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