regmap: mmio: convert some error returns to BUG()
[linux-2.6.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 <trace/events/writeback.h>
38
39 /*
40  * Sleep at most 200ms at a time in balance_dirty_pages().
41  */
42 #define MAX_PAUSE               max(HZ/5, 1)
43
44 /*
45  * Try to keep balance_dirty_pages() call intervals higher than this many pages
46  * by raising pause time to max_pause when falls below it.
47  */
48 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
49
50 /*
51  * Estimate write bandwidth at 200ms intervals.
52  */
53 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
54
55 #define RATELIMIT_CALC_SHIFT    10
56
57 /*
58  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
59  * will look to see if it needs to force writeback or throttling.
60  */
61 static long ratelimit_pages = 32;
62
63 /* The following parameters are exported via /proc/sys/vm */
64
65 /*
66  * Start background writeback (via writeback threads) at this percentage
67  */
68 int dirty_background_ratio = 10;
69
70 /*
71  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
72  * dirty_background_ratio * the amount of dirtyable memory
73  */
74 unsigned long dirty_background_bytes;
75
76 /*
77  * free highmem will not be subtracted from the total free memory
78  * for calculating free ratios if vm_highmem_is_dirtyable is true
79  */
80 int vm_highmem_is_dirtyable;
81
82 /*
83  * The generator of dirty data starts writeback at this percentage
84  */
85 int vm_dirty_ratio = 20;
86
87 /*
88  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
89  * vm_dirty_ratio * the amount of dirtyable memory
90  */
91 unsigned long vm_dirty_bytes;
92
93 /*
94  * The interval between `kupdate'-style writebacks
95  */
96 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
97
98 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
99
100 /*
101  * The longest time for which data is allowed to remain dirty
102  */
103 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
104
105 /*
106  * Flag that makes the machine dump writes/reads and block dirtyings.
107  */
108 int block_dump;
109
110 /*
111  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112  * a full sync is triggered after this time elapses without any disk activity.
113  */
114 int laptop_mode;
115
116 EXPORT_SYMBOL(laptop_mode);
117
118 /* End of sysctl-exported parameters */
119
120 unsigned long global_dirty_limit;
121
122 /*
123  * Scale the writeback cache size proportional to the relative writeout speeds.
124  *
125  * We do this by keeping a floating proportion between BDIs, based on page
126  * writeback completions [end_page_writeback()]. Those devices that write out
127  * pages fastest will get the larger share, while the slower will get a smaller
128  * share.
129  *
130  * We use page writeout completions because we are interested in getting rid of
131  * dirty pages. Having them written out is the primary goal.
132  *
133  * We introduce a concept of time, a period over which we measure these events,
134  * because demand can/will vary over time. The length of this period itself is
135  * measured in page writeback completions.
136  *
137  */
138 static struct prop_descriptor vm_completions;
139
140 /*
141  * Work out the current dirty-memory clamping and background writeout
142  * thresholds.
143  *
144  * The main aim here is to lower them aggressively if there is a lot of mapped
145  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
146  * pages.  It is better to clamp down on writers than to start swapping, and
147  * performing lots of scanning.
148  *
149  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
150  *
151  * We don't permit the clamping level to fall below 5% - that is getting rather
152  * excessive.
153  *
154  * We make sure that the background writeout level is below the adjusted
155  * clamping level.
156  */
157
158 /*
159  * In a memory zone, there is a certain amount of pages we consider
160  * available for the page cache, which is essentially the number of
161  * free and reclaimable pages, minus some zone reserves to protect
162  * lowmem and the ability to uphold the zone's watermarks without
163  * requiring writeback.
164  *
165  * This number of dirtyable pages is the base value of which the
166  * user-configurable dirty ratio is the effictive number of pages that
167  * are allowed to be actually dirtied.  Per individual zone, or
168  * globally by using the sum of dirtyable pages over all zones.
169  *
170  * Because the user is allowed to specify the dirty limit globally as
171  * absolute number of bytes, calculating the per-zone dirty limit can
172  * require translating the configured limit into a percentage of
173  * global dirtyable memory first.
174  */
175
176 static unsigned long highmem_dirtyable_memory(unsigned long total)
177 {
178 #ifdef CONFIG_HIGHMEM
179         int node;
180         unsigned long x = 0;
181
182         for_each_node_state(node, N_HIGH_MEMORY) {
183                 struct zone *z =
184                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
185
186                 x += zone_page_state(z, NR_FREE_PAGES) +
187                      zone_reclaimable_pages(z) - z->dirty_balance_reserve;
188         }
189         /*
190          * Make sure that the number of highmem pages is never larger
191          * than the number of the total dirtyable memory. This can only
192          * occur in very strange VM situations but we want to make sure
193          * that this does not occur.
194          */
195         return min(x, total);
196 #else
197         return 0;
198 #endif
199 }
200
201 /**
202  * global_dirtyable_memory - number of globally dirtyable pages
203  *
204  * Returns the global number of pages potentially available for dirty
205  * page cache.  This is the base value for the global dirty limits.
206  */
207 unsigned long global_dirtyable_memory(void)
208 {
209         unsigned long x;
210
211         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
212             dirty_balance_reserve;
213
214         if (!vm_highmem_is_dirtyable)
215                 x -= highmem_dirtyable_memory(x);
216
217         return x + 1;   /* Ensure that we never return 0 */
218 }
219
220 /*
221  * global_dirty_limits - background-writeback and dirty-throttling thresholds
222  *
223  * Calculate the dirty thresholds based on sysctl parameters
224  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
225  * - vm.dirty_ratio             or  vm.dirty_bytes
226  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
227  * real-time tasks.
228  */
229 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
230 {
231         unsigned long background;
232         unsigned long dirty;
233         unsigned long uninitialized_var(available_memory);
234         struct task_struct *tsk;
235
236         if (!vm_dirty_bytes || !dirty_background_bytes)
237                 available_memory = global_dirtyable_memory();
238
239         if (vm_dirty_bytes)
240                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
241         else
242                 dirty = (vm_dirty_ratio * available_memory) / 100;
243
244         if (dirty_background_bytes)
245                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
246         else
247                 background = (dirty_background_ratio * available_memory) / 100;
248
249         if (background >= dirty)
250                 background = dirty / 2;
251         tsk = current;
252         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
253                 background += background / 4;
254                 dirty += dirty / 4;
255         }
256         *pbackground = background;
257         *pdirty = dirty;
258         trace_global_dirty_state(background, dirty);
259 }
260
261 /**
262  * zone_dirtyable_memory - number of dirtyable pages in a zone
263  * @zone: the zone
264  *
265  * Returns the zone's number of pages potentially available for dirty
266  * page cache.  This is the base value for the per-zone dirty limits.
267  */
268 static unsigned long zone_dirtyable_memory(struct zone *zone)
269 {
270         /*
271          * The effective global number of dirtyable pages may exclude
272          * highmem as a big-picture measure to keep the ratio between
273          * dirty memory and lowmem reasonable.
274          *
275          * But this function is purely about the individual zone and a
276          * highmem zone can hold its share of dirty pages, so we don't
277          * care about vm_highmem_is_dirtyable here.
278          */
279         return zone_page_state(zone, NR_FREE_PAGES) +
280                zone_reclaimable_pages(zone) -
281                zone->dirty_balance_reserve;
282 }
283
284 /**
285  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
286  * @zone: the zone
287  *
288  * Returns the maximum number of dirty pages allowed in a zone, based
289  * on the zone's dirtyable memory.
290  */
291 static unsigned long zone_dirty_limit(struct zone *zone)
292 {
293         unsigned long zone_memory = zone_dirtyable_memory(zone);
294         struct task_struct *tsk = current;
295         unsigned long dirty;
296
297         if (vm_dirty_bytes)
298                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
299                         zone_memory / global_dirtyable_memory();
300         else
301                 dirty = vm_dirty_ratio * zone_memory / 100;
302
303         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
304                 dirty += dirty / 4;
305
306         return dirty;
307 }
308
309 /**
310  * zone_dirty_ok - tells whether a zone is within its dirty limits
311  * @zone: the zone to check
312  *
313  * Returns %true when the dirty pages in @zone are within the zone's
314  * dirty limit, %false if the limit is exceeded.
315  */
316 bool zone_dirty_ok(struct zone *zone)
317 {
318         unsigned long limit = zone_dirty_limit(zone);
319
320         return zone_page_state(zone, NR_FILE_DIRTY) +
321                zone_page_state(zone, NR_UNSTABLE_NFS) +
322                zone_page_state(zone, NR_WRITEBACK) <= limit;
323 }
324
325 /*
326  * couple the period to the dirty_ratio:
327  *
328  *   period/2 ~ roundup_pow_of_two(dirty limit)
329  */
330 static int calc_period_shift(void)
331 {
332         unsigned long dirty_total;
333
334         if (vm_dirty_bytes)
335                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
336         else
337                 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
338                                 100;
339         return 2 + ilog2(dirty_total - 1);
340 }
341
342 /*
343  * update the period when the dirty threshold changes.
344  */
345 static void update_completion_period(void)
346 {
347         int shift = calc_period_shift();
348         prop_change_shift(&vm_completions, shift);
349
350         writeback_set_ratelimit();
351 }
352
353 int dirty_background_ratio_handler(struct ctl_table *table, int write,
354                 void __user *buffer, size_t *lenp,
355                 loff_t *ppos)
356 {
357         int ret;
358
359         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
360         if (ret == 0 && write)
361                 dirty_background_bytes = 0;
362         return ret;
363 }
364
365 int dirty_background_bytes_handler(struct ctl_table *table, int write,
366                 void __user *buffer, size_t *lenp,
367                 loff_t *ppos)
368 {
369         int ret;
370
371         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
372         if (ret == 0 && write)
373                 dirty_background_ratio = 0;
374         return ret;
375 }
376
377 int dirty_ratio_handler(struct ctl_table *table, int write,
378                 void __user *buffer, size_t *lenp,
379                 loff_t *ppos)
380 {
381         int old_ratio = vm_dirty_ratio;
382         int ret;
383
384         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
385         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
386                 update_completion_period();
387                 vm_dirty_bytes = 0;
388         }
389         return ret;
390 }
391
392 int dirty_bytes_handler(struct ctl_table *table, int write,
393                 void __user *buffer, size_t *lenp,
394                 loff_t *ppos)
395 {
396         unsigned long old_bytes = vm_dirty_bytes;
397         int ret;
398
399         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
400         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
401                 update_completion_period();
402                 vm_dirty_ratio = 0;
403         }
404         return ret;
405 }
406
407 /*
408  * Increment the BDI's writeout completion count and the global writeout
409  * completion count. Called from test_clear_page_writeback().
410  */
411 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
412 {
413         __inc_bdi_stat(bdi, BDI_WRITTEN);
414         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
415                               bdi->max_prop_frac);
416 }
417
418 void bdi_writeout_inc(struct backing_dev_info *bdi)
419 {
420         unsigned long flags;
421
422         local_irq_save(flags);
423         __bdi_writeout_inc(bdi);
424         local_irq_restore(flags);
425 }
426 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
427
428 /*
429  * Obtain an accurate fraction of the BDI's portion.
430  */
431 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
432                 long *numerator, long *denominator)
433 {
434         prop_fraction_percpu(&vm_completions, &bdi->completions,
435                                 numerator, denominator);
436 }
437
438 /*
439  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
440  * registered backing devices, which, for obvious reasons, can not
441  * exceed 100%.
442  */
443 static unsigned int bdi_min_ratio;
444
445 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
446 {
447         int ret = 0;
448
449         spin_lock_bh(&bdi_lock);
450         if (min_ratio > bdi->max_ratio) {
451                 ret = -EINVAL;
452         } else {
453                 min_ratio -= bdi->min_ratio;
454                 if (bdi_min_ratio + min_ratio < 100) {
455                         bdi_min_ratio += min_ratio;
456                         bdi->min_ratio += min_ratio;
457                 } else {
458                         ret = -EINVAL;
459                 }
460         }
461         spin_unlock_bh(&bdi_lock);
462
463         return ret;
464 }
465
466 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
467 {
468         int ret = 0;
469
470         if (max_ratio > 100)
471                 return -EINVAL;
472
473         spin_lock_bh(&bdi_lock);
474         if (bdi->min_ratio > max_ratio) {
475                 ret = -EINVAL;
476         } else {
477                 bdi->max_ratio = max_ratio;
478                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
479         }
480         spin_unlock_bh(&bdi_lock);
481
482         return ret;
483 }
484 EXPORT_SYMBOL(bdi_set_max_ratio);
485
486 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
487                                            unsigned long bg_thresh)
488 {
489         return (thresh + bg_thresh) / 2;
490 }
491
492 static unsigned long hard_dirty_limit(unsigned long thresh)
493 {
494         return max(thresh, global_dirty_limit);
495 }
496
497 /**
498  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
499  * @bdi: the backing_dev_info to query
500  * @dirty: global dirty limit in pages
501  *
502  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
503  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
504  *
505  * Note that balance_dirty_pages() will only seriously take it as a hard limit
506  * when sleeping max_pause per page is not enough to keep the dirty pages under
507  * control. For example, when the device is completely stalled due to some error
508  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
509  * In the other normal situations, it acts more gently by throttling the tasks
510  * more (rather than completely block them) when the bdi dirty pages go high.
511  *
512  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
513  * - starving fast devices
514  * - piling up dirty pages (that will take long time to sync) on slow devices
515  *
516  * The bdi's share of dirty limit will be adapting to its throughput and
517  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
518  */
519 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
520 {
521         u64 bdi_dirty;
522         long numerator, denominator;
523
524         /*
525          * Calculate this BDI's share of the dirty ratio.
526          */
527         bdi_writeout_fraction(bdi, &numerator, &denominator);
528
529         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
530         bdi_dirty *= numerator;
531         do_div(bdi_dirty, denominator);
532
533         bdi_dirty += (dirty * bdi->min_ratio) / 100;
534         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
535                 bdi_dirty = dirty * bdi->max_ratio / 100;
536
537         return bdi_dirty;
538 }
539
540 /*
541  * Dirty position control.
542  *
543  * (o) global/bdi setpoints
544  *
545  * We want the dirty pages be balanced around the global/bdi setpoints.
546  * When the number of dirty pages is higher/lower than the setpoint, the
547  * dirty position control ratio (and hence task dirty ratelimit) will be
548  * decreased/increased to bring the dirty pages back to the setpoint.
549  *
550  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
551  *
552  *     if (dirty < setpoint) scale up   pos_ratio
553  *     if (dirty > setpoint) scale down pos_ratio
554  *
555  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
556  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
557  *
558  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
559  *
560  * (o) global control line
561  *
562  *     ^ pos_ratio
563  *     |
564  *     |            |<===== global dirty control scope ======>|
565  * 2.0 .............*
566  *     |            .*
567  *     |            . *
568  *     |            .   *
569  *     |            .     *
570  *     |            .        *
571  *     |            .            *
572  * 1.0 ................................*
573  *     |            .                  .     *
574  *     |            .                  .          *
575  *     |            .                  .              *
576  *     |            .                  .                 *
577  *     |            .                  .                    *
578  *   0 +------------.------------------.----------------------*------------->
579  *           freerun^          setpoint^                 limit^   dirty pages
580  *
581  * (o) bdi control line
582  *
583  *     ^ pos_ratio
584  *     |
585  *     |            *
586  *     |              *
587  *     |                *
588  *     |                  *
589  *     |                    * |<=========== span ============>|
590  * 1.0 .......................*
591  *     |                      . *
592  *     |                      .   *
593  *     |                      .     *
594  *     |                      .       *
595  *     |                      .         *
596  *     |                      .           *
597  *     |                      .             *
598  *     |                      .               *
599  *     |                      .                 *
600  *     |                      .                   *
601  *     |                      .                     *
602  * 1/4 ...............................................* * * * * * * * * * * *
603  *     |                      .                         .
604  *     |                      .                           .
605  *     |                      .                             .
606  *   0 +----------------------.-------------------------------.------------->
607  *                bdi_setpoint^                    x_intercept^
608  *
609  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
610  * be smoothly throttled down to normal if it starts high in situations like
611  * - start writing to a slow SD card and a fast disk at the same time. The SD
612  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
613  * - the bdi dirty thresh drops quickly due to change of JBOD workload
614  */
615 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
616                                         unsigned long thresh,
617                                         unsigned long bg_thresh,
618                                         unsigned long dirty,
619                                         unsigned long bdi_thresh,
620                                         unsigned long bdi_dirty)
621 {
622         unsigned long write_bw = bdi->avg_write_bandwidth;
623         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
624         unsigned long limit = hard_dirty_limit(thresh);
625         unsigned long x_intercept;
626         unsigned long setpoint;         /* dirty pages' target balance point */
627         unsigned long bdi_setpoint;
628         unsigned long span;
629         long long pos_ratio;            /* for scaling up/down the rate limit */
630         long x;
631
632         if (unlikely(dirty >= limit))
633                 return 0;
634
635         /*
636          * global setpoint
637          *
638          *                           setpoint - dirty 3
639          *        f(dirty) := 1.0 + (----------------)
640          *                           limit - setpoint
641          *
642          * it's a 3rd order polynomial that subjects to
643          *
644          * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
645          * (2) f(setpoint) = 1.0 => the balance point
646          * (3) f(limit)    = 0   => the hard limit
647          * (4) df/dx      <= 0   => negative feedback control
648          * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
649          *     => fast response on large errors; small oscillation near setpoint
650          */
651         setpoint = (freerun + limit) / 2;
652         x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
653                     limit - setpoint + 1);
654         pos_ratio = x;
655         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
656         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
657         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
658
659         /*
660          * We have computed basic pos_ratio above based on global situation. If
661          * the bdi is over/under its share of dirty pages, we want to scale
662          * pos_ratio further down/up. That is done by the following mechanism.
663          */
664
665         /*
666          * bdi setpoint
667          *
668          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
669          *
670          *                        x_intercept - bdi_dirty
671          *                     := --------------------------
672          *                        x_intercept - bdi_setpoint
673          *
674          * The main bdi control line is a linear function that subjects to
675          *
676          * (1) f(bdi_setpoint) = 1.0
677          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
678          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
679          *
680          * For single bdi case, the dirty pages are observed to fluctuate
681          * regularly within range
682          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
683          * for various filesystems, where (2) can yield in a reasonable 12.5%
684          * fluctuation range for pos_ratio.
685          *
686          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
687          * own size, so move the slope over accordingly and choose a slope that
688          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
689          */
690         if (unlikely(bdi_thresh > thresh))
691                 bdi_thresh = thresh;
692         /*
693          * It's very possible that bdi_thresh is close to 0 not because the
694          * device is slow, but that it has remained inactive for long time.
695          * Honour such devices a reasonable good (hopefully IO efficient)
696          * threshold, so that the occasional writes won't be blocked and active
697          * writes can rampup the threshold quickly.
698          */
699         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
700         /*
701          * scale global setpoint to bdi's:
702          *      bdi_setpoint = setpoint * bdi_thresh / thresh
703          */
704         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
705         bdi_setpoint = setpoint * (u64)x >> 16;
706         /*
707          * Use span=(8*write_bw) in single bdi case as indicated by
708          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
709          *
710          *        bdi_thresh                    thresh - bdi_thresh
711          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
712          *          thresh                            thresh
713          */
714         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
715         x_intercept = bdi_setpoint + span;
716
717         if (bdi_dirty < x_intercept - span / 4) {
718                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
719                                     x_intercept - bdi_setpoint + 1);
720         } else
721                 pos_ratio /= 4;
722
723         /*
724          * bdi reserve area, safeguard against dirty pool underrun and disk idle
725          * It may push the desired control point of global dirty pages higher
726          * than setpoint.
727          */
728         x_intercept = bdi_thresh / 2;
729         if (bdi_dirty < x_intercept) {
730                 if (bdi_dirty > x_intercept / 8)
731                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
732                 else
733                         pos_ratio *= 8;
734         }
735
736         return pos_ratio;
737 }
738
739 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
740                                        unsigned long elapsed,
741                                        unsigned long written)
742 {
743         const unsigned long period = roundup_pow_of_two(3 * HZ);
744         unsigned long avg = bdi->avg_write_bandwidth;
745         unsigned long old = bdi->write_bandwidth;
746         u64 bw;
747
748         /*
749          * bw = written * HZ / elapsed
750          *
751          *                   bw * elapsed + write_bandwidth * (period - elapsed)
752          * write_bandwidth = ---------------------------------------------------
753          *                                          period
754          */
755         bw = written - bdi->written_stamp;
756         bw *= HZ;
757         if (unlikely(elapsed > period)) {
758                 do_div(bw, elapsed);
759                 avg = bw;
760                 goto out;
761         }
762         bw += (u64)bdi->write_bandwidth * (period - elapsed);
763         bw >>= ilog2(period);
764
765         /*
766          * one more level of smoothing, for filtering out sudden spikes
767          */
768         if (avg > old && old >= (unsigned long)bw)
769                 avg -= (avg - old) >> 3;
770
771         if (avg < old && old <= (unsigned long)bw)
772                 avg += (old - avg) >> 3;
773
774 out:
775         bdi->write_bandwidth = bw;
776         bdi->avg_write_bandwidth = avg;
777 }
778
779 /*
780  * The global dirtyable memory and dirty threshold could be suddenly knocked
781  * down by a large amount (eg. on the startup of KVM in a swapless system).
782  * This may throw the system into deep dirty exceeded state and throttle
783  * heavy/light dirtiers alike. To retain good responsiveness, maintain
784  * global_dirty_limit for tracking slowly down to the knocked down dirty
785  * threshold.
786  */
787 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
788 {
789         unsigned long limit = global_dirty_limit;
790
791         /*
792          * Follow up in one step.
793          */
794         if (limit < thresh) {
795                 limit = thresh;
796                 goto update;
797         }
798
799         /*
800          * Follow down slowly. Use the higher one as the target, because thresh
801          * may drop below dirty. This is exactly the reason to introduce
802          * global_dirty_limit which is guaranteed to lie above the dirty pages.
803          */
804         thresh = max(thresh, dirty);
805         if (limit > thresh) {
806                 limit -= (limit - thresh) >> 5;
807                 goto update;
808         }
809         return;
810 update:
811         global_dirty_limit = limit;
812 }
813
814 static void global_update_bandwidth(unsigned long thresh,
815                                     unsigned long dirty,
816                                     unsigned long now)
817 {
818         static DEFINE_SPINLOCK(dirty_lock);
819         static unsigned long update_time;
820
821         /*
822          * check locklessly first to optimize away locking for the most time
823          */
824         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
825                 return;
826
827         spin_lock(&dirty_lock);
828         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
829                 update_dirty_limit(thresh, dirty);
830                 update_time = now;
831         }
832         spin_unlock(&dirty_lock);
833 }
834
835 /*
836  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
837  *
838  * Normal bdi tasks will be curbed at or below it in long term.
839  * Obviously it should be around (write_bw / N) when there are N dd tasks.
840  */
841 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
842                                        unsigned long thresh,
843                                        unsigned long bg_thresh,
844                                        unsigned long dirty,
845                                        unsigned long bdi_thresh,
846                                        unsigned long bdi_dirty,
847                                        unsigned long dirtied,
848                                        unsigned long elapsed)
849 {
850         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
851         unsigned long limit = hard_dirty_limit(thresh);
852         unsigned long setpoint = (freerun + limit) / 2;
853         unsigned long write_bw = bdi->avg_write_bandwidth;
854         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
855         unsigned long dirty_rate;
856         unsigned long task_ratelimit;
857         unsigned long balanced_dirty_ratelimit;
858         unsigned long pos_ratio;
859         unsigned long step;
860         unsigned long x;
861
862         /*
863          * The dirty rate will match the writeout rate in long term, except
864          * when dirty pages are truncated by userspace or re-dirtied by FS.
865          */
866         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
867
868         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
869                                        bdi_thresh, bdi_dirty);
870         /*
871          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
872          */
873         task_ratelimit = (u64)dirty_ratelimit *
874                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
875         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
876
877         /*
878          * A linear estimation of the "balanced" throttle rate. The theory is,
879          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
880          * dirty_rate will be measured to be (N * task_ratelimit). So the below
881          * formula will yield the balanced rate limit (write_bw / N).
882          *
883          * Note that the expanded form is not a pure rate feedback:
884          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
885          * but also takes pos_ratio into account:
886          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
887          *
888          * (1) is not realistic because pos_ratio also takes part in balancing
889          * the dirty rate.  Consider the state
890          *      pos_ratio = 0.5                                              (3)
891          *      rate = 2 * (write_bw / N)                                    (4)
892          * If (1) is used, it will stuck in that state! Because each dd will
893          * be throttled at
894          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
895          * yielding
896          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
897          * put (6) into (1) we get
898          *      rate_(i+1) = rate_(i)                                        (7)
899          *
900          * So we end up using (2) to always keep
901          *      rate_(i+1) ~= (write_bw / N)                                 (8)
902          * regardless of the value of pos_ratio. As long as (8) is satisfied,
903          * pos_ratio is able to drive itself to 1.0, which is not only where
904          * the dirty count meet the setpoint, but also where the slope of
905          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
906          */
907         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
908                                            dirty_rate | 1);
909         /*
910          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
911          */
912         if (unlikely(balanced_dirty_ratelimit > write_bw))
913                 balanced_dirty_ratelimit = write_bw;
914
915         /*
916          * We could safely do this and return immediately:
917          *
918          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
919          *
920          * However to get a more stable dirty_ratelimit, the below elaborated
921          * code makes use of task_ratelimit to filter out sigular points and
922          * limit the step size.
923          *
924          * The below code essentially only uses the relative value of
925          *
926          *      task_ratelimit - dirty_ratelimit
927          *      = (pos_ratio - 1) * dirty_ratelimit
928          *
929          * which reflects the direction and size of dirty position error.
930          */
931
932         /*
933          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
934          * task_ratelimit is on the same side of dirty_ratelimit, too.
935          * For example, when
936          * - dirty_ratelimit > balanced_dirty_ratelimit
937          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
938          * lowering dirty_ratelimit will help meet both the position and rate
939          * control targets. Otherwise, don't update dirty_ratelimit if it will
940          * only help meet the rate target. After all, what the users ultimately
941          * feel and care are stable dirty rate and small position error.
942          *
943          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
944          * and filter out the sigular points of balanced_dirty_ratelimit. Which
945          * keeps jumping around randomly and can even leap far away at times
946          * due to the small 200ms estimation period of dirty_rate (we want to
947          * keep that period small to reduce time lags).
948          */
949         step = 0;
950         if (dirty < setpoint) {
951                 x = min(bdi->balanced_dirty_ratelimit,
952                          min(balanced_dirty_ratelimit, task_ratelimit));
953                 if (dirty_ratelimit < x)
954                         step = x - dirty_ratelimit;
955         } else {
956                 x = max(bdi->balanced_dirty_ratelimit,
957                          max(balanced_dirty_ratelimit, task_ratelimit));
958                 if (dirty_ratelimit > x)
959                         step = dirty_ratelimit - x;
960         }
961
962         /*
963          * Don't pursue 100% rate matching. It's impossible since the balanced
964          * rate itself is constantly fluctuating. So decrease the track speed
965          * when it gets close to the target. Helps eliminate pointless tremors.
966          */
967         step >>= dirty_ratelimit / (2 * step + 1);
968         /*
969          * Limit the tracking speed to avoid overshooting.
970          */
971         step = (step + 7) / 8;
972
973         if (dirty_ratelimit < balanced_dirty_ratelimit)
974                 dirty_ratelimit += step;
975         else
976                 dirty_ratelimit -= step;
977
978         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
979         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
980
981         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
982 }
983
984 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
985                             unsigned long thresh,
986                             unsigned long bg_thresh,
987                             unsigned long dirty,
988                             unsigned long bdi_thresh,
989                             unsigned long bdi_dirty,
990                             unsigned long start_time)
991 {
992         unsigned long now = jiffies;
993         unsigned long elapsed = now - bdi->bw_time_stamp;
994         unsigned long dirtied;
995         unsigned long written;
996
997         /*
998          * rate-limit, only update once every 200ms.
999          */
1000         if (elapsed < BANDWIDTH_INTERVAL)
1001                 return;
1002
1003         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1004         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1005
1006         /*
1007          * Skip quiet periods when disk bandwidth is under-utilized.
1008          * (at least 1s idle time between two flusher runs)
1009          */
1010         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1011                 goto snapshot;
1012
1013         if (thresh) {
1014                 global_update_bandwidth(thresh, dirty, now);
1015                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1016                                            bdi_thresh, bdi_dirty,
1017                                            dirtied, elapsed);
1018         }
1019         bdi_update_write_bandwidth(bdi, elapsed, written);
1020
1021 snapshot:
1022         bdi->dirtied_stamp = dirtied;
1023         bdi->written_stamp = written;
1024         bdi->bw_time_stamp = now;
1025 }
1026
1027 static 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         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1036                 return;
1037         spin_lock(&bdi->wb.list_lock);
1038         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1039                                bdi_thresh, bdi_dirty, start_time);
1040         spin_unlock(&bdi->wb.list_lock);
1041 }
1042
1043 /*
1044  * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1045  * will look to see if it needs to start dirty throttling.
1046  *
1047  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1048  * global_page_state() too often. So scale it near-sqrt to the safety margin
1049  * (the number of pages we may dirty without exceeding the dirty limits).
1050  */
1051 static unsigned long dirty_poll_interval(unsigned long dirty,
1052                                          unsigned long thresh)
1053 {
1054         if (thresh > dirty)
1055                 return 1UL << (ilog2(thresh - dirty) >> 1);
1056
1057         return 1;
1058 }
1059
1060 static long bdi_max_pause(struct backing_dev_info *bdi,
1061                           unsigned long bdi_dirty)
1062 {
1063         long bw = bdi->avg_write_bandwidth;
1064         long t;
1065
1066         /*
1067          * Limit pause time for small memory systems. If sleeping for too long
1068          * time, a small pool of dirty/writeback pages may go empty and disk go
1069          * idle.
1070          *
1071          * 8 serves as the safety ratio.
1072          */
1073         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1074         t++;
1075
1076         return min_t(long, t, MAX_PAUSE);
1077 }
1078
1079 static long bdi_min_pause(struct backing_dev_info *bdi,
1080                           long max_pause,
1081                           unsigned long task_ratelimit,
1082                           unsigned long dirty_ratelimit,
1083                           int *nr_dirtied_pause)
1084 {
1085         long hi = ilog2(bdi->avg_write_bandwidth);
1086         long lo = ilog2(bdi->dirty_ratelimit);
1087         long t;         /* target pause */
1088         long pause;     /* estimated next pause */
1089         int pages;      /* target nr_dirtied_pause */
1090
1091         /* target for 10ms pause on 1-dd case */
1092         t = max(1, HZ / 100);
1093
1094         /*
1095          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1096          * overheads.
1097          *
1098          * (N * 10ms) on 2^N concurrent tasks.
1099          */
1100         if (hi > lo)
1101                 t += (hi - lo) * (10 * HZ) / 1024;
1102
1103         /*
1104          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1105          * on the much more stable dirty_ratelimit. However the next pause time
1106          * will be computed based on task_ratelimit and the two rate limits may
1107          * depart considerably at some time. Especially if task_ratelimit goes
1108          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1109          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1110          * result task_ratelimit won't be executed faithfully, which could
1111          * eventually bring down dirty_ratelimit.
1112          *
1113          * We apply two rules to fix it up:
1114          * 1) try to estimate the next pause time and if necessary, use a lower
1115          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1116          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1117          * 2) limit the target pause time to max_pause/2, so that the normal
1118          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1119          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1120          */
1121         t = min(t, 1 + max_pause / 2);
1122         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1123
1124         /*
1125          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1126          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1127          * When the 16 consecutive reads are often interrupted by some dirty
1128          * throttling pause during the async writes, cfq will go into idles
1129          * (deadline is fine). So push nr_dirtied_pause as high as possible
1130          * until reaches DIRTY_POLL_THRESH=32 pages.
1131          */
1132         if (pages < DIRTY_POLL_THRESH) {
1133                 t = max_pause;
1134                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1135                 if (pages > DIRTY_POLL_THRESH) {
1136                         pages = DIRTY_POLL_THRESH;
1137                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1138                 }
1139         }
1140
1141         pause = HZ * pages / (task_ratelimit + 1);
1142         if (pause > max_pause) {
1143                 t = max_pause;
1144                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1145         }
1146
1147         *nr_dirtied_pause = pages;
1148         /*
1149          * The minimal pause time will normally be half the target pause time.
1150          */
1151         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1152 }
1153
1154 /*
1155  * balance_dirty_pages() must be called by processes which are generating dirty
1156  * data.  It looks at the number of dirty pages in the machine and will force
1157  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1158  * If we're over `background_thresh' then the writeback threads are woken to
1159  * perform some writeout.
1160  */
1161 static void balance_dirty_pages(struct address_space *mapping,
1162                                 unsigned long pages_dirtied)
1163 {
1164         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1165         unsigned long bdi_reclaimable;
1166         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1167         unsigned long bdi_dirty;
1168         unsigned long freerun;
1169         unsigned long background_thresh;
1170         unsigned long dirty_thresh;
1171         unsigned long bdi_thresh;
1172         long period;
1173         long pause;
1174         long max_pause;
1175         long min_pause;
1176         int nr_dirtied_pause;
1177         bool dirty_exceeded = false;
1178         unsigned long task_ratelimit;
1179         unsigned long dirty_ratelimit;
1180         unsigned long pos_ratio;
1181         struct backing_dev_info *bdi = mapping->backing_dev_info;
1182         unsigned long start_time = jiffies;
1183
1184         for (;;) {
1185                 unsigned long now = jiffies;
1186
1187                 /*
1188                  * Unstable writes are a feature of certain networked
1189                  * filesystems (i.e. NFS) in which data may have been
1190                  * written to the server's write cache, but has not yet
1191                  * been flushed to permanent storage.
1192                  */
1193                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1194                                         global_page_state(NR_UNSTABLE_NFS);
1195                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1196
1197                 global_dirty_limits(&background_thresh, &dirty_thresh);
1198
1199                 /*
1200                  * Throttle it only when the background writeback cannot
1201                  * catch-up. This avoids (excessively) small writeouts
1202                  * when the bdi limits are ramping up.
1203                  */
1204                 freerun = dirty_freerun_ceiling(dirty_thresh,
1205                                                 background_thresh);
1206                 if (nr_dirty <= freerun) {
1207                         current->dirty_paused_when = now;
1208                         current->nr_dirtied = 0;
1209                         current->nr_dirtied_pause =
1210                                 dirty_poll_interval(nr_dirty, dirty_thresh);
1211                         break;
1212                 }
1213
1214                 if (unlikely(!writeback_in_progress(bdi)))
1215                         bdi_start_background_writeback(bdi);
1216
1217                 /*
1218                  * bdi_thresh is not treated as some limiting factor as
1219                  * dirty_thresh, due to reasons
1220                  * - in JBOD setup, bdi_thresh can fluctuate a lot
1221                  * - in a system with HDD and USB key, the USB key may somehow
1222                  *   go into state (bdi_dirty >> bdi_thresh) either because
1223                  *   bdi_dirty starts high, or because bdi_thresh drops low.
1224                  *   In this case we don't want to hard throttle the USB key
1225                  *   dirtiers for 100 seconds until bdi_dirty drops under
1226                  *   bdi_thresh. Instead the auxiliary bdi control line in
1227                  *   bdi_position_ratio() will let the dirtier task progress
1228                  *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1229                  */
1230                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1231
1232                 /*
1233                  * In order to avoid the stacked BDI deadlock we need
1234                  * to ensure we accurately count the 'dirty' pages when
1235                  * the threshold is low.
1236                  *
1237                  * Otherwise it would be possible to get thresh+n pages
1238                  * reported dirty, even though there are thresh-m pages
1239                  * actually dirty; with m+n sitting in the percpu
1240                  * deltas.
1241                  */
1242                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1243                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1244                         bdi_dirty = bdi_reclaimable +
1245                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
1246                 } else {
1247                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1248                         bdi_dirty = bdi_reclaimable +
1249                                     bdi_stat(bdi, BDI_WRITEBACK);
1250                 }
1251
1252                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1253                                   (nr_dirty > dirty_thresh);
1254                 if (dirty_exceeded && !bdi->dirty_exceeded)
1255                         bdi->dirty_exceeded = 1;
1256
1257                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1258                                      nr_dirty, bdi_thresh, bdi_dirty,
1259                                      start_time);
1260
1261                 dirty_ratelimit = bdi->dirty_ratelimit;
1262                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1263                                                background_thresh, nr_dirty,
1264                                                bdi_thresh, bdi_dirty);
1265                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1266                                                         RATELIMIT_CALC_SHIFT;
1267                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1268                 min_pause = bdi_min_pause(bdi, max_pause,
1269                                           task_ratelimit, dirty_ratelimit,
1270                                           &nr_dirtied_pause);
1271
1272                 if (unlikely(task_ratelimit == 0)) {
1273                         period = max_pause;
1274                         pause = max_pause;
1275                         goto pause;
1276                 }
1277                 period = HZ * pages_dirtied / task_ratelimit;
1278                 pause = period;
1279                 if (current->dirty_paused_when)
1280                         pause -= now - current->dirty_paused_when;
1281                 /*
1282                  * For less than 1s think time (ext3/4 may block the dirtier
1283                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1284                  * however at much less frequency), try to compensate it in
1285                  * future periods by updating the virtual time; otherwise just
1286                  * do a reset, as it may be a light dirtier.
1287                  */
1288                 if (pause < min_pause) {
1289                         trace_balance_dirty_pages(bdi,
1290                                                   dirty_thresh,
1291                                                   background_thresh,
1292                                                   nr_dirty,
1293                                                   bdi_thresh,
1294                                                   bdi_dirty,
1295                                                   dirty_ratelimit,
1296                                                   task_ratelimit,
1297                                                   pages_dirtied,
1298                                                   period,
1299                                                   min(pause, 0L),
1300                                                   start_time);
1301                         if (pause < -HZ) {
1302                                 current->dirty_paused_when = now;
1303                                 current->nr_dirtied = 0;
1304                         } else if (period) {
1305                                 current->dirty_paused_when += period;
1306                                 current->nr_dirtied = 0;
1307                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1308                                 current->nr_dirtied_pause += pages_dirtied;
1309                         break;
1310                 }
1311                 if (unlikely(pause > max_pause)) {
1312                         /* for occasional dropped task_ratelimit */
1313                         now += min(pause - max_pause, max_pause);
1314                         pause = max_pause;
1315                 }
1316
1317 pause:
1318                 trace_balance_dirty_pages(bdi,
1319                                           dirty_thresh,
1320                                           background_thresh,
1321                                           nr_dirty,
1322                                           bdi_thresh,
1323                                           bdi_dirty,
1324                                           dirty_ratelimit,
1325                                           task_ratelimit,
1326                                           pages_dirtied,
1327                                           period,
1328                                           pause,
1329                                           start_time);
1330                 __set_current_state(TASK_KILLABLE);
1331                 io_schedule_timeout(pause);
1332
1333                 current->dirty_paused_when = now + pause;
1334                 current->nr_dirtied = 0;
1335                 current->nr_dirtied_pause = nr_dirtied_pause;
1336
1337                 /*
1338                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1339                  * also keep "1000+ dd on a slow USB stick" under control.
1340                  */
1341                 if (task_ratelimit)
1342                         break;
1343
1344                 /*
1345                  * In the case of an unresponding NFS server and the NFS dirty
1346                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1347                  * to go through, so that tasks on them still remain responsive.
1348                  *
1349                  * In theory 1 page is enough to keep the comsumer-producer
1350                  * pipe going: the flusher cleans 1 page => the task dirties 1
1351                  * more page. However bdi_dirty has accounting errors.  So use
1352                  * the larger and more IO friendly bdi_stat_error.
1353                  */
1354                 if (bdi_dirty <= bdi_stat_error(bdi))
1355                         break;
1356
1357                 if (fatal_signal_pending(current))
1358                         break;
1359         }
1360
1361         if (!dirty_exceeded && bdi->dirty_exceeded)
1362                 bdi->dirty_exceeded = 0;
1363
1364         if (writeback_in_progress(bdi))
1365                 return;
1366
1367         /*
1368          * In laptop mode, we wait until hitting the higher threshold before
1369          * starting background writeout, and then write out all the way down
1370          * to the lower threshold.  So slow writers cause minimal disk activity.
1371          *
1372          * In normal mode, we start background writeout at the lower
1373          * background_thresh, to keep the amount of dirty memory low.
1374          */
1375         if (laptop_mode)
1376                 return;
1377
1378         if (nr_reclaimable > background_thresh)
1379                 bdi_start_background_writeback(bdi);
1380 }
1381
1382 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1383 {
1384         if (set_page_dirty(page) || page_mkwrite) {
1385                 struct address_space *mapping = page_mapping(page);
1386
1387                 if (mapping)
1388                         balance_dirty_pages_ratelimited(mapping);
1389         }
1390 }
1391
1392 static DEFINE_PER_CPU(int, bdp_ratelimits);
1393
1394 /*
1395  * Normal tasks are throttled by
1396  *      loop {
1397  *              dirty tsk->nr_dirtied_pause pages;
1398  *              take a snap in balance_dirty_pages();
1399  *      }
1400  * However there is a worst case. If every task exit immediately when dirtied
1401  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1402  * called to throttle the page dirties. The solution is to save the not yet
1403  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1404  * randomly into the running tasks. This works well for the above worst case,
1405  * as the new task will pick up and accumulate the old task's leaked dirty
1406  * count and eventually get throttled.
1407  */
1408 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1409
1410 /**
1411  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1412  * @mapping: address_space which was dirtied
1413  * @nr_pages_dirtied: number of pages which the caller has just dirtied
1414  *
1415  * Processes which are dirtying memory should call in here once for each page
1416  * which was newly dirtied.  The function will periodically check the system's
1417  * dirty state and will initiate writeback if needed.
1418  *
1419  * On really big machines, get_writeback_state is expensive, so try to avoid
1420  * calling it too often (ratelimiting).  But once we're over the dirty memory
1421  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1422  * from overshooting the limit by (ratelimit_pages) each.
1423  */
1424 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1425                                         unsigned long nr_pages_dirtied)
1426 {
1427         struct backing_dev_info *bdi = mapping->backing_dev_info;
1428         int ratelimit;
1429         int *p;
1430
1431         if (!bdi_cap_account_dirty(bdi))
1432                 return;
1433
1434         ratelimit = current->nr_dirtied_pause;
1435         if (bdi->dirty_exceeded)
1436                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1437
1438         preempt_disable();
1439         /*
1440          * This prevents one CPU to accumulate too many dirtied pages without
1441          * calling into balance_dirty_pages(), which can happen when there are
1442          * 1000+ tasks, all of them start dirtying pages at exactly the same
1443          * time, hence all honoured too large initial task->nr_dirtied_pause.
1444          */
1445         p =  &__get_cpu_var(bdp_ratelimits);
1446         if (unlikely(current->nr_dirtied >= ratelimit))
1447                 *p = 0;
1448         else if (unlikely(*p >= ratelimit_pages)) {
1449                 *p = 0;
1450                 ratelimit = 0;
1451         }
1452         /*
1453          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1454          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1455          * the dirty throttling and livelock other long-run dirtiers.
1456          */
1457         p = &__get_cpu_var(dirty_throttle_leaks);
1458         if (*p > 0 && current->nr_dirtied < ratelimit) {
1459                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1460                 *p -= nr_pages_dirtied;
1461                 current->nr_dirtied += nr_pages_dirtied;
1462         }
1463         preempt_enable();
1464
1465         if (unlikely(current->nr_dirtied >= ratelimit))
1466                 balance_dirty_pages(mapping, current->nr_dirtied);
1467 }
1468 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1469
1470 void throttle_vm_writeout(gfp_t gfp_mask)
1471 {
1472         unsigned long background_thresh;
1473         unsigned long dirty_thresh;
1474
1475         for ( ; ; ) {
1476                 global_dirty_limits(&background_thresh, &dirty_thresh);
1477                 dirty_thresh = hard_dirty_limit(dirty_thresh);
1478
1479                 /*
1480                  * Boost the allowable dirty threshold a bit for page
1481                  * allocators so they don't get DoS'ed by heavy writers
1482                  */
1483                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1484
1485                 if (global_page_state(NR_UNSTABLE_NFS) +
1486                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1487                                 break;
1488                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1489
1490                 /*
1491                  * The caller might hold locks which can prevent IO completion
1492                  * or progress in the filesystem.  So we cannot just sit here
1493                  * waiting for IO to complete.
1494                  */
1495                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1496                         break;
1497         }
1498 }
1499
1500 /*
1501  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1502  */
1503 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1504         void __user *buffer, size_t *length, loff_t *ppos)
1505 {
1506         proc_dointvec(table, write, buffer, length, ppos);
1507         bdi_arm_supers_timer();
1508         return 0;
1509 }
1510
1511 #ifdef CONFIG_BLOCK
1512 void laptop_mode_timer_fn(unsigned long data)
1513 {
1514         struct request_queue *q = (struct request_queue *)data;
1515         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1516                 global_page_state(NR_UNSTABLE_NFS);
1517
1518         /*
1519          * We want to write everything out, not just down to the dirty
1520          * threshold
1521          */
1522         if (bdi_has_dirty_io(&q->backing_dev_info))
1523                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1524                                         WB_REASON_LAPTOP_TIMER);
1525 }
1526
1527 /*
1528  * We've spun up the disk and we're in laptop mode: schedule writeback
1529  * of all dirty data a few seconds from now.  If the flush is already scheduled
1530  * then push it back - the user is still using the disk.
1531  */
1532 void laptop_io_completion(struct backing_dev_info *info)
1533 {
1534         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1535 }
1536
1537 /*
1538  * We're in laptop mode and we've just synced. The sync's writes will have
1539  * caused another writeback to be scheduled by laptop_io_completion.
1540  * Nothing needs to be written back anymore, so we unschedule the writeback.
1541  */
1542 void laptop_sync_completion(void)
1543 {
1544         struct backing_dev_info *bdi;
1545
1546         rcu_read_lock();
1547
1548         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1549                 del_timer(&bdi->laptop_mode_wb_timer);
1550
1551         rcu_read_unlock();
1552 }
1553 #endif
1554
1555 /*
1556  * If ratelimit_pages is too high then we can get into dirty-data overload
1557  * if a large number of processes all perform writes at the same time.
1558  * If it is too low then SMP machines will call the (expensive)
1559  * get_writeback_state too often.
1560  *
1561  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1562  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1563  * thresholds.
1564  */
1565
1566 void writeback_set_ratelimit(void)
1567 {
1568         unsigned long background_thresh;
1569         unsigned long dirty_thresh;
1570         global_dirty_limits(&background_thresh, &dirty_thresh);
1571         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1572         if (ratelimit_pages < 16)
1573                 ratelimit_pages = 16;
1574 }
1575
1576 static int __cpuinit
1577 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1578 {
1579         writeback_set_ratelimit();
1580         return NOTIFY_DONE;
1581 }
1582
1583 static struct notifier_block __cpuinitdata ratelimit_nb = {
1584         .notifier_call  = ratelimit_handler,
1585         .next           = NULL,
1586 };
1587
1588 /*
1589  * Called early on to tune the page writeback dirty limits.
1590  *
1591  * We used to scale dirty pages according to how total memory
1592  * related to pages that could be allocated for buffers (by
1593  * comparing nr_free_buffer_pages() to vm_total_pages.
1594  *
1595  * However, that was when we used "dirty_ratio" to scale with
1596  * all memory, and we don't do that any more. "dirty_ratio"
1597  * is now applied to total non-HIGHPAGE memory (by subtracting
1598  * totalhigh_pages from vm_total_pages), and as such we can't
1599  * get into the old insane situation any more where we had
1600  * large amounts of dirty pages compared to a small amount of
1601  * non-HIGHMEM memory.
1602  *
1603  * But we might still want to scale the dirty_ratio by how
1604  * much memory the box has..
1605  */
1606 void __init page_writeback_init(void)
1607 {
1608         int shift;
1609
1610         writeback_set_ratelimit();
1611         register_cpu_notifier(&ratelimit_nb);
1612
1613         shift = calc_period_shift();
1614         prop_descriptor_init(&vm_completions, shift);
1615 }
1616
1617 /**
1618  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1619  * @mapping: address space structure to write
1620  * @start: starting page index
1621  * @end: ending page index (inclusive)
1622  *
1623  * This function scans the page range from @start to @end (inclusive) and tags
1624  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1625  * that write_cache_pages (or whoever calls this function) will then use
1626  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1627  * used to avoid livelocking of writeback by a process steadily creating new
1628  * dirty pages in the file (thus it is important for this function to be quick
1629  * so that it can tag pages faster than a dirtying process can create them).
1630  */
1631 /*
1632  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1633  */
1634 void tag_pages_for_writeback(struct address_space *mapping,
1635                              pgoff_t start, pgoff_t end)
1636 {
1637 #define WRITEBACK_TAG_BATCH 4096
1638         unsigned long tagged;
1639
1640         do {
1641                 spin_lock_irq(&mapping->tree_lock);
1642                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1643                                 &start, end, WRITEBACK_TAG_BATCH,
1644                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1645                 spin_unlock_irq(&mapping->tree_lock);
1646                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1647                 cond_resched();
1648                 /* We check 'start' to handle wrapping when end == ~0UL */
1649         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1650 }
1651 EXPORT_SYMBOL(tag_pages_for_writeback);
1652
1653 /**
1654  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1655  * @mapping: address space structure to write
1656  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1657  * @writepage: function called for each page
1658  * @data: data passed to writepage function
1659  *
1660  * If a page is already under I/O, write_cache_pages() skips it, even
1661  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1662  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1663  * and msync() need to guarantee that all the data which was dirty at the time
1664  * the call was made get new I/O started against them.  If wbc->sync_mode is
1665  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1666  * existing IO to complete.
1667  *
1668  * To avoid livelocks (when other process dirties new pages), we first tag
1669  * pages which should be written back with TOWRITE tag and only then start
1670  * writing them. For data-integrity sync we have to be careful so that we do
1671  * not miss some pages (e.g., because some other process has cleared TOWRITE
1672  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1673  * by the process clearing the DIRTY tag (and submitting the page for IO).
1674  */
1675 int write_cache_pages(struct address_space *mapping,
1676                       struct writeback_control *wbc, writepage_t writepage,
1677                       void *data)
1678 {
1679         int ret = 0;
1680         int done = 0;
1681         struct pagevec pvec;
1682         int nr_pages;
1683         pgoff_t uninitialized_var(writeback_index);
1684         pgoff_t index;
1685         pgoff_t end;            /* Inclusive */
1686         pgoff_t done_index;
1687         int cycled;
1688         int range_whole = 0;
1689         int tag;
1690
1691         pagevec_init(&pvec, 0);
1692         if (wbc->range_cyclic) {
1693                 writeback_index = mapping->writeback_index; /* prev offset */
1694                 index = writeback_index;
1695                 if (index == 0)
1696                         cycled = 1;
1697                 else
1698                         cycled = 0;
1699                 end = -1;
1700         } else {
1701                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1702                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1703                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1704                         range_whole = 1;
1705                 cycled = 1; /* ignore range_cyclic tests */
1706         }
1707         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1708                 tag = PAGECACHE_TAG_TOWRITE;
1709         else
1710                 tag = PAGECACHE_TAG_DIRTY;
1711 retry:
1712         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1713                 tag_pages_for_writeback(mapping, index, end);
1714         done_index = index;
1715         while (!done && (index <= end)) {
1716                 int i;
1717
1718                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1719                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1720                 if (nr_pages == 0)
1721                         break;
1722
1723                 for (i = 0; i < nr_pages; i++) {
1724                         struct page *page = pvec.pages[i];
1725
1726                         /*
1727                          * At this point, the page may be truncated or
1728                          * invalidated (changing page->mapping to NULL), or
1729                          * even swizzled back from swapper_space to tmpfs file
1730                          * mapping. However, page->index will not change
1731                          * because we have a reference on the page.
1732                          */
1733                         if (page->index > end) {
1734                                 /*
1735                                  * can't be range_cyclic (1st pass) because
1736                                  * end == -1 in that case.
1737                                  */
1738                                 done = 1;
1739                                 break;
1740                         }
1741
1742                         done_index = page->index;
1743
1744                         lock_page(page);
1745
1746                         /*
1747                          * Page truncated or invalidated. We can freely skip it
1748                          * then, even for data integrity operations: the page
1749                          * has disappeared concurrently, so there could be no
1750                          * real expectation of this data interity operation
1751                          * even if there is now a new, dirty page at the same
1752                          * pagecache address.
1753                          */
1754                         if (unlikely(page->mapping != mapping)) {
1755 continue_unlock:
1756                                 unlock_page(page);
1757                                 continue;
1758                         }
1759
1760                         if (!PageDirty(page)) {
1761                                 /* someone wrote it for us */
1762                                 goto continue_unlock;
1763                         }
1764
1765                         if (PageWriteback(page)) {
1766                                 if (wbc->sync_mode != WB_SYNC_NONE)
1767                                         wait_on_page_writeback(page);
1768                                 else
1769                                         goto continue_unlock;
1770                         }
1771
1772                         BUG_ON(PageWriteback(page));
1773                         if (!clear_page_dirty_for_io(page))
1774                                 goto continue_unlock;
1775
1776                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1777                         ret = (*writepage)(page, wbc, data);
1778                         if (unlikely(ret)) {
1779                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1780                                         unlock_page(page);
1781                                         ret = 0;
1782                                 } else {
1783                                         /*
1784                                          * done_index is set past this page,
1785                                          * so media errors will not choke
1786                                          * background writeout for the entire
1787                                          * file. This has consequences for
1788                                          * range_cyclic semantics (ie. it may
1789                                          * not be suitable for data integrity
1790                                          * writeout).
1791                                          */
1792                                         done_index = page->index + 1;
1793                                         done = 1;
1794                                         break;
1795                                 }
1796                         }
1797
1798                         /*
1799                          * We stop writing back only if we are not doing
1800                          * integrity sync. In case of integrity sync we have to
1801                          * keep going until we have written all the pages
1802                          * we tagged for writeback prior to entering this loop.
1803                          */
1804                         if (--wbc->nr_to_write <= 0 &&
1805                             wbc->sync_mode == WB_SYNC_NONE) {
1806                                 done = 1;
1807                                 break;
1808                         }
1809                 }
1810                 pagevec_release(&pvec);
1811                 cond_resched();
1812         }
1813         if (!cycled && !done) {
1814                 /*
1815                  * range_cyclic:
1816                  * We hit the last page and there is more work to be done: wrap
1817                  * back to the start of the file
1818                  */
1819                 cycled = 1;
1820                 index = 0;
1821                 end = writeback_index - 1;
1822                 goto retry;
1823         }
1824         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1825                 mapping->writeback_index = done_index;
1826
1827         return ret;
1828 }
1829 EXPORT_SYMBOL(write_cache_pages);
1830
1831 /*
1832  * Function used by generic_writepages to call the real writepage
1833  * function and set the mapping flags on error
1834  */
1835 static int __writepage(struct page *page, struct writeback_control *wbc,
1836                        void *data)
1837 {
1838         struct address_space *mapping = data;
1839         int ret = mapping->a_ops->writepage(page, wbc);
1840         mapping_set_error(mapping, ret);
1841         return ret;
1842 }
1843
1844 /**
1845  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1846  * @mapping: address space structure to write
1847  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1848  *
1849  * This is a library function, which implements the writepages()
1850  * address_space_operation.
1851  */
1852 int generic_writepages(struct address_space *mapping,
1853                        struct writeback_control *wbc)
1854 {
1855         struct blk_plug plug;
1856         int ret;
1857
1858         /* deal with chardevs and other special file */
1859         if (!mapping->a_ops->writepage)
1860                 return 0;
1861
1862         blk_start_plug(&plug);
1863         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1864         blk_finish_plug(&plug);
1865         return ret;
1866 }
1867
1868 EXPORT_SYMBOL(generic_writepages);
1869
1870 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1871 {
1872         int ret;
1873
1874         if (wbc->nr_to_write <= 0)
1875                 return 0;
1876         if (mapping->a_ops->writepages)
1877                 ret = mapping->a_ops->writepages(mapping, wbc);
1878         else
1879                 ret = generic_writepages(mapping, wbc);
1880         return ret;
1881 }
1882
1883 /**
1884  * write_one_page - write out a single page and optionally wait on I/O
1885  * @page: the page to write
1886  * @wait: if true, wait on writeout
1887  *
1888  * The page must be locked by the caller and will be unlocked upon return.
1889  *
1890  * write_one_page() returns a negative error code if I/O failed.
1891  */
1892 int write_one_page(struct page *page, int wait)
1893 {
1894         struct address_space *mapping = page->mapping;
1895         int ret = 0;
1896         struct writeback_control wbc = {
1897                 .sync_mode = WB_SYNC_ALL,
1898                 .nr_to_write = 1,
1899         };
1900
1901         BUG_ON(!PageLocked(page));
1902
1903         if (wait)
1904                 wait_on_page_writeback(page);
1905
1906         if (clear_page_dirty_for_io(page)) {
1907                 page_cache_get(page);
1908                 ret = mapping->a_ops->writepage(page, &wbc);
1909                 if (ret == 0 && wait) {
1910                         wait_on_page_writeback(page);
1911                         if (PageError(page))
1912                                 ret = -EIO;
1913                 }
1914                 page_cache_release(page);
1915         } else {
1916                 unlock_page(page);
1917         }
1918         return ret;
1919 }
1920 EXPORT_SYMBOL(write_one_page);
1921
1922 /*
1923  * For address_spaces which do not use buffers nor write back.
1924  */
1925 int __set_page_dirty_no_writeback(struct page *page)
1926 {
1927         if (!PageDirty(page))
1928                 return !TestSetPageDirty(page);
1929         return 0;
1930 }
1931
1932 /*
1933  * Helper function for set_page_dirty family.
1934  * NOTE: This relies on being atomic wrt interrupts.
1935  */
1936 void account_page_dirtied(struct page *page, struct address_space *mapping)
1937 {
1938         if (mapping_cap_account_dirty(mapping)) {
1939                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1940                 __inc_zone_page_state(page, NR_DIRTIED);
1941                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1942                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1943                 task_io_account_write(PAGE_CACHE_SIZE);
1944                 current->nr_dirtied++;
1945                 this_cpu_inc(bdp_ratelimits);
1946         }
1947 }
1948 EXPORT_SYMBOL(account_page_dirtied);
1949
1950 /*
1951  * Helper function for set_page_writeback family.
1952  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1953  * wrt interrupts.
1954  */
1955 void account_page_writeback(struct page *page)
1956 {
1957         inc_zone_page_state(page, NR_WRITEBACK);
1958 }
1959 EXPORT_SYMBOL(account_page_writeback);
1960
1961 /*
1962  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1963  * its radix tree.
1964  *
1965  * This is also used when a single buffer is being dirtied: we want to set the
1966  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1967  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1968  *
1969  * Most callers have locked the page, which pins the address_space in memory.
1970  * But zap_pte_range() does not lock the page, however in that case the
1971  * mapping is pinned by the vma's ->vm_file reference.
1972  *
1973  * We take care to handle the case where the page was truncated from the
1974  * mapping by re-checking page_mapping() inside tree_lock.
1975  */
1976 int __set_page_dirty_nobuffers(struct page *page)
1977 {
1978         if (!TestSetPageDirty(page)) {
1979                 struct address_space *mapping = page_mapping(page);
1980                 struct address_space *mapping2;
1981
1982                 if (!mapping)
1983                         return 1;
1984
1985                 spin_lock_irq(&mapping->tree_lock);
1986                 mapping2 = page_mapping(page);
1987                 if (mapping2) { /* Race with truncate? */
1988                         BUG_ON(mapping2 != mapping);
1989                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1990                         account_page_dirtied(page, mapping);
1991                         radix_tree_tag_set(&mapping->page_tree,
1992                                 page_index(page), PAGECACHE_TAG_DIRTY);
1993                 }
1994                 spin_unlock_irq(&mapping->tree_lock);
1995                 if (mapping->host) {
1996                         /* !PageAnon && !swapper_space */
1997                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1998                 }
1999                 return 1;
2000         }
2001         return 0;
2002 }
2003 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2004
2005 /*
2006  * Call this whenever redirtying a page, to de-account the dirty counters
2007  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2008  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2009  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2010  * control.
2011  */
2012 void account_page_redirty(struct page *page)
2013 {
2014         struct address_space *mapping = page->mapping;
2015         if (mapping && mapping_cap_account_dirty(mapping)) {
2016                 current->nr_dirtied--;
2017                 dec_zone_page_state(page, NR_DIRTIED);
2018                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2019         }
2020 }
2021 EXPORT_SYMBOL(account_page_redirty);
2022
2023 /*
2024  * When a writepage implementation decides that it doesn't want to write this
2025  * page for some reason, it should redirty the locked page via
2026  * redirty_page_for_writepage() and it should then unlock the page and return 0
2027  */
2028 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2029 {
2030         wbc->pages_skipped++;
2031         account_page_redirty(page);
2032         return __set_page_dirty_nobuffers(page);
2033 }
2034 EXPORT_SYMBOL(redirty_page_for_writepage);
2035
2036 /*
2037  * Dirty a page.
2038  *
2039  * For pages with a mapping this should be done under the page lock
2040  * for the benefit of asynchronous memory errors who prefer a consistent
2041  * dirty state. This rule can be broken in some special cases,
2042  * but should be better not to.
2043  *
2044  * If the mapping doesn't provide a set_page_dirty a_op, then
2045  * just fall through and assume that it wants buffer_heads.
2046  */
2047 int set_page_dirty(struct page *page)
2048 {
2049         struct address_space *mapping = page_mapping(page);
2050
2051         if (likely(mapping)) {
2052                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2053                 /*
2054                  * readahead/lru_deactivate_page could remain
2055                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2056                  * About readahead, if the page is written, the flags would be
2057                  * reset. So no problem.
2058                  * About lru_deactivate_page, if the page is redirty, the flag
2059                  * will be reset. So no problem. but if the page is used by readahead
2060                  * it will confuse readahead and make it restart the size rampup
2061                  * process. But it's a trivial problem.
2062                  */
2063                 ClearPageReclaim(page);
2064 #ifdef CONFIG_BLOCK
2065                 if (!spd)
2066                         spd = __set_page_dirty_buffers;
2067 #endif
2068                 return (*spd)(page);
2069         }
2070         if (!PageDirty(page)) {
2071                 if (!TestSetPageDirty(page))
2072                         return 1;
2073         }
2074         return 0;
2075 }
2076 EXPORT_SYMBOL(set_page_dirty);
2077
2078 /*
2079  * set_page_dirty() is racy if the caller has no reference against
2080  * page->mapping->host, and if the page is unlocked.  This is because another
2081  * CPU could truncate the page off the mapping and then free the mapping.
2082  *
2083  * Usually, the page _is_ locked, or the caller is a user-space process which
2084  * holds a reference on the inode by having an open file.
2085  *
2086  * In other cases, the page should be locked before running set_page_dirty().
2087  */
2088 int set_page_dirty_lock(struct page *page)
2089 {
2090         int ret;
2091
2092         lock_page(page);
2093         ret = set_page_dirty(page);
2094         unlock_page(page);
2095         return ret;
2096 }
2097 EXPORT_SYMBOL(set_page_dirty_lock);
2098
2099 /*
2100  * Clear a page's dirty flag, while caring for dirty memory accounting.
2101  * Returns true if the page was previously dirty.
2102  *
2103  * This is for preparing to put the page under writeout.  We leave the page
2104  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2105  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2106  * implementation will run either set_page_writeback() or set_page_dirty(),
2107  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2108  * back into sync.
2109  *
2110  * This incoherency between the page's dirty flag and radix-tree tag is
2111  * unfortunate, but it only exists while the page is locked.
2112  */
2113 int clear_page_dirty_for_io(struct page *page)
2114 {
2115         struct address_space *mapping = page_mapping(page);
2116
2117         BUG_ON(!PageLocked(page));
2118
2119         if (mapping && mapping_cap_account_dirty(mapping)) {
2120                 /*
2121                  * Yes, Virginia, this is indeed insane.
2122                  *
2123                  * We use this sequence to make sure that
2124                  *  (a) we account for dirty stats properly
2125                  *  (b) we tell the low-level filesystem to
2126                  *      mark the whole page dirty if it was
2127                  *      dirty in a pagetable. Only to then
2128                  *  (c) clean the page again and return 1 to
2129                  *      cause the writeback.
2130                  *
2131                  * This way we avoid all nasty races with the
2132                  * dirty bit in multiple places and clearing
2133                  * them concurrently from different threads.
2134                  *
2135                  * Note! Normally the "set_page_dirty(page)"
2136                  * has no effect on the actual dirty bit - since
2137                  * that will already usually be set. But we
2138                  * need the side effects, and it can help us
2139                  * avoid races.
2140                  *
2141                  * We basically use the page "master dirty bit"
2142                  * as a serialization point for all the different
2143                  * threads doing their things.
2144                  */
2145                 if (page_mkclean(page))
2146                         set_page_dirty(page);
2147                 /*
2148                  * We carefully synchronise fault handlers against
2149                  * installing a dirty pte and marking the page dirty
2150                  * at this point. We do this by having them hold the
2151                  * page lock at some point after installing their
2152                  * pte, but before marking the page dirty.
2153                  * Pages are always locked coming in here, so we get
2154                  * the desired exclusion. See mm/memory.c:do_wp_page()
2155                  * for more comments.
2156                  */
2157                 if (TestClearPageDirty(page)) {
2158                         dec_zone_page_state(page, NR_FILE_DIRTY);
2159                         dec_bdi_stat(mapping->backing_dev_info,
2160                                         BDI_RECLAIMABLE);
2161                         return 1;
2162                 }
2163                 return 0;
2164         }
2165         return TestClearPageDirty(page);
2166 }
2167 EXPORT_SYMBOL(clear_page_dirty_for_io);
2168
2169 int test_clear_page_writeback(struct page *page)
2170 {
2171         struct address_space *mapping = page_mapping(page);
2172         int ret;
2173
2174         if (mapping) {
2175                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2176                 unsigned long flags;
2177
2178                 spin_lock_irqsave(&mapping->tree_lock, flags);
2179                 ret = TestClearPageWriteback(page);
2180                 if (ret) {
2181                         radix_tree_tag_clear(&mapping->page_tree,
2182                                                 page_index(page),
2183                                                 PAGECACHE_TAG_WRITEBACK);
2184                         if (bdi_cap_account_writeback(bdi)) {
2185                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2186                                 __bdi_writeout_inc(bdi);
2187                         }
2188                 }
2189                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2190         } else {
2191                 ret = TestClearPageWriteback(page);
2192         }
2193         if (ret) {
2194                 dec_zone_page_state(page, NR_WRITEBACK);
2195                 inc_zone_page_state(page, NR_WRITTEN);
2196         }
2197         return ret;
2198 }
2199
2200 int test_set_page_writeback(struct page *page)
2201 {
2202         struct address_space *mapping = page_mapping(page);
2203         int ret;
2204
2205         if (mapping) {
2206                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2207                 unsigned long flags;
2208
2209                 spin_lock_irqsave(&mapping->tree_lock, flags);
2210                 ret = TestSetPageWriteback(page);
2211                 if (!ret) {
2212                         radix_tree_tag_set(&mapping->page_tree,
2213                                                 page_index(page),
2214                                                 PAGECACHE_TAG_WRITEBACK);
2215                         if (bdi_cap_account_writeback(bdi))
2216                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2217                 }
2218                 if (!PageDirty(page))
2219                         radix_tree_tag_clear(&mapping->page_tree,
2220                                                 page_index(page),
2221                                                 PAGECACHE_TAG_DIRTY);
2222                 radix_tree_tag_clear(&mapping->page_tree,
2223                                      page_index(page),
2224                                      PAGECACHE_TAG_TOWRITE);
2225                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2226         } else {
2227                 ret = TestSetPageWriteback(page);
2228         }
2229         if (!ret)
2230                 account_page_writeback(page);
2231         return ret;
2232
2233 }
2234 EXPORT_SYMBOL(test_set_page_writeback);
2235
2236 /*
2237  * Return true if any of the pages in the mapping are marked with the
2238  * passed tag.
2239  */
2240 int mapping_tagged(struct address_space *mapping, int tag)
2241 {
2242         return radix_tree_tagged(&mapping->page_tree, tag);
2243 }
2244 EXPORT_SYMBOL(mapping_tagged);