25e7770309b871e42dbaf7ccedf0132b265eacc2
[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 RATELIMIT_PAGES to ensure that reasonably
48  * large amounts of I/O are submitted.
49  */
50 static inline long sync_writeback_pages(void)
51 {
52         return ratelimit_pages + ratelimit_pages / 2;
53 }
54
55 /* The following parameters are exported via /proc/sys/vm */
56
57 /*
58  * Start background writeback (via pdflush) at this percentage
59  */
60 int dirty_background_ratio = 10;
61
62 /*
63  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
64  * dirty_background_ratio * the amount of dirtyable memory
65  */
66 unsigned long dirty_background_bytes;
67
68 /*
69  * free highmem will not be subtracted from the total free memory
70  * for calculating free ratios if vm_highmem_is_dirtyable is true
71  */
72 int vm_highmem_is_dirtyable;
73
74 /*
75  * The generator of dirty data starts writeback at this percentage
76  */
77 int vm_dirty_ratio = 20;
78
79 /*
80  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
81  * vm_dirty_ratio * the amount of dirtyable memory
82  */
83 unsigned long vm_dirty_bytes;
84
85 /*
86  * The interval between `kupdate'-style writebacks
87  */
88 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
89
90 /*
91  * The longest time for which data is allowed to remain dirty
92  */
93 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
94
95 /*
96  * Flag that makes the machine dump writes/reads and block dirtyings.
97  */
98 int block_dump;
99
100 /*
101  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
102  * a full sync is triggered after this time elapses without any disk activity.
103  */
104 int laptop_mode;
105
106 EXPORT_SYMBOL(laptop_mode);
107
108 /* End of sysctl-exported parameters */
109
110
111 /*
112  * Scale the writeback cache size proportional to the relative writeout speeds.
113  *
114  * We do this by keeping a floating proportion between BDIs, based on page
115  * writeback completions [end_page_writeback()]. Those devices that write out
116  * pages fastest will get the larger share, while the slower will get a smaller
117  * share.
118  *
119  * We use page writeout completions because we are interested in getting rid of
120  * dirty pages. Having them written out is the primary goal.
121  *
122  * We introduce a concept of time, a period over which we measure these events,
123  * because demand can/will vary over time. The length of this period itself is
124  * measured in page writeback completions.
125  *
126  */
127 static struct prop_descriptor vm_completions;
128 static struct prop_descriptor vm_dirties;
129
130 /*
131  * couple the period to the dirty_ratio:
132  *
133  *   period/2 ~ roundup_pow_of_two(dirty limit)
134  */
135 static int calc_period_shift(void)
136 {
137         unsigned long dirty_total;
138
139         if (vm_dirty_bytes)
140                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
141         else
142                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
143                                 100;
144         return 2 + ilog2(dirty_total - 1);
145 }
146
147 /*
148  * update the period when the dirty threshold changes.
149  */
150 static void update_completion_period(void)
151 {
152         int shift = calc_period_shift();
153         prop_change_shift(&vm_completions, shift);
154         prop_change_shift(&vm_dirties, shift);
155 }
156
157 int dirty_background_ratio_handler(struct ctl_table *table, int write,
158                 struct file *filp, void __user *buffer, size_t *lenp,
159                 loff_t *ppos)
160 {
161         int ret;
162
163         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
164         if (ret == 0 && write)
165                 dirty_background_bytes = 0;
166         return ret;
167 }
168
169 int dirty_background_bytes_handler(struct ctl_table *table, int write,
170                 struct file *filp, void __user *buffer, size_t *lenp,
171                 loff_t *ppos)
172 {
173         int ret;
174
175         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
176         if (ret == 0 && write)
177                 dirty_background_ratio = 0;
178         return ret;
179 }
180
181 int dirty_ratio_handler(struct ctl_table *table, int write,
182                 struct file *filp, void __user *buffer, size_t *lenp,
183                 loff_t *ppos)
184 {
185         int old_ratio = vm_dirty_ratio;
186         int ret;
187
188         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
189         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
190                 update_completion_period();
191                 vm_dirty_bytes = 0;
192         }
193         return ret;
194 }
195
196
197 int dirty_bytes_handler(struct ctl_table *table, int write,
198                 struct file *filp, void __user *buffer, size_t *lenp,
199                 loff_t *ppos)
200 {
201         unsigned long old_bytes = vm_dirty_bytes;
202         int ret;
203
204         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
205         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
206                 update_completion_period();
207                 vm_dirty_ratio = 0;
208         }
209         return ret;
210 }
211
212 /*
213  * Increment the BDI's writeout completion count and the global writeout
214  * completion count. Called from test_clear_page_writeback().
215  */
216 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
217 {
218         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
219                               bdi->max_prop_frac);
220 }
221
222 void bdi_writeout_inc(struct backing_dev_info *bdi)
223 {
224         unsigned long flags;
225
226         local_irq_save(flags);
227         __bdi_writeout_inc(bdi);
228         local_irq_restore(flags);
229 }
230 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
231
232 void task_dirty_inc(struct task_struct *tsk)
233 {
234         prop_inc_single(&vm_dirties, &tsk->dirties);
235 }
236
237 /*
238  * Obtain an accurate fraction of the BDI's portion.
239  */
240 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
241                 long *numerator, long *denominator)
242 {
243         if (bdi_cap_writeback_dirty(bdi)) {
244                 prop_fraction_percpu(&vm_completions, &bdi->completions,
245                                 numerator, denominator);
246         } else {
247                 *numerator = 0;
248                 *denominator = 1;
249         }
250 }
251
252 /*
253  * Clip the earned share of dirty pages to that which is actually available.
254  * This avoids exceeding the total dirty_limit when the floating averages
255  * fluctuate too quickly.
256  */
257 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
258                 unsigned long dirty, unsigned long *pbdi_dirty)
259 {
260         unsigned long avail_dirty;
261
262         avail_dirty = global_page_state(NR_FILE_DIRTY) +
263                  global_page_state(NR_WRITEBACK) +
264                  global_page_state(NR_UNSTABLE_NFS) +
265                  global_page_state(NR_WRITEBACK_TEMP);
266
267         if (avail_dirty < dirty)
268                 avail_dirty = dirty - avail_dirty;
269         else
270                 avail_dirty = 0;
271
272         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
273                 bdi_stat(bdi, BDI_WRITEBACK);
274
275         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
276 }
277
278 static inline void task_dirties_fraction(struct task_struct *tsk,
279                 long *numerator, long *denominator)
280 {
281         prop_fraction_single(&vm_dirties, &tsk->dirties,
282                                 numerator, denominator);
283 }
284
285 /*
286  * scale the dirty limit
287  *
288  * task specific dirty limit:
289  *
290  *   dirty -= (dirty/8) * p_{t}
291  */
292 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
293 {
294         long numerator, denominator;
295         unsigned long dirty = *pdirty;
296         u64 inv = dirty >> 3;
297
298         task_dirties_fraction(tsk, &numerator, &denominator);
299         inv *= numerator;
300         do_div(inv, denominator);
301
302         dirty -= inv;
303         if (dirty < *pdirty/2)
304                 dirty = *pdirty/2;
305
306         *pdirty = dirty;
307 }
308
309 /*
310  *
311  */
312 static unsigned int bdi_min_ratio;
313
314 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315 {
316         int ret = 0;
317
318         spin_lock(&bdi_lock);
319         if (min_ratio > bdi->max_ratio) {
320                 ret = -EINVAL;
321         } else {
322                 min_ratio -= bdi->min_ratio;
323                 if (bdi_min_ratio + min_ratio < 100) {
324                         bdi_min_ratio += min_ratio;
325                         bdi->min_ratio += min_ratio;
326                 } else {
327                         ret = -EINVAL;
328                 }
329         }
330         spin_unlock(&bdi_lock);
331
332         return ret;
333 }
334
335 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
336 {
337         int ret = 0;
338
339         if (max_ratio > 100)
340                 return -EINVAL;
341
342         spin_lock(&bdi_lock);
343         if (bdi->min_ratio > max_ratio) {
344                 ret = -EINVAL;
345         } else {
346                 bdi->max_ratio = max_ratio;
347                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348         }
349         spin_unlock(&bdi_lock);
350
351         return ret;
352 }
353 EXPORT_SYMBOL(bdi_set_max_ratio);
354
355 /*
356  * Work out the current dirty-memory clamping and background writeout
357  * thresholds.
358  *
359  * The main aim here is to lower them aggressively if there is a lot of mapped
360  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
361  * pages.  It is better to clamp down on writers than to start swapping, and
362  * performing lots of scanning.
363  *
364  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
365  *
366  * We don't permit the clamping level to fall below 5% - that is getting rather
367  * excessive.
368  *
369  * We make sure that the background writeout level is below the adjusted
370  * clamping level.
371  */
372
373 static unsigned long highmem_dirtyable_memory(unsigned long total)
374 {
375 #ifdef CONFIG_HIGHMEM
376         int node;
377         unsigned long x = 0;
378
379         for_each_node_state(node, N_HIGH_MEMORY) {
380                 struct zone *z =
381                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
382
383                 x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
384         }
385         /*
386          * Make sure that the number of highmem pages is never larger
387          * than the number of the total dirtyable memory. This can only
388          * occur in very strange VM situations but we want to make sure
389          * that this does not occur.
390          */
391         return min(x, total);
392 #else
393         return 0;
394 #endif
395 }
396
397 /**
398  * determine_dirtyable_memory - amount of memory that may be used
399  *
400  * Returns the numebr of pages that can currently be freed and used
401  * by the kernel for direct mappings.
402  */
403 unsigned long determine_dirtyable_memory(void)
404 {
405         unsigned long x;
406
407         x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
408
409         if (!vm_highmem_is_dirtyable)
410                 x -= highmem_dirtyable_memory(x);
411
412         return x + 1;   /* Ensure that we never return 0 */
413 }
414
415 void
416 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
417                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
418 {
419         unsigned long background;
420         unsigned long dirty;
421         unsigned long available_memory = determine_dirtyable_memory();
422         struct task_struct *tsk;
423
424         if (vm_dirty_bytes)
425                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
426         else {
427                 int dirty_ratio;
428
429                 dirty_ratio = vm_dirty_ratio;
430                 if (dirty_ratio < 5)
431                         dirty_ratio = 5;
432                 dirty = (dirty_ratio * available_memory) / 100;
433         }
434
435         if (dirty_background_bytes)
436                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
437         else
438                 background = (dirty_background_ratio * available_memory) / 100;
439
440         if (background >= dirty)
441                 background = dirty / 2;
442         tsk = current;
443         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
444                 background += background / 4;
445                 dirty += dirty / 4;
446         }
447         *pbackground = background;
448         *pdirty = dirty;
449
450         if (bdi) {
451                 u64 bdi_dirty;
452                 long numerator, denominator;
453
454                 /*
455                  * Calculate this BDI's share of the dirty ratio.
456                  */
457                 bdi_writeout_fraction(bdi, &numerator, &denominator);
458
459                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
460                 bdi_dirty *= numerator;
461                 do_div(bdi_dirty, denominator);
462                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
463                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
464                         bdi_dirty = dirty * bdi->max_ratio / 100;
465
466                 *pbdi_dirty = bdi_dirty;
467                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
468                 task_dirty_limit(current, pbdi_dirty);
469         }
470 }
471
472 /*
473  * balance_dirty_pages() must be called by processes which are generating dirty
474  * data.  It looks at the number of dirty pages in the machine and will force
475  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
476  * If we're over `background_thresh' then pdflush is woken to perform some
477  * writeout.
478  */
479 static void balance_dirty_pages(struct address_space *mapping)
480 {
481         long nr_reclaimable, bdi_nr_reclaimable;
482         long nr_writeback, bdi_nr_writeback;
483         unsigned long background_thresh;
484         unsigned long dirty_thresh;
485         unsigned long bdi_thresh;
486         unsigned long pages_written = 0;
487         unsigned long write_chunk = sync_writeback_pages();
488
489         struct backing_dev_info *bdi = mapping->backing_dev_info;
490
491         for (;;) {
492                 struct writeback_control wbc = {
493                         .bdi            = bdi,
494                         .sync_mode      = WB_SYNC_NONE,
495                         .older_than_this = NULL,
496                         .nr_to_write    = write_chunk,
497                         .range_cyclic   = 1,
498                 };
499
500                 get_dirty_limits(&background_thresh, &dirty_thresh,
501                                 &bdi_thresh, bdi);
502
503                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
504                                         global_page_state(NR_UNSTABLE_NFS);
505                 nr_writeback = global_page_state(NR_WRITEBACK);
506
507                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
508                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
509
510                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
511                         break;
512
513                 /*
514                  * Throttle it only when the background writeback cannot
515                  * catch-up. This avoids (excessively) small writeouts
516                  * when the bdi limits are ramping up.
517                  */
518                 if (nr_reclaimable + nr_writeback <
519                                 (background_thresh + dirty_thresh) / 2)
520                         break;
521
522                 if (!bdi->dirty_exceeded)
523                         bdi->dirty_exceeded = 1;
524
525                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
526                  * Unstable writes are a feature of certain networked
527                  * filesystems (i.e. NFS) in which data may have been
528                  * written to the server's write cache, but has not yet
529                  * been flushed to permanent storage.
530                  * Only move pages to writeback if this bdi is over its
531                  * threshold otherwise wait until the disk writes catch
532                  * up.
533                  */
534                 if (bdi_nr_reclaimable > bdi_thresh) {
535                         writeback_inodes_wbc(&wbc);
536                         pages_written += write_chunk - wbc.nr_to_write;
537                         get_dirty_limits(&background_thresh, &dirty_thresh,
538                                        &bdi_thresh, bdi);
539                 }
540
541                 /*
542                  * In order to avoid the stacked BDI deadlock we need
543                  * to ensure we accurately count the 'dirty' pages when
544                  * the threshold is low.
545                  *
546                  * Otherwise it would be possible to get thresh+n pages
547                  * reported dirty, even though there are thresh-m pages
548                  * actually dirty; with m+n sitting in the percpu
549                  * deltas.
550                  */
551                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
552                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
553                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
554                 } else if (bdi_nr_reclaimable) {
555                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
556                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
557                 }
558
559                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
560                         break;
561                 if (pages_written >= write_chunk)
562                         break;          /* We've done our duty */
563
564                 schedule_timeout(1);
565         }
566
567         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
568                         bdi->dirty_exceeded)
569                 bdi->dirty_exceeded = 0;
570
571         if (writeback_in_progress(bdi))
572                 return;         /* pdflush is already working this queue */
573
574         /*
575          * In laptop mode, we wait until hitting the higher threshold before
576          * starting background writeout, and then write out all the way down
577          * to the lower threshold.  So slow writers cause minimal disk activity.
578          *
579          * In normal mode, we start background writeout at the lower
580          * background_thresh, to keep the amount of dirty memory low.
581          */
582         if ((laptop_mode && pages_written) ||
583             (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY)
584                                           + global_page_state(NR_UNSTABLE_NFS))
585                                           > background_thresh))) {
586                 struct writeback_control wbc = {
587                         .bdi            = bdi,
588                         .sync_mode      = WB_SYNC_NONE,
589                         .nr_to_write    = nr_writeback,
590                 };
591
592
593                 bdi_start_writeback(&wbc);
594         }
595 }
596
597 void set_page_dirty_balance(struct page *page, int page_mkwrite)
598 {
599         if (set_page_dirty(page) || page_mkwrite) {
600                 struct address_space *mapping = page_mapping(page);
601
602                 if (mapping)
603                         balance_dirty_pages_ratelimited(mapping);
604         }
605 }
606
607 /**
608  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
609  * @mapping: address_space which was dirtied
610  * @nr_pages_dirtied: number of pages which the caller has just dirtied
611  *
612  * Processes which are dirtying memory should call in here once for each page
613  * which was newly dirtied.  The function will periodically check the system's
614  * dirty state and will initiate writeback if needed.
615  *
616  * On really big machines, get_writeback_state is expensive, so try to avoid
617  * calling it too often (ratelimiting).  But once we're over the dirty memory
618  * limit we decrease the ratelimiting by a lot, to prevent individual processes
619  * from overshooting the limit by (ratelimit_pages) each.
620  */
621 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
622                                         unsigned long nr_pages_dirtied)
623 {
624         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
625         unsigned long ratelimit;
626         unsigned long *p;
627
628         ratelimit = ratelimit_pages;
629         if (mapping->backing_dev_info->dirty_exceeded)
630                 ratelimit = 8;
631
632         /*
633          * Check the rate limiting. Also, we do not want to throttle real-time
634          * tasks in balance_dirty_pages(). Period.
635          */
636         preempt_disable();
637         p =  &__get_cpu_var(ratelimits);
638         *p += nr_pages_dirtied;
639         if (unlikely(*p >= ratelimit)) {
640                 *p = 0;
641                 preempt_enable();
642                 balance_dirty_pages(mapping);
643                 return;
644         }
645         preempt_enable();
646 }
647 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
648
649 void throttle_vm_writeout(gfp_t gfp_mask)
650 {
651         unsigned long background_thresh;
652         unsigned long dirty_thresh;
653
654         for ( ; ; ) {
655                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
656
657                 /*
658                  * Boost the allowable dirty threshold a bit for page
659                  * allocators so they don't get DoS'ed by heavy writers
660                  */
661                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
662
663                 if (global_page_state(NR_UNSTABLE_NFS) +
664                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
665                                 break;
666                 congestion_wait(BLK_RW_ASYNC, HZ/10);
667
668                 /*
669                  * The caller might hold locks which can prevent IO completion
670                  * or progress in the filesystem.  So we cannot just sit here
671                  * waiting for IO to complete.
672                  */
673                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
674                         break;
675         }
676 }
677
678 static void laptop_timer_fn(unsigned long unused);
679
680 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
681
682 /*
683  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
684  */
685 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
686         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
687 {
688         proc_dointvec(table, write, file, buffer, length, ppos);
689         return 0;
690 }
691
692 static void do_laptop_sync(struct work_struct *work)
693 {
694         wakeup_flusher_threads(0);
695         kfree(work);
696 }
697
698 static void laptop_timer_fn(unsigned long unused)
699 {
700         struct work_struct *work;
701
702         work = kmalloc(sizeof(*work), GFP_ATOMIC);
703         if (work) {
704                 INIT_WORK(work, do_laptop_sync);
705                 schedule_work(work);
706         }
707 }
708
709 /*
710  * We've spun up the disk and we're in laptop mode: schedule writeback
711  * of all dirty data a few seconds from now.  If the flush is already scheduled
712  * then push it back - the user is still using the disk.
713  */
714 void laptop_io_completion(void)
715 {
716         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
717 }
718
719 /*
720  * We're in laptop mode and we've just synced. The sync's writes will have
721  * caused another writeback to be scheduled by laptop_io_completion.
722  * Nothing needs to be written back anymore, so we unschedule the writeback.
723  */
724 void laptop_sync_completion(void)
725 {
726         del_timer(&laptop_mode_wb_timer);
727 }
728
729 /*
730  * If ratelimit_pages is too high then we can get into dirty-data overload
731  * if a large number of processes all perform writes at the same time.
732  * If it is too low then SMP machines will call the (expensive)
733  * get_writeback_state too often.
734  *
735  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
736  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
737  * thresholds before writeback cuts in.
738  *
739  * But the limit should not be set too high.  Because it also controls the
740  * amount of memory which the balance_dirty_pages() caller has to write back.
741  * If this is too large then the caller will block on the IO queue all the
742  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
743  * will write six megabyte chunks, max.
744  */
745
746 void writeback_set_ratelimit(void)
747 {
748         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
749         if (ratelimit_pages < 16)
750                 ratelimit_pages = 16;
751         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
752                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
753 }
754
755 static int __cpuinit
756 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
757 {
758         writeback_set_ratelimit();
759         return NOTIFY_DONE;
760 }
761
762 static struct notifier_block __cpuinitdata ratelimit_nb = {
763         .notifier_call  = ratelimit_handler,
764         .next           = NULL,
765 };
766
767 /*
768  * Called early on to tune the page writeback dirty limits.
769  *
770  * We used to scale dirty pages according to how total memory
771  * related to pages that could be allocated for buffers (by
772  * comparing nr_free_buffer_pages() to vm_total_pages.
773  *
774  * However, that was when we used "dirty_ratio" to scale with
775  * all memory, and we don't do that any more. "dirty_ratio"
776  * is now applied to total non-HIGHPAGE memory (by subtracting
777  * totalhigh_pages from vm_total_pages), and as such we can't
778  * get into the old insane situation any more where we had
779  * large amounts of dirty pages compared to a small amount of
780  * non-HIGHMEM memory.
781  *
782  * But we might still want to scale the dirty_ratio by how
783  * much memory the box has..
784  */
785 void __init page_writeback_init(void)
786 {
787         int shift;
788
789         writeback_set_ratelimit();
790         register_cpu_notifier(&ratelimit_nb);
791
792         shift = calc_period_shift();
793         prop_descriptor_init(&vm_completions, shift);
794         prop_descriptor_init(&vm_dirties, shift);
795 }
796
797 /**
798  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
799  * @mapping: address space structure to write
800  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
801  * @writepage: function called for each page
802  * @data: data passed to writepage function
803  *
804  * If a page is already under I/O, write_cache_pages() skips it, even
805  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
806  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
807  * and msync() need to guarantee that all the data which was dirty at the time
808  * the call was made get new I/O started against them.  If wbc->sync_mode is
809  * WB_SYNC_ALL then we were called for data integrity and we must wait for
810  * existing IO to complete.
811  */
812 int write_cache_pages(struct address_space *mapping,
813                       struct writeback_control *wbc, writepage_t writepage,
814                       void *data)
815 {
816         struct backing_dev_info *bdi = mapping->backing_dev_info;
817         int ret = 0;
818         int done = 0;
819         struct pagevec pvec;
820         int nr_pages;
821         pgoff_t uninitialized_var(writeback_index);
822         pgoff_t index;
823         pgoff_t end;            /* Inclusive */
824         pgoff_t done_index;
825         int cycled;
826         int range_whole = 0;
827         long nr_to_write = wbc->nr_to_write;
828
829         if (wbc->nonblocking && bdi_write_congested(bdi)) {
830                 wbc->encountered_congestion = 1;
831                 return 0;
832         }
833
834         pagevec_init(&pvec, 0);
835         if (wbc->range_cyclic) {
836                 writeback_index = mapping->writeback_index; /* prev offset */
837                 index = writeback_index;
838                 if (index == 0)
839                         cycled = 1;
840                 else
841                         cycled = 0;
842                 end = -1;
843         } else {
844                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
845                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
846                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
847                         range_whole = 1;
848                 cycled = 1; /* ignore range_cyclic tests */
849         }
850 retry:
851         done_index = index;
852         while (!done && (index <= end)) {
853                 int i;
854
855                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
856                               PAGECACHE_TAG_DIRTY,
857                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
858                 if (nr_pages == 0)
859                         break;
860
861                 for (i = 0; i < nr_pages; i++) {
862                         struct page *page = pvec.pages[i];
863
864                         /*
865                          * At this point, the page may be truncated or
866                          * invalidated (changing page->mapping to NULL), or
867                          * even swizzled back from swapper_space to tmpfs file
868                          * mapping. However, page->index will not change
869                          * because we have a reference on the page.
870                          */
871                         if (page->index > end) {
872                                 /*
873                                  * can't be range_cyclic (1st pass) because
874                                  * end == -1 in that case.
875                                  */
876                                 done = 1;
877                                 break;
878                         }
879
880                         done_index = page->index + 1;
881
882                         lock_page(page);
883
884                         /*
885                          * Page truncated or invalidated. We can freely skip it
886                          * then, even for data integrity operations: the page
887                          * has disappeared concurrently, so there could be no
888                          * real expectation of this data interity operation
889                          * even if there is now a new, dirty page at the same
890                          * pagecache address.
891                          */
892                         if (unlikely(page->mapping != mapping)) {
893 continue_unlock:
894                                 unlock_page(page);
895                                 continue;
896                         }
897
898                         if (!PageDirty(page)) {
899                                 /* someone wrote it for us */
900                                 goto continue_unlock;
901                         }
902
903                         if (PageWriteback(page)) {
904                                 if (wbc->sync_mode != WB_SYNC_NONE)
905                                         wait_on_page_writeback(page);
906                                 else
907                                         goto continue_unlock;
908                         }
909
910                         BUG_ON(PageWriteback(page));
911                         if (!clear_page_dirty_for_io(page))
912                                 goto continue_unlock;
913
914                         ret = (*writepage)(page, wbc, data);
915                         if (unlikely(ret)) {
916                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
917                                         unlock_page(page);
918                                         ret = 0;
919                                 } else {
920                                         /*
921                                          * done_index is set past this page,
922                                          * so media errors will not choke
923                                          * background writeout for the entire
924                                          * file. This has consequences for
925                                          * range_cyclic semantics (ie. it may
926                                          * not be suitable for data integrity
927                                          * writeout).
928                                          */
929                                         done = 1;
930                                         break;
931                                 }
932                         }
933
934                         if (nr_to_write > 0) {
935                                 nr_to_write--;
936                                 if (nr_to_write == 0 &&
937                                     wbc->sync_mode == WB_SYNC_NONE) {
938                                         /*
939                                          * We stop writing back only if we are
940                                          * not doing integrity sync. In case of
941                                          * integrity sync we have to keep going
942                                          * because someone may be concurrently
943                                          * dirtying pages, and we might have
944                                          * synced a lot of newly appeared dirty
945                                          * pages, but have not synced all of the
946                                          * old dirty pages.
947                                          */
948                                         done = 1;
949                                         break;
950                                 }
951                         }
952
953                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
954                                 wbc->encountered_congestion = 1;
955                                 done = 1;
956                                 break;
957                         }
958                 }
959                 pagevec_release(&pvec);
960                 cond_resched();
961         }
962         if (!cycled && !done) {
963                 /*
964                  * range_cyclic:
965                  * We hit the last page and there is more work to be done: wrap
966                  * back to the start of the file
967                  */
968                 cycled = 1;
969                 index = 0;
970                 end = writeback_index - 1;
971                 goto retry;
972         }
973         if (!wbc->no_nrwrite_index_update) {
974                 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
975                         mapping->writeback_index = done_index;
976                 wbc->nr_to_write = nr_to_write;
977         }
978
979         return ret;
980 }
981 EXPORT_SYMBOL(write_cache_pages);
982
983 /*
984  * Function used by generic_writepages to call the real writepage
985  * function and set the mapping flags on error
986  */
987 static int __writepage(struct page *page, struct writeback_control *wbc,
988                        void *data)
989 {
990         struct address_space *mapping = data;
991         int ret = mapping->a_ops->writepage(page, wbc);
992         mapping_set_error(mapping, ret);
993         return ret;
994 }
995
996 /**
997  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
998  * @mapping: address space structure to write
999  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1000  *
1001  * This is a library function, which implements the writepages()
1002  * address_space_operation.
1003  */
1004 int generic_writepages(struct address_space *mapping,
1005                        struct writeback_control *wbc)
1006 {
1007         /* deal with chardevs and other special file */
1008         if (!mapping->a_ops->writepage)
1009                 return 0;
1010
1011         return write_cache_pages(mapping, wbc, __writepage, mapping);
1012 }
1013
1014 EXPORT_SYMBOL(generic_writepages);
1015
1016 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1017 {
1018         int ret;
1019
1020         if (wbc->nr_to_write <= 0)
1021                 return 0;
1022         wbc->for_writepages = 1;
1023         if (mapping->a_ops->writepages)
1024                 ret = mapping->a_ops->writepages(mapping, wbc);
1025         else
1026                 ret = generic_writepages(mapping, wbc);
1027         wbc->for_writepages = 0;
1028         return ret;
1029 }
1030
1031 /**
1032  * write_one_page - write out a single page and optionally wait on I/O
1033  * @page: the page to write
1034  * @wait: if true, wait on writeout
1035  *
1036  * The page must be locked by the caller and will be unlocked upon return.
1037  *
1038  * write_one_page() returns a negative error code if I/O failed.
1039  */
1040 int write_one_page(struct page *page, int wait)
1041 {
1042         struct address_space *mapping = page->mapping;
1043         int ret = 0;
1044         struct writeback_control wbc = {
1045                 .sync_mode = WB_SYNC_ALL,
1046                 .nr_to_write = 1,
1047         };
1048
1049         BUG_ON(!PageLocked(page));
1050
1051         if (wait)
1052                 wait_on_page_writeback(page);
1053
1054         if (clear_page_dirty_for_io(page)) {
1055                 page_cache_get(page);
1056                 ret = mapping->a_ops->writepage(page, &wbc);
1057                 if (ret == 0 && wait) {
1058                         wait_on_page_writeback(page);
1059                         if (PageError(page))
1060                                 ret = -EIO;
1061                 }
1062                 page_cache_release(page);
1063         } else {
1064                 unlock_page(page);
1065         }
1066         return ret;
1067 }
1068 EXPORT_SYMBOL(write_one_page);
1069
1070 /*
1071  * For address_spaces which do not use buffers nor write back.
1072  */
1073 int __set_page_dirty_no_writeback(struct page *page)
1074 {
1075         if (!PageDirty(page))
1076                 SetPageDirty(page);
1077         return 0;
1078 }
1079
1080 /*
1081  * Helper function for set_page_dirty family.
1082  * NOTE: This relies on being atomic wrt interrupts.
1083  */
1084 void account_page_dirtied(struct page *page, struct address_space *mapping)
1085 {
1086         if (mapping_cap_account_dirty(mapping)) {
1087                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1088                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1089                 task_dirty_inc(current);
1090                 task_io_account_write(PAGE_CACHE_SIZE);
1091         }
1092 }
1093
1094 /*
1095  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1096  * its radix tree.
1097  *
1098  * This is also used when a single buffer is being dirtied: we want to set the
1099  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1100  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1101  *
1102  * Most callers have locked the page, which pins the address_space in memory.
1103  * But zap_pte_range() does not lock the page, however in that case the
1104  * mapping is pinned by the vma's ->vm_file reference.
1105  *
1106  * We take care to handle the case where the page was truncated from the
1107  * mapping by re-checking page_mapping() inside tree_lock.
1108  */
1109 int __set_page_dirty_nobuffers(struct page *page)
1110 {
1111         if (!TestSetPageDirty(page)) {
1112                 struct address_space *mapping = page_mapping(page);
1113                 struct address_space *mapping2;
1114
1115                 if (!mapping)
1116                         return 1;
1117
1118                 spin_lock_irq(&mapping->tree_lock);
1119                 mapping2 = page_mapping(page);
1120                 if (mapping2) { /* Race with truncate? */
1121                         BUG_ON(mapping2 != mapping);
1122                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1123                         account_page_dirtied(page, mapping);
1124                         radix_tree_tag_set(&mapping->page_tree,
1125                                 page_index(page), PAGECACHE_TAG_DIRTY);
1126                 }
1127                 spin_unlock_irq(&mapping->tree_lock);
1128                 if (mapping->host) {
1129                         /* !PageAnon && !swapper_space */
1130                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1131                 }
1132                 return 1;
1133         }
1134         return 0;
1135 }
1136 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1137
1138 /*
1139  * When a writepage implementation decides that it doesn't want to write this
1140  * page for some reason, it should redirty the locked page via
1141  * redirty_page_for_writepage() and it should then unlock the page and return 0
1142  */
1143 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1144 {
1145         wbc->pages_skipped++;
1146         return __set_page_dirty_nobuffers(page);
1147 }
1148 EXPORT_SYMBOL(redirty_page_for_writepage);
1149
1150 /*
1151  * If the mapping doesn't provide a set_page_dirty a_op, then
1152  * just fall through and assume that it wants buffer_heads.
1153  */
1154 int set_page_dirty(struct page *page)
1155 {
1156         struct address_space *mapping = page_mapping(page);
1157
1158         if (likely(mapping)) {
1159                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1160 #ifdef CONFIG_BLOCK
1161                 if (!spd)
1162                         spd = __set_page_dirty_buffers;
1163 #endif
1164                 return (*spd)(page);
1165         }
1166         if (!PageDirty(page)) {
1167                 if (!TestSetPageDirty(page))
1168                         return 1;
1169         }
1170         return 0;
1171 }
1172 EXPORT_SYMBOL(set_page_dirty);
1173
1174 /*
1175  * set_page_dirty() is racy if the caller has no reference against
1176  * page->mapping->host, and if the page is unlocked.  This is because another
1177  * CPU could truncate the page off the mapping and then free the mapping.
1178  *
1179  * Usually, the page _is_ locked, or the caller is a user-space process which
1180  * holds a reference on the inode by having an open file.
1181  *
1182  * In other cases, the page should be locked before running set_page_dirty().
1183  */
1184 int set_page_dirty_lock(struct page *page)
1185 {
1186         int ret;
1187
1188         lock_page_nosync(page);
1189         ret = set_page_dirty(page);
1190         unlock_page(page);
1191         return ret;
1192 }
1193 EXPORT_SYMBOL(set_page_dirty_lock);
1194
1195 /*
1196  * Clear a page's dirty flag, while caring for dirty memory accounting.
1197  * Returns true if the page was previously dirty.
1198  *
1199  * This is for preparing to put the page under writeout.  We leave the page
1200  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1201  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1202  * implementation will run either set_page_writeback() or set_page_dirty(),
1203  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1204  * back into sync.
1205  *
1206  * This incoherency between the page's dirty flag and radix-tree tag is
1207  * unfortunate, but it only exists while the page is locked.
1208  */
1209 int clear_page_dirty_for_io(struct page *page)
1210 {
1211         struct address_space *mapping = page_mapping(page);
1212
1213         BUG_ON(!PageLocked(page));
1214
1215         ClearPageReclaim(page);
1216         if (mapping && mapping_cap_account_dirty(mapping)) {
1217                 /*
1218                  * Yes, Virginia, this is indeed insane.
1219                  *
1220                  * We use this sequence to make sure that
1221                  *  (a) we account for dirty stats properly
1222                  *  (b) we tell the low-level filesystem to
1223                  *      mark the whole page dirty if it was
1224                  *      dirty in a pagetable. Only to then
1225                  *  (c) clean the page again and return 1 to
1226                  *      cause the writeback.
1227                  *
1228                  * This way we avoid all nasty races with the
1229                  * dirty bit in multiple places and clearing
1230                  * them concurrently from different threads.
1231                  *
1232                  * Note! Normally the "set_page_dirty(page)"
1233                  * has no effect on the actual dirty bit - since
1234                  * that will already usually be set. But we
1235                  * need the side effects, and it can help us
1236                  * avoid races.
1237                  *
1238                  * We basically use the page "master dirty bit"
1239                  * as a serialization point for all the different
1240                  * threads doing their things.
1241                  */
1242                 if (page_mkclean(page))
1243                         set_page_dirty(page);
1244                 /*
1245                  * We carefully synchronise fault handlers against
1246                  * installing a dirty pte and marking the page dirty
1247                  * at this point. We do this by having them hold the
1248                  * page lock at some point after installing their
1249                  * pte, but before marking the page dirty.
1250                  * Pages are always locked coming in here, so we get
1251                  * the desired exclusion. See mm/memory.c:do_wp_page()
1252                  * for more comments.
1253                  */
1254                 if (TestClearPageDirty(page)) {
1255                         dec_zone_page_state(page, NR_FILE_DIRTY);
1256                         dec_bdi_stat(mapping->backing_dev_info,
1257                                         BDI_RECLAIMABLE);
1258                         return 1;
1259                 }
1260                 return 0;
1261         }
1262         return TestClearPageDirty(page);
1263 }
1264 EXPORT_SYMBOL(clear_page_dirty_for_io);
1265
1266 int test_clear_page_writeback(struct page *page)
1267 {
1268         struct address_space *mapping = page_mapping(page);
1269         int ret;
1270
1271         if (mapping) {
1272                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1273                 unsigned long flags;
1274
1275                 spin_lock_irqsave(&mapping->tree_lock, flags);
1276                 ret = TestClearPageWriteback(page);
1277                 if (ret) {
1278                         radix_tree_tag_clear(&mapping->page_tree,
1279                                                 page_index(page),
1280                                                 PAGECACHE_TAG_WRITEBACK);
1281                         if (bdi_cap_account_writeback(bdi)) {
1282                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1283                                 __bdi_writeout_inc(bdi);
1284                         }
1285                 }
1286                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1287         } else {
1288                 ret = TestClearPageWriteback(page);
1289         }
1290         if (ret)
1291                 dec_zone_page_state(page, NR_WRITEBACK);
1292         return ret;
1293 }
1294
1295 int test_set_page_writeback(struct page *page)
1296 {
1297         struct address_space *mapping = page_mapping(page);
1298         int ret;
1299
1300         if (mapping) {
1301                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1302                 unsigned long flags;
1303
1304                 spin_lock_irqsave(&mapping->tree_lock, flags);
1305                 ret = TestSetPageWriteback(page);
1306                 if (!ret) {
1307                         radix_tree_tag_set(&mapping->page_tree,
1308                                                 page_index(page),
1309                                                 PAGECACHE_TAG_WRITEBACK);
1310                         if (bdi_cap_account_writeback(bdi))
1311                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1312                 }
1313                 if (!PageDirty(page))
1314                         radix_tree_tag_clear(&mapping->page_tree,
1315                                                 page_index(page),
1316                                                 PAGECACHE_TAG_DIRTY);
1317                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1318         } else {
1319                 ret = TestSetPageWriteback(page);
1320         }
1321         if (!ret)
1322                 inc_zone_page_state(page, NR_WRITEBACK);
1323         return ret;
1324
1325 }
1326 EXPORT_SYMBOL(test_set_page_writeback);
1327
1328 /*
1329  * Return true if any of the pages in the mapping are marked with the
1330  * passed tag.
1331  */
1332 int mapping_tagged(struct address_space *mapping, int tag)
1333 {
1334         int ret;
1335         rcu_read_lock();
1336         ret = radix_tree_tagged(&mapping->page_tree, tag);
1337         rcu_read_unlock();
1338         return ret;
1339 }
1340 EXPORT_SYMBOL(mapping_tagged);