mm/page-writeback: highmem_is_dirtyable option
[linux-2.6.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    akpm@zip.com.au
11  *              Initial version
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h>
36 #include <linux/pagevec.h>
37
38 /*
39  * The maximum number of pages to writeout in a single bdflush/kupdate
40  * operation.  We do this so we don't hold I_SYNC against an inode for
41  * enormous amounts of time, which would block a userspace task which has
42  * been forced to throttle against that inode.  Also, the code reevaluates
43  * the dirty each time it has written this many pages.
44  */
45 #define MAX_WRITEBACK_PAGES     1024
46
47 /*
48  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
49  * will look to see if it needs to force writeback or throttling.
50  */
51 static long ratelimit_pages = 32;
52
53 /*
54  * When balance_dirty_pages decides that the caller needs to perform some
55  * non-background writeback, this is how many pages it will attempt to write.
56  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
57  * large amounts of I/O are submitted.
58  */
59 static inline long sync_writeback_pages(void)
60 {
61         return ratelimit_pages + ratelimit_pages / 2;
62 }
63
64 /* The following parameters are exported via /proc/sys/vm */
65
66 /*
67  * Start background writeback (via pdflush) at this percentage
68  */
69 int dirty_background_ratio = 5;
70
71 /*
72  * free highmem will not be subtracted from the total free memory
73  * for calculating free ratios if vm_highmem_is_dirtyable is true
74  */
75 int vm_highmem_is_dirtyable;
76
77 /*
78  * The generator of dirty data starts writeback at this percentage
79  */
80 int vm_dirty_ratio = 10;
81
82 /*
83  * The interval between `kupdate'-style writebacks, in jiffies
84  */
85 int dirty_writeback_interval = 5 * HZ;
86
87 /*
88  * The longest number of jiffies for which data is allowed to remain dirty
89  */
90 int dirty_expire_interval = 30 * HZ;
91
92 /*
93  * Flag that makes the machine dump writes/reads and block dirtyings.
94  */
95 int block_dump;
96
97 /*
98  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
99  * a full sync is triggered after this time elapses without any disk activity.
100  */
101 int laptop_mode;
102
103 EXPORT_SYMBOL(laptop_mode);
104
105 /* End of sysctl-exported parameters */
106
107
108 static void background_writeout(unsigned long _min_pages);
109
110 /*
111  * Scale the writeback cache size proportional to the relative writeout speeds.
112  *
113  * We do this by keeping a floating proportion between BDIs, based on page
114  * writeback completions [end_page_writeback()]. Those devices that write out
115  * pages fastest will get the larger share, while the slower will get a smaller
116  * share.
117  *
118  * We use page writeout completions because we are interested in getting rid of
119  * dirty pages. Having them written out is the primary goal.
120  *
121  * We introduce a concept of time, a period over which we measure these events,
122  * because demand can/will vary over time. The length of this period itself is
123  * measured in page writeback completions.
124  *
125  */
126 static struct prop_descriptor vm_completions;
127 static struct prop_descriptor vm_dirties;
128
129 static unsigned long determine_dirtyable_memory(void);
130
131 /*
132  * couple the period to the dirty_ratio:
133  *
134  *   period/2 ~ roundup_pow_of_two(dirty limit)
135  */
136 static int calc_period_shift(void)
137 {
138         unsigned long dirty_total;
139
140         dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
141         return 2 + ilog2(dirty_total - 1);
142 }
143
144 /*
145  * update the period when the dirty ratio changes.
146  */
147 int dirty_ratio_handler(struct ctl_table *table, int write,
148                 struct file *filp, void __user *buffer, size_t *lenp,
149                 loff_t *ppos)
150 {
151         int old_ratio = vm_dirty_ratio;
152         int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
153         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
154                 int shift = calc_period_shift();
155                 prop_change_shift(&vm_completions, shift);
156                 prop_change_shift(&vm_dirties, shift);
157         }
158         return ret;
159 }
160
161 /*
162  * Increment the BDI's writeout completion count and the global writeout
163  * completion count. Called from test_clear_page_writeback().
164  */
165 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
166 {
167         __prop_inc_percpu(&vm_completions, &bdi->completions);
168 }
169
170 static inline void task_dirty_inc(struct task_struct *tsk)
171 {
172         prop_inc_single(&vm_dirties, &tsk->dirties);
173 }
174
175 /*
176  * Obtain an accurate fraction of the BDI's portion.
177  */
178 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
179                 long *numerator, long *denominator)
180 {
181         if (bdi_cap_writeback_dirty(bdi)) {
182                 prop_fraction_percpu(&vm_completions, &bdi->completions,
183                                 numerator, denominator);
184         } else {
185                 *numerator = 0;
186                 *denominator = 1;
187         }
188 }
189
190 /*
191  * Clip the earned share of dirty pages to that which is actually available.
192  * This avoids exceeding the total dirty_limit when the floating averages
193  * fluctuate too quickly.
194  */
195 static void
196 clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
197 {
198         long avail_dirty;
199
200         avail_dirty = dirty -
201                 (global_page_state(NR_FILE_DIRTY) +
202                  global_page_state(NR_WRITEBACK) +
203                  global_page_state(NR_UNSTABLE_NFS));
204
205         if (avail_dirty < 0)
206                 avail_dirty = 0;
207
208         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
209                 bdi_stat(bdi, BDI_WRITEBACK);
210
211         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
212 }
213
214 static inline void task_dirties_fraction(struct task_struct *tsk,
215                 long *numerator, long *denominator)
216 {
217         prop_fraction_single(&vm_dirties, &tsk->dirties,
218                                 numerator, denominator);
219 }
220
221 /*
222  * scale the dirty limit
223  *
224  * task specific dirty limit:
225  *
226  *   dirty -= (dirty/8) * p_{t}
227  */
228 static void task_dirty_limit(struct task_struct *tsk, long *pdirty)
229 {
230         long numerator, denominator;
231         long dirty = *pdirty;
232         u64 inv = dirty >> 3;
233
234         task_dirties_fraction(tsk, &numerator, &denominator);
235         inv *= numerator;
236         do_div(inv, denominator);
237
238         dirty -= inv;
239         if (dirty < *pdirty/2)
240                 dirty = *pdirty/2;
241
242         *pdirty = dirty;
243 }
244
245 /*
246  * Work out the current dirty-memory clamping and background writeout
247  * thresholds.
248  *
249  * The main aim here is to lower them aggressively if there is a lot of mapped
250  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
251  * pages.  It is better to clamp down on writers than to start swapping, and
252  * performing lots of scanning.
253  *
254  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
255  *
256  * We don't permit the clamping level to fall below 5% - that is getting rather
257  * excessive.
258  *
259  * We make sure that the background writeout level is below the adjusted
260  * clamping level.
261  */
262
263 static unsigned long highmem_dirtyable_memory(unsigned long total)
264 {
265 #ifdef CONFIG_HIGHMEM
266         int node;
267         unsigned long x = 0;
268
269         for_each_node_state(node, N_HIGH_MEMORY) {
270                 struct zone *z =
271                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
272
273                 x += zone_page_state(z, NR_FREE_PAGES)
274                         + zone_page_state(z, NR_INACTIVE)
275                         + zone_page_state(z, NR_ACTIVE);
276         }
277         /*
278          * Make sure that the number of highmem pages is never larger
279          * than the number of the total dirtyable memory. This can only
280          * occur in very strange VM situations but we want to make sure
281          * that this does not occur.
282          */
283         return min(x, total);
284 #else
285         return 0;
286 #endif
287 }
288
289 static unsigned long determine_dirtyable_memory(void)
290 {
291         unsigned long x;
292
293         x = global_page_state(NR_FREE_PAGES)
294                 + global_page_state(NR_INACTIVE)
295                 + global_page_state(NR_ACTIVE);
296
297         if (!vm_highmem_is_dirtyable)
298                 x -= highmem_dirtyable_memory(x);
299
300         return x + 1;   /* Ensure that we never return 0 */
301 }
302
303 static void
304 get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
305                  struct backing_dev_info *bdi)
306 {
307         int background_ratio;           /* Percentages */
308         int dirty_ratio;
309         long background;
310         long dirty;
311         unsigned long available_memory = determine_dirtyable_memory();
312         struct task_struct *tsk;
313
314         dirty_ratio = vm_dirty_ratio;
315         if (dirty_ratio < 5)
316                 dirty_ratio = 5;
317
318         background_ratio = dirty_background_ratio;
319         if (background_ratio >= dirty_ratio)
320                 background_ratio = dirty_ratio / 2;
321
322         background = (background_ratio * available_memory) / 100;
323         dirty = (dirty_ratio * available_memory) / 100;
324         tsk = current;
325         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
326                 background += background / 4;
327                 dirty += dirty / 4;
328         }
329         *pbackground = background;
330         *pdirty = dirty;
331
332         if (bdi) {
333                 u64 bdi_dirty = dirty;
334                 long numerator, denominator;
335
336                 /*
337                  * Calculate this BDI's share of the dirty ratio.
338                  */
339                 bdi_writeout_fraction(bdi, &numerator, &denominator);
340
341                 bdi_dirty *= numerator;
342                 do_div(bdi_dirty, denominator);
343
344                 *pbdi_dirty = bdi_dirty;
345                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
346                 task_dirty_limit(current, pbdi_dirty);
347         }
348 }
349
350 /*
351  * balance_dirty_pages() must be called by processes which are generating dirty
352  * data.  It looks at the number of dirty pages in the machine and will force
353  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
354  * If we're over `background_thresh' then pdflush is woken to perform some
355  * writeout.
356  */
357 static void balance_dirty_pages(struct address_space *mapping)
358 {
359         long nr_reclaimable, bdi_nr_reclaimable;
360         long nr_writeback, bdi_nr_writeback;
361         long background_thresh;
362         long dirty_thresh;
363         long bdi_thresh;
364         unsigned long pages_written = 0;
365         unsigned long write_chunk = sync_writeback_pages();
366
367         struct backing_dev_info *bdi = mapping->backing_dev_info;
368
369         for (;;) {
370                 struct writeback_control wbc = {
371                         .bdi            = bdi,
372                         .sync_mode      = WB_SYNC_NONE,
373                         .older_than_this = NULL,
374                         .nr_to_write    = write_chunk,
375                         .range_cyclic   = 1,
376                 };
377
378                 get_dirty_limits(&background_thresh, &dirty_thresh,
379                                 &bdi_thresh, bdi);
380
381                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
382                                         global_page_state(NR_UNSTABLE_NFS);
383                 nr_writeback = global_page_state(NR_WRITEBACK);
384
385                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
386                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
387
388                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
389                         break;
390
391                 /*
392                  * Throttle it only when the background writeback cannot
393                  * catch-up. This avoids (excessively) small writeouts
394                  * when the bdi limits are ramping up.
395                  */
396                 if (nr_reclaimable + nr_writeback <
397                                 (background_thresh + dirty_thresh) / 2)
398                         break;
399
400                 if (!bdi->dirty_exceeded)
401                         bdi->dirty_exceeded = 1;
402
403                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
404                  * Unstable writes are a feature of certain networked
405                  * filesystems (i.e. NFS) in which data may have been
406                  * written to the server's write cache, but has not yet
407                  * been flushed to permanent storage.
408                  */
409                 if (bdi_nr_reclaimable) {
410                         writeback_inodes(&wbc);
411                         pages_written += write_chunk - wbc.nr_to_write;
412                         get_dirty_limits(&background_thresh, &dirty_thresh,
413                                        &bdi_thresh, bdi);
414                 }
415
416                 /*
417                  * In order to avoid the stacked BDI deadlock we need
418                  * to ensure we accurately count the 'dirty' pages when
419                  * the threshold is low.
420                  *
421                  * Otherwise it would be possible to get thresh+n pages
422                  * reported dirty, even though there are thresh-m pages
423                  * actually dirty; with m+n sitting in the percpu
424                  * deltas.
425                  */
426                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
427                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
428                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
429                 } else if (bdi_nr_reclaimable) {
430                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
431                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
432                 }
433
434                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
435                         break;
436                 if (pages_written >= write_chunk)
437                         break;          /* We've done our duty */
438
439                 congestion_wait(WRITE, HZ/10);
440         }
441
442         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
443                         bdi->dirty_exceeded)
444                 bdi->dirty_exceeded = 0;
445
446         if (writeback_in_progress(bdi))
447                 return;         /* pdflush is already working this queue */
448
449         /*
450          * In laptop mode, we wait until hitting the higher threshold before
451          * starting background writeout, and then write out all the way down
452          * to the lower threshold.  So slow writers cause minimal disk activity.
453          *
454          * In normal mode, we start background writeout at the lower
455          * background_thresh, to keep the amount of dirty memory low.
456          */
457         if ((laptop_mode && pages_written) ||
458                         (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
459                                           + global_page_state(NR_UNSTABLE_NFS)
460                                           > background_thresh)))
461                 pdflush_operation(background_writeout, 0);
462 }
463
464 void set_page_dirty_balance(struct page *page, int page_mkwrite)
465 {
466         if (set_page_dirty(page) || page_mkwrite) {
467                 struct address_space *mapping = page_mapping(page);
468
469                 if (mapping)
470                         balance_dirty_pages_ratelimited(mapping);
471         }
472 }
473
474 /**
475  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
476  * @mapping: address_space which was dirtied
477  * @nr_pages_dirtied: number of pages which the caller has just dirtied
478  *
479  * Processes which are dirtying memory should call in here once for each page
480  * which was newly dirtied.  The function will periodically check the system's
481  * dirty state and will initiate writeback if needed.
482  *
483  * On really big machines, get_writeback_state is expensive, so try to avoid
484  * calling it too often (ratelimiting).  But once we're over the dirty memory
485  * limit we decrease the ratelimiting by a lot, to prevent individual processes
486  * from overshooting the limit by (ratelimit_pages) each.
487  */
488 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
489                                         unsigned long nr_pages_dirtied)
490 {
491         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
492         unsigned long ratelimit;
493         unsigned long *p;
494
495         ratelimit = ratelimit_pages;
496         if (mapping->backing_dev_info->dirty_exceeded)
497                 ratelimit = 8;
498
499         /*
500          * Check the rate limiting. Also, we do not want to throttle real-time
501          * tasks in balance_dirty_pages(). Period.
502          */
503         preempt_disable();
504         p =  &__get_cpu_var(ratelimits);
505         *p += nr_pages_dirtied;
506         if (unlikely(*p >= ratelimit)) {
507                 *p = 0;
508                 preempt_enable();
509                 balance_dirty_pages(mapping);
510                 return;
511         }
512         preempt_enable();
513 }
514 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
515
516 void throttle_vm_writeout(gfp_t gfp_mask)
517 {
518         long background_thresh;
519         long dirty_thresh;
520
521         for ( ; ; ) {
522                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
523
524                 /*
525                  * Boost the allowable dirty threshold a bit for page
526                  * allocators so they don't get DoS'ed by heavy writers
527                  */
528                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
529
530                 if (global_page_state(NR_UNSTABLE_NFS) +
531                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
532                                 break;
533                 congestion_wait(WRITE, HZ/10);
534
535                 /*
536                  * The caller might hold locks which can prevent IO completion
537                  * or progress in the filesystem.  So we cannot just sit here
538                  * waiting for IO to complete.
539                  */
540                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
541                         break;
542         }
543 }
544
545 /*
546  * writeback at least _min_pages, and keep writing until the amount of dirty
547  * memory is less than the background threshold, or until we're all clean.
548  */
549 static void background_writeout(unsigned long _min_pages)
550 {
551         long min_pages = _min_pages;
552         struct writeback_control wbc = {
553                 .bdi            = NULL,
554                 .sync_mode      = WB_SYNC_NONE,
555                 .older_than_this = NULL,
556                 .nr_to_write    = 0,
557                 .nonblocking    = 1,
558                 .range_cyclic   = 1,
559         };
560
561         for ( ; ; ) {
562                 long background_thresh;
563                 long dirty_thresh;
564
565                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
566                 if (global_page_state(NR_FILE_DIRTY) +
567                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
568                                 && min_pages <= 0)
569                         break;
570                 wbc.encountered_congestion = 0;
571                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
572                 wbc.pages_skipped = 0;
573                 writeback_inodes(&wbc);
574                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
575                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
576                         /* Wrote less than expected */
577                         congestion_wait(WRITE, HZ/10);
578                         if (!wbc.encountered_congestion)
579                                 break;
580                 }
581         }
582 }
583
584 /*
585  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
586  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
587  * -1 if all pdflush threads were busy.
588  */
589 int wakeup_pdflush(long nr_pages)
590 {
591         if (nr_pages == 0)
592                 nr_pages = global_page_state(NR_FILE_DIRTY) +
593                                 global_page_state(NR_UNSTABLE_NFS);
594         return pdflush_operation(background_writeout, nr_pages);
595 }
596
597 static void wb_timer_fn(unsigned long unused);
598 static void laptop_timer_fn(unsigned long unused);
599
600 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
601 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
602
603 /*
604  * Periodic writeback of "old" data.
605  *
606  * Define "old": the first time one of an inode's pages is dirtied, we mark the
607  * dirtying-time in the inode's address_space.  So this periodic writeback code
608  * just walks the superblock inode list, writing back any inodes which are
609  * older than a specific point in time.
610  *
611  * Try to run once per dirty_writeback_interval.  But if a writeback event
612  * takes longer than a dirty_writeback_interval interval, then leave a
613  * one-second gap.
614  *
615  * older_than_this takes precedence over nr_to_write.  So we'll only write back
616  * all dirty pages if they are all attached to "old" mappings.
617  */
618 static void wb_kupdate(unsigned long arg)
619 {
620         unsigned long oldest_jif;
621         unsigned long start_jif;
622         unsigned long next_jif;
623         long nr_to_write;
624         struct writeback_control wbc = {
625                 .bdi            = NULL,
626                 .sync_mode      = WB_SYNC_NONE,
627                 .older_than_this = &oldest_jif,
628                 .nr_to_write    = 0,
629                 .nonblocking    = 1,
630                 .for_kupdate    = 1,
631                 .range_cyclic   = 1,
632         };
633
634         sync_supers();
635
636         oldest_jif = jiffies - dirty_expire_interval;
637         start_jif = jiffies;
638         next_jif = start_jif + dirty_writeback_interval;
639         nr_to_write = global_page_state(NR_FILE_DIRTY) +
640                         global_page_state(NR_UNSTABLE_NFS) +
641                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
642         while (nr_to_write > 0) {
643                 wbc.encountered_congestion = 0;
644                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
645                 writeback_inodes(&wbc);
646                 if (wbc.nr_to_write > 0) {
647                         if (wbc.encountered_congestion)
648                                 congestion_wait(WRITE, HZ/10);
649                         else
650                                 break;  /* All the old data is written */
651                 }
652                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
653         }
654         if (time_before(next_jif, jiffies + HZ))
655                 next_jif = jiffies + HZ;
656         if (dirty_writeback_interval)
657                 mod_timer(&wb_timer, next_jif);
658 }
659
660 /*
661  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
662  */
663 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
664         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
665 {
666         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
667         if (dirty_writeback_interval)
668                 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
669         else
670                 del_timer(&wb_timer);
671         return 0;
672 }
673
674 static void wb_timer_fn(unsigned long unused)
675 {
676         if (pdflush_operation(wb_kupdate, 0) < 0)
677                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
678 }
679
680 static void laptop_flush(unsigned long unused)
681 {
682         sys_sync();
683 }
684
685 static void laptop_timer_fn(unsigned long unused)
686 {
687         pdflush_operation(laptop_flush, 0);
688 }
689
690 /*
691  * We've spun up the disk and we're in laptop mode: schedule writeback
692  * of all dirty data a few seconds from now.  If the flush is already scheduled
693  * then push it back - the user is still using the disk.
694  */
695 void laptop_io_completion(void)
696 {
697         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
698 }
699
700 /*
701  * We're in laptop mode and we've just synced. The sync's writes will have
702  * caused another writeback to be scheduled by laptop_io_completion.
703  * Nothing needs to be written back anymore, so we unschedule the writeback.
704  */
705 void laptop_sync_completion(void)
706 {
707         del_timer(&laptop_mode_wb_timer);
708 }
709
710 /*
711  * If ratelimit_pages is too high then we can get into dirty-data overload
712  * if a large number of processes all perform writes at the same time.
713  * If it is too low then SMP machines will call the (expensive)
714  * get_writeback_state too often.
715  *
716  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
717  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
718  * thresholds before writeback cuts in.
719  *
720  * But the limit should not be set too high.  Because it also controls the
721  * amount of memory which the balance_dirty_pages() caller has to write back.
722  * If this is too large then the caller will block on the IO queue all the
723  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
724  * will write six megabyte chunks, max.
725  */
726
727 void writeback_set_ratelimit(void)
728 {
729         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
730         if (ratelimit_pages < 16)
731                 ratelimit_pages = 16;
732         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
733                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
734 }
735
736 static int __cpuinit
737 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
738 {
739         writeback_set_ratelimit();
740         return NOTIFY_DONE;
741 }
742
743 static struct notifier_block __cpuinitdata ratelimit_nb = {
744         .notifier_call  = ratelimit_handler,
745         .next           = NULL,
746 };
747
748 /*
749  * Called early on to tune the page writeback dirty limits.
750  *
751  * We used to scale dirty pages according to how total memory
752  * related to pages that could be allocated for buffers (by
753  * comparing nr_free_buffer_pages() to vm_total_pages.
754  *
755  * However, that was when we used "dirty_ratio" to scale with
756  * all memory, and we don't do that any more. "dirty_ratio"
757  * is now applied to total non-HIGHPAGE memory (by subtracting
758  * totalhigh_pages from vm_total_pages), and as such we can't
759  * get into the old insane situation any more where we had
760  * large amounts of dirty pages compared to a small amount of
761  * non-HIGHMEM memory.
762  *
763  * But we might still want to scale the dirty_ratio by how
764  * much memory the box has..
765  */
766 void __init page_writeback_init(void)
767 {
768         int shift;
769
770         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
771         writeback_set_ratelimit();
772         register_cpu_notifier(&ratelimit_nb);
773
774         shift = calc_period_shift();
775         prop_descriptor_init(&vm_completions, shift);
776         prop_descriptor_init(&vm_dirties, shift);
777 }
778
779 /**
780  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
781  * @mapping: address space structure to write
782  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
783  * @writepage: function called for each page
784  * @data: data passed to writepage function
785  *
786  * If a page is already under I/O, write_cache_pages() skips it, even
787  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
788  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
789  * and msync() need to guarantee that all the data which was dirty at the time
790  * the call was made get new I/O started against them.  If wbc->sync_mode is
791  * WB_SYNC_ALL then we were called for data integrity and we must wait for
792  * existing IO to complete.
793  */
794 int write_cache_pages(struct address_space *mapping,
795                       struct writeback_control *wbc, writepage_t writepage,
796                       void *data)
797 {
798         struct backing_dev_info *bdi = mapping->backing_dev_info;
799         int ret = 0;
800         int done = 0;
801         struct pagevec pvec;
802         int nr_pages;
803         pgoff_t index;
804         pgoff_t end;            /* Inclusive */
805         int scanned = 0;
806         int range_whole = 0;
807
808         if (wbc->nonblocking && bdi_write_congested(bdi)) {
809                 wbc->encountered_congestion = 1;
810                 return 0;
811         }
812
813         pagevec_init(&pvec, 0);
814         if (wbc->range_cyclic) {
815                 index = mapping->writeback_index; /* Start from prev offset */
816                 end = -1;
817         } else {
818                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
819                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
820                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
821                         range_whole = 1;
822                 scanned = 1;
823         }
824 retry:
825         while (!done && (index <= end) &&
826                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
827                                               PAGECACHE_TAG_DIRTY,
828                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
829                 unsigned i;
830
831                 scanned = 1;
832                 for (i = 0; i < nr_pages; i++) {
833                         struct page *page = pvec.pages[i];
834
835                         /*
836                          * At this point we hold neither mapping->tree_lock nor
837                          * lock on the page itself: the page may be truncated or
838                          * invalidated (changing page->mapping to NULL), or even
839                          * swizzled back from swapper_space to tmpfs file
840                          * mapping
841                          */
842                         lock_page(page);
843
844                         if (unlikely(page->mapping != mapping)) {
845                                 unlock_page(page);
846                                 continue;
847                         }
848
849                         if (!wbc->range_cyclic && page->index > end) {
850                                 done = 1;
851                                 unlock_page(page);
852                                 continue;
853                         }
854
855                         if (wbc->sync_mode != WB_SYNC_NONE)
856                                 wait_on_page_writeback(page);
857
858                         if (PageWriteback(page) ||
859                             !clear_page_dirty_for_io(page)) {
860                                 unlock_page(page);
861                                 continue;
862                         }
863
864                         ret = (*writepage)(page, wbc, data);
865
866                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
867                                 unlock_page(page);
868                                 ret = 0;
869                         }
870                         if (ret || (--(wbc->nr_to_write) <= 0))
871                                 done = 1;
872                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
873                                 wbc->encountered_congestion = 1;
874                                 done = 1;
875                         }
876                 }
877                 pagevec_release(&pvec);
878                 cond_resched();
879         }
880         if (!scanned && !done) {
881                 /*
882                  * We hit the last page and there is more work to be done: wrap
883                  * back to the start of the file
884                  */
885                 scanned = 1;
886                 index = 0;
887                 goto retry;
888         }
889         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
890                 mapping->writeback_index = index;
891         return ret;
892 }
893 EXPORT_SYMBOL(write_cache_pages);
894
895 /*
896  * Function used by generic_writepages to call the real writepage
897  * function and set the mapping flags on error
898  */
899 static int __writepage(struct page *page, struct writeback_control *wbc,
900                        void *data)
901 {
902         struct address_space *mapping = data;
903         int ret = mapping->a_ops->writepage(page, wbc);
904         mapping_set_error(mapping, ret);
905         return ret;
906 }
907
908 /**
909  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
910  * @mapping: address space structure to write
911  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
912  *
913  * This is a library function, which implements the writepages()
914  * address_space_operation.
915  */
916 int generic_writepages(struct address_space *mapping,
917                        struct writeback_control *wbc)
918 {
919         /* deal with chardevs and other special file */
920         if (!mapping->a_ops->writepage)
921                 return 0;
922
923         return write_cache_pages(mapping, wbc, __writepage, mapping);
924 }
925
926 EXPORT_SYMBOL(generic_writepages);
927
928 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
929 {
930         int ret;
931
932         if (wbc->nr_to_write <= 0)
933                 return 0;
934         wbc->for_writepages = 1;
935         if (mapping->a_ops->writepages)
936                 ret = mapping->a_ops->writepages(mapping, wbc);
937         else
938                 ret = generic_writepages(mapping, wbc);
939         wbc->for_writepages = 0;
940         return ret;
941 }
942
943 /**
944  * write_one_page - write out a single page and optionally wait on I/O
945  * @page: the page to write
946  * @wait: if true, wait on writeout
947  *
948  * The page must be locked by the caller and will be unlocked upon return.
949  *
950  * write_one_page() returns a negative error code if I/O failed.
951  */
952 int write_one_page(struct page *page, int wait)
953 {
954         struct address_space *mapping = page->mapping;
955         int ret = 0;
956         struct writeback_control wbc = {
957                 .sync_mode = WB_SYNC_ALL,
958                 .nr_to_write = 1,
959         };
960
961         BUG_ON(!PageLocked(page));
962
963         if (wait)
964                 wait_on_page_writeback(page);
965
966         if (clear_page_dirty_for_io(page)) {
967                 page_cache_get(page);
968                 ret = mapping->a_ops->writepage(page, &wbc);
969                 if (ret == 0 && wait) {
970                         wait_on_page_writeback(page);
971                         if (PageError(page))
972                                 ret = -EIO;
973                 }
974                 page_cache_release(page);
975         } else {
976                 unlock_page(page);
977         }
978         return ret;
979 }
980 EXPORT_SYMBOL(write_one_page);
981
982 /*
983  * For address_spaces which do not use buffers nor write back.
984  */
985 int __set_page_dirty_no_writeback(struct page *page)
986 {
987         if (!PageDirty(page))
988                 SetPageDirty(page);
989         return 0;
990 }
991
992 /*
993  * For address_spaces which do not use buffers.  Just tag the page as dirty in
994  * its radix tree.
995  *
996  * This is also used when a single buffer is being dirtied: we want to set the
997  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
998  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
999  *
1000  * Most callers have locked the page, which pins the address_space in memory.
1001  * But zap_pte_range() does not lock the page, however in that case the
1002  * mapping is pinned by the vma's ->vm_file reference.
1003  *
1004  * We take care to handle the case where the page was truncated from the
1005  * mapping by re-checking page_mapping() inside tree_lock.
1006  */
1007 int __set_page_dirty_nobuffers(struct page *page)
1008 {
1009         if (!TestSetPageDirty(page)) {
1010                 struct address_space *mapping = page_mapping(page);
1011                 struct address_space *mapping2;
1012
1013                 if (!mapping)
1014                         return 1;
1015
1016                 write_lock_irq(&mapping->tree_lock);
1017                 mapping2 = page_mapping(page);
1018                 if (mapping2) { /* Race with truncate? */
1019                         BUG_ON(mapping2 != mapping);
1020                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1021                         if (mapping_cap_account_dirty(mapping)) {
1022                                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1023                                 __inc_bdi_stat(mapping->backing_dev_info,
1024                                                 BDI_RECLAIMABLE);
1025                                 task_io_account_write(PAGE_CACHE_SIZE);
1026                         }
1027                         radix_tree_tag_set(&mapping->page_tree,
1028                                 page_index(page), PAGECACHE_TAG_DIRTY);
1029                 }
1030                 write_unlock_irq(&mapping->tree_lock);
1031                 if (mapping->host) {
1032                         /* !PageAnon && !swapper_space */
1033                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1034                 }
1035                 return 1;
1036         }
1037         return 0;
1038 }
1039 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1040
1041 /*
1042  * When a writepage implementation decides that it doesn't want to write this
1043  * page for some reason, it should redirty the locked page via
1044  * redirty_page_for_writepage() and it should then unlock the page and return 0
1045  */
1046 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1047 {
1048         wbc->pages_skipped++;
1049         return __set_page_dirty_nobuffers(page);
1050 }
1051 EXPORT_SYMBOL(redirty_page_for_writepage);
1052
1053 /*
1054  * If the mapping doesn't provide a set_page_dirty a_op, then
1055  * just fall through and assume that it wants buffer_heads.
1056  */
1057 static int __set_page_dirty(struct page *page)
1058 {
1059         struct address_space *mapping = page_mapping(page);
1060
1061         if (likely(mapping)) {
1062                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1063 #ifdef CONFIG_BLOCK
1064                 if (!spd)
1065                         spd = __set_page_dirty_buffers;
1066 #endif
1067                 return (*spd)(page);
1068         }
1069         if (!PageDirty(page)) {
1070                 if (!TestSetPageDirty(page))
1071                         return 1;
1072         }
1073         return 0;
1074 }
1075
1076 int fastcall set_page_dirty(struct page *page)
1077 {
1078         int ret = __set_page_dirty(page);
1079         if (ret)
1080                 task_dirty_inc(current);
1081         return ret;
1082 }
1083 EXPORT_SYMBOL(set_page_dirty);
1084
1085 /*
1086  * set_page_dirty() is racy if the caller has no reference against
1087  * page->mapping->host, and if the page is unlocked.  This is because another
1088  * CPU could truncate the page off the mapping and then free the mapping.
1089  *
1090  * Usually, the page _is_ locked, or the caller is a user-space process which
1091  * holds a reference on the inode by having an open file.
1092  *
1093  * In other cases, the page should be locked before running set_page_dirty().
1094  */
1095 int set_page_dirty_lock(struct page *page)
1096 {
1097         int ret;
1098
1099         lock_page_nosync(page);
1100         ret = set_page_dirty(page);
1101         unlock_page(page);
1102         return ret;
1103 }
1104 EXPORT_SYMBOL(set_page_dirty_lock);
1105
1106 /*
1107  * Clear a page's dirty flag, while caring for dirty memory accounting.
1108  * Returns true if the page was previously dirty.
1109  *
1110  * This is for preparing to put the page under writeout.  We leave the page
1111  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1112  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1113  * implementation will run either set_page_writeback() or set_page_dirty(),
1114  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1115  * back into sync.
1116  *
1117  * This incoherency between the page's dirty flag and radix-tree tag is
1118  * unfortunate, but it only exists while the page is locked.
1119  */
1120 int clear_page_dirty_for_io(struct page *page)
1121 {
1122         struct address_space *mapping = page_mapping(page);
1123
1124         BUG_ON(!PageLocked(page));
1125
1126         ClearPageReclaim(page);
1127         if (mapping && mapping_cap_account_dirty(mapping)) {
1128                 /*
1129                  * Yes, Virginia, this is indeed insane.
1130                  *
1131                  * We use this sequence to make sure that
1132                  *  (a) we account for dirty stats properly
1133                  *  (b) we tell the low-level filesystem to
1134                  *      mark the whole page dirty if it was
1135                  *      dirty in a pagetable. Only to then
1136                  *  (c) clean the page again and return 1 to
1137                  *      cause the writeback.
1138                  *
1139                  * This way we avoid all nasty races with the
1140                  * dirty bit in multiple places and clearing
1141                  * them concurrently from different threads.
1142                  *
1143                  * Note! Normally the "set_page_dirty(page)"
1144                  * has no effect on the actual dirty bit - since
1145                  * that will already usually be set. But we
1146                  * need the side effects, and it can help us
1147                  * avoid races.
1148                  *
1149                  * We basically use the page "master dirty bit"
1150                  * as a serialization point for all the different
1151                  * threads doing their things.
1152                  */
1153                 if (page_mkclean(page))
1154                         set_page_dirty(page);
1155                 /*
1156                  * We carefully synchronise fault handlers against
1157                  * installing a dirty pte and marking the page dirty
1158                  * at this point. We do this by having them hold the
1159                  * page lock at some point after installing their
1160                  * pte, but before marking the page dirty.
1161                  * Pages are always locked coming in here, so we get
1162                  * the desired exclusion. See mm/memory.c:do_wp_page()
1163                  * for more comments.
1164                  */
1165                 if (TestClearPageDirty(page)) {
1166                         dec_zone_page_state(page, NR_FILE_DIRTY);
1167                         dec_bdi_stat(mapping->backing_dev_info,
1168                                         BDI_RECLAIMABLE);
1169                         return 1;
1170                 }
1171                 return 0;
1172         }
1173         return TestClearPageDirty(page);
1174 }
1175 EXPORT_SYMBOL(clear_page_dirty_for_io);
1176
1177 int test_clear_page_writeback(struct page *page)
1178 {
1179         struct address_space *mapping = page_mapping(page);
1180         int ret;
1181
1182         if (mapping) {
1183                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1184                 unsigned long flags;
1185
1186                 write_lock_irqsave(&mapping->tree_lock, flags);
1187                 ret = TestClearPageWriteback(page);
1188                 if (ret) {
1189                         radix_tree_tag_clear(&mapping->page_tree,
1190                                                 page_index(page),
1191                                                 PAGECACHE_TAG_WRITEBACK);
1192                         if (bdi_cap_writeback_dirty(bdi)) {
1193                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1194                                 __bdi_writeout_inc(bdi);
1195                         }
1196                 }
1197                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1198         } else {
1199                 ret = TestClearPageWriteback(page);
1200         }
1201         if (ret)
1202                 dec_zone_page_state(page, NR_WRITEBACK);
1203         return ret;
1204 }
1205
1206 int test_set_page_writeback(struct page *page)
1207 {
1208         struct address_space *mapping = page_mapping(page);
1209         int ret;
1210
1211         if (mapping) {
1212                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1213                 unsigned long flags;
1214
1215                 write_lock_irqsave(&mapping->tree_lock, flags);
1216                 ret = TestSetPageWriteback(page);
1217                 if (!ret) {
1218                         radix_tree_tag_set(&mapping->page_tree,
1219                                                 page_index(page),
1220                                                 PAGECACHE_TAG_WRITEBACK);
1221                         if (bdi_cap_writeback_dirty(bdi))
1222                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1223                 }
1224                 if (!PageDirty(page))
1225                         radix_tree_tag_clear(&mapping->page_tree,
1226                                                 page_index(page),
1227                                                 PAGECACHE_TAG_DIRTY);
1228                 write_unlock_irqrestore(&mapping->tree_lock, flags);
1229         } else {
1230                 ret = TestSetPageWriteback(page);
1231         }
1232         if (!ret)
1233                 inc_zone_page_state(page, NR_WRITEBACK);
1234         return ret;
1235
1236 }
1237 EXPORT_SYMBOL(test_set_page_writeback);
1238
1239 /*
1240  * Return true if any of the pages in the mapping are marked with the
1241  * passed tag.
1242  */
1243 int mapping_tagged(struct address_space *mapping, int tag)
1244 {
1245         int ret;
1246         rcu_read_lock();
1247         ret = radix_tree_tagged(&mapping->page_tree, tag);
1248         rcu_read_unlock();
1249         return ret;
1250 }
1251 EXPORT_SYMBOL(mapping_tagged);