mm: spinlock tree_lock
[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 /*
130  * couple the period to the dirty_ratio:
131  *
132  *   period/2 ~ roundup_pow_of_two(dirty limit)
133  */
134 static int calc_period_shift(void)
135 {
136         unsigned long dirty_total;
137
138         dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
139         return 2 + ilog2(dirty_total - 1);
140 }
141
142 /*
143  * update the period when the dirty ratio changes.
144  */
145 int dirty_ratio_handler(struct ctl_table *table, int write,
146                 struct file *filp, void __user *buffer, size_t *lenp,
147                 loff_t *ppos)
148 {
149         int old_ratio = vm_dirty_ratio;
150         int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
151         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
152                 int shift = calc_period_shift();
153                 prop_change_shift(&vm_completions, shift);
154                 prop_change_shift(&vm_dirties, shift);
155         }
156         return ret;
157 }
158
159 /*
160  * Increment the BDI's writeout completion count and the global writeout
161  * completion count. Called from test_clear_page_writeback().
162  */
163 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
164 {
165         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
166                               bdi->max_prop_frac);
167 }
168
169 void bdi_writeout_inc(struct backing_dev_info *bdi)
170 {
171         unsigned long flags;
172
173         local_irq_save(flags);
174         __bdi_writeout_inc(bdi);
175         local_irq_restore(flags);
176 }
177 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
178
179 static inline void task_dirty_inc(struct task_struct *tsk)
180 {
181         prop_inc_single(&vm_dirties, &tsk->dirties);
182 }
183
184 /*
185  * Obtain an accurate fraction of the BDI's portion.
186  */
187 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
188                 long *numerator, long *denominator)
189 {
190         if (bdi_cap_writeback_dirty(bdi)) {
191                 prop_fraction_percpu(&vm_completions, &bdi->completions,
192                                 numerator, denominator);
193         } else {
194                 *numerator = 0;
195                 *denominator = 1;
196         }
197 }
198
199 /*
200  * Clip the earned share of dirty pages to that which is actually available.
201  * This avoids exceeding the total dirty_limit when the floating averages
202  * fluctuate too quickly.
203  */
204 static void
205 clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
206 {
207         long avail_dirty;
208
209         avail_dirty = dirty -
210                 (global_page_state(NR_FILE_DIRTY) +
211                  global_page_state(NR_WRITEBACK) +
212                  global_page_state(NR_UNSTABLE_NFS) +
213                  global_page_state(NR_WRITEBACK_TEMP));
214
215         if (avail_dirty < 0)
216                 avail_dirty = 0;
217
218         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
219                 bdi_stat(bdi, BDI_WRITEBACK);
220
221         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
222 }
223
224 static inline void task_dirties_fraction(struct task_struct *tsk,
225                 long *numerator, long *denominator)
226 {
227         prop_fraction_single(&vm_dirties, &tsk->dirties,
228                                 numerator, denominator);
229 }
230
231 /*
232  * scale the dirty limit
233  *
234  * task specific dirty limit:
235  *
236  *   dirty -= (dirty/8) * p_{t}
237  */
238 static void task_dirty_limit(struct task_struct *tsk, long *pdirty)
239 {
240         long numerator, denominator;
241         long dirty = *pdirty;
242         u64 inv = dirty >> 3;
243
244         task_dirties_fraction(tsk, &numerator, &denominator);
245         inv *= numerator;
246         do_div(inv, denominator);
247
248         dirty -= inv;
249         if (dirty < *pdirty/2)
250                 dirty = *pdirty/2;
251
252         *pdirty = dirty;
253 }
254
255 /*
256  *
257  */
258 static DEFINE_SPINLOCK(bdi_lock);
259 static unsigned int bdi_min_ratio;
260
261 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
262 {
263         int ret = 0;
264         unsigned long flags;
265
266         spin_lock_irqsave(&bdi_lock, flags);
267         if (min_ratio > bdi->max_ratio) {
268                 ret = -EINVAL;
269         } else {
270                 min_ratio -= bdi->min_ratio;
271                 if (bdi_min_ratio + min_ratio < 100) {
272                         bdi_min_ratio += min_ratio;
273                         bdi->min_ratio += min_ratio;
274                 } else {
275                         ret = -EINVAL;
276                 }
277         }
278         spin_unlock_irqrestore(&bdi_lock, flags);
279
280         return ret;
281 }
282
283 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
284 {
285         unsigned long flags;
286         int ret = 0;
287
288         if (max_ratio > 100)
289                 return -EINVAL;
290
291         spin_lock_irqsave(&bdi_lock, flags);
292         if (bdi->min_ratio > max_ratio) {
293                 ret = -EINVAL;
294         } else {
295                 bdi->max_ratio = max_ratio;
296                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
297         }
298         spin_unlock_irqrestore(&bdi_lock, flags);
299
300         return ret;
301 }
302 EXPORT_SYMBOL(bdi_set_max_ratio);
303
304 /*
305  * Work out the current dirty-memory clamping and background writeout
306  * thresholds.
307  *
308  * The main aim here is to lower them aggressively if there is a lot of mapped
309  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
310  * pages.  It is better to clamp down on writers than to start swapping, and
311  * performing lots of scanning.
312  *
313  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
314  *
315  * We don't permit the clamping level to fall below 5% - that is getting rather
316  * excessive.
317  *
318  * We make sure that the background writeout level is below the adjusted
319  * clamping level.
320  */
321
322 static unsigned long highmem_dirtyable_memory(unsigned long total)
323 {
324 #ifdef CONFIG_HIGHMEM
325         int node;
326         unsigned long x = 0;
327
328         for_each_node_state(node, N_HIGH_MEMORY) {
329                 struct zone *z =
330                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
331
332                 x += zone_page_state(z, NR_FREE_PAGES)
333                         + zone_page_state(z, NR_INACTIVE)
334                         + zone_page_state(z, NR_ACTIVE);
335         }
336         /*
337          * Make sure that the number of highmem pages is never larger
338          * than the number of the total dirtyable memory. This can only
339          * occur in very strange VM situations but we want to make sure
340          * that this does not occur.
341          */
342         return min(x, total);
343 #else
344         return 0;
345 #endif
346 }
347
348 /**
349  * determine_dirtyable_memory - amount of memory that may be used
350  *
351  * Returns the numebr of pages that can currently be freed and used
352  * by the kernel for direct mappings.
353  */
354 unsigned long determine_dirtyable_memory(void)
355 {
356         unsigned long x;
357
358         x = global_page_state(NR_FREE_PAGES)
359                 + global_page_state(NR_INACTIVE)
360                 + global_page_state(NR_ACTIVE);
361
362         if (!vm_highmem_is_dirtyable)
363                 x -= highmem_dirtyable_memory(x);
364
365         return x + 1;   /* Ensure that we never return 0 */
366 }
367
368 void
369 get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
370                  struct backing_dev_info *bdi)
371 {
372         int background_ratio;           /* Percentages */
373         int dirty_ratio;
374         long background;
375         long dirty;
376         unsigned long available_memory = determine_dirtyable_memory();
377         struct task_struct *tsk;
378
379         dirty_ratio = vm_dirty_ratio;
380         if (dirty_ratio < 5)
381                 dirty_ratio = 5;
382
383         background_ratio = dirty_background_ratio;
384         if (background_ratio >= dirty_ratio)
385                 background_ratio = dirty_ratio / 2;
386
387         background = (background_ratio * available_memory) / 100;
388         dirty = (dirty_ratio * available_memory) / 100;
389         tsk = current;
390         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
391                 background += background / 4;
392                 dirty += dirty / 4;
393         }
394         *pbackground = background;
395         *pdirty = dirty;
396
397         if (bdi) {
398                 u64 bdi_dirty;
399                 long numerator, denominator;
400
401                 /*
402                  * Calculate this BDI's share of the dirty ratio.
403                  */
404                 bdi_writeout_fraction(bdi, &numerator, &denominator);
405
406                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
407                 bdi_dirty *= numerator;
408                 do_div(bdi_dirty, denominator);
409                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
410                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
411                         bdi_dirty = dirty * bdi->max_ratio / 100;
412
413                 *pbdi_dirty = bdi_dirty;
414                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
415                 task_dirty_limit(current, pbdi_dirty);
416         }
417 }
418
419 /*
420  * balance_dirty_pages() must be called by processes which are generating dirty
421  * data.  It looks at the number of dirty pages in the machine and will force
422  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
423  * If we're over `background_thresh' then pdflush is woken to perform some
424  * writeout.
425  */
426 static void balance_dirty_pages(struct address_space *mapping)
427 {
428         long nr_reclaimable, bdi_nr_reclaimable;
429         long nr_writeback, bdi_nr_writeback;
430         long background_thresh;
431         long dirty_thresh;
432         long bdi_thresh;
433         unsigned long pages_written = 0;
434         unsigned long write_chunk = sync_writeback_pages();
435
436         struct backing_dev_info *bdi = mapping->backing_dev_info;
437
438         for (;;) {
439                 struct writeback_control wbc = {
440                         .bdi            = bdi,
441                         .sync_mode      = WB_SYNC_NONE,
442                         .older_than_this = NULL,
443                         .nr_to_write    = write_chunk,
444                         .range_cyclic   = 1,
445                 };
446
447                 get_dirty_limits(&background_thresh, &dirty_thresh,
448                                 &bdi_thresh, bdi);
449
450                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
451                                         global_page_state(NR_UNSTABLE_NFS);
452                 nr_writeback = global_page_state(NR_WRITEBACK);
453
454                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
455                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
456
457                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
458                         break;
459
460                 /*
461                  * Throttle it only when the background writeback cannot
462                  * catch-up. This avoids (excessively) small writeouts
463                  * when the bdi limits are ramping up.
464                  */
465                 if (nr_reclaimable + nr_writeback <
466                                 (background_thresh + dirty_thresh) / 2)
467                         break;
468
469                 if (!bdi->dirty_exceeded)
470                         bdi->dirty_exceeded = 1;
471
472                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
473                  * Unstable writes are a feature of certain networked
474                  * filesystems (i.e. NFS) in which data may have been
475                  * written to the server's write cache, but has not yet
476                  * been flushed to permanent storage.
477                  */
478                 if (bdi_nr_reclaimable) {
479                         writeback_inodes(&wbc);
480                         pages_written += write_chunk - wbc.nr_to_write;
481                         get_dirty_limits(&background_thresh, &dirty_thresh,
482                                        &bdi_thresh, bdi);
483                 }
484
485                 /*
486                  * In order to avoid the stacked BDI deadlock we need
487                  * to ensure we accurately count the 'dirty' pages when
488                  * the threshold is low.
489                  *
490                  * Otherwise it would be possible to get thresh+n pages
491                  * reported dirty, even though there are thresh-m pages
492                  * actually dirty; with m+n sitting in the percpu
493                  * deltas.
494                  */
495                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
496                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
497                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
498                 } else if (bdi_nr_reclaimable) {
499                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
500                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
501                 }
502
503                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
504                         break;
505                 if (pages_written >= write_chunk)
506                         break;          /* We've done our duty */
507
508                 congestion_wait(WRITE, HZ/10);
509         }
510
511         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
512                         bdi->dirty_exceeded)
513                 bdi->dirty_exceeded = 0;
514
515         if (writeback_in_progress(bdi))
516                 return;         /* pdflush is already working this queue */
517
518         /*
519          * In laptop mode, we wait until hitting the higher threshold before
520          * starting background writeout, and then write out all the way down
521          * to the lower threshold.  So slow writers cause minimal disk activity.
522          *
523          * In normal mode, we start background writeout at the lower
524          * background_thresh, to keep the amount of dirty memory low.
525          */
526         if ((laptop_mode && pages_written) ||
527                         (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
528                                           + global_page_state(NR_UNSTABLE_NFS)
529                                           > background_thresh)))
530                 pdflush_operation(background_writeout, 0);
531 }
532
533 void set_page_dirty_balance(struct page *page, int page_mkwrite)
534 {
535         if (set_page_dirty(page) || page_mkwrite) {
536                 struct address_space *mapping = page_mapping(page);
537
538                 if (mapping)
539                         balance_dirty_pages_ratelimited(mapping);
540         }
541 }
542
543 /**
544  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
545  * @mapping: address_space which was dirtied
546  * @nr_pages_dirtied: number of pages which the caller has just dirtied
547  *
548  * Processes which are dirtying memory should call in here once for each page
549  * which was newly dirtied.  The function will periodically check the system's
550  * dirty state and will initiate writeback if needed.
551  *
552  * On really big machines, get_writeback_state is expensive, so try to avoid
553  * calling it too often (ratelimiting).  But once we're over the dirty memory
554  * limit we decrease the ratelimiting by a lot, to prevent individual processes
555  * from overshooting the limit by (ratelimit_pages) each.
556  */
557 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
558                                         unsigned long nr_pages_dirtied)
559 {
560         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
561         unsigned long ratelimit;
562         unsigned long *p;
563
564         ratelimit = ratelimit_pages;
565         if (mapping->backing_dev_info->dirty_exceeded)
566                 ratelimit = 8;
567
568         /*
569          * Check the rate limiting. Also, we do not want to throttle real-time
570          * tasks in balance_dirty_pages(). Period.
571          */
572         preempt_disable();
573         p =  &__get_cpu_var(ratelimits);
574         *p += nr_pages_dirtied;
575         if (unlikely(*p >= ratelimit)) {
576                 *p = 0;
577                 preempt_enable();
578                 balance_dirty_pages(mapping);
579                 return;
580         }
581         preempt_enable();
582 }
583 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
584
585 void throttle_vm_writeout(gfp_t gfp_mask)
586 {
587         long background_thresh;
588         long dirty_thresh;
589
590         for ( ; ; ) {
591                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
592
593                 /*
594                  * Boost the allowable dirty threshold a bit for page
595                  * allocators so they don't get DoS'ed by heavy writers
596                  */
597                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
598
599                 if (global_page_state(NR_UNSTABLE_NFS) +
600                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
601                                 break;
602                 congestion_wait(WRITE, HZ/10);
603
604                 /*
605                  * The caller might hold locks which can prevent IO completion
606                  * or progress in the filesystem.  So we cannot just sit here
607                  * waiting for IO to complete.
608                  */
609                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
610                         break;
611         }
612 }
613
614 /*
615  * writeback at least _min_pages, and keep writing until the amount of dirty
616  * memory is less than the background threshold, or until we're all clean.
617  */
618 static void background_writeout(unsigned long _min_pages)
619 {
620         long min_pages = _min_pages;
621         struct writeback_control wbc = {
622                 .bdi            = NULL,
623                 .sync_mode      = WB_SYNC_NONE,
624                 .older_than_this = NULL,
625                 .nr_to_write    = 0,
626                 .nonblocking    = 1,
627                 .range_cyclic   = 1,
628         };
629
630         for ( ; ; ) {
631                 long background_thresh;
632                 long dirty_thresh;
633
634                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
635                 if (global_page_state(NR_FILE_DIRTY) +
636                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
637                                 && min_pages <= 0)
638                         break;
639                 wbc.more_io = 0;
640                 wbc.encountered_congestion = 0;
641                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
642                 wbc.pages_skipped = 0;
643                 writeback_inodes(&wbc);
644                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
645                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
646                         /* Wrote less than expected */
647                         if (wbc.encountered_congestion || wbc.more_io)
648                                 congestion_wait(WRITE, HZ/10);
649                         else
650                                 break;
651                 }
652         }
653 }
654
655 /*
656  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
657  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
658  * -1 if all pdflush threads were busy.
659  */
660 int wakeup_pdflush(long nr_pages)
661 {
662         if (nr_pages == 0)
663                 nr_pages = global_page_state(NR_FILE_DIRTY) +
664                                 global_page_state(NR_UNSTABLE_NFS);
665         return pdflush_operation(background_writeout, nr_pages);
666 }
667
668 static void wb_timer_fn(unsigned long unused);
669 static void laptop_timer_fn(unsigned long unused);
670
671 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
672 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
673
674 /*
675  * Periodic writeback of "old" data.
676  *
677  * Define "old": the first time one of an inode's pages is dirtied, we mark the
678  * dirtying-time in the inode's address_space.  So this periodic writeback code
679  * just walks the superblock inode list, writing back any inodes which are
680  * older than a specific point in time.
681  *
682  * Try to run once per dirty_writeback_interval.  But if a writeback event
683  * takes longer than a dirty_writeback_interval interval, then leave a
684  * one-second gap.
685  *
686  * older_than_this takes precedence over nr_to_write.  So we'll only write back
687  * all dirty pages if they are all attached to "old" mappings.
688  */
689 static void wb_kupdate(unsigned long arg)
690 {
691         unsigned long oldest_jif;
692         unsigned long start_jif;
693         unsigned long next_jif;
694         long nr_to_write;
695         struct writeback_control wbc = {
696                 .bdi            = NULL,
697                 .sync_mode      = WB_SYNC_NONE,
698                 .older_than_this = &oldest_jif,
699                 .nr_to_write    = 0,
700                 .nonblocking    = 1,
701                 .for_kupdate    = 1,
702                 .range_cyclic   = 1,
703         };
704
705         sync_supers();
706
707         oldest_jif = jiffies - dirty_expire_interval;
708         start_jif = jiffies;
709         next_jif = start_jif + dirty_writeback_interval;
710         nr_to_write = global_page_state(NR_FILE_DIRTY) +
711                         global_page_state(NR_UNSTABLE_NFS) +
712                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
713         while (nr_to_write > 0) {
714                 wbc.more_io = 0;
715                 wbc.encountered_congestion = 0;
716                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
717                 writeback_inodes(&wbc);
718                 if (wbc.nr_to_write > 0) {
719                         if (wbc.encountered_congestion || wbc.more_io)
720                                 congestion_wait(WRITE, HZ/10);
721                         else
722                                 break;  /* All the old data is written */
723                 }
724                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
725         }
726         if (time_before(next_jif, jiffies + HZ))
727                 next_jif = jiffies + HZ;
728         if (dirty_writeback_interval)
729                 mod_timer(&wb_timer, next_jif);
730 }
731
732 /*
733  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
734  */
735 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
736         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
737 {
738         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
739         if (dirty_writeback_interval)
740                 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
741         else
742                 del_timer(&wb_timer);
743         return 0;
744 }
745
746 static void wb_timer_fn(unsigned long unused)
747 {
748         if (pdflush_operation(wb_kupdate, 0) < 0)
749                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
750 }
751
752 static void laptop_flush(unsigned long unused)
753 {
754         sys_sync();
755 }
756
757 static void laptop_timer_fn(unsigned long unused)
758 {
759         pdflush_operation(laptop_flush, 0);
760 }
761
762 /*
763  * We've spun up the disk and we're in laptop mode: schedule writeback
764  * of all dirty data a few seconds from now.  If the flush is already scheduled
765  * then push it back - the user is still using the disk.
766  */
767 void laptop_io_completion(void)
768 {
769         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
770 }
771
772 /*
773  * We're in laptop mode and we've just synced. The sync's writes will have
774  * caused another writeback to be scheduled by laptop_io_completion.
775  * Nothing needs to be written back anymore, so we unschedule the writeback.
776  */
777 void laptop_sync_completion(void)
778 {
779         del_timer(&laptop_mode_wb_timer);
780 }
781
782 /*
783  * If ratelimit_pages is too high then we can get into dirty-data overload
784  * if a large number of processes all perform writes at the same time.
785  * If it is too low then SMP machines will call the (expensive)
786  * get_writeback_state too often.
787  *
788  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
789  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
790  * thresholds before writeback cuts in.
791  *
792  * But the limit should not be set too high.  Because it also controls the
793  * amount of memory which the balance_dirty_pages() caller has to write back.
794  * If this is too large then the caller will block on the IO queue all the
795  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
796  * will write six megabyte chunks, max.
797  */
798
799 void writeback_set_ratelimit(void)
800 {
801         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
802         if (ratelimit_pages < 16)
803                 ratelimit_pages = 16;
804         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
805                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
806 }
807
808 static int __cpuinit
809 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
810 {
811         writeback_set_ratelimit();
812         return NOTIFY_DONE;
813 }
814
815 static struct notifier_block __cpuinitdata ratelimit_nb = {
816         .notifier_call  = ratelimit_handler,
817         .next           = NULL,
818 };
819
820 /*
821  * Called early on to tune the page writeback dirty limits.
822  *
823  * We used to scale dirty pages according to how total memory
824  * related to pages that could be allocated for buffers (by
825  * comparing nr_free_buffer_pages() to vm_total_pages.
826  *
827  * However, that was when we used "dirty_ratio" to scale with
828  * all memory, and we don't do that any more. "dirty_ratio"
829  * is now applied to total non-HIGHPAGE memory (by subtracting
830  * totalhigh_pages from vm_total_pages), and as such we can't
831  * get into the old insane situation any more where we had
832  * large amounts of dirty pages compared to a small amount of
833  * non-HIGHMEM memory.
834  *
835  * But we might still want to scale the dirty_ratio by how
836  * much memory the box has..
837  */
838 void __init page_writeback_init(void)
839 {
840         int shift;
841
842         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
843         writeback_set_ratelimit();
844         register_cpu_notifier(&ratelimit_nb);
845
846         shift = calc_period_shift();
847         prop_descriptor_init(&vm_completions, shift);
848         prop_descriptor_init(&vm_dirties, shift);
849 }
850
851 /**
852  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
853  * @mapping: address space structure to write
854  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
855  * @writepage: function called for each page
856  * @data: data passed to writepage function
857  *
858  * If a page is already under I/O, write_cache_pages() skips it, even
859  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
860  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
861  * and msync() need to guarantee that all the data which was dirty at the time
862  * the call was made get new I/O started against them.  If wbc->sync_mode is
863  * WB_SYNC_ALL then we were called for data integrity and we must wait for
864  * existing IO to complete.
865  */
866 int write_cache_pages(struct address_space *mapping,
867                       struct writeback_control *wbc, writepage_t writepage,
868                       void *data)
869 {
870         struct backing_dev_info *bdi = mapping->backing_dev_info;
871         int ret = 0;
872         int done = 0;
873         struct pagevec pvec;
874         int nr_pages;
875         pgoff_t index;
876         pgoff_t end;            /* Inclusive */
877         int scanned = 0;
878         int range_whole = 0;
879
880         if (wbc->nonblocking && bdi_write_congested(bdi)) {
881                 wbc->encountered_congestion = 1;
882                 return 0;
883         }
884
885         pagevec_init(&pvec, 0);
886         if (wbc->range_cyclic) {
887                 index = mapping->writeback_index; /* Start from prev offset */
888                 end = -1;
889         } else {
890                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
891                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
892                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
893                         range_whole = 1;
894                 scanned = 1;
895         }
896 retry:
897         while (!done && (index <= end) &&
898                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
899                                               PAGECACHE_TAG_DIRTY,
900                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
901                 unsigned i;
902
903                 scanned = 1;
904                 for (i = 0; i < nr_pages; i++) {
905                         struct page *page = pvec.pages[i];
906
907                         /*
908                          * At this point we hold neither mapping->tree_lock nor
909                          * lock on the page itself: the page may be truncated or
910                          * invalidated (changing page->mapping to NULL), or even
911                          * swizzled back from swapper_space to tmpfs file
912                          * mapping
913                          */
914                         lock_page(page);
915
916                         if (unlikely(page->mapping != mapping)) {
917                                 unlock_page(page);
918                                 continue;
919                         }
920
921                         if (!wbc->range_cyclic && page->index > end) {
922                                 done = 1;
923                                 unlock_page(page);
924                                 continue;
925                         }
926
927                         if (wbc->sync_mode != WB_SYNC_NONE)
928                                 wait_on_page_writeback(page);
929
930                         if (PageWriteback(page) ||
931                             !clear_page_dirty_for_io(page)) {
932                                 unlock_page(page);
933                                 continue;
934                         }
935
936                         ret = (*writepage)(page, wbc, data);
937
938                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
939                                 unlock_page(page);
940                                 ret = 0;
941                         }
942                         if (ret || (--(wbc->nr_to_write) <= 0))
943                                 done = 1;
944                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
945                                 wbc->encountered_congestion = 1;
946                                 done = 1;
947                         }
948                 }
949                 pagevec_release(&pvec);
950                 cond_resched();
951         }
952         if (!scanned && !done) {
953                 /*
954                  * We hit the last page and there is more work to be done: wrap
955                  * back to the start of the file
956                  */
957                 scanned = 1;
958                 index = 0;
959                 goto retry;
960         }
961         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
962                 mapping->writeback_index = index;
963
964         if (wbc->range_cont)
965                 wbc->range_start = index << PAGE_CACHE_SHIFT;
966         return ret;
967 }
968 EXPORT_SYMBOL(write_cache_pages);
969
970 /*
971  * Function used by generic_writepages to call the real writepage
972  * function and set the mapping flags on error
973  */
974 static int __writepage(struct page *page, struct writeback_control *wbc,
975                        void *data)
976 {
977         struct address_space *mapping = data;
978         int ret = mapping->a_ops->writepage(page, wbc);
979         mapping_set_error(mapping, ret);
980         return ret;
981 }
982
983 /**
984  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
985  * @mapping: address space structure to write
986  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
987  *
988  * This is a library function, which implements the writepages()
989  * address_space_operation.
990  */
991 int generic_writepages(struct address_space *mapping,
992                        struct writeback_control *wbc)
993 {
994         /* deal with chardevs and other special file */
995         if (!mapping->a_ops->writepage)
996                 return 0;
997
998         return write_cache_pages(mapping, wbc, __writepage, mapping);
999 }
1000
1001 EXPORT_SYMBOL(generic_writepages);
1002
1003 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1004 {
1005         int ret;
1006
1007         if (wbc->nr_to_write <= 0)
1008                 return 0;
1009         wbc->for_writepages = 1;
1010         if (mapping->a_ops->writepages)
1011                 ret = mapping->a_ops->writepages(mapping, wbc);
1012         else
1013                 ret = generic_writepages(mapping, wbc);
1014         wbc->for_writepages = 0;
1015         return ret;
1016 }
1017
1018 /**
1019  * write_one_page - write out a single page and optionally wait on I/O
1020  * @page: the page to write
1021  * @wait: if true, wait on writeout
1022  *
1023  * The page must be locked by the caller and will be unlocked upon return.
1024  *
1025  * write_one_page() returns a negative error code if I/O failed.
1026  */
1027 int write_one_page(struct page *page, int wait)
1028 {
1029         struct address_space *mapping = page->mapping;
1030         int ret = 0;
1031         struct writeback_control wbc = {
1032                 .sync_mode = WB_SYNC_ALL,
1033                 .nr_to_write = 1,
1034         };
1035
1036         BUG_ON(!PageLocked(page));
1037
1038         if (wait)
1039                 wait_on_page_writeback(page);
1040
1041         if (clear_page_dirty_for_io(page)) {
1042                 page_cache_get(page);
1043                 ret = mapping->a_ops->writepage(page, &wbc);
1044                 if (ret == 0 && wait) {
1045                         wait_on_page_writeback(page);
1046                         if (PageError(page))
1047                                 ret = -EIO;
1048                 }
1049                 page_cache_release(page);
1050         } else {
1051                 unlock_page(page);
1052         }
1053         return ret;
1054 }
1055 EXPORT_SYMBOL(write_one_page);
1056
1057 /*
1058  * For address_spaces which do not use buffers nor write back.
1059  */
1060 int __set_page_dirty_no_writeback(struct page *page)
1061 {
1062         if (!PageDirty(page))
1063                 SetPageDirty(page);
1064         return 0;
1065 }
1066
1067 /*
1068  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1069  * its radix tree.
1070  *
1071  * This is also used when a single buffer is being dirtied: we want to set the
1072  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1073  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1074  *
1075  * Most callers have locked the page, which pins the address_space in memory.
1076  * But zap_pte_range() does not lock the page, however in that case the
1077  * mapping is pinned by the vma's ->vm_file reference.
1078  *
1079  * We take care to handle the case where the page was truncated from the
1080  * mapping by re-checking page_mapping() inside tree_lock.
1081  */
1082 int __set_page_dirty_nobuffers(struct page *page)
1083 {
1084         if (!TestSetPageDirty(page)) {
1085                 struct address_space *mapping = page_mapping(page);
1086                 struct address_space *mapping2;
1087
1088                 if (!mapping)
1089                         return 1;
1090
1091                 spin_lock_irq(&mapping->tree_lock);
1092                 mapping2 = page_mapping(page);
1093                 if (mapping2) { /* Race with truncate? */
1094                         BUG_ON(mapping2 != mapping);
1095                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1096                         if (mapping_cap_account_dirty(mapping)) {
1097                                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1098                                 __inc_bdi_stat(mapping->backing_dev_info,
1099                                                 BDI_RECLAIMABLE);
1100                                 task_io_account_write(PAGE_CACHE_SIZE);
1101                         }
1102                         radix_tree_tag_set(&mapping->page_tree,
1103                                 page_index(page), PAGECACHE_TAG_DIRTY);
1104                 }
1105                 spin_unlock_irq(&mapping->tree_lock);
1106                 if (mapping->host) {
1107                         /* !PageAnon && !swapper_space */
1108                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1109                 }
1110                 return 1;
1111         }
1112         return 0;
1113 }
1114 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1115
1116 /*
1117  * When a writepage implementation decides that it doesn't want to write this
1118  * page for some reason, it should redirty the locked page via
1119  * redirty_page_for_writepage() and it should then unlock the page and return 0
1120  */
1121 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1122 {
1123         wbc->pages_skipped++;
1124         return __set_page_dirty_nobuffers(page);
1125 }
1126 EXPORT_SYMBOL(redirty_page_for_writepage);
1127
1128 /*
1129  * If the mapping doesn't provide a set_page_dirty a_op, then
1130  * just fall through and assume that it wants buffer_heads.
1131  */
1132 static int __set_page_dirty(struct page *page)
1133 {
1134         struct address_space *mapping = page_mapping(page);
1135
1136         if (likely(mapping)) {
1137                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1138 #ifdef CONFIG_BLOCK
1139                 if (!spd)
1140                         spd = __set_page_dirty_buffers;
1141 #endif
1142                 return (*spd)(page);
1143         }
1144         if (!PageDirty(page)) {
1145                 if (!TestSetPageDirty(page))
1146                         return 1;
1147         }
1148         return 0;
1149 }
1150
1151 int set_page_dirty(struct page *page)
1152 {
1153         int ret = __set_page_dirty(page);
1154         if (ret)
1155                 task_dirty_inc(current);
1156         return ret;
1157 }
1158 EXPORT_SYMBOL(set_page_dirty);
1159
1160 /*
1161  * set_page_dirty() is racy if the caller has no reference against
1162  * page->mapping->host, and if the page is unlocked.  This is because another
1163  * CPU could truncate the page off the mapping and then free the mapping.
1164  *
1165  * Usually, the page _is_ locked, or the caller is a user-space process which
1166  * holds a reference on the inode by having an open file.
1167  *
1168  * In other cases, the page should be locked before running set_page_dirty().
1169  */
1170 int set_page_dirty_lock(struct page *page)
1171 {
1172         int ret;
1173
1174         lock_page_nosync(page);
1175         ret = set_page_dirty(page);
1176         unlock_page(page);
1177         return ret;
1178 }
1179 EXPORT_SYMBOL(set_page_dirty_lock);
1180
1181 /*
1182  * Clear a page's dirty flag, while caring for dirty memory accounting.
1183  * Returns true if the page was previously dirty.
1184  *
1185  * This is for preparing to put the page under writeout.  We leave the page
1186  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1187  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1188  * implementation will run either set_page_writeback() or set_page_dirty(),
1189  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1190  * back into sync.
1191  *
1192  * This incoherency between the page's dirty flag and radix-tree tag is
1193  * unfortunate, but it only exists while the page is locked.
1194  */
1195 int clear_page_dirty_for_io(struct page *page)
1196 {
1197         struct address_space *mapping = page_mapping(page);
1198
1199         BUG_ON(!PageLocked(page));
1200
1201         ClearPageReclaim(page);
1202         if (mapping && mapping_cap_account_dirty(mapping)) {
1203                 /*
1204                  * Yes, Virginia, this is indeed insane.
1205                  *
1206                  * We use this sequence to make sure that
1207                  *  (a) we account for dirty stats properly
1208                  *  (b) we tell the low-level filesystem to
1209                  *      mark the whole page dirty if it was
1210                  *      dirty in a pagetable. Only to then
1211                  *  (c) clean the page again and return 1 to
1212                  *      cause the writeback.
1213                  *
1214                  * This way we avoid all nasty races with the
1215                  * dirty bit in multiple places and clearing
1216                  * them concurrently from different threads.
1217                  *
1218                  * Note! Normally the "set_page_dirty(page)"
1219                  * has no effect on the actual dirty bit - since
1220                  * that will already usually be set. But we
1221                  * need the side effects, and it can help us
1222                  * avoid races.
1223                  *
1224                  * We basically use the page "master dirty bit"
1225                  * as a serialization point for all the different
1226                  * threads doing their things.
1227                  */
1228                 if (page_mkclean(page))
1229                         set_page_dirty(page);
1230                 /*
1231                  * We carefully synchronise fault handlers against
1232                  * installing a dirty pte and marking the page dirty
1233                  * at this point. We do this by having them hold the
1234                  * page lock at some point after installing their
1235                  * pte, but before marking the page dirty.
1236                  * Pages are always locked coming in here, so we get
1237                  * the desired exclusion. See mm/memory.c:do_wp_page()
1238                  * for more comments.
1239                  */
1240                 if (TestClearPageDirty(page)) {
1241                         dec_zone_page_state(page, NR_FILE_DIRTY);
1242                         dec_bdi_stat(mapping->backing_dev_info,
1243                                         BDI_RECLAIMABLE);
1244                         return 1;
1245                 }
1246                 return 0;
1247         }
1248         return TestClearPageDirty(page);
1249 }
1250 EXPORT_SYMBOL(clear_page_dirty_for_io);
1251
1252 int test_clear_page_writeback(struct page *page)
1253 {
1254         struct address_space *mapping = page_mapping(page);
1255         int ret;
1256
1257         if (mapping) {
1258                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1259                 unsigned long flags;
1260
1261                 spin_lock_irqsave(&mapping->tree_lock, flags);
1262                 ret = TestClearPageWriteback(page);
1263                 if (ret) {
1264                         radix_tree_tag_clear(&mapping->page_tree,
1265                                                 page_index(page),
1266                                                 PAGECACHE_TAG_WRITEBACK);
1267                         if (bdi_cap_account_writeback(bdi)) {
1268                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1269                                 __bdi_writeout_inc(bdi);
1270                         }
1271                 }
1272                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1273         } else {
1274                 ret = TestClearPageWriteback(page);
1275         }
1276         if (ret)
1277                 dec_zone_page_state(page, NR_WRITEBACK);
1278         return ret;
1279 }
1280
1281 int test_set_page_writeback(struct page *page)
1282 {
1283         struct address_space *mapping = page_mapping(page);
1284         int ret;
1285
1286         if (mapping) {
1287                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1288                 unsigned long flags;
1289
1290                 spin_lock_irqsave(&mapping->tree_lock, flags);
1291                 ret = TestSetPageWriteback(page);
1292                 if (!ret) {
1293                         radix_tree_tag_set(&mapping->page_tree,
1294                                                 page_index(page),
1295                                                 PAGECACHE_TAG_WRITEBACK);
1296                         if (bdi_cap_account_writeback(bdi))
1297                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1298                 }
1299                 if (!PageDirty(page))
1300                         radix_tree_tag_clear(&mapping->page_tree,
1301                                                 page_index(page),
1302                                                 PAGECACHE_TAG_DIRTY);
1303                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1304         } else {
1305                 ret = TestSetPageWriteback(page);
1306         }
1307         if (!ret)
1308                 inc_zone_page_state(page, NR_WRITEBACK);
1309         return ret;
1310
1311 }
1312 EXPORT_SYMBOL(test_set_page_writeback);
1313
1314 /*
1315  * Return true if any of the pages in the mapping are marked with the
1316  * passed tag.
1317  */
1318 int mapping_tagged(struct address_space *mapping, int tag)
1319 {
1320         int ret;
1321         rcu_read_lock();
1322         ret = radix_tree_tagged(&mapping->page_tree, tag);
1323         rcu_read_unlock();
1324         return ret;
1325 }
1326 EXPORT_SYMBOL(mapping_tagged);