writeback: don't use schedule_timeout() without setting runstate
[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_bh(&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_bh(&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_bh(&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_bh(&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_interruptible(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                 bdi_start_writeback(bdi, nr_writeback);
587 }
588
589 void set_page_dirty_balance(struct page *page, int page_mkwrite)
590 {
591         if (set_page_dirty(page) || page_mkwrite) {
592                 struct address_space *mapping = page_mapping(page);
593
594                 if (mapping)
595                         balance_dirty_pages_ratelimited(mapping);
596         }
597 }
598
599 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
600
601 /**
602  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
603  * @mapping: address_space which was dirtied
604  * @nr_pages_dirtied: number of pages which the caller has just dirtied
605  *
606  * Processes which are dirtying memory should call in here once for each page
607  * which was newly dirtied.  The function will periodically check the system's
608  * dirty state and will initiate writeback if needed.
609  *
610  * On really big machines, get_writeback_state is expensive, so try to avoid
611  * calling it too often (ratelimiting).  But once we're over the dirty memory
612  * limit we decrease the ratelimiting by a lot, to prevent individual processes
613  * from overshooting the limit by (ratelimit_pages) each.
614  */
615 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
616                                         unsigned long nr_pages_dirtied)
617 {
618         unsigned long ratelimit;
619         unsigned long *p;
620
621         ratelimit = ratelimit_pages;
622         if (mapping->backing_dev_info->dirty_exceeded)
623                 ratelimit = 8;
624
625         /*
626          * Check the rate limiting. Also, we do not want to throttle real-time
627          * tasks in balance_dirty_pages(). Period.
628          */
629         preempt_disable();
630         p =  &__get_cpu_var(bdp_ratelimits);
631         *p += nr_pages_dirtied;
632         if (unlikely(*p >= ratelimit)) {
633                 *p = 0;
634                 preempt_enable();
635                 balance_dirty_pages(mapping);
636                 return;
637         }
638         preempt_enable();
639 }
640 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
641
642 void throttle_vm_writeout(gfp_t gfp_mask)
643 {
644         unsigned long background_thresh;
645         unsigned long dirty_thresh;
646
647         for ( ; ; ) {
648                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
649
650                 /*
651                  * Boost the allowable dirty threshold a bit for page
652                  * allocators so they don't get DoS'ed by heavy writers
653                  */
654                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
655
656                 if (global_page_state(NR_UNSTABLE_NFS) +
657                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
658                                 break;
659                 congestion_wait(BLK_RW_ASYNC, HZ/10);
660
661                 /*
662                  * The caller might hold locks which can prevent IO completion
663                  * or progress in the filesystem.  So we cannot just sit here
664                  * waiting for IO to complete.
665                  */
666                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
667                         break;
668         }
669 }
670
671 static void laptop_timer_fn(unsigned long unused);
672
673 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
674
675 /*
676  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
677  */
678 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
679         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
680 {
681         proc_dointvec(table, write, file, buffer, length, ppos);
682         return 0;
683 }
684
685 static void do_laptop_sync(struct work_struct *work)
686 {
687         wakeup_flusher_threads(0);
688         kfree(work);
689 }
690
691 static void laptop_timer_fn(unsigned long unused)
692 {
693         struct work_struct *work;
694
695         work = kmalloc(sizeof(*work), GFP_ATOMIC);
696         if (work) {
697                 INIT_WORK(work, do_laptop_sync);
698                 schedule_work(work);
699         }
700 }
701
702 /*
703  * We've spun up the disk and we're in laptop mode: schedule writeback
704  * of all dirty data a few seconds from now.  If the flush is already scheduled
705  * then push it back - the user is still using the disk.
706  */
707 void laptop_io_completion(void)
708 {
709         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
710 }
711
712 /*
713  * We're in laptop mode and we've just synced. The sync's writes will have
714  * caused another writeback to be scheduled by laptop_io_completion.
715  * Nothing needs to be written back anymore, so we unschedule the writeback.
716  */
717 void laptop_sync_completion(void)
718 {
719         del_timer(&laptop_mode_wb_timer);
720 }
721
722 /*
723  * If ratelimit_pages is too high then we can get into dirty-data overload
724  * if a large number of processes all perform writes at the same time.
725  * If it is too low then SMP machines will call the (expensive)
726  * get_writeback_state too often.
727  *
728  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
729  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
730  * thresholds before writeback cuts in.
731  *
732  * But the limit should not be set too high.  Because it also controls the
733  * amount of memory which the balance_dirty_pages() caller has to write back.
734  * If this is too large then the caller will block on the IO queue all the
735  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
736  * will write six megabyte chunks, max.
737  */
738
739 void writeback_set_ratelimit(void)
740 {
741         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
742         if (ratelimit_pages < 16)
743                 ratelimit_pages = 16;
744         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
745                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
746 }
747
748 static int __cpuinit
749 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
750 {
751         writeback_set_ratelimit();
752         return NOTIFY_DONE;
753 }
754
755 static struct notifier_block __cpuinitdata ratelimit_nb = {
756         .notifier_call  = ratelimit_handler,
757         .next           = NULL,
758 };
759
760 /*
761  * Called early on to tune the page writeback dirty limits.
762  *
763  * We used to scale dirty pages according to how total memory
764  * related to pages that could be allocated for buffers (by
765  * comparing nr_free_buffer_pages() to vm_total_pages.
766  *
767  * However, that was when we used "dirty_ratio" to scale with
768  * all memory, and we don't do that any more. "dirty_ratio"
769  * is now applied to total non-HIGHPAGE memory (by subtracting
770  * totalhigh_pages from vm_total_pages), and as such we can't
771  * get into the old insane situation any more where we had
772  * large amounts of dirty pages compared to a small amount of
773  * non-HIGHMEM memory.
774  *
775  * But we might still want to scale the dirty_ratio by how
776  * much memory the box has..
777  */
778 void __init page_writeback_init(void)
779 {
780         int shift;
781
782         writeback_set_ratelimit();
783         register_cpu_notifier(&ratelimit_nb);
784
785         shift = calc_period_shift();
786         prop_descriptor_init(&vm_completions, shift);
787         prop_descriptor_init(&vm_dirties, shift);
788 }
789
790 /**
791  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
792  * @mapping: address space structure to write
793  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
794  * @writepage: function called for each page
795  * @data: data passed to writepage function
796  *
797  * If a page is already under I/O, write_cache_pages() skips it, even
798  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
799  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
800  * and msync() need to guarantee that all the data which was dirty at the time
801  * the call was made get new I/O started against them.  If wbc->sync_mode is
802  * WB_SYNC_ALL then we were called for data integrity and we must wait for
803  * existing IO to complete.
804  */
805 int write_cache_pages(struct address_space *mapping,
806                       struct writeback_control *wbc, writepage_t writepage,
807                       void *data)
808 {
809         struct backing_dev_info *bdi = mapping->backing_dev_info;
810         int ret = 0;
811         int done = 0;
812         struct pagevec pvec;
813         int nr_pages;
814         pgoff_t uninitialized_var(writeback_index);
815         pgoff_t index;
816         pgoff_t end;            /* Inclusive */
817         pgoff_t done_index;
818         int cycled;
819         int range_whole = 0;
820         long nr_to_write = wbc->nr_to_write;
821
822         if (wbc->nonblocking && bdi_write_congested(bdi)) {
823                 wbc->encountered_congestion = 1;
824                 return 0;
825         }
826
827         pagevec_init(&pvec, 0);
828         if (wbc->range_cyclic) {
829                 writeback_index = mapping->writeback_index; /* prev offset */
830                 index = writeback_index;
831                 if (index == 0)
832                         cycled = 1;
833                 else
834                         cycled = 0;
835                 end = -1;
836         } else {
837                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
838                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
839                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
840                         range_whole = 1;
841                 cycled = 1; /* ignore range_cyclic tests */
842         }
843 retry:
844         done_index = index;
845         while (!done && (index <= end)) {
846                 int i;
847
848                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
849                               PAGECACHE_TAG_DIRTY,
850                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
851                 if (nr_pages == 0)
852                         break;
853
854                 for (i = 0; i < nr_pages; i++) {
855                         struct page *page = pvec.pages[i];
856
857                         /*
858                          * At this point, the page may be truncated or
859                          * invalidated (changing page->mapping to NULL), or
860                          * even swizzled back from swapper_space to tmpfs file
861                          * mapping. However, page->index will not change
862                          * because we have a reference on the page.
863                          */
864                         if (page->index > end) {
865                                 /*
866                                  * can't be range_cyclic (1st pass) because
867                                  * end == -1 in that case.
868                                  */
869                                 done = 1;
870                                 break;
871                         }
872
873                         done_index = page->index + 1;
874
875                         lock_page(page);
876
877                         /*
878                          * Page truncated or invalidated. We can freely skip it
879                          * then, even for data integrity operations: the page
880                          * has disappeared concurrently, so there could be no
881                          * real expectation of this data interity operation
882                          * even if there is now a new, dirty page at the same
883                          * pagecache address.
884                          */
885                         if (unlikely(page->mapping != mapping)) {
886 continue_unlock:
887                                 unlock_page(page);
888                                 continue;
889                         }
890
891                         if (!PageDirty(page)) {
892                                 /* someone wrote it for us */
893                                 goto continue_unlock;
894                         }
895
896                         if (PageWriteback(page)) {
897                                 if (wbc->sync_mode != WB_SYNC_NONE)
898                                         wait_on_page_writeback(page);
899                                 else
900                                         goto continue_unlock;
901                         }
902
903                         BUG_ON(PageWriteback(page));
904                         if (!clear_page_dirty_for_io(page))
905                                 goto continue_unlock;
906
907                         ret = (*writepage)(page, wbc, data);
908                         if (unlikely(ret)) {
909                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
910                                         unlock_page(page);
911                                         ret = 0;
912                                 } else {
913                                         /*
914                                          * done_index is set past this page,
915                                          * so media errors will not choke
916                                          * background writeout for the entire
917                                          * file. This has consequences for
918                                          * range_cyclic semantics (ie. it may
919                                          * not be suitable for data integrity
920                                          * writeout).
921                                          */
922                                         done = 1;
923                                         break;
924                                 }
925                         }
926
927                         if (nr_to_write > 0) {
928                                 nr_to_write--;
929                                 if (nr_to_write == 0 &&
930                                     wbc->sync_mode == WB_SYNC_NONE) {
931                                         /*
932                                          * We stop writing back only if we are
933                                          * not doing integrity sync. In case of
934                                          * integrity sync we have to keep going
935                                          * because someone may be concurrently
936                                          * dirtying pages, and we might have
937                                          * synced a lot of newly appeared dirty
938                                          * pages, but have not synced all of the
939                                          * old dirty pages.
940                                          */
941                                         done = 1;
942                                         break;
943                                 }
944                         }
945
946                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
947                                 wbc->encountered_congestion = 1;
948                                 done = 1;
949                                 break;
950                         }
951                 }
952                 pagevec_release(&pvec);
953                 cond_resched();
954         }
955         if (!cycled && !done) {
956                 /*
957                  * range_cyclic:
958                  * We hit the last page and there is more work to be done: wrap
959                  * back to the start of the file
960                  */
961                 cycled = 1;
962                 index = 0;
963                 end = writeback_index - 1;
964                 goto retry;
965         }
966         if (!wbc->no_nrwrite_index_update) {
967                 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
968                         mapping->writeback_index = done_index;
969                 wbc->nr_to_write = nr_to_write;
970         }
971
972         return ret;
973 }
974 EXPORT_SYMBOL(write_cache_pages);
975
976 /*
977  * Function used by generic_writepages to call the real writepage
978  * function and set the mapping flags on error
979  */
980 static int __writepage(struct page *page, struct writeback_control *wbc,
981                        void *data)
982 {
983         struct address_space *mapping = data;
984         int ret = mapping->a_ops->writepage(page, wbc);
985         mapping_set_error(mapping, ret);
986         return ret;
987 }
988
989 /**
990  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
991  * @mapping: address space structure to write
992  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
993  *
994  * This is a library function, which implements the writepages()
995  * address_space_operation.
996  */
997 int generic_writepages(struct address_space *mapping,
998                        struct writeback_control *wbc)
999 {
1000         /* deal with chardevs and other special file */
1001         if (!mapping->a_ops->writepage)
1002                 return 0;
1003
1004         return write_cache_pages(mapping, wbc, __writepage, mapping);
1005 }
1006
1007 EXPORT_SYMBOL(generic_writepages);
1008
1009 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1010 {
1011         int ret;
1012
1013         if (wbc->nr_to_write <= 0)
1014                 return 0;
1015         if (mapping->a_ops->writepages)
1016                 ret = mapping->a_ops->writepages(mapping, wbc);
1017         else
1018                 ret = generic_writepages(mapping, wbc);
1019         return ret;
1020 }
1021
1022 /**
1023  * write_one_page - write out a single page and optionally wait on I/O
1024  * @page: the page to write
1025  * @wait: if true, wait on writeout
1026  *
1027  * The page must be locked by the caller and will be unlocked upon return.
1028  *
1029  * write_one_page() returns a negative error code if I/O failed.
1030  */
1031 int write_one_page(struct page *page, int wait)
1032 {
1033         struct address_space *mapping = page->mapping;
1034         int ret = 0;
1035         struct writeback_control wbc = {
1036                 .sync_mode = WB_SYNC_ALL,
1037                 .nr_to_write = 1,
1038         };
1039
1040         BUG_ON(!PageLocked(page));
1041
1042         if (wait)
1043                 wait_on_page_writeback(page);
1044
1045         if (clear_page_dirty_for_io(page)) {
1046                 page_cache_get(page);
1047                 ret = mapping->a_ops->writepage(page, &wbc);
1048                 if (ret == 0 && wait) {
1049                         wait_on_page_writeback(page);
1050                         if (PageError(page))
1051                                 ret = -EIO;
1052                 }
1053                 page_cache_release(page);
1054         } else {
1055                 unlock_page(page);
1056         }
1057         return ret;
1058 }
1059 EXPORT_SYMBOL(write_one_page);
1060
1061 /*
1062  * For address_spaces which do not use buffers nor write back.
1063  */
1064 int __set_page_dirty_no_writeback(struct page *page)
1065 {
1066         if (!PageDirty(page))
1067                 SetPageDirty(page);
1068         return 0;
1069 }
1070
1071 /*
1072  * Helper function for set_page_dirty family.
1073  * NOTE: This relies on being atomic wrt interrupts.
1074  */
1075 void account_page_dirtied(struct page *page, struct address_space *mapping)
1076 {
1077         if (mapping_cap_account_dirty(mapping)) {
1078                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1079                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1080                 task_dirty_inc(current);
1081                 task_io_account_write(PAGE_CACHE_SIZE);
1082         }
1083 }
1084
1085 /*
1086  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1087  * its radix tree.
1088  *
1089  * This is also used when a single buffer is being dirtied: we want to set the
1090  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1091  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1092  *
1093  * Most callers have locked the page, which pins the address_space in memory.
1094  * But zap_pte_range() does not lock the page, however in that case the
1095  * mapping is pinned by the vma's ->vm_file reference.
1096  *
1097  * We take care to handle the case where the page was truncated from the
1098  * mapping by re-checking page_mapping() inside tree_lock.
1099  */
1100 int __set_page_dirty_nobuffers(struct page *page)
1101 {
1102         if (!TestSetPageDirty(page)) {
1103                 struct address_space *mapping = page_mapping(page);
1104                 struct address_space *mapping2;
1105
1106                 if (!mapping)
1107                         return 1;
1108
1109                 spin_lock_irq(&mapping->tree_lock);
1110                 mapping2 = page_mapping(page);
1111                 if (mapping2) { /* Race with truncate? */
1112                         BUG_ON(mapping2 != mapping);
1113                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1114                         account_page_dirtied(page, mapping);
1115                         radix_tree_tag_set(&mapping->page_tree,
1116                                 page_index(page), PAGECACHE_TAG_DIRTY);
1117                 }
1118                 spin_unlock_irq(&mapping->tree_lock);
1119                 if (mapping->host) {
1120                         /* !PageAnon && !swapper_space */
1121                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1122                 }
1123                 return 1;
1124         }
1125         return 0;
1126 }
1127 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1128
1129 /*
1130  * When a writepage implementation decides that it doesn't want to write this
1131  * page for some reason, it should redirty the locked page via
1132  * redirty_page_for_writepage() and it should then unlock the page and return 0
1133  */
1134 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1135 {
1136         wbc->pages_skipped++;
1137         return __set_page_dirty_nobuffers(page);
1138 }
1139 EXPORT_SYMBOL(redirty_page_for_writepage);
1140
1141 /*
1142  * If the mapping doesn't provide a set_page_dirty a_op, then
1143  * just fall through and assume that it wants buffer_heads.
1144  */
1145 int set_page_dirty(struct page *page)
1146 {
1147         struct address_space *mapping = page_mapping(page);
1148
1149         if (likely(mapping)) {
1150                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1151 #ifdef CONFIG_BLOCK
1152                 if (!spd)
1153                         spd = __set_page_dirty_buffers;
1154 #endif
1155                 return (*spd)(page);
1156         }
1157         if (!PageDirty(page)) {
1158                 if (!TestSetPageDirty(page))
1159                         return 1;
1160         }
1161         return 0;
1162 }
1163 EXPORT_SYMBOL(set_page_dirty);
1164
1165 /*
1166  * set_page_dirty() is racy if the caller has no reference against
1167  * page->mapping->host, and if the page is unlocked.  This is because another
1168  * CPU could truncate the page off the mapping and then free the mapping.
1169  *
1170  * Usually, the page _is_ locked, or the caller is a user-space process which
1171  * holds a reference on the inode by having an open file.
1172  *
1173  * In other cases, the page should be locked before running set_page_dirty().
1174  */
1175 int set_page_dirty_lock(struct page *page)
1176 {
1177         int ret;
1178
1179         lock_page_nosync(page);
1180         ret = set_page_dirty(page);
1181         unlock_page(page);
1182         return ret;
1183 }
1184 EXPORT_SYMBOL(set_page_dirty_lock);
1185
1186 /*
1187  * Clear a page's dirty flag, while caring for dirty memory accounting.
1188  * Returns true if the page was previously dirty.
1189  *
1190  * This is for preparing to put the page under writeout.  We leave the page
1191  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1192  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1193  * implementation will run either set_page_writeback() or set_page_dirty(),
1194  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1195  * back into sync.
1196  *
1197  * This incoherency between the page's dirty flag and radix-tree tag is
1198  * unfortunate, but it only exists while the page is locked.
1199  */
1200 int clear_page_dirty_for_io(struct page *page)
1201 {
1202         struct address_space *mapping = page_mapping(page);
1203
1204         BUG_ON(!PageLocked(page));
1205
1206         ClearPageReclaim(page);
1207         if (mapping && mapping_cap_account_dirty(mapping)) {
1208                 /*
1209                  * Yes, Virginia, this is indeed insane.
1210                  *
1211                  * We use this sequence to make sure that
1212                  *  (a) we account for dirty stats properly
1213                  *  (b) we tell the low-level filesystem to
1214                  *      mark the whole page dirty if it was
1215                  *      dirty in a pagetable. Only to then
1216                  *  (c) clean the page again and return 1 to
1217                  *      cause the writeback.
1218                  *
1219                  * This way we avoid all nasty races with the
1220                  * dirty bit in multiple places and clearing
1221                  * them concurrently from different threads.
1222                  *
1223                  * Note! Normally the "set_page_dirty(page)"
1224                  * has no effect on the actual dirty bit - since
1225                  * that will already usually be set. But we
1226                  * need the side effects, and it can help us
1227                  * avoid races.
1228                  *
1229                  * We basically use the page "master dirty bit"
1230                  * as a serialization point for all the different
1231                  * threads doing their things.
1232                  */
1233                 if (page_mkclean(page))
1234                         set_page_dirty(page);
1235                 /*
1236                  * We carefully synchronise fault handlers against
1237                  * installing a dirty pte and marking the page dirty
1238                  * at this point. We do this by having them hold the
1239                  * page lock at some point after installing their
1240                  * pte, but before marking the page dirty.
1241                  * Pages are always locked coming in here, so we get
1242                  * the desired exclusion. See mm/memory.c:do_wp_page()
1243                  * for more comments.
1244                  */
1245                 if (TestClearPageDirty(page)) {
1246                         dec_zone_page_state(page, NR_FILE_DIRTY);
1247                         dec_bdi_stat(mapping->backing_dev_info,
1248                                         BDI_RECLAIMABLE);
1249                         return 1;
1250                 }
1251                 return 0;
1252         }
1253         return TestClearPageDirty(page);
1254 }
1255 EXPORT_SYMBOL(clear_page_dirty_for_io);
1256
1257 int test_clear_page_writeback(struct page *page)
1258 {
1259         struct address_space *mapping = page_mapping(page);
1260         int ret;
1261
1262         if (mapping) {
1263                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1264                 unsigned long flags;
1265
1266                 spin_lock_irqsave(&mapping->tree_lock, flags);
1267                 ret = TestClearPageWriteback(page);
1268                 if (ret) {
1269                         radix_tree_tag_clear(&mapping->page_tree,
1270                                                 page_index(page),
1271                                                 PAGECACHE_TAG_WRITEBACK);
1272                         if (bdi_cap_account_writeback(bdi)) {
1273                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1274                                 __bdi_writeout_inc(bdi);
1275                         }
1276                 }
1277                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1278         } else {
1279                 ret = TestClearPageWriteback(page);
1280         }
1281         if (ret)
1282                 dec_zone_page_state(page, NR_WRITEBACK);
1283         return ret;
1284 }
1285
1286 int test_set_page_writeback(struct page *page)
1287 {
1288         struct address_space *mapping = page_mapping(page);
1289         int ret;
1290
1291         if (mapping) {
1292                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1293                 unsigned long flags;
1294
1295                 spin_lock_irqsave(&mapping->tree_lock, flags);
1296                 ret = TestSetPageWriteback(page);
1297                 if (!ret) {
1298                         radix_tree_tag_set(&mapping->page_tree,
1299                                                 page_index(page),
1300                                                 PAGECACHE_TAG_WRITEBACK);
1301                         if (bdi_cap_account_writeback(bdi))
1302                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1303                 }
1304                 if (!PageDirty(page))
1305                         radix_tree_tag_clear(&mapping->page_tree,
1306                                                 page_index(page),
1307                                                 PAGECACHE_TAG_DIRTY);
1308                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1309         } else {
1310                 ret = TestSetPageWriteback(page);
1311         }
1312         if (!ret)
1313                 inc_zone_page_state(page, NR_WRITEBACK);
1314         return ret;
1315
1316 }
1317 EXPORT_SYMBOL(test_set_page_writeback);
1318
1319 /*
1320  * Return true if any of the pages in the mapping are marked with the
1321  * passed tag.
1322  */
1323 int mapping_tagged(struct address_space *mapping, int tag)
1324 {
1325         int ret;
1326         rcu_read_lock();
1327         ret = radix_tree_tagged(&mapping->page_tree, tag);
1328         rcu_read_unlock();
1329         return ret;
1330 }
1331 EXPORT_SYMBOL(mapping_tagged);