Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / 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    akpm@zip.com.au
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  * The maximum number of pages to writeout in a single bdflush/kupdate
40  * operation.  We do this so we don't hold I_SYNC against an inode for
41  * enormous amounts of time, which would block a userspace task which has
42  * been forced to throttle against that inode.  Also, the code reevaluates
43  * the dirty each time it has written this many pages.
44  */
45 #define MAX_WRITEBACK_PAGES     1024
46
47 /*
48  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
49  * will look to see if it needs to force writeback or throttling.
50  */
51 static long ratelimit_pages = 32;
52
53 /*
54  * When balance_dirty_pages decides that the caller needs to perform some
55  * non-background writeback, this is how many pages it will attempt to write.
56  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
57  * large amounts of I/O are submitted.
58  */
59 static inline long sync_writeback_pages(void)
60 {
61         return ratelimit_pages + ratelimit_pages / 2;
62 }
63
64 /* The following parameters are exported via /proc/sys/vm */
65
66 /*
67  * Start background writeback (via pdflush) at this percentage
68  */
69 int dirty_background_ratio = 5;
70
71 /*
72  * free highmem will not be subtracted from the total free memory
73  * for calculating free ratios if vm_highmem_is_dirtyable is true
74  */
75 int vm_highmem_is_dirtyable;
76
77 /*
78  * The generator of dirty data starts writeback at this percentage
79  */
80 int vm_dirty_ratio = 10;
81
82 /*
83  * The interval between `kupdate'-style writebacks, in jiffies
84  */
85 int dirty_writeback_interval = 5 * HZ;
86
87 /*
88  * The longest number of jiffies for which data is allowed to remain dirty
89  */
90 int dirty_expire_interval = 30 * HZ;
91
92 /*
93  * Flag that makes the machine dump writes/reads and block dirtyings.
94  */
95 int block_dump;
96
97 /*
98  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
99  * a full sync is triggered after this time elapses without any disk activity.
100  */
101 int laptop_mode;
102
103 EXPORT_SYMBOL(laptop_mode);
104
105 /* End of sysctl-exported parameters */
106
107
108 static void background_writeout(unsigned long _min_pages);
109
110 /*
111  * Scale the writeback cache size proportional to the relative writeout speeds.
112  *
113  * We do this by keeping a floating proportion between BDIs, based on page
114  * writeback completions [end_page_writeback()]. Those devices that write out
115  * pages fastest will get the larger share, while the slower will get a smaller
116  * share.
117  *
118  * We use page writeout completions because we are interested in getting rid of
119  * dirty pages. Having them written out is the primary goal.
120  *
121  * We introduce a concept of time, a period over which we measure these events,
122  * because demand can/will vary over time. The length of this period itself is
123  * measured in page writeback completions.
124  *
125  */
126 static struct prop_descriptor vm_completions;
127 static struct prop_descriptor vm_dirties;
128
129 static unsigned long determine_dirtyable_memory(void);
130
131 /*
132  * couple the period to the dirty_ratio:
133  *
134  *   period/2 ~ roundup_pow_of_two(dirty limit)
135  */
136 static int calc_period_shift(void)
137 {
138         unsigned long dirty_total;
139
140         dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
141         return 2 + ilog2(dirty_total - 1);
142 }
143
144 /*
145  * update the period when the dirty ratio changes.
146  */
147 int dirty_ratio_handler(struct ctl_table *table, int write,
148                 struct file *filp, void __user *buffer, size_t *lenp,
149                 loff_t *ppos)
150 {
151         int old_ratio = vm_dirty_ratio;
152         int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
153         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
154                 int shift = calc_period_shift();
155                 prop_change_shift(&vm_completions, shift);
156                 prop_change_shift(&vm_dirties, shift);
157         }
158         return ret;
159 }
160
161 /*
162  * Increment the BDI's writeout completion count and the global writeout
163  * completion count. Called from test_clear_page_writeback().
164  */
165 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
166 {
167         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
168                               bdi->max_prop_frac);
169 }
170
171 void bdi_writeout_inc(struct backing_dev_info *bdi)
172 {
173         unsigned long flags;
174
175         local_irq_save(flags);
176         __bdi_writeout_inc(bdi);
177         local_irq_restore(flags);
178 }
179 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
180
181 static inline void task_dirty_inc(struct task_struct *tsk)
182 {
183         prop_inc_single(&vm_dirties, &tsk->dirties);
184 }
185
186 /*
187  * Obtain an accurate fraction of the BDI's portion.
188  */
189 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
190                 long *numerator, long *denominator)
191 {
192         if (bdi_cap_writeback_dirty(bdi)) {
193                 prop_fraction_percpu(&vm_completions, &bdi->completions,
194                                 numerator, denominator);
195         } else {
196                 *numerator = 0;
197                 *denominator = 1;
198         }
199 }
200
201 /*
202  * Clip the earned share of dirty pages to that which is actually available.
203  * This avoids exceeding the total dirty_limit when the floating averages
204  * fluctuate too quickly.
205  */
206 static void
207 clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
208 {
209         long avail_dirty;
210
211         avail_dirty = dirty -
212                 (global_page_state(NR_FILE_DIRTY) +
213                  global_page_state(NR_WRITEBACK) +
214                  global_page_state(NR_UNSTABLE_NFS) +
215                  global_page_state(NR_WRITEBACK_TEMP));
216
217         if (avail_dirty < 0)
218                 avail_dirty = 0;
219
220         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
221                 bdi_stat(bdi, BDI_WRITEBACK);
222
223         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
224 }
225
226 static inline void task_dirties_fraction(struct task_struct *tsk,
227                 long *numerator, long *denominator)
228 {
229         prop_fraction_single(&vm_dirties, &tsk->dirties,
230                                 numerator, denominator);
231 }
232
233 /*
234  * scale the dirty limit
235  *
236  * task specific dirty limit:
237  *
238  *   dirty -= (dirty/8) * p_{t}
239  */
240 static void task_dirty_limit(struct task_struct *tsk, long *pdirty)
241 {
242         long numerator, denominator;
243         long dirty = *pdirty;
244         u64 inv = dirty >> 3;
245
246         task_dirties_fraction(tsk, &numerator, &denominator);
247         inv *= numerator;
248         do_div(inv, denominator);
249
250         dirty -= inv;
251         if (dirty < *pdirty/2)
252                 dirty = *pdirty/2;
253
254         *pdirty = dirty;
255 }
256
257 /*
258  *
259  */
260 static DEFINE_SPINLOCK(bdi_lock);
261 static unsigned int bdi_min_ratio;
262
263 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
264 {
265         int ret = 0;
266         unsigned long flags;
267
268         spin_lock_irqsave(&bdi_lock, flags);
269         if (min_ratio > bdi->max_ratio) {
270                 ret = -EINVAL;
271         } else {
272                 min_ratio -= bdi->min_ratio;
273                 if (bdi_min_ratio + min_ratio < 100) {
274                         bdi_min_ratio += min_ratio;
275                         bdi->min_ratio += min_ratio;
276                 } else {
277                         ret = -EINVAL;
278                 }
279         }
280         spin_unlock_irqrestore(&bdi_lock, flags);
281
282         return ret;
283 }
284
285 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
286 {
287         unsigned long flags;
288         int ret = 0;
289
290         if (max_ratio > 100)
291                 return -EINVAL;
292
293         spin_lock_irqsave(&bdi_lock, flags);
294         if (bdi->min_ratio > max_ratio) {
295                 ret = -EINVAL;
296         } else {
297                 bdi->max_ratio = max_ratio;
298                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
299         }
300         spin_unlock_irqrestore(&bdi_lock, flags);
301
302         return ret;
303 }
304 EXPORT_SYMBOL(bdi_set_max_ratio);
305
306 /*
307  * Work out the current dirty-memory clamping and background writeout
308  * thresholds.
309  *
310  * The main aim here is to lower them aggressively if there is a lot of mapped
311  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
312  * pages.  It is better to clamp down on writers than to start swapping, and
313  * performing lots of scanning.
314  *
315  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
316  *
317  * We don't permit the clamping level to fall below 5% - that is getting rather
318  * excessive.
319  *
320  * We make sure that the background writeout level is below the adjusted
321  * clamping level.
322  */
323
324 static unsigned long highmem_dirtyable_memory(unsigned long total)
325 {
326 #ifdef CONFIG_HIGHMEM
327         int node;
328         unsigned long x = 0;
329
330         for_each_node_state(node, N_HIGH_MEMORY) {
331                 struct zone *z =
332                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
333
334                 x += zone_page_state(z, NR_FREE_PAGES)
335                         + zone_page_state(z, NR_INACTIVE)
336                         + zone_page_state(z, NR_ACTIVE);
337         }
338         /*
339          * Make sure that the number of highmem pages is never larger
340          * than the number of the total dirtyable memory. This can only
341          * occur in very strange VM situations but we want to make sure
342          * that this does not occur.
343          */
344         return min(x, total);
345 #else
346         return 0;
347 #endif
348 }
349
350 static unsigned long determine_dirtyable_memory(void)
351 {
352         unsigned long x;
353
354         x = global_page_state(NR_FREE_PAGES)
355                 + global_page_state(NR_INACTIVE)
356                 + global_page_state(NR_ACTIVE);
357
358         if (!vm_highmem_is_dirtyable)
359                 x -= highmem_dirtyable_memory(x);
360
361         return x + 1;   /* Ensure that we never return 0 */
362 }
363
364 void
365 get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
366                  struct backing_dev_info *bdi)
367 {
368         int background_ratio;           /* Percentages */
369         int dirty_ratio;
370         long background;
371         long dirty;
372         unsigned long available_memory = determine_dirtyable_memory();
373         struct task_struct *tsk;
374
375         dirty_ratio = vm_dirty_ratio;
376         if (dirty_ratio < 5)
377                 dirty_ratio = 5;
378
379         background_ratio = dirty_background_ratio;
380         if (background_ratio >= dirty_ratio)
381                 background_ratio = dirty_ratio / 2;
382
383         background = (background_ratio * available_memory) / 100;
384         dirty = (dirty_ratio * available_memory) / 100;
385         tsk = current;
386         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
387                 background += background / 4;
388                 dirty += dirty / 4;
389         }
390         *pbackground = background;
391         *pdirty = dirty;
392
393         if (bdi) {
394                 u64 bdi_dirty;
395                 long numerator, denominator;
396
397                 /*
398                  * Calculate this BDI's share of the dirty ratio.
399                  */
400                 bdi_writeout_fraction(bdi, &numerator, &denominator);
401
402                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
403                 bdi_dirty *= numerator;
404                 do_div(bdi_dirty, denominator);
405                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
406                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
407                         bdi_dirty = dirty * bdi->max_ratio / 100;
408
409                 *pbdi_dirty = bdi_dirty;
410                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
411                 task_dirty_limit(current, pbdi_dirty);
412         }
413 }
414
415 /*
416  * balance_dirty_pages() must be called by processes which are generating dirty
417  * data.  It looks at the number of dirty pages in the machine and will force
418  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
419  * If we're over `background_thresh' then pdflush is woken to perform some
420  * writeout.
421  */
422 static void balance_dirty_pages(struct address_space *mapping)
423 {
424         long nr_reclaimable, bdi_nr_reclaimable;
425         long nr_writeback, bdi_nr_writeback;
426         long background_thresh;
427         long dirty_thresh;
428         long bdi_thresh;
429         unsigned long pages_written = 0;
430         unsigned long write_chunk = sync_writeback_pages();
431
432         struct backing_dev_info *bdi = mapping->backing_dev_info;
433
434         for (;;) {
435                 struct writeback_control wbc = {
436                         .bdi            = bdi,
437                         .sync_mode      = WB_SYNC_NONE,
438                         .older_than_this = NULL,
439                         .nr_to_write    = write_chunk,
440                         .range_cyclic   = 1,
441                 };
442
443                 get_dirty_limits(&background_thresh, &dirty_thresh,
444                                 &bdi_thresh, bdi);
445
446                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
447                                         global_page_state(NR_UNSTABLE_NFS);
448                 nr_writeback = global_page_state(NR_WRITEBACK);
449
450                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
451                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
452
453                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
454                         break;
455
456                 /*
457                  * Throttle it only when the background writeback cannot
458                  * catch-up. This avoids (excessively) small writeouts
459                  * when the bdi limits are ramping up.
460                  */
461                 if (nr_reclaimable + nr_writeback <
462                                 (background_thresh + dirty_thresh) / 2)
463                         break;
464
465                 if (!bdi->dirty_exceeded)
466                         bdi->dirty_exceeded = 1;
467
468                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
469                  * Unstable writes are a feature of certain networked
470                  * filesystems (i.e. NFS) in which data may have been
471                  * written to the server's write cache, but has not yet
472                  * been flushed to permanent storage.
473                  */
474                 if (bdi_nr_reclaimable) {
475                         writeback_inodes(&wbc);
476                         pages_written += write_chunk - wbc.nr_to_write;
477                         get_dirty_limits(&background_thresh, &dirty_thresh,
478                                        &bdi_thresh, bdi);
479                 }
480
481                 /*
482                  * In order to avoid the stacked BDI deadlock we need
483                  * to ensure we accurately count the 'dirty' pages when
484                  * the threshold is low.
485                  *
486                  * Otherwise it would be possible to get thresh+n pages
487                  * reported dirty, even though there are thresh-m pages
488                  * actually dirty; with m+n sitting in the percpu
489                  * deltas.
490                  */
491                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
492                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
493                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
494                 } else if (bdi_nr_reclaimable) {
495                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
496                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
497                 }
498
499                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
500                         break;
501                 if (pages_written >= write_chunk)
502                         break;          /* We've done our duty */
503
504                 congestion_wait(WRITE, HZ/10);
505         }
506
507         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
508                         bdi->dirty_exceeded)
509                 bdi->dirty_exceeded = 0;
510
511         if (writeback_in_progress(bdi))
512                 return;         /* pdflush is already working this queue */
513
514         /*
515          * In laptop mode, we wait until hitting the higher threshold before
516          * starting background writeout, and then write out all the way down
517          * to the lower threshold.  So slow writers cause minimal disk activity.
518          *
519          * In normal mode, we start background writeout at the lower
520          * background_thresh, to keep the amount of dirty memory low.
521          */
522         if ((laptop_mode && pages_written) ||
523                         (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
524                                           + global_page_state(NR_UNSTABLE_NFS)
525                                           > background_thresh)))
526                 pdflush_operation(background_writeout, 0);
527 }
528
529 void set_page_dirty_balance(struct page *page, int page_mkwrite)
530 {
531         if (set_page_dirty(page) || page_mkwrite) {
532                 struct address_space *mapping = page_mapping(page);
533
534                 if (mapping)
535                         balance_dirty_pages_ratelimited(mapping);
536         }
537 }
538
539 /**
540  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
541  * @mapping: address_space which was dirtied
542  * @nr_pages_dirtied: number of pages which the caller has just dirtied
543  *
544  * Processes which are dirtying memory should call in here once for each page
545  * which was newly dirtied.  The function will periodically check the system's
546  * dirty state and will initiate writeback if needed.
547  *
548  * On really big machines, get_writeback_state is expensive, so try to avoid
549  * calling it too often (ratelimiting).  But once we're over the dirty memory
550  * limit we decrease the ratelimiting by a lot, to prevent individual processes
551  * from overshooting the limit by (ratelimit_pages) each.
552  */
553 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
554                                         unsigned long nr_pages_dirtied)
555 {
556         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
557         unsigned long ratelimit;
558         unsigned long *p;
559
560         ratelimit = ratelimit_pages;
561         if (mapping->backing_dev_info->dirty_exceeded)
562                 ratelimit = 8;
563
564         /*
565          * Check the rate limiting. Also, we do not want to throttle real-time
566          * tasks in balance_dirty_pages(). Period.
567          */
568         preempt_disable();
569         p =  &__get_cpu_var(ratelimits);
570         *p += nr_pages_dirtied;
571         if (unlikely(*p >= ratelimit)) {
572                 *p = 0;
573                 preempt_enable();
574                 balance_dirty_pages(mapping);
575                 return;
576         }
577         preempt_enable();
578 }
579 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
580
581 void throttle_vm_writeout(gfp_t gfp_mask)
582 {
583         long background_thresh;
584         long dirty_thresh;
585
586         for ( ; ; ) {
587                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
588
589                 /*
590                  * Boost the allowable dirty threshold a bit for page
591                  * allocators so they don't get DoS'ed by heavy writers
592                  */
593                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
594
595                 if (global_page_state(NR_UNSTABLE_NFS) +
596                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
597                                 break;
598                 congestion_wait(WRITE, HZ/10);
599
600                 /*
601                  * The caller might hold locks which can prevent IO completion
602                  * or progress in the filesystem.  So we cannot just sit here
603                  * waiting for IO to complete.
604                  */
605                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
606                         break;
607         }
608 }
609
610 /*
611  * writeback at least _min_pages, and keep writing until the amount of dirty
612  * memory is less than the background threshold, or until we're all clean.
613  */
614 static void background_writeout(unsigned long _min_pages)
615 {
616         long min_pages = _min_pages;
617         struct writeback_control wbc = {
618                 .bdi            = NULL,
619                 .sync_mode      = WB_SYNC_NONE,
620                 .older_than_this = NULL,
621                 .nr_to_write    = 0,
622                 .nonblocking    = 1,
623                 .range_cyclic   = 1,
624         };
625
626         for ( ; ; ) {
627                 long background_thresh;
628                 long dirty_thresh;
629
630                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
631                 if (global_page_state(NR_FILE_DIRTY) +
632                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
633                                 && min_pages <= 0)
634                         break;
635                 wbc.more_io = 0;
636                 wbc.encountered_congestion = 0;
637                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
638                 wbc.pages_skipped = 0;
639                 writeback_inodes(&wbc);
640                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
641                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
642                         /* Wrote less than expected */
643                         if (wbc.encountered_congestion || wbc.more_io)
644                                 congestion_wait(WRITE, HZ/10);
645                         else
646                                 break;
647                 }
648         }
649 }
650
651 /*
652  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
653  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
654  * -1 if all pdflush threads were busy.
655  */
656 int wakeup_pdflush(long nr_pages)
657 {
658         if (nr_pages == 0)
659                 nr_pages = global_page_state(NR_FILE_DIRTY) +
660                                 global_page_state(NR_UNSTABLE_NFS);
661         return pdflush_operation(background_writeout, nr_pages);
662 }
663
664 static void wb_timer_fn(unsigned long unused);
665 static void laptop_timer_fn(unsigned long unused);
666
667 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
668 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
669
670 /*
671  * Periodic writeback of "old" data.
672  *
673  * Define "old": the first time one of an inode's pages is dirtied, we mark the
674  * dirtying-time in the inode's address_space.  So this periodic writeback code
675  * just walks the superblock inode list, writing back any inodes which are
676  * older than a specific point in time.
677  *
678  * Try to run once per dirty_writeback_interval.  But if a writeback event
679  * takes longer than a dirty_writeback_interval interval, then leave a
680  * one-second gap.
681  *
682  * older_than_this takes precedence over nr_to_write.  So we'll only write back
683  * all dirty pages if they are all attached to "old" mappings.
684  */
685 static void wb_kupdate(unsigned long arg)
686 {
687         unsigned long oldest_jif;
688         unsigned long start_jif;
689         unsigned long next_jif;
690         long nr_to_write;
691         struct writeback_control wbc = {
692                 .bdi            = NULL,
693                 .sync_mode      = WB_SYNC_NONE,
694                 .older_than_this = &oldest_jif,
695                 .nr_to_write    = 0,
696                 .nonblocking    = 1,
697                 .for_kupdate    = 1,
698                 .range_cyclic   = 1,
699         };
700
701         sync_supers();
702
703         oldest_jif = jiffies - dirty_expire_interval;
704         start_jif = jiffies;
705         next_jif = start_jif + dirty_writeback_interval;
706         nr_to_write = global_page_state(NR_FILE_DIRTY) +
707                         global_page_state(NR_UNSTABLE_NFS) +
708                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
709         while (nr_to_write > 0) {
710                 wbc.more_io = 0;
711                 wbc.encountered_congestion = 0;
712                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
713                 writeback_inodes(&wbc);
714                 if (wbc.nr_to_write > 0) {
715                         if (wbc.encountered_congestion || wbc.more_io)
716                                 congestion_wait(WRITE, HZ/10);
717                         else
718                                 break;  /* All the old data is written */
719                 }
720                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
721         }
722         if (time_before(next_jif, jiffies + HZ))
723                 next_jif = jiffies + HZ;
724         if (dirty_writeback_interval)
725                 mod_timer(&wb_timer, next_jif);
726 }
727
728 /*
729  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
730  */
731 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
732         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
733 {
734         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
735         if (dirty_writeback_interval)
736                 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
737         else
738                 del_timer(&wb_timer);
739         return 0;
740 }
741
742 static void wb_timer_fn(unsigned long unused)
743 {
744         if (pdflush_operation(wb_kupdate, 0) < 0)
745                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
746 }
747
748 static void laptop_flush(unsigned long unused)
749 {
750         sys_sync();
751 }
752
753 static void laptop_timer_fn(unsigned long unused)
754 {
755         pdflush_operation(laptop_flush, 0);
756 }
757
758 /*
759  * We've spun up the disk and we're in laptop mode: schedule writeback
760  * of all dirty data a few seconds from now.  If the flush is already scheduled
761  * then push it back - the user is still using the disk.
762  */
763 void laptop_io_completion(void)
764 {
765         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
766 }
767
768 /*
769  * We're in laptop mode and we've just synced. The sync's writes will have
770  * caused another writeback to be scheduled by laptop_io_completion.
771  * Nothing needs to be written back anymore, so we unschedule the writeback.
772  */
773 void laptop_sync_completion(void)
774 {
775         del_timer(&laptop_mode_wb_timer);
776 }
777
778 /*
779  * If ratelimit_pages is too high then we can get into dirty-data overload
780  * if a large number of processes all perform writes at the same time.
781  * If it is too low then SMP machines will call the (expensive)
782  * get_writeback_state too often.
783  *
784  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
785  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
786  * thresholds before writeback cuts in.
787  *
788  * But the limit should not be set too high.  Because it also controls the
789  * amount of memory which the balance_dirty_pages() caller has to write back.
790  * If this is too large then the caller will block on the IO queue all the
791  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
792  * will write six megabyte chunks, max.
793  */
794
795 void writeback_set_ratelimit(void)
796 {
797         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
798         if (ratelimit_pages < 16)
799                 ratelimit_pages = 16;
800         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
801                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
802 }
803
804 static int __cpuinit
805 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
806 {
807         writeback_set_ratelimit();
808         return NOTIFY_DONE;
809 }
810
811 static struct notifier_block __cpuinitdata ratelimit_nb = {
812         .notifier_call  = ratelimit_handler,
813         .next           = NULL,
814 };
815
816 /*
817  * Called early on to tune the page writeback dirty limits.
818  *
819  * We used to scale dirty pages according to how total memory
820  * related to pages that could be allocated for buffers (by
821  * comparing nr_free_buffer_pages() to vm_total_pages.
822  *
823  * However, that was when we used "dirty_ratio" to scale with
824  * all memory, and we don't do that any more. "dirty_ratio"
825  * is now applied to total non-HIGHPAGE memory (by subtracting
826  * totalhigh_pages from vm_total_pages), and as such we can't
827  * get into the old insane situation any more where we had
828  * large amounts of dirty pages compared to a small amount of
829  * non-HIGHMEM memory.
830  *
831  * But we might still want to scale the dirty_ratio by how
832  * much memory the box has..
833  */
834 void __init page_writeback_init(void)
835 {
836         int shift;
837
838         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
839         writeback_set_ratelimit();
840         register_cpu_notifier(&ratelimit_nb);
841
842         shift = calc_period_shift();
843         prop_descriptor_init(&vm_completions, shift);
844         prop_descriptor_init(&vm_dirties, shift);
845 }
846
847 /**
848  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
849  * @mapping: address space structure to write
850  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
851  * @writepage: function called for each page
852  * @data: data passed to writepage function
853  *
854  * If a page is already under I/O, write_cache_pages() skips it, even
855  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
856  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
857  * and msync() need to guarantee that all the data which was dirty at the time
858  * the call was made get new I/O started against them.  If wbc->sync_mode is
859  * WB_SYNC_ALL then we were called for data integrity and we must wait for
860  * existing IO to complete.
861  */
862 int write_cache_pages(struct address_space *mapping,
863                       struct writeback_control *wbc, writepage_t writepage,
864                       void *data)
865 {
866         struct backing_dev_info *bdi = mapping->backing_dev_info;
867         int ret = 0;
868         int done = 0;
869         struct pagevec pvec;
870         int nr_pages;
871         pgoff_t index;
872         pgoff_t end;            /* Inclusive */
873         int scanned = 0;
874         int range_whole = 0;
875
876         if (wbc->nonblocking && bdi_write_congested(bdi)) {
877                 wbc->encountered_congestion = 1;
878                 return 0;
879         }
880
881         pagevec_init(&pvec, 0);
882         if (wbc->range_cyclic) {
883                 index = mapping->writeback_index; /* Start from prev offset */
884                 end = -1;
885         } else {
886                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
887                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
888                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
889                         range_whole = 1;
890                 scanned = 1;
891         }
892 retry:
893         while (!done && (index <= end) &&
894                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
895                                               PAGECACHE_TAG_DIRTY,
896                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
897                 unsigned i;
898
899                 scanned = 1;
900                 for (i = 0; i < nr_pages; i++) {
901                         struct page *page = pvec.pages[i];
902
903                         /*
904                          * At this point we hold neither mapping->tree_lock nor
905                          * lock on the page itself: the page may be truncated or
906                          * invalidated (changing page->mapping to NULL), or even
907                          * swizzled back from swapper_space to tmpfs file
908                          * mapping
909                          */
910                         lock_page(page);
911
912                         if (unlikely(page->mapping != mapping)) {
913                                 unlock_page(page);
914                                 continue;
915                         }
916
917                         if (!wbc->range_cyclic && page->index > end) {
918                                 done = 1;
919                                 unlock_page(page);
920                                 continue;
921                         }
922
923                         if (wbc->sync_mode != WB_SYNC_NONE)
924                                 wait_on_page_writeback(page);
925
926                         if (PageWriteback(page) ||
927                             !clear_page_dirty_for_io(page)) {
928                                 unlock_page(page);
929                                 continue;
930                         }
931
932                         ret = (*writepage)(page, wbc, data);
933
934                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
935                                 unlock_page(page);
936                                 ret = 0;
937                         }
938                         if (ret || (--(wbc->nr_to_write) <= 0))
939                                 done = 1;
940                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
941                                 wbc->encountered_congestion = 1;
942                                 done = 1;
943                         }
944                 }
945                 pagevec_release(&pvec);
946                 cond_resched();
947         }
948         if (!scanned && !done) {
949                 /*
950                  * We hit the last page and there is more work to be done: wrap
951                  * back to the start of the file
952                  */
953                 scanned = 1;
954                 index = 0;
955                 goto retry;
956         }
957         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
958                 mapping->writeback_index = index;
959         return ret;
960 }
961 EXPORT_SYMBOL(write_cache_pages);
962
963 /*
964  * Function used by generic_writepages to call the real writepage
965  * function and set the mapping flags on error
966  */
967 static int __writepage(struct page *page, struct writeback_control *wbc,
968                        void *data)
969 {
970         struct address_space *mapping = data;
971         int ret = mapping->a_ops->writepage(page, wbc);
972         mapping_set_error(mapping, ret);
973         return ret;
974 }
975
976 /**
977  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
978  * @mapping: address space structure to write
979  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
980  *
981  * This is a library function, which implements the writepages()
982  * address_space_operation.
983  */
984 int generic_writepages(struct address_space *mapping,
985                        struct writeback_control *wbc)
986 {
987         /* deal with chardevs and other special file */
988         if (!mapping->a_ops->writepage)
989                 return 0;
990
991         return write_cache_pages(mapping, wbc, __writepage, mapping);
992 }
993
994 EXPORT_SYMBOL(generic_writepages);
995
996 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
997 {
998         int ret;
999
1000         if (wbc->nr_to_write <= 0)
1001                 return 0;
1002         wbc->for_writepages = 1;
1003         if (mapping->a_ops->writepages)
1004                 ret = mapping->a_ops->writepages(mapping, wbc);
1005         else
1006                 ret = generic_writepages(mapping, wbc);
1007         wbc->for_writepages = 0;
1008         return ret;
1009 }
1010
1011 /**
1012  * write_one_page - write out a single page and optionally wait on I/O
1013  * @page: the page to write
1014  * @wait: if true, wait on writeout
1015  *
1016  * The page must be locked by the caller and will be unlocked upon return.
1017  *
1018  * write_one_page() returns a negative error code if I/O failed.
1019  */
1020 int write_one_page(struct page *page, int wait)
1021 {
1022         struct address_space *mapping = page->mapping;
1023         int ret = 0;
1024         struct writeback_control wbc = {
1025                 .sync_mode = WB_SYNC_ALL,
1026                 .nr_to_write = 1,
1027         };
1028
1029         BUG_ON(!PageLocked(page));
1030
1031         if (wait)
1032                 wait_on_page_writeback(page);
1033
1034         if (clear_page_dirty_for_io(page)) {
1035                 page_cache_get(page);
1036                 ret = mapping->a_ops->writepage(page, &wbc);
1037                 if (ret == 0 && wait) {
1038                         wait_on_page_writeback(page);
1039                         if (PageError(page))
1040                                 ret = -EIO;
1041                 }
1042                 page_cache_release(page);
1043         } else {
1044                 unlock_page(page);
1045         }
1046         return ret;
1047 }
1048 EXPORT_SYMBOL(write_one_page);
1049
1050 /*
1051  * For address_spaces which do not use buffers nor write back.
1052  */
1053 int __set_page_dirty_no_writeback(struct page *page)
1054 {
1055         if (!PageDirty(page))
1056                 SetPageDirty(page);
1057         return 0;
1058 }
1059
1060 /*
1061  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1062  * its radix tree.
1063  *
1064  * This is also used when a single buffer is being dirtied: we want to set the
1065  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1066  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1067  *
1068  * Most callers have locked the page, which pins the address_space in memory.
1069  * But zap_pte_range() does not lock the page, however in that case the
1070  * mapping is pinned by the vma's ->vm_file reference.
1071  *
1072  * We take care to handle the case where the page was truncated from the
1073  * mapping by re-checking page_mapping() inside tree_lock.
1074  */
1075 int __set_page_dirty_nobuffers(struct page *page)
1076 {
1077         if (!TestSetPageDirty(page)) {
1078                 struct address_space *mapping = page_mapping(page);
1079                 struct address_space *mapping2;
1080
1081                 if (!mapping)
1082                         return 1;
1083
1084                 write_lock_irq(&mapping->tree_lock);
1085                 mapping2 = page_mapping(page);
1086                 if (mapping2) { /* Race with truncate? */
1087                         BUG_ON(mapping2 != mapping);
1088                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1089                         if (mapping_cap_account_dirty(mapping)) {
1090                                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1091                                 __inc_bdi_stat(mapping->backing_dev_info,
1092                                                 BDI_RECLAIMABLE);
1093                                 task_io_account_write(PAGE_CACHE_SIZE);
1094                         }
1095                         radix_tree_tag_set(&mapping->page_tree,
1096                                 page_index(page), PAGECACHE_TAG_DIRTY);
1097                 }
1098                 write_unlock_irq(&mapping->tree_lock);
1099                 if (mapping->host) {
1100                         /* !PageAnon && !swapper_space */
1101                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1102                 }
1103                 return 1;
1104         }
1105         return 0;
1106 }
1107 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1108
1109 /*
1110  * When a writepage implementation decides that it doesn't want to write this
1111  * page for some reason, it should redirty the locked page via
1112  * redirty_page_for_writepage() and it should then unlock the page and return 0
1113  */
1114 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1115 {
1116         wbc->pages_skipped++;
1117         return __set_page_dirty_nobuffers(page);
1118 }
1119 EXPORT_SYMBOL(redirty_page_for_writepage);
1120
1121 /*
1122  * If the mapping doesn't provide a set_page_dirty a_op, then
1123  * just fall through and assume that it wants buffer_heads.
1124  */
1125 static int __set_page_dirty(struct page *page)
1126 {
1127         struct address_space *mapping = page_mapping(page);
1128
1129         if (likely(mapping)) {
1130                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1131 #ifdef CONFIG_BLOCK
1132                 if (!spd)
1133                         spd = __set_page_dirty_buffers;
1134 #endif
1135                 return (*spd)(page);
1136         }
1137         if (!PageDirty(page)) {
1138                 if (!TestSetPageDirty(page))
1139                         return 1;
1140         }
1141         return 0;
1142 }
1143
1144 int set_page_dirty(struct page *page)
1145 {
1146         int ret = __set_page_dirty(page);
1147         if (ret)
1148                 task_dirty_inc(current);
1149         return ret;
1150 }
1151 EXPORT_SYMBOL(set_page_dirty);
1152
1153 /*
1154  * set_page_dirty() is racy if the caller has no reference against
1155  * page->mapping->host, and if the page is unlocked.  This is because another
1156  * CPU could truncate the page off the mapping and then free the mapping.
1157  *
1158  * Usually, the page _is_ locked, or the caller is a user-space process which
1159  * holds a reference on the inode by having an open file.
1160  *
1161  * In other cases, the page should be locked before running set_page_dirty().
1162  */
1163 int set_page_dirty_lock(struct page *page)
1164 {
1165         int ret;
1166
1167         lock_page_nosync(page);
1168         ret = set_page_dirty(page);
1169         unlock_page(page);
1170         return ret;
1171 }
1172 EXPORT_SYMBOL(set_page_dirty_lock);
1173
1174 /*
1175  * Clear a page's dirty flag, while caring for dirty memory accounting.
1176  * Returns true if the page was previously dirty.
1177  *
1178  * This is for preparing to put the page under writeout.  We leave the page
1179  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1180  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1181  * implementation will run either set_page_writeback() or set_page_dirty(),
1182  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1183  * back into sync.
1184  *
1185  * This incoherency between the page's dirty flag and radix-tree tag is
1186  * unfortunate, but it only exists while the page is locked.
1187  */
1188 int clear_page_dirty_for_io(struct page *page)
1189 {
1190         struct address_space *mapping = page_mapping(page);
1191
1192         BUG_ON(!PageLocked(page));
1193
1194         ClearPageReclaim(page);
1195         if (mapping && mapping_cap_account_dirty(mapping)) {
1196                 /*
1197                  * Yes, Virginia, this is indeed insane.
1198                  *
1199                  * We use this sequence to make sure that
1200                  *  (a) we account for dirty stats properly
1201                  *  (b) we tell the low-level filesystem to
1202                  *      mark the whole page dirty if it was
1203                  *      dirty in a pagetable. Only to then
1204                  *  (c) clean the page again and return 1 to
1205                  *      cause the writeback.
1206                  *
1207                  * This way we avoid all nasty races with the
1208                  * dirty bit in multiple places and clearing
1209                  * them concurrently from different threads.
1210                  *
1211                  * Note! Normally the "set_page_dirty(page)"
1212                  * has no effect on the actual dirty bit - since
1213                  * that will already usually be set. But we
1214                  * need the side effects, and it can help us
1215                  * avoid races.
1216                  *
1217                  * We basically use the page "master dirty bit"
1218                  * as a serialization point for all the different
1219                  * threads doing their things.
1220                  */
1221                 if (page_mkclean(page))
1222                         set_page_dirty(page);
1223                 /*
1224                  * We carefully synchronise fault handlers against
1225                  * installing a dirty pte and marking the page dirty
1226                  * at this point. We do this by having them hold the
1227                  * page lock at some point after installing their
1228                  * pte, but before marking the page dirty.
1229                  * Pages are always locked coming in here, so we get
1230                  * the desired exclusion. See mm/memory.c:do_wp_page()
1231                  * for more comments.
1232                  */
1233                 if (TestClearPageDirty(page)) {
1234                         dec_zone_page_state(page, NR_FILE_DIRTY);
1235                         dec_bdi_stat(mapping->backing_dev_info,
1236                                         BDI_RECLAIMABLE);
1237                         return 1;
1238                 }
1239                 return 0;
1240         }
1241         return TestClearPageDirty(page);
1242 }
1243 EXPORT_SYMBOL(clear_page_dirty_for_io);
1244
1245 int test_clear_page_writeback(struct page *page)
1246 {
1247         struct address_space *mapping = page_mapping(page);
1248         int ret;
1249
1250         if (mapping) {
1251                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1252                 unsigned long flags;
1253
1254                 write_lock_irqsave(&mapping->tree_lock, flags);
1255                 ret = TestClearPageWriteback(page);
1256                 if (ret) {
1257                         radix_tree_tag_clear(&mapping->page_tree,
1258                                                 page_index(page),
1259                                                 PAGECACHE_TAG_WRITEBACK);
1260                         if (bdi_cap_account_writeback(bdi)) {
1261                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1262                                 __bdi_writeout_inc(bdi);
1263                         }
1264                 }
1265                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1266         } else {
1267                 ret = TestClearPageWriteback(page);
1268         }
1269         if (ret)
1270                 dec_zone_page_state(page, NR_WRITEBACK);
1271         return ret;
1272 }
1273
1274 int test_set_page_writeback(struct page *page)
1275 {
1276         struct address_space *mapping = page_mapping(page);
1277         int ret;
1278
1279         if (mapping) {
1280                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1281                 unsigned long flags;
1282
1283                 write_lock_irqsave(&mapping->tree_lock, flags);
1284                 ret = TestSetPageWriteback(page);
1285                 if (!ret) {
1286                         radix_tree_tag_set(&mapping->page_tree,
1287                                                 page_index(page),
1288                                                 PAGECACHE_TAG_WRITEBACK);
1289                         if (bdi_cap_account_writeback(bdi))
1290                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1291                 }
1292                 if (!PageDirty(page))
1293                         radix_tree_tag_clear(&mapping->page_tree,
1294                                                 page_index(page),
1295                                                 PAGECACHE_TAG_DIRTY);
1296                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1297         } else {
1298                 ret = TestSetPageWriteback(page);
1299         }
1300         if (!ret)
1301                 inc_zone_page_state(page, NR_WRITEBACK);
1302         return ret;
1303
1304 }
1305 EXPORT_SYMBOL(test_set_page_writeback);
1306
1307 /*
1308  * Return true if any of the pages in the mapping are marked with the
1309  * passed tag.
1310  */
1311 int mapping_tagged(struct address_space *mapping, int tag)
1312 {
1313         int ret;
1314         rcu_read_lock();
1315         ret = radix_tree_tagged(&mapping->page_tree, tag);
1316         rcu_read_unlock();
1317         return ret;
1318 }
1319 EXPORT_SYMBOL(mapping_tagged);