Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[linux-2.6] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    akpm@zip.com.au
11  *              Initial version
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h>
36 #include <linux/pagevec.h>
37
38 /*
39  * The maximum number of pages to writeout in a single bdflush/kupdate
40  * operation.  We do this so we don't hold I_SYNC against an inode for
41  * enormous amounts of time, which would block a userspace task which has
42  * been forced to throttle against that inode.  Also, the code reevaluates
43  * the dirty each time it has written this many pages.
44  */
45 #define MAX_WRITEBACK_PAGES     1024
46
47 /*
48  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
49  * will look to see if it needs to force writeback or throttling.
50  */
51 static long ratelimit_pages = 32;
52
53 /*
54  * When balance_dirty_pages decides that the caller needs to perform some
55  * non-background writeback, this is how many pages it will attempt to write.
56  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
57  * large amounts of I/O are submitted.
58  */
59 static inline long sync_writeback_pages(void)
60 {
61         return ratelimit_pages + ratelimit_pages / 2;
62 }
63
64 /* The following parameters are exported via /proc/sys/vm */
65
66 /*
67  * Start background writeback (via pdflush) at this percentage
68  */
69 int dirty_background_ratio = 5;
70
71 /*
72  * 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         for ( ; ; ) {
506                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
507
508                 /*
509                  * Boost the allowable dirty threshold a bit for page
510                  * allocators so they don't get DoS'ed by heavy writers
511                  */
512                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
513
514                 if (global_page_state(NR_UNSTABLE_NFS) +
515                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
516                                 break;
517                 congestion_wait(WRITE, HZ/10);
518
519                 /*
520                  * The caller might hold locks which can prevent IO completion
521                  * or progress in the filesystem.  So we cannot just sit here
522                  * waiting for IO to complete.
523                  */
524                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
525                         break;
526         }
527 }
528
529 /*
530  * writeback at least _min_pages, and keep writing until the amount of dirty
531  * memory is less than the background threshold, or until we're all clean.
532  */
533 static void background_writeout(unsigned long _min_pages)
534 {
535         long min_pages = _min_pages;
536         struct writeback_control wbc = {
537                 .bdi            = NULL,
538                 .sync_mode      = WB_SYNC_NONE,
539                 .older_than_this = NULL,
540                 .nr_to_write    = 0,
541                 .nonblocking    = 1,
542                 .range_cyclic   = 1,
543         };
544
545         for ( ; ; ) {
546                 long background_thresh;
547                 long dirty_thresh;
548
549                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
550                 if (global_page_state(NR_FILE_DIRTY) +
551                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
552                                 && min_pages <= 0)
553                         break;
554                 wbc.more_io = 0;
555                 wbc.encountered_congestion = 0;
556                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
557                 wbc.pages_skipped = 0;
558                 writeback_inodes(&wbc);
559                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
560                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
561                         /* Wrote less than expected */
562                         if (wbc.encountered_congestion || wbc.more_io)
563                                 congestion_wait(WRITE, HZ/10);
564                         else
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.more_io = 0;
630                 wbc.encountered_congestion = 0;
631                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
632                 writeback_inodes(&wbc);
633                 if (wbc.nr_to_write > 0) {
634                         if (wbc.encountered_congestion || wbc.more_io)
635                                 congestion_wait(WRITE, HZ/10);
636                         else
637                                 break;  /* All the old data is written */
638                 }
639                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
640         }
641         if (time_before(next_jif, jiffies + HZ))
642                 next_jif = jiffies + HZ;
643         if (dirty_writeback_interval)
644                 mod_timer(&wb_timer, next_jif);
645 }
646
647 /*
648  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
649  */
650 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
651         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
652 {
653         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
654         if (dirty_writeback_interval)
655                 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
656         else
657                 del_timer(&wb_timer);
658         return 0;
659 }
660
661 static void wb_timer_fn(unsigned long unused)
662 {
663         if (pdflush_operation(wb_kupdate, 0) < 0)
664                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
665 }
666
667 static void laptop_flush(unsigned long unused)
668 {
669         sys_sync();
670 }
671
672 static void laptop_timer_fn(unsigned long unused)
673 {
674         pdflush_operation(laptop_flush, 0);
675 }
676
677 /*
678  * We've spun up the disk and we're in laptop mode: schedule writeback
679  * of all dirty data a few seconds from now.  If the flush is already scheduled
680  * then push it back - the user is still using the disk.
681  */
682 void laptop_io_completion(void)
683 {
684         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
685 }
686
687 /*
688  * We're in laptop mode and we've just synced. The sync's writes will have
689  * caused another writeback to be scheduled by laptop_io_completion.
690  * Nothing needs to be written back anymore, so we unschedule the writeback.
691  */
692 void laptop_sync_completion(void)
693 {
694         del_timer(&laptop_mode_wb_timer);
695 }
696
697 /*
698  * If ratelimit_pages is too high then we can get into dirty-data overload
699  * if a large number of processes all perform writes at the same time.
700  * If it is too low then SMP machines will call the (expensive)
701  * get_writeback_state too often.
702  *
703  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
704  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
705  * thresholds before writeback cuts in.
706  *
707  * But the limit should not be set too high.  Because it also controls the
708  * amount of memory which the balance_dirty_pages() caller has to write back.
709  * If this is too large then the caller will block on the IO queue all the
710  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
711  * will write six megabyte chunks, max.
712  */
713
714 void writeback_set_ratelimit(void)
715 {
716         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
717         if (ratelimit_pages < 16)
718                 ratelimit_pages = 16;
719         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
720                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
721 }
722
723 static int __cpuinit
724 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
725 {
726         writeback_set_ratelimit();
727         return NOTIFY_DONE;
728 }
729
730 static struct notifier_block __cpuinitdata ratelimit_nb = {
731         .notifier_call  = ratelimit_handler,
732         .next           = NULL,
733 };
734
735 /*
736  * Called early on to tune the page writeback dirty limits.
737  *
738  * We used to scale dirty pages according to how total memory
739  * related to pages that could be allocated for buffers (by
740  * comparing nr_free_buffer_pages() to vm_total_pages.
741  *
742  * However, that was when we used "dirty_ratio" to scale with
743  * all memory, and we don't do that any more. "dirty_ratio"
744  * is now applied to total non-HIGHPAGE memory (by subtracting
745  * totalhigh_pages from vm_total_pages), and as such we can't
746  * get into the old insane situation any more where we had
747  * large amounts of dirty pages compared to a small amount of
748  * non-HIGHMEM memory.
749  *
750  * But we might still want to scale the dirty_ratio by how
751  * much memory the box has..
752  */
753 void __init page_writeback_init(void)
754 {
755         int shift;
756
757         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
758         writeback_set_ratelimit();
759         register_cpu_notifier(&ratelimit_nb);
760
761         shift = calc_period_shift();
762         prop_descriptor_init(&vm_completions, shift);
763         prop_descriptor_init(&vm_dirties, shift);
764 }
765
766 /**
767  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
768  * @mapping: address space structure to write
769  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
770  * @writepage: function called for each page
771  * @data: data passed to writepage function
772  *
773  * If a page is already under I/O, write_cache_pages() skips it, even
774  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
775  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
776  * and msync() need to guarantee that all the data which was dirty at the time
777  * the call was made get new I/O started against them.  If wbc->sync_mode is
778  * WB_SYNC_ALL then we were called for data integrity and we must wait for
779  * existing IO to complete.
780  */
781 int write_cache_pages(struct address_space *mapping,
782                       struct writeback_control *wbc, writepage_t writepage,
783                       void *data)
784 {
785         struct backing_dev_info *bdi = mapping->backing_dev_info;
786         int ret = 0;
787         int done = 0;
788         struct pagevec pvec;
789         int nr_pages;
790         pgoff_t index;
791         pgoff_t end;            /* Inclusive */
792         int scanned = 0;
793         int range_whole = 0;
794
795         if (wbc->nonblocking && bdi_write_congested(bdi)) {
796                 wbc->encountered_congestion = 1;
797                 return 0;
798         }
799
800         pagevec_init(&pvec, 0);
801         if (wbc->range_cyclic) {
802                 index = mapping->writeback_index; /* Start from prev offset */
803                 end = -1;
804         } else {
805                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
806                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
807                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
808                         range_whole = 1;
809                 scanned = 1;
810         }
811 retry:
812         while (!done && (index <= end) &&
813                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
814                                               PAGECACHE_TAG_DIRTY,
815                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
816                 unsigned i;
817
818                 scanned = 1;
819                 for (i = 0; i < nr_pages; i++) {
820                         struct page *page = pvec.pages[i];
821
822                         /*
823                          * At this point we hold neither mapping->tree_lock nor
824                          * lock on the page itself: the page may be truncated or
825                          * invalidated (changing page->mapping to NULL), or even
826                          * swizzled back from swapper_space to tmpfs file
827                          * mapping
828                          */
829                         lock_page(page);
830
831                         if (unlikely(page->mapping != mapping)) {
832                                 unlock_page(page);
833                                 continue;
834                         }
835
836                         if (!wbc->range_cyclic && page->index > end) {
837                                 done = 1;
838                                 unlock_page(page);
839                                 continue;
840                         }
841
842                         if (wbc->sync_mode != WB_SYNC_NONE)
843                                 wait_on_page_writeback(page);
844
845                         if (PageWriteback(page) ||
846                             !clear_page_dirty_for_io(page)) {
847                                 unlock_page(page);
848                                 continue;
849                         }
850
851                         ret = (*writepage)(page, wbc, data);
852
853                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
854                                 unlock_page(page);
855                                 ret = 0;
856                         }
857                         if (ret || (--(wbc->nr_to_write) <= 0))
858                                 done = 1;
859                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
860                                 wbc->encountered_congestion = 1;
861                                 done = 1;
862                         }
863                 }
864                 pagevec_release(&pvec);
865                 cond_resched();
866         }
867         if (!scanned && !done) {
868                 /*
869                  * We hit the last page and there is more work to be done: wrap
870                  * back to the start of the file
871                  */
872                 scanned = 1;
873                 index = 0;
874                 goto retry;
875         }
876         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
877                 mapping->writeback_index = index;
878         return ret;
879 }
880 EXPORT_SYMBOL(write_cache_pages);
881
882 /*
883  * Function used by generic_writepages to call the real writepage
884  * function and set the mapping flags on error
885  */
886 static int __writepage(struct page *page, struct writeback_control *wbc,
887                        void *data)
888 {
889         struct address_space *mapping = data;
890         int ret = mapping->a_ops->writepage(page, wbc);
891         mapping_set_error(mapping, ret);
892         return ret;
893 }
894
895 /**
896  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
897  * @mapping: address space structure to write
898  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
899  *
900  * This is a library function, which implements the writepages()
901  * address_space_operation.
902  */
903 int generic_writepages(struct address_space *mapping,
904                        struct writeback_control *wbc)
905 {
906         /* deal with chardevs and other special file */
907         if (!mapping->a_ops->writepage)
908                 return 0;
909
910         return write_cache_pages(mapping, wbc, __writepage, mapping);
911 }
912
913 EXPORT_SYMBOL(generic_writepages);
914
915 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
916 {
917         int ret;
918
919         if (wbc->nr_to_write <= 0)
920                 return 0;
921         wbc->for_writepages = 1;
922         if (mapping->a_ops->writepages)
923                 ret = mapping->a_ops->writepages(mapping, wbc);
924         else
925                 ret = generic_writepages(mapping, wbc);
926         wbc->for_writepages = 0;
927         return ret;
928 }
929
930 /**
931  * write_one_page - write out a single page and optionally wait on I/O
932  * @page: the page to write
933  * @wait: if true, wait on writeout
934  *
935  * The page must be locked by the caller and will be unlocked upon return.
936  *
937  * write_one_page() returns a negative error code if I/O failed.
938  */
939 int write_one_page(struct page *page, int wait)
940 {
941         struct address_space *mapping = page->mapping;
942         int ret = 0;
943         struct writeback_control wbc = {
944                 .sync_mode = WB_SYNC_ALL,
945                 .nr_to_write = 1,
946         };
947
948         BUG_ON(!PageLocked(page));
949
950         if (wait)
951                 wait_on_page_writeback(page);
952
953         if (clear_page_dirty_for_io(page)) {
954                 page_cache_get(page);
955                 ret = mapping->a_ops->writepage(page, &wbc);
956                 if (ret == 0 && wait) {
957                         wait_on_page_writeback(page);
958                         if (PageError(page))
959                                 ret = -EIO;
960                 }
961                 page_cache_release(page);
962         } else {
963                 unlock_page(page);
964         }
965         return ret;
966 }
967 EXPORT_SYMBOL(write_one_page);
968
969 /*
970  * For address_spaces which do not use buffers nor write back.
971  */
972 int __set_page_dirty_no_writeback(struct page *page)
973 {
974         if (!PageDirty(page))
975                 SetPageDirty(page);
976         return 0;
977 }
978
979 /*
980  * For address_spaces which do not use buffers.  Just tag the page as dirty in
981  * its radix tree.
982  *
983  * This is also used when a single buffer is being dirtied: we want to set the
984  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
985  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
986  *
987  * Most callers have locked the page, which pins the address_space in memory.
988  * But zap_pte_range() does not lock the page, however in that case the
989  * mapping is pinned by the vma's ->vm_file reference.
990  *
991  * We take care to handle the case where the page was truncated from the
992  * mapping by re-checking page_mapping() inside tree_lock.
993  */
994 int __set_page_dirty_nobuffers(struct page *page)
995 {
996         if (!TestSetPageDirty(page)) {
997                 struct address_space *mapping = page_mapping(page);
998                 struct address_space *mapping2;
999
1000                 if (!mapping)
1001                         return 1;
1002
1003                 write_lock_irq(&mapping->tree_lock);
1004                 mapping2 = page_mapping(page);
1005                 if (mapping2) { /* Race with truncate? */
1006                         BUG_ON(mapping2 != mapping);
1007                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1008                         if (mapping_cap_account_dirty(mapping)) {
1009                                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1010                                 __inc_bdi_stat(mapping->backing_dev_info,
1011                                                 BDI_RECLAIMABLE);
1012                                 task_io_account_write(PAGE_CACHE_SIZE);
1013                         }
1014                         radix_tree_tag_set(&mapping->page_tree,
1015                                 page_index(page), PAGECACHE_TAG_DIRTY);
1016                 }
1017                 write_unlock_irq(&mapping->tree_lock);
1018                 if (mapping->host) {
1019                         /* !PageAnon && !swapper_space */
1020                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1021                 }
1022                 return 1;
1023         }
1024         return 0;
1025 }
1026 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1027
1028 /*
1029  * When a writepage implementation decides that it doesn't want to write this
1030  * page for some reason, it should redirty the locked page via
1031  * redirty_page_for_writepage() and it should then unlock the page and return 0
1032  */
1033 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1034 {
1035         wbc->pages_skipped++;
1036         return __set_page_dirty_nobuffers(page);
1037 }
1038 EXPORT_SYMBOL(redirty_page_for_writepage);
1039
1040 /*
1041  * If the mapping doesn't provide a set_page_dirty a_op, then
1042  * just fall through and assume that it wants buffer_heads.
1043  */
1044 static int __set_page_dirty(struct page *page)
1045 {
1046         struct address_space *mapping = page_mapping(page);
1047
1048         if (likely(mapping)) {
1049                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1050 #ifdef CONFIG_BLOCK
1051                 if (!spd)
1052                         spd = __set_page_dirty_buffers;
1053 #endif
1054                 return (*spd)(page);
1055         }
1056         if (!PageDirty(page)) {
1057                 if (!TestSetPageDirty(page))
1058                         return 1;
1059         }
1060         return 0;
1061 }
1062
1063 int fastcall set_page_dirty(struct page *page)
1064 {
1065         int ret = __set_page_dirty(page);
1066         if (ret)
1067                 task_dirty_inc(current);
1068         return ret;
1069 }
1070 EXPORT_SYMBOL(set_page_dirty);
1071
1072 /*
1073  * set_page_dirty() is racy if the caller has no reference against
1074  * page->mapping->host, and if the page is unlocked.  This is because another
1075  * CPU could truncate the page off the mapping and then free the mapping.
1076  *
1077  * Usually, the page _is_ locked, or the caller is a user-space process which
1078  * holds a reference on the inode by having an open file.
1079  *
1080  * In other cases, the page should be locked before running set_page_dirty().
1081  */
1082 int set_page_dirty_lock(struct page *page)
1083 {
1084         int ret;
1085
1086         lock_page_nosync(page);
1087         ret = set_page_dirty(page);
1088         unlock_page(page);
1089         return ret;
1090 }
1091 EXPORT_SYMBOL(set_page_dirty_lock);
1092
1093 /*
1094  * Clear a page's dirty flag, while caring for dirty memory accounting.
1095  * Returns true if the page was previously dirty.
1096  *
1097  * This is for preparing to put the page under writeout.  We leave the page
1098  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1099  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1100  * implementation will run either set_page_writeback() or set_page_dirty(),
1101  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1102  * back into sync.
1103  *
1104  * This incoherency between the page's dirty flag and radix-tree tag is
1105  * unfortunate, but it only exists while the page is locked.
1106  */
1107 int clear_page_dirty_for_io(struct page *page)
1108 {
1109         struct address_space *mapping = page_mapping(page);
1110
1111         BUG_ON(!PageLocked(page));
1112
1113         ClearPageReclaim(page);
1114         if (mapping && mapping_cap_account_dirty(mapping)) {
1115                 /*
1116                  * Yes, Virginia, this is indeed insane.
1117                  *
1118                  * We use this sequence to make sure that
1119                  *  (a) we account for dirty stats properly
1120                  *  (b) we tell the low-level filesystem to
1121                  *      mark the whole page dirty if it was
1122                  *      dirty in a pagetable. Only to then
1123                  *  (c) clean the page again and return 1 to
1124                  *      cause the writeback.
1125                  *
1126                  * This way we avoid all nasty races with the
1127                  * dirty bit in multiple places and clearing
1128                  * them concurrently from different threads.
1129                  *
1130                  * Note! Normally the "set_page_dirty(page)"
1131                  * has no effect on the actual dirty bit - since
1132                  * that will already usually be set. But we
1133                  * need the side effects, and it can help us
1134                  * avoid races.
1135                  *
1136                  * We basically use the page "master dirty bit"
1137                  * as a serialization point for all the different
1138                  * threads doing their things.
1139                  */
1140                 if (page_mkclean(page))
1141                         set_page_dirty(page);
1142                 /*
1143                  * We carefully synchronise fault handlers against
1144                  * installing a dirty pte and marking the page dirty
1145                  * at this point. We do this by having them hold the
1146                  * page lock at some point after installing their
1147                  * pte, but before marking the page dirty.
1148                  * Pages are always locked coming in here, so we get
1149                  * the desired exclusion. See mm/memory.c:do_wp_page()
1150                  * for more comments.
1151                  */
1152                 if (TestClearPageDirty(page)) {
1153                         dec_zone_page_state(page, NR_FILE_DIRTY);
1154                         dec_bdi_stat(mapping->backing_dev_info,
1155                                         BDI_RECLAIMABLE);
1156                         return 1;
1157                 }
1158                 return 0;
1159         }
1160         return TestClearPageDirty(page);
1161 }
1162 EXPORT_SYMBOL(clear_page_dirty_for_io);
1163
1164 int test_clear_page_writeback(struct page *page)
1165 {
1166         struct address_space *mapping = page_mapping(page);
1167         int ret;
1168
1169         if (mapping) {
1170                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1171                 unsigned long flags;
1172
1173                 write_lock_irqsave(&mapping->tree_lock, flags);
1174                 ret = TestClearPageWriteback(page);
1175                 if (ret) {
1176                         radix_tree_tag_clear(&mapping->page_tree,
1177                                                 page_index(page),
1178                                                 PAGECACHE_TAG_WRITEBACK);
1179                         if (bdi_cap_writeback_dirty(bdi)) {
1180                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1181                                 __bdi_writeout_inc(bdi);
1182                         }
1183                 }
1184                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1185         } else {
1186                 ret = TestClearPageWriteback(page);
1187         }
1188         if (ret)
1189                 dec_zone_page_state(page, NR_WRITEBACK);
1190         return ret;
1191 }
1192
1193 int test_set_page_writeback(struct page *page)
1194 {
1195         struct address_space *mapping = page_mapping(page);
1196         int ret;
1197
1198         if (mapping) {
1199                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1200                 unsigned long flags;
1201
1202                 write_lock_irqsave(&mapping->tree_lock, flags);
1203                 ret = TestSetPageWriteback(page);
1204                 if (!ret) {
1205                         radix_tree_tag_set(&mapping->page_tree,
1206                                                 page_index(page),
1207                                                 PAGECACHE_TAG_WRITEBACK);
1208                         if (bdi_cap_writeback_dirty(bdi))
1209                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1210                 }
1211                 if (!PageDirty(page))
1212                         radix_tree_tag_clear(&mapping->page_tree,
1213                                                 page_index(page),
1214                                                 PAGECACHE_TAG_DIRTY);
1215                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1216         } else {
1217                 ret = TestSetPageWriteback(page);
1218         }
1219         if (!ret)
1220                 inc_zone_page_state(page, NR_WRITEBACK);
1221         return ret;
1222
1223 }
1224 EXPORT_SYMBOL(test_set_page_writeback);
1225
1226 /*
1227  * Return true if any of the pages in the mapping are marked with the
1228  * passed tag.
1229  */
1230 int mapping_tagged(struct address_space *mapping, int tag)
1231 {
1232         int ret;
1233         rcu_read_lock();
1234         ret = radix_tree_tagged(&mapping->page_tree, tag);
1235         rcu_read_unlock();
1236         return ret;
1237 }
1238 EXPORT_SYMBOL(mapping_tagged);