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