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