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