writeback: stop background writeback when below background threshold
[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/module.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>
36 #include <linux/pagevec.h>
37
38 /*
39  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
40  * will look to see if it needs to force writeback or throttling.
41  */
42 static long ratelimit_pages = 32;
43
44 /*
45  * When balance_dirty_pages decides that the caller needs to perform some
46  * non-background writeback, this is how many pages it will attempt to write.
47  * It should be somewhat larger than dirtied pages to ensure that reasonably
48  * large amounts of I/O are submitted.
49  */
50 static inline long sync_writeback_pages(unsigned long dirtied)
51 {
52         if (dirtied < ratelimit_pages)
53                 dirtied = ratelimit_pages;
54
55         return dirtied + dirtied / 2;
56 }
57
58 /* The following parameters are exported via /proc/sys/vm */
59
60 /*
61  * Start background writeback (via pdflush) at this percentage
62  */
63 int dirty_background_ratio = 10;
64
65 /*
66  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
67  * dirty_background_ratio * the amount of dirtyable memory
68  */
69 unsigned long dirty_background_bytes;
70
71 /*
72  * free highmem will not be subtracted from the total free memory
73  * for calculating free ratios if vm_highmem_is_dirtyable is true
74  */
75 int vm_highmem_is_dirtyable;
76
77 /*
78  * The generator of dirty data starts writeback at this percentage
79  */
80 int vm_dirty_ratio = 20;
81
82 /*
83  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
84  * vm_dirty_ratio * the amount of dirtyable memory
85  */
86 unsigned long vm_dirty_bytes;
87
88 /*
89  * The interval between `kupdate'-style writebacks
90  */
91 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
92
93 /*
94  * The longest time for which data is allowed to remain dirty
95  */
96 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
97
98 /*
99  * Flag that makes the machine dump writes/reads and block dirtyings.
100  */
101 int block_dump;
102
103 /*
104  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
105  * a full sync is triggered after this time elapses without any disk activity.
106  */
107 int laptop_mode;
108
109 EXPORT_SYMBOL(laptop_mode);
110
111 /* End of sysctl-exported parameters */
112
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 static struct prop_descriptor vm_dirties;
132
133 /*
134  * couple the period to the dirty_ratio:
135  *
136  *   period/2 ~ roundup_pow_of_two(dirty limit)
137  */
138 static int calc_period_shift(void)
139 {
140         unsigned long dirty_total;
141
142         if (vm_dirty_bytes)
143                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
144         else
145                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146                                 100;
147         return 2 + ilog2(dirty_total - 1);
148 }
149
150 /*
151  * update the period when the dirty threshold changes.
152  */
153 static void update_completion_period(void)
154 {
155         int shift = calc_period_shift();
156         prop_change_shift(&vm_completions, shift);
157         prop_change_shift(&vm_dirties, shift);
158 }
159
160 int dirty_background_ratio_handler(struct ctl_table *table, int write,
161                 struct file *filp, void __user *buffer, size_t *lenp,
162                 loff_t *ppos)
163 {
164         int ret;
165
166         ret = proc_dointvec_minmax(table, write, filp, 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                 struct file *filp, void __user *buffer, size_t *lenp,
174                 loff_t *ppos)
175 {
176         int ret;
177
178         ret = proc_doulongvec_minmax(table, write, filp, 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                 struct file *filp, 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, filp, 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                 struct file *filp, 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, filp, 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         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
222                               bdi->max_prop_frac);
223 }
224
225 void bdi_writeout_inc(struct backing_dev_info *bdi)
226 {
227         unsigned long flags;
228
229         local_irq_save(flags);
230         __bdi_writeout_inc(bdi);
231         local_irq_restore(flags);
232 }
233 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
234
235 void task_dirty_inc(struct task_struct *tsk)
236 {
237         prop_inc_single(&vm_dirties, &tsk->dirties);
238 }
239
240 /*
241  * Obtain an accurate fraction of the BDI's portion.
242  */
243 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
244                 long *numerator, long *denominator)
245 {
246         if (bdi_cap_writeback_dirty(bdi)) {
247                 prop_fraction_percpu(&vm_completions, &bdi->completions,
248                                 numerator, denominator);
249         } else {
250                 *numerator = 0;
251                 *denominator = 1;
252         }
253 }
254
255 /*
256  * Clip the earned share of dirty pages to that which is actually available.
257  * This avoids exceeding the total dirty_limit when the floating averages
258  * fluctuate too quickly.
259  */
260 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
261                 unsigned long dirty, unsigned long *pbdi_dirty)
262 {
263         unsigned long avail_dirty;
264
265         avail_dirty = global_page_state(NR_FILE_DIRTY) +
266                  global_page_state(NR_WRITEBACK) +
267                  global_page_state(NR_UNSTABLE_NFS) +
268                  global_page_state(NR_WRITEBACK_TEMP);
269
270         if (avail_dirty < dirty)
271                 avail_dirty = dirty - avail_dirty;
272         else
273                 avail_dirty = 0;
274
275         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
276                 bdi_stat(bdi, BDI_WRITEBACK);
277
278         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
279 }
280
281 static inline void task_dirties_fraction(struct task_struct *tsk,
282                 long *numerator, long *denominator)
283 {
284         prop_fraction_single(&vm_dirties, &tsk->dirties,
285                                 numerator, denominator);
286 }
287
288 /*
289  * scale the dirty limit
290  *
291  * task specific dirty limit:
292  *
293  *   dirty -= (dirty/8) * p_{t}
294  */
295 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
296 {
297         long numerator, denominator;
298         unsigned long dirty = *pdirty;
299         u64 inv = dirty >> 3;
300
301         task_dirties_fraction(tsk, &numerator, &denominator);
302         inv *= numerator;
303         do_div(inv, denominator);
304
305         dirty -= inv;
306         if (dirty < *pdirty/2)
307                 dirty = *pdirty/2;
308
309         *pdirty = dirty;
310 }
311
312 /*
313  *
314  */
315 static unsigned int bdi_min_ratio;
316
317 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
318 {
319         int ret = 0;
320
321         spin_lock_bh(&bdi_lock);
322         if (min_ratio > bdi->max_ratio) {
323                 ret = -EINVAL;
324         } else {
325                 min_ratio -= bdi->min_ratio;
326                 if (bdi_min_ratio + min_ratio < 100) {
327                         bdi_min_ratio += min_ratio;
328                         bdi->min_ratio += min_ratio;
329                 } else {
330                         ret = -EINVAL;
331                 }
332         }
333         spin_unlock_bh(&bdi_lock);
334
335         return ret;
336 }
337
338 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
339 {
340         int ret = 0;
341
342         if (max_ratio > 100)
343                 return -EINVAL;
344
345         spin_lock_bh(&bdi_lock);
346         if (bdi->min_ratio > max_ratio) {
347                 ret = -EINVAL;
348         } else {
349                 bdi->max_ratio = max_ratio;
350                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
351         }
352         spin_unlock_bh(&bdi_lock);
353
354         return ret;
355 }
356 EXPORT_SYMBOL(bdi_set_max_ratio);
357
358 /*
359  * Work out the current dirty-memory clamping and background writeout
360  * thresholds.
361  *
362  * The main aim here is to lower them aggressively if there is a lot of mapped
363  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
364  * pages.  It is better to clamp down on writers than to start swapping, and
365  * performing lots of scanning.
366  *
367  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
368  *
369  * We don't permit the clamping level to fall below 5% - that is getting rather
370  * excessive.
371  *
372  * We make sure that the background writeout level is below the adjusted
373  * clamping level.
374  */
375
376 static unsigned long highmem_dirtyable_memory(unsigned long total)
377 {
378 #ifdef CONFIG_HIGHMEM
379         int node;
380         unsigned long x = 0;
381
382         for_each_node_state(node, N_HIGH_MEMORY) {
383                 struct zone *z =
384                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
385
386                 x += zone_page_state(z, NR_FREE_PAGES) +
387                      zone_reclaimable_pages(z);
388         }
389         /*
390          * Make sure that the number of highmem pages is never larger
391          * than the number of the total dirtyable memory. This can only
392          * occur in very strange VM situations but we want to make sure
393          * that this does not occur.
394          */
395         return min(x, total);
396 #else
397         return 0;
398 #endif
399 }
400
401 /**
402  * determine_dirtyable_memory - amount of memory that may be used
403  *
404  * Returns the numebr of pages that can currently be freed and used
405  * by the kernel for direct mappings.
406  */
407 unsigned long determine_dirtyable_memory(void)
408 {
409         unsigned long x;
410
411         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
412
413         if (!vm_highmem_is_dirtyable)
414                 x -= highmem_dirtyable_memory(x);
415
416         return x + 1;   /* Ensure that we never return 0 */
417 }
418
419 void
420 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
421                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
422 {
423         unsigned long background;
424         unsigned long dirty;
425         unsigned long available_memory = determine_dirtyable_memory();
426         struct task_struct *tsk;
427
428         if (vm_dirty_bytes)
429                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
430         else {
431                 int dirty_ratio;
432
433                 dirty_ratio = vm_dirty_ratio;
434                 if (dirty_ratio < 5)
435                         dirty_ratio = 5;
436                 dirty = (dirty_ratio * available_memory) / 100;
437         }
438
439         if (dirty_background_bytes)
440                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
441         else
442                 background = (dirty_background_ratio * available_memory) / 100;
443
444         if (background >= dirty)
445                 background = dirty / 2;
446         tsk = current;
447         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
448                 background += background / 4;
449                 dirty += dirty / 4;
450         }
451         *pbackground = background;
452         *pdirty = dirty;
453
454         if (bdi) {
455                 u64 bdi_dirty;
456                 long numerator, denominator;
457
458                 /*
459                  * Calculate this BDI's share of the dirty ratio.
460                  */
461                 bdi_writeout_fraction(bdi, &numerator, &denominator);
462
463                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
464                 bdi_dirty *= numerator;
465                 do_div(bdi_dirty, denominator);
466                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
467                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
468                         bdi_dirty = dirty * bdi->max_ratio / 100;
469
470                 *pbdi_dirty = bdi_dirty;
471                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
472                 task_dirty_limit(current, pbdi_dirty);
473         }
474 }
475
476 /*
477  * balance_dirty_pages() must be called by processes which are generating dirty
478  * data.  It looks at the number of dirty pages in the machine and will force
479  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
480  * If we're over `background_thresh' then pdflush is woken to perform some
481  * writeout.
482  */
483 static void balance_dirty_pages(struct address_space *mapping,
484                                 unsigned long write_chunk)
485 {
486         long nr_reclaimable, bdi_nr_reclaimable;
487         long nr_writeback, bdi_nr_writeback;
488         unsigned long background_thresh;
489         unsigned long dirty_thresh;
490         unsigned long bdi_thresh;
491         unsigned long pages_written = 0;
492         unsigned long pause = 1;
493
494         struct backing_dev_info *bdi = mapping->backing_dev_info;
495
496         for (;;) {
497                 struct writeback_control wbc = {
498                         .bdi            = bdi,
499                         .sync_mode      = WB_SYNC_NONE,
500                         .older_than_this = NULL,
501                         .nr_to_write    = write_chunk,
502                         .range_cyclic   = 1,
503                 };
504
505                 get_dirty_limits(&background_thresh, &dirty_thresh,
506                                 &bdi_thresh, bdi);
507
508                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
509                                         global_page_state(NR_UNSTABLE_NFS);
510                 nr_writeback = global_page_state(NR_WRITEBACK);
511
512                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
513                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
514
515                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
516                         break;
517
518                 /*
519                  * Throttle it only when the background writeback cannot
520                  * catch-up. This avoids (excessively) small writeouts
521                  * when the bdi limits are ramping up.
522                  */
523                 if (nr_reclaimable + nr_writeback <
524                                 (background_thresh + dirty_thresh) / 2)
525                         break;
526
527                 if (!bdi->dirty_exceeded)
528                         bdi->dirty_exceeded = 1;
529
530                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
531                  * Unstable writes are a feature of certain networked
532                  * filesystems (i.e. NFS) in which data may have been
533                  * written to the server's write cache, but has not yet
534                  * been flushed to permanent storage.
535                  * Only move pages to writeback if this bdi is over its
536                  * threshold otherwise wait until the disk writes catch
537                  * up.
538                  */
539                 if (bdi_nr_reclaimable > bdi_thresh) {
540                         writeback_inodes_wbc(&wbc);
541                         pages_written += write_chunk - wbc.nr_to_write;
542                         get_dirty_limits(&background_thresh, &dirty_thresh,
543                                        &bdi_thresh, bdi);
544                 }
545
546                 /*
547                  * In order to avoid the stacked BDI deadlock we need
548                  * to ensure we accurately count the 'dirty' pages when
549                  * the threshold is low.
550                  *
551                  * Otherwise it would be possible to get thresh+n pages
552                  * reported dirty, even though there are thresh-m pages
553                  * actually dirty; with m+n sitting in the percpu
554                  * deltas.
555                  */
556                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
557                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
558                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
559                 } else if (bdi_nr_reclaimable) {
560                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
561                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
562                 }
563
564                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
565                         break;
566                 if (pages_written >= write_chunk)
567                         break;          /* We've done our duty */
568
569                 schedule_timeout_interruptible(pause);
570
571                 /*
572                  * Increase the delay for each loop, up to our previous
573                  * default of taking a 100ms nap.
574                  */
575                 pause <<= 1;
576                 if (pause > HZ / 10)
577                         pause = HZ / 10;
578         }
579
580         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
581                         bdi->dirty_exceeded)
582                 bdi->dirty_exceeded = 0;
583
584         if (writeback_in_progress(bdi))
585                 return;         /* pdflush is already working this queue */
586
587         /*
588          * In laptop mode, we wait until hitting the higher threshold before
589          * starting background writeout, and then write out all the way down
590          * to the lower threshold.  So slow writers cause minimal disk activity.
591          *
592          * In normal mode, we start background writeout at the lower
593          * background_thresh, to keep the amount of dirty memory low.
594          */
595         if ((laptop_mode && pages_written) ||
596             (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
597                                + global_page_state(NR_UNSTABLE_NFS))
598                                           > background_thresh)))
599                 bdi_start_writeback(bdi, 0);
600 }
601
602 void set_page_dirty_balance(struct page *page, int page_mkwrite)
603 {
604         if (set_page_dirty(page) || page_mkwrite) {
605                 struct address_space *mapping = page_mapping(page);
606
607                 if (mapping)
608                         balance_dirty_pages_ratelimited(mapping);
609         }
610 }
611
612 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
613
614 /**
615  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
616  * @mapping: address_space which was dirtied
617  * @nr_pages_dirtied: number of pages which the caller has just dirtied
618  *
619  * Processes which are dirtying memory should call in here once for each page
620  * which was newly dirtied.  The function will periodically check the system's
621  * dirty state and will initiate writeback if needed.
622  *
623  * On really big machines, get_writeback_state is expensive, so try to avoid
624  * calling it too often (ratelimiting).  But once we're over the dirty memory
625  * limit we decrease the ratelimiting by a lot, to prevent individual processes
626  * from overshooting the limit by (ratelimit_pages) each.
627  */
628 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
629                                         unsigned long nr_pages_dirtied)
630 {
631         unsigned long ratelimit;
632         unsigned long *p;
633
634         ratelimit = ratelimit_pages;
635         if (mapping->backing_dev_info->dirty_exceeded)
636                 ratelimit = 8;
637
638         /*
639          * Check the rate limiting. Also, we do not want to throttle real-time
640          * tasks in balance_dirty_pages(). Period.
641          */
642         preempt_disable();
643         p =  &__get_cpu_var(bdp_ratelimits);
644         *p += nr_pages_dirtied;
645         if (unlikely(*p >= ratelimit)) {
646                 ratelimit = sync_writeback_pages(*p);
647                 *p = 0;
648                 preempt_enable();
649                 balance_dirty_pages(mapping, ratelimit);
650                 return;
651         }
652         preempt_enable();
653 }
654 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
655
656 void throttle_vm_writeout(gfp_t gfp_mask)
657 {
658         unsigned long background_thresh;
659         unsigned long dirty_thresh;
660
661         for ( ; ; ) {
662                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
663
664                 /*
665                  * Boost the allowable dirty threshold a bit for page
666                  * allocators so they don't get DoS'ed by heavy writers
667                  */
668                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
669
670                 if (global_page_state(NR_UNSTABLE_NFS) +
671                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
672                                 break;
673                 congestion_wait(BLK_RW_ASYNC, HZ/10);
674
675                 /*
676                  * The caller might hold locks which can prevent IO completion
677                  * or progress in the filesystem.  So we cannot just sit here
678                  * waiting for IO to complete.
679                  */
680                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
681                         break;
682         }
683 }
684
685 static void laptop_timer_fn(unsigned long unused);
686
687 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
688
689 /*
690  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
691  */
692 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
693         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
694 {
695         proc_dointvec(table, write, file, buffer, length, ppos);
696         return 0;
697 }
698
699 static void do_laptop_sync(struct work_struct *work)
700 {
701         wakeup_flusher_threads(0);
702         kfree(work);
703 }
704
705 static void laptop_timer_fn(unsigned long unused)
706 {
707         struct work_struct *work;
708
709         work = kmalloc(sizeof(*work), GFP_ATOMIC);
710         if (work) {
711                 INIT_WORK(work, do_laptop_sync);
712                 schedule_work(work);
713         }
714 }
715
716 /*
717  * We've spun up the disk and we're in laptop mode: schedule writeback
718  * of all dirty data a few seconds from now.  If the flush is already scheduled
719  * then push it back - the user is still using the disk.
720  */
721 void laptop_io_completion(void)
722 {
723         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
724 }
725
726 /*
727  * We're in laptop mode and we've just synced. The sync's writes will have
728  * caused another writeback to be scheduled by laptop_io_completion.
729  * Nothing needs to be written back anymore, so we unschedule the writeback.
730  */
731 void laptop_sync_completion(void)
732 {
733         del_timer(&laptop_mode_wb_timer);
734 }
735
736 /*
737  * If ratelimit_pages is too high then we can get into dirty-data overload
738  * if a large number of processes all perform writes at the same time.
739  * If it is too low then SMP machines will call the (expensive)
740  * get_writeback_state too often.
741  *
742  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
743  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
744  * thresholds before writeback cuts in.
745  *
746  * But the limit should not be set too high.  Because it also controls the
747  * amount of memory which the balance_dirty_pages() caller has to write back.
748  * If this is too large then the caller will block on the IO queue all the
749  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
750  * will write six megabyte chunks, max.
751  */
752
753 void writeback_set_ratelimit(void)
754 {
755         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
756         if (ratelimit_pages < 16)
757                 ratelimit_pages = 16;
758         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
759                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
760 }
761
762 static int __cpuinit
763 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
764 {
765         writeback_set_ratelimit();
766         return NOTIFY_DONE;
767 }
768
769 static struct notifier_block __cpuinitdata ratelimit_nb = {
770         .notifier_call  = ratelimit_handler,
771         .next           = NULL,
772 };
773
774 /*
775  * Called early on to tune the page writeback dirty limits.
776  *
777  * We used to scale dirty pages according to how total memory
778  * related to pages that could be allocated for buffers (by
779  * comparing nr_free_buffer_pages() to vm_total_pages.
780  *
781  * However, that was when we used "dirty_ratio" to scale with
782  * all memory, and we don't do that any more. "dirty_ratio"
783  * is now applied to total non-HIGHPAGE memory (by subtracting
784  * totalhigh_pages from vm_total_pages), and as such we can't
785  * get into the old insane situation any more where we had
786  * large amounts of dirty pages compared to a small amount of
787  * non-HIGHMEM memory.
788  *
789  * But we might still want to scale the dirty_ratio by how
790  * much memory the box has..
791  */
792 void __init page_writeback_init(void)
793 {
794         int shift;
795
796         writeback_set_ratelimit();
797         register_cpu_notifier(&ratelimit_nb);
798
799         shift = calc_period_shift();
800         prop_descriptor_init(&vm_completions, shift);
801         prop_descriptor_init(&vm_dirties, shift);
802 }
803
804 /**
805  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
806  * @mapping: address space structure to write
807  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
808  * @writepage: function called for each page
809  * @data: data passed to writepage function
810  *
811  * If a page is already under I/O, write_cache_pages() skips it, even
812  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
813  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
814  * and msync() need to guarantee that all the data which was dirty at the time
815  * the call was made get new I/O started against them.  If wbc->sync_mode is
816  * WB_SYNC_ALL then we were called for data integrity and we must wait for
817  * existing IO to complete.
818  */
819 int write_cache_pages(struct address_space *mapping,
820                       struct writeback_control *wbc, writepage_t writepage,
821                       void *data)
822 {
823         struct backing_dev_info *bdi = mapping->backing_dev_info;
824         int ret = 0;
825         int done = 0;
826         struct pagevec pvec;
827         int nr_pages;
828         pgoff_t uninitialized_var(writeback_index);
829         pgoff_t index;
830         pgoff_t end;            /* Inclusive */
831         pgoff_t done_index;
832         int cycled;
833         int range_whole = 0;
834         long nr_to_write = wbc->nr_to_write;
835
836         if (wbc->nonblocking && bdi_write_congested(bdi)) {
837                 wbc->encountered_congestion = 1;
838                 return 0;
839         }
840
841         pagevec_init(&pvec, 0);
842         if (wbc->range_cyclic) {
843                 writeback_index = mapping->writeback_index; /* prev offset */
844                 index = writeback_index;
845                 if (index == 0)
846                         cycled = 1;
847                 else
848                         cycled = 0;
849                 end = -1;
850         } else {
851                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
852                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
853                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
854                         range_whole = 1;
855                 cycled = 1; /* ignore range_cyclic tests */
856         }
857 retry:
858         done_index = index;
859         while (!done && (index <= end)) {
860                 int i;
861
862                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
863                               PAGECACHE_TAG_DIRTY,
864                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
865                 if (nr_pages == 0)
866                         break;
867
868                 for (i = 0; i < nr_pages; i++) {
869                         struct page *page = pvec.pages[i];
870
871                         /*
872                          * At this point, the page may be truncated or
873                          * invalidated (changing page->mapping to NULL), or
874                          * even swizzled back from swapper_space to tmpfs file
875                          * mapping. However, page->index will not change
876                          * because we have a reference on the page.
877                          */
878                         if (page->index > end) {
879                                 /*
880                                  * can't be range_cyclic (1st pass) because
881                                  * end == -1 in that case.
882                                  */
883                                 done = 1;
884                                 break;
885                         }
886
887                         done_index = page->index + 1;
888
889                         lock_page(page);
890
891                         /*
892                          * Page truncated or invalidated. We can freely skip it
893                          * then, even for data integrity operations: the page
894                          * has disappeared concurrently, so there could be no
895                          * real expectation of this data interity operation
896                          * even if there is now a new, dirty page at the same
897                          * pagecache address.
898                          */
899                         if (unlikely(page->mapping != mapping)) {
900 continue_unlock:
901                                 unlock_page(page);
902                                 continue;
903                         }
904
905                         if (!PageDirty(page)) {
906                                 /* someone wrote it for us */
907                                 goto continue_unlock;
908                         }
909
910                         if (PageWriteback(page)) {
911                                 if (wbc->sync_mode != WB_SYNC_NONE)
912                                         wait_on_page_writeback(page);
913                                 else
914                                         goto continue_unlock;
915                         }
916
917                         BUG_ON(PageWriteback(page));
918                         if (!clear_page_dirty_for_io(page))
919                                 goto continue_unlock;
920
921                         ret = (*writepage)(page, wbc, data);
922                         if (unlikely(ret)) {
923                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
924                                         unlock_page(page);
925                                         ret = 0;
926                                 } else {
927                                         /*
928                                          * done_index is set past this page,
929                                          * so media errors will not choke
930                                          * background writeout for the entire
931                                          * file. This has consequences for
932                                          * range_cyclic semantics (ie. it may
933                                          * not be suitable for data integrity
934                                          * writeout).
935                                          */
936                                         done = 1;
937                                         break;
938                                 }
939                         }
940
941                         if (nr_to_write > 0) {
942                                 nr_to_write--;
943                                 if (nr_to_write == 0 &&
944                                     wbc->sync_mode == WB_SYNC_NONE) {
945                                         /*
946                                          * We stop writing back only if we are
947                                          * not doing integrity sync. In case of
948                                          * integrity sync we have to keep going
949                                          * because someone may be concurrently
950                                          * dirtying pages, and we might have
951                                          * synced a lot of newly appeared dirty
952                                          * pages, but have not synced all of the
953                                          * old dirty pages.
954                                          */
955                                         done = 1;
956                                         break;
957                                 }
958                         }
959
960                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
961                                 wbc->encountered_congestion = 1;
962                                 done = 1;
963                                 break;
964                         }
965                 }
966                 pagevec_release(&pvec);
967                 cond_resched();
968         }
969         if (!cycled && !done) {
970                 /*
971                  * range_cyclic:
972                  * We hit the last page and there is more work to be done: wrap
973                  * back to the start of the file
974                  */
975                 cycled = 1;
976                 index = 0;
977                 end = writeback_index - 1;
978                 goto retry;
979         }
980         if (!wbc->no_nrwrite_index_update) {
981                 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
982                         mapping->writeback_index = done_index;
983                 wbc->nr_to_write = nr_to_write;
984         }
985
986         return ret;
987 }
988 EXPORT_SYMBOL(write_cache_pages);
989
990 /*
991  * Function used by generic_writepages to call the real writepage
992  * function and set the mapping flags on error
993  */
994 static int __writepage(struct page *page, struct writeback_control *wbc,
995                        void *data)
996 {
997         struct address_space *mapping = data;
998         int ret = mapping->a_ops->writepage(page, wbc);
999         mapping_set_error(mapping, ret);
1000         return ret;
1001 }
1002
1003 /**
1004  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1005  * @mapping: address space structure to write
1006  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1007  *
1008  * This is a library function, which implements the writepages()
1009  * address_space_operation.
1010  */
1011 int generic_writepages(struct address_space *mapping,
1012                        struct writeback_control *wbc)
1013 {
1014         /* deal with chardevs and other special file */
1015         if (!mapping->a_ops->writepage)
1016                 return 0;
1017
1018         return write_cache_pages(mapping, wbc, __writepage, mapping);
1019 }
1020
1021 EXPORT_SYMBOL(generic_writepages);
1022
1023 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1024 {
1025         int ret;
1026
1027         if (wbc->nr_to_write <= 0)
1028                 return 0;
1029         if (mapping->a_ops->writepages)
1030                 ret = mapping->a_ops->writepages(mapping, wbc);
1031         else
1032                 ret = generic_writepages(mapping, wbc);
1033         return ret;
1034 }
1035
1036 /**
1037  * write_one_page - write out a single page and optionally wait on I/O
1038  * @page: the page to write
1039  * @wait: if true, wait on writeout
1040  *
1041  * The page must be locked by the caller and will be unlocked upon return.
1042  *
1043  * write_one_page() returns a negative error code if I/O failed.
1044  */
1045 int write_one_page(struct page *page, int wait)
1046 {
1047         struct address_space *mapping = page->mapping;
1048         int ret = 0;
1049         struct writeback_control wbc = {
1050                 .sync_mode = WB_SYNC_ALL,
1051                 .nr_to_write = 1,
1052         };
1053
1054         BUG_ON(!PageLocked(page));
1055
1056         if (wait)
1057                 wait_on_page_writeback(page);
1058
1059         if (clear_page_dirty_for_io(page)) {
1060                 page_cache_get(page);
1061                 ret = mapping->a_ops->writepage(page, &wbc);
1062                 if (ret == 0 && wait) {
1063                         wait_on_page_writeback(page);
1064                         if (PageError(page))
1065                                 ret = -EIO;
1066                 }
1067                 page_cache_release(page);
1068         } else {
1069                 unlock_page(page);
1070         }
1071         return ret;
1072 }
1073 EXPORT_SYMBOL(write_one_page);
1074
1075 /*
1076  * For address_spaces which do not use buffers nor write back.
1077  */
1078 int __set_page_dirty_no_writeback(struct page *page)
1079 {
1080         if (!PageDirty(page))
1081                 SetPageDirty(page);
1082         return 0;
1083 }
1084
1085 /*
1086  * Helper function for set_page_dirty family.
1087  * NOTE: This relies on being atomic wrt interrupts.
1088  */
1089 void account_page_dirtied(struct page *page, struct address_space *mapping)
1090 {
1091         if (mapping_cap_account_dirty(mapping)) {
1092                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1093                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1094                 task_dirty_inc(current);
1095                 task_io_account_write(PAGE_CACHE_SIZE);
1096         }
1097 }
1098
1099 /*
1100  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1101  * its radix tree.
1102  *
1103  * This is also used when a single buffer is being dirtied: we want to set the
1104  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1105  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1106  *
1107  * Most callers have locked the page, which pins the address_space in memory.
1108  * But zap_pte_range() does not lock the page, however in that case the
1109  * mapping is pinned by the vma's ->vm_file reference.
1110  *
1111  * We take care to handle the case where the page was truncated from the
1112  * mapping by re-checking page_mapping() inside tree_lock.
1113  */
1114 int __set_page_dirty_nobuffers(struct page *page)
1115 {
1116         if (!TestSetPageDirty(page)) {
1117                 struct address_space *mapping = page_mapping(page);
1118                 struct address_space *mapping2;
1119
1120                 if (!mapping)
1121                         return 1;
1122
1123                 spin_lock_irq(&mapping->tree_lock);
1124                 mapping2 = page_mapping(page);
1125                 if (mapping2) { /* Race with truncate? */
1126                         BUG_ON(mapping2 != mapping);
1127                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1128                         account_page_dirtied(page, mapping);
1129                         radix_tree_tag_set(&mapping->page_tree,
1130                                 page_index(page), PAGECACHE_TAG_DIRTY);
1131                 }
1132                 spin_unlock_irq(&mapping->tree_lock);
1133                 if (mapping->host) {
1134                         /* !PageAnon && !swapper_space */
1135                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1136                 }
1137                 return 1;
1138         }
1139         return 0;
1140 }
1141 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1142
1143 /*
1144  * When a writepage implementation decides that it doesn't want to write this
1145  * page for some reason, it should redirty the locked page via
1146  * redirty_page_for_writepage() and it should then unlock the page and return 0
1147  */
1148 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1149 {
1150         wbc->pages_skipped++;
1151         return __set_page_dirty_nobuffers(page);
1152 }
1153 EXPORT_SYMBOL(redirty_page_for_writepage);
1154
1155 /*
1156  * If the mapping doesn't provide a set_page_dirty a_op, then
1157  * just fall through and assume that it wants buffer_heads.
1158  */
1159 int set_page_dirty(struct page *page)
1160 {
1161         struct address_space *mapping = page_mapping(page);
1162
1163         if (likely(mapping)) {
1164                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1165 #ifdef CONFIG_BLOCK
1166                 if (!spd)
1167                         spd = __set_page_dirty_buffers;
1168 #endif
1169                 return (*spd)(page);
1170         }
1171         if (!PageDirty(page)) {
1172                 if (!TestSetPageDirty(page))
1173                         return 1;
1174         }
1175         return 0;
1176 }
1177 EXPORT_SYMBOL(set_page_dirty);
1178
1179 /*
1180  * set_page_dirty() is racy if the caller has no reference against
1181  * page->mapping->host, and if the page is unlocked.  This is because another
1182  * CPU could truncate the page off the mapping and then free the mapping.
1183  *
1184  * Usually, the page _is_ locked, or the caller is a user-space process which
1185  * holds a reference on the inode by having an open file.
1186  *
1187  * In other cases, the page should be locked before running set_page_dirty().
1188  */
1189 int set_page_dirty_lock(struct page *page)
1190 {
1191         int ret;
1192
1193         lock_page_nosync(page);
1194         ret = set_page_dirty(page);
1195         unlock_page(page);
1196         return ret;
1197 }
1198 EXPORT_SYMBOL(set_page_dirty_lock);
1199
1200 /*
1201  * Clear a page's dirty flag, while caring for dirty memory accounting.
1202  * Returns true if the page was previously dirty.
1203  *
1204  * This is for preparing to put the page under writeout.  We leave the page
1205  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1206  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1207  * implementation will run either set_page_writeback() or set_page_dirty(),
1208  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1209  * back into sync.
1210  *
1211  * This incoherency between the page's dirty flag and radix-tree tag is
1212  * unfortunate, but it only exists while the page is locked.
1213  */
1214 int clear_page_dirty_for_io(struct page *page)
1215 {
1216         struct address_space *mapping = page_mapping(page);
1217
1218         BUG_ON(!PageLocked(page));
1219
1220         ClearPageReclaim(page);
1221         if (mapping && mapping_cap_account_dirty(mapping)) {
1222                 /*
1223                  * Yes, Virginia, this is indeed insane.
1224                  *
1225                  * We use this sequence to make sure that
1226                  *  (a) we account for dirty stats properly
1227                  *  (b) we tell the low-level filesystem to
1228                  *      mark the whole page dirty if it was
1229                  *      dirty in a pagetable. Only to then
1230                  *  (c) clean the page again and return 1 to
1231                  *      cause the writeback.
1232                  *
1233                  * This way we avoid all nasty races with the
1234                  * dirty bit in multiple places and clearing
1235                  * them concurrently from different threads.
1236                  *
1237                  * Note! Normally the "set_page_dirty(page)"
1238                  * has no effect on the actual dirty bit - since
1239                  * that will already usually be set. But we
1240                  * need the side effects, and it can help us
1241                  * avoid races.
1242                  *
1243                  * We basically use the page "master dirty bit"
1244                  * as a serialization point for all the different
1245                  * threads doing their things.
1246                  */
1247                 if (page_mkclean(page))
1248                         set_page_dirty(page);
1249                 /*
1250                  * We carefully synchronise fault handlers against
1251                  * installing a dirty pte and marking the page dirty
1252                  * at this point. We do this by having them hold the
1253                  * page lock at some point after installing their
1254                  * pte, but before marking the page dirty.
1255                  * Pages are always locked coming in here, so we get
1256                  * the desired exclusion. See mm/memory.c:do_wp_page()
1257                  * for more comments.
1258                  */
1259                 if (TestClearPageDirty(page)) {
1260                         dec_zone_page_state(page, NR_FILE_DIRTY);
1261                         dec_bdi_stat(mapping->backing_dev_info,
1262                                         BDI_RECLAIMABLE);
1263                         return 1;
1264                 }
1265                 return 0;
1266         }
1267         return TestClearPageDirty(page);
1268 }
1269 EXPORT_SYMBOL(clear_page_dirty_for_io);
1270
1271 int test_clear_page_writeback(struct page *page)
1272 {
1273         struct address_space *mapping = page_mapping(page);
1274         int ret;
1275
1276         if (mapping) {
1277                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1278                 unsigned long flags;
1279
1280                 spin_lock_irqsave(&mapping->tree_lock, flags);
1281                 ret = TestClearPageWriteback(page);
1282                 if (ret) {
1283                         radix_tree_tag_clear(&mapping->page_tree,
1284                                                 page_index(page),
1285                                                 PAGECACHE_TAG_WRITEBACK);
1286                         if (bdi_cap_account_writeback(bdi)) {
1287                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1288                                 __bdi_writeout_inc(bdi);
1289                         }
1290                 }
1291                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1292         } else {
1293                 ret = TestClearPageWriteback(page);
1294         }
1295         if (ret)
1296                 dec_zone_page_state(page, NR_WRITEBACK);
1297         return ret;
1298 }
1299
1300 int test_set_page_writeback(struct page *page)
1301 {
1302         struct address_space *mapping = page_mapping(page);
1303         int ret;
1304
1305         if (mapping) {
1306                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1307                 unsigned long flags;
1308
1309                 spin_lock_irqsave(&mapping->tree_lock, flags);
1310                 ret = TestSetPageWriteback(page);
1311                 if (!ret) {
1312                         radix_tree_tag_set(&mapping->page_tree,
1313                                                 page_index(page),
1314                                                 PAGECACHE_TAG_WRITEBACK);
1315                         if (bdi_cap_account_writeback(bdi))
1316                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1317                 }
1318                 if (!PageDirty(page))
1319                         radix_tree_tag_clear(&mapping->page_tree,
1320                                                 page_index(page),
1321                                                 PAGECACHE_TAG_DIRTY);
1322                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1323         } else {
1324                 ret = TestSetPageWriteback(page);
1325         }
1326         if (!ret)
1327                 inc_zone_page_state(page, NR_WRITEBACK);
1328         return ret;
1329
1330 }
1331 EXPORT_SYMBOL(test_set_page_writeback);
1332
1333 /*
1334  * Return true if any of the pages in the mapping are marked with the
1335  * passed tag.
1336  */
1337 int mapping_tagged(struct address_space *mapping, int tag)
1338 {
1339         int ret;
1340         rcu_read_lock();
1341         ret = radix_tree_tagged(&mapping->page_tree, tag);
1342         rcu_read_unlock();
1343         return ret;
1344 }
1345 EXPORT_SYMBOL(mapping_tagged);