 io_schedule_timeout(pause);
1375
1376                 current->dirty_paused_when = now + pause;
1377                 current->nr_dirtied = 0;
1378                 current->nr_dirtied_pause = nr_dirtied_pause;
1379
1380                 /*
1381                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1382                  * also keep "1000+ dd on a slow USB stick" under control.
1383                  */
1384                 if (task_ratelimit)
1385                         break;
1386
1387                 /*
1388                  * In the case of an unresponding NFS server and the NFS dirty
1389                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1390                  * to go through, so that tasks on them still remain responsive.
1391                  *
1392                  * In theory 1 page is enough to keep the comsumer-producer
1393                  * pipe going: the flusher cleans 1 page => the task dirties 1
1394                  * more page. However bdi_dirty has accounting errors.  So use
1395                  * the larger and more IO friendly bdi_stat_error.
1396                  */
1397                 if (bdi_dirty <= bdi_stat_error(bdi))
1398                         break;
1399
1400                 if (fatal_signal_pending(current))
1401                         break;
1402         }
1403
1404         if (!dirty_exceeded && bdi->dirty_exceeded)
1405                 bdi->dirty_exceeded = 0;
1406
1407         if (writeback_in_progress(bdi))
1408                 return;
1409
1410         /*
1411          * In laptop mode, we wait until hitting the higher threshold before
1412          * starting background writeout, and then write out all the way down
1413          * to the lower threshold.  So slow writers cause minimal disk activity.
1414          *
1415          * In normal mode, we start background writeout at the lower
1416          * background_thresh, to keep the amount of dirty memory low.
1417          */
1418         if (laptop_mode)
1419                 return;
1420
1421         if (nr_reclaimable > background_thresh)
1422                 bdi_start_background_writeback(bdi);
1423 }
1424
1425 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1426 {
1427         if (set_page_dirty(page) || page_mkwrite) {
1428                 struct address_space *mapping = page_mapping(page);
1429
1430                 if (mapping)
1431                         balance_dirty_pages_ratelimited(mapping);
1432         }
1433 }
1434
1435 static DEFINE_PER_CPU(int, bdp_ratelimits);
1436
1437 /*
1438  * Normal tasks are throttled by
1439  *      loop {
1440  *              dirty tsk->nr_dirtied_pause pages;
1441  *              take a snap in balance_dirty_pages();
1442  *      }
1443  * However there is a worst case. If every task exit immediately when dirtied
1444  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1445  * called to throttle the page dirties. The solution is to save the not yet
1446  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1447  * randomly into the running tasks. This works well for the above worst case,
1448  * as the new task will pick up and accumulate the old task's leaked dirty
1449  * count and eventually get throttled.
1450  */
1451 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1452
1453 /**
1454  * balance_dirty_pages_ratelimited - balance dirty memory state
1455  * @mapping: address_space which was dirtied
1456  *
1457  * Processes which are dirtying memory should call in here once for each page
1458  * which was newly dirtied.  The function will periodically check the system's
1459  * dirty state and will initiate writeback if needed.
1460  *
1461  * On really big machines, get_writeback_state is expensive, so try to avoid
1462  * calling it too often (ratelimiting).  But once we're over the dirty memory
1463  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1464  * from overshooting the limit by (ratelimit_pages) each.
1465  */
1466 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1467 {
1468         struct backing_dev_info *bdi = mapping->backing_dev_info;
1469         int ratelimit;
1470         int *p;
1471
1472         if (!bdi_cap_account_dirty(bdi))
1473                 return;
1474
1475         ratelimit = current->nr_dirtied_pause;
1476         if (bdi->dirty_exceeded)
1477                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1478
1479         preempt_disable();
1480         /*
1481          * This prevents one CPU to accumulate too many dirtied pages without
1482          * calling into balance_dirty_pages(), which can happen when there are
1483          * 1000+ tasks, all of them start dirtying pages at exactly the same
1484          * time, hence all honoured too large initial task->nr_dirtied_pause.
1485          */
1486         p =  &__get_cpu_var(bdp_ratelimits);
1487         if (unlikely(current->nr_dirtied >= ratelimit))
1488                 *p = 0;
1489         else if (unlikely(*p >= ratelimit_pages)) {
1490                 *p = 0;
1491                 ratelimit = 0;
1492         }
1493         /*
1494          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1495          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1496          * the dirty throttling and livelock other long-run dirtiers.
1497          */
1498         p = &__get_cpu_var(dirty_throttle_leaks);
1499         if (*p > 0 && current->nr_dirtied < ratelimit) {
1500                 unsigned long nr_pages_dirtied;
1501                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1502                 *p -= nr_pages_dirtied;
1503                 current->nr_dirtied += nr_pages_dirtied;
1504         }
1505         preempt_enable();
1506
1507         if (unlikely(current->nr_dirtied >= ratelimit))
1508                 balance_dirty_pages(mapping, current->nr_dirtied);
1509 }
1510 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1511
1512 void throttle_vm_writeout(gfp_t gfp_mask)
1513 {
1514         unsigned long background_thresh;
1515         unsigned long dirty_thresh;
1516
1517         for ( ; ; ) {
1518                 global_dirty_limits(&background_thresh, &dirty_thresh);
1519                 dirty_thresh = hard_dirty_limit(dirty_thresh);
1520
1521                 /*
1522                  * Boost the allowable dirty threshold a bit for page
1523                  * allocators so they don't get DoS'ed by heavy writers
1524                  */
1525                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1526
1527                 if (global_page_state(NR_UNSTABLE_NFS) +
1528                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1529                                 break;
1530                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1531
1532                 /*
1533                  * The caller might hold locks which can prevent IO completion
1534                  * or progress in the filesystem.  So we cannot just sit here
1535                  * waiting for IO to complete.
1536                  */
1537                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1538                         break;
1539         }
1540 }
1541
1542 /*
1543  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1544  */
1545 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1546         void __user *buffer, size_t *length, loff_t *ppos)
1547 {
1548         proc_dointvec(table, write, buffer, length, ppos);
1549         return 0;
1550 }
1551
1552 #ifdef CONFIG_BLOCK
1553 void laptop_mode_timer_fn(unsigned long data)
1554 {
1555         struct request_queue *q = (struct request_queue *)data;
1556         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1557                 global_page_state(NR_UNSTABLE_NFS);
1558
1559         /*
1560          * We want to write everything out, not just down to the dirty
1561          * threshold
1562          */
1563         if (bdi_has_dirty_io(&q->backing_dev_info))
1564                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1565                                         WB_REASON_LAPTOP_TIMER);
1566 }
1567
1568 /*
1569  * We've spun up the disk and we're in laptop mode: schedule writeback
1570  * of all dirty data a few seconds from now.  If the flush is already scheduled
1571  * then push it back - the user is still using the disk.
1572  */
1573 void laptop_io_completion(struct backing_dev_info *info)
1574 {
1575         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1576 }
1577
1578 /*
1579  * We're in laptop mode and we've just synced. The sync's writes will have
1580  * caused another writeback to be scheduled by laptop_io_completion.
1581  * Nothing needs to be written back anymore, so we unschedule the writeback.
1582  */
1583 void laptop_sync_completion(void)
1584 {
1585         struct backing_dev_info *bdi;
1586
1587         rcu_read_lock();
1588
1589         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1590                 del_timer(&bdi->laptop_mode_wb_timer);
1591
1592         rcu_read_unlock();
1593 }
1594 #endif
1595
1596 /*
1597  * If ratelimit_pages is too high then we can get into dirty-data overload
1598  * if a large number of processes all perform writes at the same time.
1599  * If it is too low then SMP machines will call the (expensive)
1600  * get_writeback_state too often.
1601  *
1602  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1603  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1604  * thresholds.
1605  */
1606
1607 void writeback_set_ratelimit(void)
1608 {
1609         unsigned long background_thresh;
1610         unsigned long dirty_thresh;
1611         global_dirty_limits(&background_thresh, &dirty_thresh);
1612         global_dirty_limit = dirty_thresh;
1613         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1614         if (ratelimit_pages < 16)
1615                 ratelimit_pages = 16;
1616 }
1617
1618 static int __cpuinit
1619 ratelimit_handler(struct notifier_block *self, unsigned long action,
1620                   void *hcpu)
1621 {
1622
1623         switch (action & ~CPU_TASKS_FROZEN) {
1624         case CPU_ONLINE:
1625         case CPU_DEAD:
1626                 writeback_set_ratelimit();
1627                 return NOTIFY_OK;
1628         default:
1629                 return NOTIFY_DONE;
1630         }
1631 }
1632
1633 static struct notifier_block __cpuinitdata ratelimit_nb = {
1634         .notifier_call  = ratelimit_handler,
1635         .next           = NULL,
1636 };
1637
1638 /*
1639  * Called early on to tune the page writeback dirty limits.
1640  *
1641  * We used to scale dirty pages according to how total memory
1642  * related to pages that could be allocated for buffers (by
1643  * comparing nr_free_buffer_pages() to vm_total_pages.
1644  *
1645  * However, that was when we used "dirty_ratio" to scale with
1646  * all memory, and we don't do that any more. "dirty_ratio"
1647  * is now applied to total non-HIGHPAGE memory (by subtracting
1648  * totalhigh_pages from vm_total_pages), and as such we can't
1649  * get into the old insane situation any more where we had
1650  * large amounts of dirty pages compared to a small amount of
1651  * non-HIGHMEM memory.
1652  *
1653  * But we might still want to scale the dirty_ratio by how
1654  * much memory the box has..
1655  */
1656 void __init page_writeback_init(void)
1657 {
1658         writeback_set_ratelimit();
1659         register_cpu_notifier(&ratelimit_nb);
1660
1661         fprop_global_init(&writeout_completions);
1662 }
1663
1664 /**
1665  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1666  * @mapping: address space structure to write
1667  * @start: starting page index
1668  * @end: ending page index (inclusive)
1669  *
1670  * This function scans the page range from @start to @end (inclusive) and tags
1671  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1672  * that write_cache_pages (or whoever calls this function) will then use
1673  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1674  * used to avoid livelocking of writeback by a process steadily creating new
1675  * dirty pages in the file (thus it is important for this function to be quick
1676  * so that it can tag pages faster than a dirtying process can create them).
1677  */
1678 /*
1679  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1680  */
1681 void tag_pages_for_writeback(struct address_space *mapping,
1682                              pgoff_t start, pgoff_t end)
1683 {
1684 #define WRITEBACK_TAG_BATCH 4096
1685         unsigned long tagged;
1686
1687         do {
1688                 spin_lock_irq(&mapping->tree_lock);
1689                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1690                                 &start, end, WRITEBACK_TAG_BATCH,
1691                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1692                 spin_unlock_irq(&mapping->tree_lock);
1693                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1694                 cond_resched();
1695                 /* We check 'start' to handle wrapping when end == ~0UL */
1696         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1697 }
1698 EXPORT_SYMBOL(tag_pages_for_writeback);
1699
1700 /**
1701  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1702  * @mapping: address space structure to write
1703  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1704  * @writepage: function called for each page
1705  * @data: data passed to writepage function
1706  *
1707  * If a page is already under I/O, write_cache_pages() skips it, even
1708  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1709  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1710  * and msync() need to guarantee that all the data which was dirty at the time
1711  * the call was made get new I/O started against them.  If wbc->sync_mode is
1712  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1713  * existing IO to complete.
1714  *
1715  * To avoid livelocks (when other process dirties new pages), we first tag
1716  * pages which should be written back with TOWRITE tag and only then start
1717  * writing them. For data-integrity sync we have to be careful so that we do
1718  * not miss some pages (e.g., because some other process has cleared TOWRITE
1719  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1720  * by the process clearing the DIRTY tag (and submitting the page for IO).
1721  */
1722 int write_cache_pages(struct address_space *mapping,
1723                       struct writeback_control *wbc, writepage_t writepage,
1724                       void *data)
1725 {
1726         int ret = 0;
1727         int done = 0;
1728         struct pagevec pvec;
1729         int nr_pages;
1730         pgoff_t uninitialized_var(writeback_index);
1731         pgoff_t index;
1732         pgoff_t end;            /* Inclusive */
1733         pgoff_t done_index;
1734         int cycled;
1735         int range_whole = 0;
1736         int tag;
1737
1738         pagevec_init(&pvec, 0);
1739         if (wbc->range_cyclic) {
1740                 writeback_index = mapping->writeback_index; /* prev offset */
1741                 index = writeback_index;
1742                 if (index == 0)
1743                         cycled = 1;
1744                 else
1745                         cycled = 0;
1746                 end = -1;
1747         } else {
1748                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1749                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1750                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1751                         range_whole = 1;
1752                 cycled = 1; /* ignore range_cyclic tests */
1753         }
1754         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1755                 tag = PAGECACHE_TAG_TOWRITE;
1756         else
1757                 tag = PAGECACHE_TAG_DIRTY;
1758 retry:
1759         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1760                 tag_pages_for_writeback(mapping, index, end);
1761         done_index = index;
1762         while (!done && (index <= end)) {
1763                 int i;
1764
1765                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1766                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1767                 if (nr_pages == 0)
1768                         break;
1769
1770                 for (i = 0; i < nr_pages; i++) {
1771                         struct page *page = pvec.pages[i];
1772
1773                         /*
1774                          * At this point, the page may be truncated or
1775                          * invalidated (changing page->mapping to NULL), or
1776                          * even swizzled back from swapper_space to tmpfs file
1777                          * mapping. However, page->index will not change
1778                          * because we have a reference on the page.
1779                          */
1780                         if (page->index > end) {
1781                                 /*
1782                                  * can't be range_cyclic (1st pass) because
1783                                  * end == -1 in that case.
1784                                  */
1785                                 done = 1;
1786                                 break;
1787                         }
1788
1789                         done_index = page->index;
1790
1791                         lock_page(page);
1792
1793                         /*
1794                          * Page truncated or invalidated. We can freely skip it
1795                          * then, even for data integrity operations: the page
1796                          * has disappeared concurrently, so there could be no
1797                          * real expectation of this data interity operation
1798                          * even if there is now a new, dirty page at the same
1799                          * pagecache address.
1800                          */
1801                         if (unlikely(page->mapping != mapping)) {
1802 continue_unlock:
1803                                 unlock_page(page);
1804                                 continue;
1805                         }
1806
1807                         if (!PageDirty(page)) {
1808                                 /* someone wrote it for us */
1809                                 goto continue_unlock;
1810                         }
1811
1812                         if (PageWriteback(page)) {
1813                                 if (wbc->sync_mode != WB_SYNC_NONE)
1814                                         wait_on_page_writeback(page);
1815                                 else
1816                                         goto continue_unlock;
1817                         }
1818
1819                         BUG_ON(PageWriteback(page));
1820                         if (!clear_page_dirty_for_io(page))
1821                                 goto continue_unlock;
1822
1823                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1824                         ret = (*writepage)(page, wbc, data);
1825                         if (unlikely(ret)) {
1826                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1827                                         unlock_page(page);
1828                                         ret = 0;
1829                                 } else {
1830                                         /*
1831                                          * done_index is set past this page,
1832                                          * so media errors will not choke
1833                                          * background writeout for the entire
1834                                          * file. This has consequences for
1835                                          * range_cyclic semantics (ie. it may
1836                                          * not be suitable for data integrity
1837                                          * writeout).
1838                                          */
1839                                         done_index = page->index + 1;
1840                                         done = 1;
1841                                         break;
1842                                 }
1843                         }
1844
1845                         /*
1846                          * We stop writing back only if we are not doing
1847                          * integrity sync. In case of integrity sync we have to
1848                          * keep going until we have written all the pages
1849                          * we tagged for writeback prior to entering this loop.
1850                          */
1851                         if (--wbc->nr_to_write <= 0 &&
1852                             wbc->sync_mode == WB_SYNC_NONE) {
1853                                 done = 1;
1854                                 break;
1855                         }
1856                 }
1857                 pagevec_release(&pvec);
1858                 cond_resched();
1859         }
1860         if (!cycled && !done) {
1861                 /*
1862                  * range_cyclic:
1863                  * We hit the last page and there is more work to be done: wrap
1864                  * back to the start of the file
1865                  */
1866                 cycled = 1;
1867                 index = 0;
1868                 end = writeback_index - 1;
1869                 goto retry;
1870         }
1871         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1872                 mapping->writeback_index = done_index;
1873
1874         return ret;
1875 }
1876 EXPORT_SYMBOL(write_cache_pages);
1877
1878 /*
1879  * Function used by generic_writepages to call the real writepage
1880  * function and set the mapping flags on error
1881  */
1882 static int __writepage(struct page *page, struct writeback_control *wbc,
1883                        void *data)
1884 {
1885         struct address_space *mapping = data;
1886         int ret = mapping->a_ops->writepage(page, wbc);
1887         mapping_set_error(mapping, ret);
1888         return ret;
1889 }
1890
1891 /**
1892  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1893  * @mapping: address space structure to write
1894  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1895  *
1896  * This is a library function, which implements the writepages()
1897  * address_space_operation.
1898  */
1899 int generic_writepages(struct address_space *mapping,
1900                        struct writeback_control *wbc)
1901 {
1902         struct blk_plug plug;
1903         int ret;
1904
1905         /* deal with chardevs and other special file */
1906         if (!mapping->a_ops->writepage)
1907                 return 0;
1908
1909         blk_start_plug(&plug);
1910         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1911         blk_finish_plug(&plug);
1912         return ret;
1913 }
1914
1915 EXPORT_SYMBOL(generic_writepages);
1916
1917 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1918 {
1919         int ret;
1920
1921         if (wbc->nr_to_write <= 0)
1922                 return 0;
1923         if (mapping->a_ops->writepages)
1924                 ret = mapping->a_ops->writepages(mapping, wbc);
1925         else
1926                 ret = generic_writepages(mapping, wbc);
1927         return ret;
1928 }
1929
1930 /**
1931  * write_one_page - write out a single page and optionally wait on I/O
1932  * @page: the page to write
1933  * @wait: if true, wait on writeout
1934  *
1935  * The page must be locked by the caller and will be unlocked upon return.
1936  *
1937  * write_one_page() returns a negative error code if I/O failed.
1938  */
1939 int write_one_page(struct page *page, int wait)
1940 {
1941         struct address_space *mapping = page->mapping;
1942         int ret = 0;
1943         struct writeback_control wbc = {
1944                 .sync_mode = WB_SYNC_ALL,
1945                 .nr_to_write = 1,
1946         };
1947
1948         BUG_ON(!PageLocked(page));
1949
1950         if (wait)
1951                 wait_on_page_writeback(page);
1952
1953         if (clear_page_dirty_for_io(page)) {
1954                 page_cache_get(page);
1955                 ret = mapping->a_ops->writepage(page, &wbc);
1956                 if (ret == 0 && wait) {
1957                         wait_on_page_writeback(page);
1958                         if (PageError(page))
1959                                 ret = -EIO;
1960                 }
1961                 page_cache_release(page);
1962         } else {
1963                 unlock_page(page);
1964         }
1965         return ret;
1966 }
1967 EXPORT_SYMBOL(write_one_page);
1968
1969 /*
1970  * For address_spaces which do not use buffers nor write back.
1971  */
1972 int __set_page_dirty_no_writeback(struct page *page)
1973 {
1974         if (!PageDirty(page))
1975                 return !TestSetPageDirty(page);
1976         return 0;
1977 }
1978
1979 /*
1980  * Helper function for set_page_dirty family.
1981  * NOTE: This relies on being atomic wrt interrupts.
1982  */
1983 void account_page_dirtied(struct page *page, struct address_space *mapping)
1984 {
1985         if (mapping_cap_account_dirty(mapping)) {
1986                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1987                 __inc_zone_page_state(page, NR_DIRTIED);
1988                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1989                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1990                 task_io_account_write(PAGE_CACHE_SIZE);
1991                 current->nr_dirtied++;
1992                 this_cpu_inc(bdp_ratelimits);
1993         }
1994 }
1995 EXPORT_SYMBOL(account_page_dirtied);
1996
1997 /*
1998  * Helper function for set_page_writeback family.
1999  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2000  * wrt interrupts.
2001  */
2002 void account_page_writeback(struct page *page)
2003 {
2004         inc_zone_page_state(page, NR_WRITEBACK);
2005 }
2006 EXPORT_SYMBOL(account_page_writeback);
2007
2008 /*
2009  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2010  * its radix tree.
2011  *
2012  * This is also used when a single buffer is being dirtied: we want to set the
2013  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2014  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2015  *
2016  * Most callers have locked the page, which pins the address_space in memory.
2017  * But zap_pte_range() does not lock the page, however in that case the
2018  * mapping is pinned by the vma's ->vm_file reference.
2019  *
2020  * We take care to handle the case where the page was truncated from the
2021  * mapping by re-checking page_mapping() inside tree_lock.
2022  */
2023 int __set_page_dirty_nobuffers(struct page *page)
2024 {
2025         if (!TestSetPageDirty(page)) {
2026                 struct address_space *mapping = page_mapping(page);
2027                 struct address_space *mapping2;
2028
2029                 if (!mapping)
2030                         return 1;
2031
2032                 spin_lock_irq(&mapping->tree_lock);
2033                 mapping2 = page_mapping(page);
2034                 if (mapping2) { /* Race with truncate? */
2035                         BUG_ON(mapping2 != mapping);
2036                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2037                         account_page_dirtied(page, mapping);
2038                         radix_tree_tag_set(&mapping->page_tree,
2039                                 page_index(page), PAGECACHE_TAG_DIRTY);
2040                 }
2041                 spin_unlock_irq(&mapping->tree_lock);
2042                 if (mapping->host) {
2043                         /* !PageAnon && !swapper_space */
2044                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2045                 }
2046                 return 1;
2047         }
2048         return 0;
2049 }
2050 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2051
2052 /*
2053  * Call this whenever redirtying a page, to de-account the dirty counters
2054  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2055  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2056  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2057  * control.
2058  */
2059 void account_page_redirty(struct page *page)
2060 {
2061         struct address_space *mapping = page->mapping;
2062         if (mapping && mapping_cap_account_dirty(mapping)) {
2063                 current->nr_dirtied--;
2064                 dec_zone_page_state(page, NR_DIRTIED);
2065                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2066         }
2067 }
2068 EXPORT_SYMBOL(account_page_redirty);
2069
2070 /*
2071  * When a writepage implementation decides that it doesn't want to write this
2072  * page for some reason, it should redirty the locked page via
2073  * redirty_page_for_writepage() and it should then unlock the page and return 0
2074  */
2075 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2076 {
2077         wbc->pages_skipped++;
2078         account_page_redirty(page);
2079         return __set_page_dirty_nobuffers(page);
2080 }
2081 EXPORT_SYMBOL(redirty_page_for_writepage);
2082
2083 /*
2084  * Dirty a page.
2085  *
2086  * For pages with a mapping this should be done under the page lock
2087  * for the benefit of asynchronous memory errors who prefer a consistent
2088  * dirty state. This rule can be broken in some special cases,
2089  * but should be better not to.
2090  *
2091  * If the mapping doesn't provide a set_page_dirty a_op, then
2092  * just fall through and assume that it wants buffer_heads.
2093  */
2094 int set_page_dirty(struct page *page)
2095 {
2096         struct address_space *mapping = page_mapping(page);
2097
2098         if (likely(mapping)) {
2099                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2100                 /*
2101                  * readahead/lru_deactivate_page could remain
2102                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2103                  * About readahead, if the page is written, the flags would be
2104                  * reset. So no problem.
2105                  * About lru_deactivate_page, if the page is redirty, the flag
2106                  * will be reset. So no problem. but if the page is used by readahead
2107                  * it will confuse readahead and make it restart the size rampup
2108                  * process. But it's a trivial problem.
2109                  */
2110                 ClearPageReclaim(page);
2111 #ifdef CONFIG_BLOCK
2112                 if (!spd)
2113                         spd = __set_page_dirty_buffers;
2114 #endif
2115                 return (*spd)(page);
2116         }
2117         if (!PageDirty(page)) {
2118                 if (!TestSetPageDirty(page))
2119                         return 1;
2120         }
2121         return 0;
2122 }
2123 EXPORT_SYMBOL(set_page_dirty);
2124
2125 /*
2126  * set_page_dirty() is racy if the caller has no reference against
2127  * page->mapping->host, and if the page is unlocked.  This is because another
2128  * CPU could truncate the page off the mapping and then free the mapping.
2129  *
2130  * Usually, the page _is_ locked, or the caller is a user-space process which
2131  * holds a reference on the inode by having an open file.
2132  *
2133  * In other cases, the page should be locked before running set_page_dirty().
2134  */
2135 int set_page_dirty_lock(struct page *page)
2136 {
2137         int ret;
2138
2139         lock_page(page);
2140         ret = set_page_dirty(page);
2141         unlock_page(page);
2142         return ret;
2143 }
2144 EXPORT_SYMBOL(set_page_dirty_lock);
2145
2146 /*
2147  * Clear a page's dirty flag, while caring for dirty memory accounting.
2148  * Returns true if the page was previously dirty.
2149  *
2150  * This is for preparing to put the page under writeout.  We leave the page
2151  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2152  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2153  * implementation will run either set_page_writeback() or set_page_dirty(),
2154  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2155  * back into sync.
2156  *
2157  * This incoherency between the page's dirty flag and radix-tree tag is
2158  * unfortunate, but it only exists while the page is locked.
2159  */
2160 int clear_page_dirty_for_io(struct page *page)
2161 {
2162         struct address_space *mapping = page_mapping(page);
2163
2164         BUG_ON(!PageLocked(page));
2165
2166         if (mapping && mapping_cap_account_dirty(mapping)) {
2167                 /*
2168                  * Yes, Virginia, this is indeed insane.
2169                  *
2170                  * We use this sequence to make sure that
2171                  *  (a) we account for dirty stats properly
2172                  *  (b) we tell the low-level filesystem to
2173                  *      mark the whole page dirty if it was
2174                  *      dirty in a pagetable. Only to then
2175                  *  (c) clean the page again and return 1 to
2176                  *      cause the writeback.
2177                  *
2178                  * This way we avoid all nasty races with the
2179                  * dirty bit in multiple places and clearing
2180                  * them concurrently from different threads.
2181                  *
2182                  * Note! Normally the "set_page_dirty(page)"
2183                  * has no effect on the actual dirty bit - since
2184                  * that will already usually be set. But we
2185                  * need the side effects, and it can help us
2186                  * avoid races.
2187                  *
2188                  * We basically use the page "master dirty bit"
2189                  * as a serialization point for all the different
2190                  * threads doing their things.
2191                  */
2192                 if (page_mkclean(page))
2193                         set_page_dirty(page);
2194                 /*
2195                  * We carefully synchronise fault handlers against
2196                  * installing a dirty pte and marking the page dirty
2197                  * at this point. We do this by having them hold the
2198                  * page lock at some point after installing their
2199                  * pte, but before marking the page dirty.
2200                  * Pages are always locked coming in here, so we get
2201                  * the desired exclusion. See mm/memory.c:do_wp_page()
2202                  * for more comments.
2203                  */
2204                 if (TestClearPageDirty(page)) {
2205                         dec_zone_page_state(page, NR_FILE_DIRTY);
2206                         dec_bdi_stat(mapping->backing_dev_info,
2207                                         BDI_RECLAIMABLE);
2208                         return 1;
2209                 }
2210                 return 0;
2211         }
2212         return TestClearPageDirty(page);
2213 }
2214 EXPORT_SYMBOL(clear_page_dirty_for_io);
2215
2216 int test_clear_page_writeback(struct page *page)
2217 {
2218         struct address_space *mapping = page_mapping(page);
2219         int ret;
2220
2221         if (mapping) {
2222                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2223                 unsigned long flags;
2224
2225                 spin_lock_irqsave(&mapping->tree_lock, flags);
2226                 ret = TestClearPageWriteback(page);
2227                 if (ret) {
2228                         radix_tree_tag_clear(&mapping->page_tree,
2229                                                 page_index(page),
2230                                                 PAGECACHE_TAG_WRITEBACK);
2231                         if (bdi_cap_account_writeback(bdi)) {
2232                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2233                                 __bdi_writeout_inc(bdi);
2234                         }
2235                 }
2236                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2237         } else {
2238                 ret = TestClearPageWriteback(page);
2239         }
2240         if (ret) {
2241                 dec_zone_page_state(page, NR_WRITEBACK);
2242                 inc_zone_page_state(page, NR_WRITTEN);
2243         }
2244         return ret;
2245 }
2246
2247 int test_set_page_writeback(struct page *page)
2248 {
2249         struct address_space *mapping = page_mapping(page);
2250         int ret;
2251
2252         if (mapping) {
2253                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2254                 unsigned long flags;
2255
2256                 spin_lock_irqsave(&mapping->tree_lock, flags);
2257                 ret = TestSetPageWriteback(page);
2258                 if (!ret) {
2259                         radix_tree_tag_set(&mapping->page_tree,
2260                                                 page_index(page),
2261                                                 PAGECACHE_TAG_WRITEBACK);
2262                         if (bdi_cap_account_writeback(bdi))
2263                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2264                 }
2265                 if (!PageDirty(page))
2266                         radix_tree_tag_clear(&mapping->page_tree,
2267                                                 page_index(page),
2268                                                 PAGECACHE_TAG_DIRTY);
2269                 radix_tree_tag_clear(&mapping->page_tree,
2270                                      page_index(page),
2271                                      PAGECACHE_TAG_TOWRITE);
2272                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2273         } else {
2274                 ret = TestSetPageWriteback(page);
2275         }
2276         if (!ret)
2277                 account_page_writeback(page);
2278         return ret;
2279
2280 }
2281 EXPORT_SYMBOL(test_set_page_writeback);
2282
2283 /*
2284  * Return true if any of the pages in the mapping are marked with the
2285  * passed tag.
2286  */
2287 int mapping_tagged(struct address_space *mapping, int tag)
2288 {
2289         return radix_tree_tagged(&mapping->page_tree, tag);
2290 }
2291 EXPORT_SYMBOL(mapping_tagged);