writeback: use RCU to protect bdi_list
[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    Andrew Morton
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  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
40  * will look to see if it needs to force writeback or throttling.
41  */
42 static long ratelimit_pages = 32;
43
44 /*
45  * When balance_dirty_pages decides that the caller needs to perform some
46  * non-background writeback, this is how many pages it will attempt to write.
47  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
48  * large amounts of I/O are submitted.
49  */
50 static inline long sync_writeback_pages(void)
51 {
52         return ratelimit_pages + ratelimit_pages / 2;
53 }
54
55 /* The following parameters are exported via /proc/sys/vm */
56
57 /*
58  * Start background writeback (via pdflush) at this percentage
59  */
60 int dirty_background_ratio = 10;
61
62 /*
63  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
64  * dirty_background_ratio * the amount of dirtyable memory
65  */
66 unsigned long dirty_background_bytes;
67
68 /*
69  * free highmem will not be subtracted from the total free memory
70  * for calculating free ratios if vm_highmem_is_dirtyable is true
71  */
72 int vm_highmem_is_dirtyable;
73
74 /*
75  * The generator of dirty data starts writeback at this percentage
76  */
77 int vm_dirty_ratio = 20;
78
79 /*
80  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
81  * vm_dirty_ratio * the amount of dirtyable memory
82  */
83 unsigned long vm_dirty_bytes;
84
85 /*
86  * The interval between `kupdate'-style writebacks
87  */
88 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
89
90 /*
91  * The longest time for which data is allowed to remain dirty
92  */
93 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
94
95 /*
96  * Flag that makes the machine dump writes/reads and block dirtyings.
97  */
98 int block_dump;
99
100 /*
101  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
102  * a full sync is triggered after this time elapses without any disk activity.
103  */
104 int laptop_mode;
105
106 EXPORT_SYMBOL(laptop_mode);
107
108 /* End of sysctl-exported parameters */
109
110
111 /*
112  * Scale the writeback cache size proportional to the relative writeout speeds.
113  *
114  * We do this by keeping a floating proportion between BDIs, based on page
115  * writeback completions [end_page_writeback()]. Those devices that write out
116  * pages fastest will get the larger share, while the slower will get a smaller
117  * share.
118  *
119  * We use page writeout completions because we are interested in getting rid of
120  * dirty pages. Having them written out is the primary goal.
121  *
122  * We introduce a concept of time, a period over which we measure these events,
123  * because demand can/will vary over time. The length of this period itself is
124  * measured in page writeback completions.
125  *
126  */
127 static struct prop_descriptor vm_completions;
128 static struct prop_descriptor vm_dirties;
129
130 /*
131  * couple the period to the dirty_ratio:
132  *
133  *   period/2 ~ roundup_pow_of_two(dirty limit)
134  */
135 static int calc_period_shift(void)
136 {
137         unsigned long dirty_total;
138
139         if (vm_dirty_bytes)
140                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
141         else
142                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
143                                 100;
144         return 2 + ilog2(dirty_total - 1);
145 }
146
147 /*
148  * update the period when the dirty threshold changes.
149  */
150 static void update_completion_period(void)
151 {
152         int shift = calc_period_shift();
153         prop_change_shift(&vm_completions, shift);
154         prop_change_shift(&vm_dirties, shift);
155 }
156
157 int dirty_background_ratio_handler(struct ctl_table *table, int write,
158                 struct file *filp, void __user *buffer, size_t *lenp,
159                 loff_t *ppos)
160 {
161         int ret;
162
163         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
164         if (ret == 0 && write)
165                 dirty_background_bytes = 0;
166         return ret;
167 }
168
169 int dirty_background_bytes_handler(struct ctl_table *table, int write,
170                 struct file *filp, void __user *buffer, size_t *lenp,
171                 loff_t *ppos)
172 {
173         int ret;
174
175         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
176         if (ret == 0 && write)
177                 dirty_background_ratio = 0;
178         return ret;
179 }
180
181 int dirty_ratio_handler(struct ctl_table *table, int write,
182                 struct file *filp, void __user *buffer, size_t *lenp,
183                 loff_t *ppos)
184 {
185         int old_ratio = vm_dirty_ratio;
186         int ret;
187
188         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
189         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
190                 update_completion_period();
191                 vm_dirty_bytes = 0;
192         }
193         return ret;
194 }
195
196
197 int dirty_bytes_handler(struct ctl_table *table, int write,
198                 struct file *filp, void __user *buffer, size_t *lenp,
199                 loff_t *ppos)
200 {
201         unsigned long old_bytes = vm_dirty_bytes;
202         int ret;
203
204         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
205         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
206                 update_completion_period();
207                 vm_dirty_ratio = 0;
208         }
209         return ret;
210 }
211
212 /*
213  * Increment the BDI's writeout completion count and the global writeout
214  * completion count. Called from test_clear_page_writeback().
215  */
216 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
217 {
218         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
219                               bdi->max_prop_frac);
220 }
221
222 void bdi_writeout_inc(struct backing_dev_info *bdi)
223 {
224         unsigned long flags;
225
226         local_irq_save(flags);
227         __bdi_writeout_inc(bdi);
228         local_irq_restore(flags);
229 }
230 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
231
232 void task_dirty_inc(struct task_struct *tsk)
233 {
234         prop_inc_single(&vm_dirties, &tsk->dirties);
235 }
236
237 /*
238  * Obtain an accurate fraction of the BDI's portion.
239  */
240 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
241                 long *numerator, long *denominator)
242 {
243         if (bdi_cap_writeback_dirty(bdi)) {
244                 prop_fraction_percpu(&vm_completions, &bdi->completions,
245                                 numerator, denominator);
246         } else {
247                 *numerator = 0;
248                 *denominator = 1;
249         }
250 }
251
252 /*
253  * Clip the earned share of dirty pages to that which is actually available.
254  * This avoids exceeding the total dirty_limit when the floating averages
255  * fluctuate too quickly.
256  */
257 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
258                 unsigned long dirty, unsigned long *pbdi_dirty)
259 {
260         unsigned long avail_dirty;
261
262         avail_dirty = global_page_state(NR_FILE_DIRTY) +
263                  global_page_state(NR_WRITEBACK) +
264                  global_page_state(NR_UNSTABLE_NFS) +
265                  global_page_state(NR_WRITEBACK_TEMP);
266
267         if (avail_dirty < dirty)
268                 avail_dirty = dirty - avail_dirty;
269         else
270                 avail_dirty = 0;
271
272         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
273                 bdi_stat(bdi, BDI_WRITEBACK);
274
275         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
276 }
277
278 static inline void task_dirties_fraction(struct task_struct *tsk,
279                 long *numerator, long *denominator)
280 {
281         prop_fraction_single(&vm_dirties, &tsk->dirties,
282                                 numerator, denominator);
283 }
284
285 /*
286  * scale the dirty limit
287  *
288  * task specific dirty limit:
289  *
290  *   dirty -= (dirty/8) * p_{t}
291  */
292 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
293 {
294         long numerator, denominator;
295         unsigned long dirty = *pdirty;
296         u64 inv = dirty >> 3;
297
298         task_dirties_fraction(tsk, &numerator, &denominator);
299         inv *= numerator;
300         do_div(inv, denominator);
301
302         dirty -= inv;
303         if (dirty < *pdirty/2)
304                 dirty = *pdirty/2;
305
306         *pdirty = dirty;
307 }
308
309 /*
310  *
311  */
312 static unsigned int bdi_min_ratio;
313
314 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315 {
316         int ret = 0;
317
318         spin_lock_bh(&bdi_lock);
319         if (min_ratio > bdi->max_ratio) {
320                 ret = -EINVAL;
321         } else {
322                 min_ratio -= bdi->min_ratio;
323                 if (bdi_min_ratio + min_ratio < 100) {
324                         bdi_min_ratio += min_ratio;
325                         bdi->min_ratio += min_ratio;
326                 } else {
327                         ret = -EINVAL;
328                 }
329         }
330         spin_unlock_bh(&bdi_lock);
331
332         return ret;
333 }
334
335 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
336 {
337         int ret = 0;
338
339         if (max_ratio > 100)
340                 return -EINVAL;
341
342         spin_lock_bh(&bdi_lock);
343         if (bdi->min_ratio > max_ratio) {
344                 ret = -EINVAL;
345         } else {
346                 bdi->max_ratio = max_ratio;
347                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348         }
349         spin_unlock_bh(&bdi_lock);
350
351         return ret;
352 }
353 EXPORT_SYMBOL(bdi_set_max_ratio);
354
355 /*
356  * Work out the current dirty-memory clamping and background writeout
357  * thresholds.
358  *
359  * The main aim here is to lower them aggressively if there is a lot of mapped
360  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
361  * pages.  It is better to clamp down on writers than to start swapping, and
362  * performing lots of scanning.
363  *
364  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
365  *
366  * We don't permit the clamping level to fall below 5% - that is getting rather
367  * excessive.
368  *
369  * We make sure that the background writeout level is below the adjusted
370  * clamping level.
371  */
372
373 static unsigned long highmem_dirtyable_memory(unsigned long total)
374 {
375 #ifdef CONFIG_HIGHMEM
376         int node;
377         unsigned long x = 0;
378
379         for_each_node_state(node, N_HIGH_MEMORY) {
380                 struct zone *z =
381                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
382
383                 x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
384         }
385         /*
386          * Make sure that the number of highmem pages is never larger
387          * than the number of the total dirtyable memory. This can only
388          * occur in very strange VM situations but we want to make sure
389          * that this does not occur.
390          */
391         return min(x, total);
392 #else
393         return 0;
394 #endif
395 }
396
397 /**
398  * determine_dirtyable_memory - amount of memory that may be used
399  *
400  * Returns the numebr of pages that can currently be freed and used
401  * by the kernel for direct mappings.
402  */
403 unsigned long determine_dirtyable_memory(void)
404 {
405         unsigned long x;
406
407         x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
408
409         if (!vm_highmem_is_dirtyable)
410                 x -= highmem_dirtyable_memory(x);
411
412         return x + 1;   /* Ensure that we never return 0 */
413 }
414
415 void
416 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
417                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
418 {
419         unsigned long background;
420         unsigned long dirty;
421         unsigned long available_memory = determine_dirtyable_memory();
422         struct task_struct *tsk;
423
424         if (vm_dirty_bytes)
425                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
426         else {
427                 int dirty_ratio;
428
429                 dirty_ratio = vm_dirty_ratio;
430                 if (dirty_ratio < 5)
431                         dirty_ratio = 5;
432                 dirty = (dirty_ratio * available_memory) / 100;
433         }
434
435         if (dirty_background_bytes)
436                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
437         else
438                 background = (dirty_background_ratio * available_memory) / 100;
439
440         if (background >= dirty)
441                 background = dirty / 2;
442         tsk = current;
443         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
444                 background += background / 4;
445                 dirty += dirty / 4;
446         }
447         *pbackground = background;
448         *pdirty = dirty;
449
450         if (bdi) {
451                 u64 bdi_dirty;
452                 long numerator, denominator;
453
454                 /*
455                  * Calculate this BDI's share of the dirty ratio.
456                  */
457                 bdi_writeout_fraction(bdi, &numerator, &denominator);
458
459                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
460                 bdi_dirty *= numerator;
461                 do_div(bdi_dirty, denominator);
462                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
463                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
464                         bdi_dirty = dirty * bdi->max_ratio / 100;
465
466                 *pbdi_dirty = bdi_dirty;
467                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
468                 task_dirty_limit(current, pbdi_dirty);
469         }
470 }
471
472 /*
473  * balance_dirty_pages() must be called by processes which are generating dirty
474  * data.  It looks at the number of dirty pages in the machine and will force
475  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
476  * If we're over `background_thresh' then pdflush is woken to perform some
477  * writeout.
478  */
479 static void balance_dirty_pages(struct address_space *mapping)
480 {
481         long nr_reclaimable, bdi_nr_reclaimable;
482         long nr_writeback, bdi_nr_writeback;
483         unsigned long background_thresh;
484         unsigned long dirty_thresh;
485         unsigned long bdi_thresh;
486         unsigned long pages_written = 0;
487         unsigned long write_chunk = sync_writeback_pages();
488
489         struct backing_dev_info *bdi = mapping->backing_dev_info;
490
491         for (;;) {
492                 struct writeback_control wbc = {
493                         .bdi            = bdi,
494                         .sync_mode      = WB_SYNC_NONE,
495                         .older_than_this = NULL,
496                         .nr_to_write    = write_chunk,
497                         .range_cyclic   = 1,
498                 };
499
500                 get_dirty_limits(&background_thresh, &dirty_thresh,
501                                 &bdi_thresh, bdi);
502
503                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
504                                         global_page_state(NR_UNSTABLE_NFS);
505                 nr_writeback = global_page_state(NR_WRITEBACK);
506
507                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
508                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
509
510                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
511                         break;
512
513                 /*
514                  * Throttle it only when the background writeback cannot
515                  * catch-up. This avoids (excessively) small writeouts
516                  * when the bdi limits are ramping up.
517                  */
518                 if (nr_reclaimable + nr_writeback <
519                                 (background_thresh + dirty_thresh) / 2)
520                         break;
521
522                 if (!bdi->dirty_exceeded)
523                         bdi->dirty_exceeded = 1;
524
525                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
526                  * Unstable writes are a feature of certain networked
527                  * filesystems (i.e. NFS) in which data may have been
528                  * written to the server's write cache, but has not yet
529                  * been flushed to permanent storage.
530                  * Only move pages to writeback if this bdi is over its
531                  * threshold otherwise wait until the disk writes catch
532                  * up.
533                  */
534                 if (bdi_nr_reclaimable > bdi_thresh) {
535                         writeback_inodes_wbc(&wbc);
536                         pages_written += write_chunk - wbc.nr_to_write;
537                         get_dirty_limits(&background_thresh, &dirty_thresh,
538                                        &bdi_thresh, bdi);
539                 }
540
541                 /*
542                  * In order to avoid the stacked BDI deadlock we need
543                  * to ensure we accurately count the 'dirty' pages when
544                  * the threshold is low.
545                  *
546                  * Otherwise it would be possible to get thresh+n pages
547                  * reported dirty, even though there are thresh-m pages
548                  * actually dirty; with m+n sitting in the percpu
549                  * deltas.
550                  */
551                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
552                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
553                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
554                 } else if (bdi_nr_reclaimable) {
555                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
556                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
557                 }
558
559                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
560                         break;
561                 if (pages_written >= write_chunk)
562                         break;          /* We've done our duty */
563
564                 schedule_timeout(1);
565         }
566
567         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
568                         bdi->dirty_exceeded)
569                 bdi->dirty_exceeded = 0;
570
571         if (writeback_in_progress(bdi))
572                 return;         /* pdflush is already working this queue */
573
574         /*
575          * In laptop mode, we wait until hitting the higher threshold before
576          * starting background writeout, and then write out all the way down
577          * to the lower threshold.  So slow writers cause minimal disk activity.
578          *
579          * In normal mode, we start background writeout at the lower
580          * background_thresh, to keep the amount of dirty memory low.
581          */
582         if ((laptop_mode && pages_written) ||
583             (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY)
584                                           + global_page_state(NR_UNSTABLE_NFS))
585                                           > background_thresh))) {
586                 struct writeback_control wbc = {
587                         .bdi            = bdi,
588                         .sync_mode      = WB_SYNC_NONE,
589                         .nr_to_write    = nr_writeback,
590                 };
591
592
593                 bdi_start_writeback(&wbc);
594         }
595 }
596
597 void set_page_dirty_balance(struct page *page, int page_mkwrite)
598 {
599         if (set_page_dirty(page) || page_mkwrite) {
600                 struct address_space *mapping = page_mapping(page);
601
602                 if (mapping)
603                         balance_dirty_pages_ratelimited(mapping);
604         }
605 }
606
607 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
608
609 /**
610  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
611  * @mapping: address_space which was dirtied
612  * @nr_pages_dirtied: number of pages which the caller has just dirtied
613  *
614  * Processes which are dirtying memory should call in here once for each page
615  * which was newly dirtied.  The function will periodically check the system's
616  * dirty state and will initiate writeback if needed.
617  *
618  * On really big machines, get_writeback_state is expensive, so try to avoid
619  * calling it too often (ratelimiting).  But once we're over the dirty memory
620  * limit we decrease the ratelimiting by a lot, to prevent individual processes
621  * from overshooting the limit by (ratelimit_pages) each.
622  */
623 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
624                                         unsigned long nr_pages_dirtied)
625 {
626         unsigned long ratelimit;
627         unsigned long *p;
628
629         ratelimit = ratelimit_pages;
630         if (mapping->backing_dev_info->dirty_exceeded)
631                 ratelimit = 8;
632
633         /*
634          * Check the rate limiting. Also, we do not want to throttle real-time
635          * tasks in balance_dirty_pages(). Period.
636          */
637         preempt_disable();
638         p =  &__get_cpu_var(bdp_ratelimits);
639         *p += nr_pages_dirtied;
640         if (unlikely(*p >= ratelimit)) {
641                 *p = 0;
642                 preempt_enable();
643                 balance_dirty_pages(mapping);
644                 return;
645         }
646         preempt_enable();
647 }
648 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
649
650 void throttle_vm_writeout(gfp_t gfp_mask)
651 {
652         unsigned long background_thresh;
653         unsigned long dirty_thresh;
654
655         for ( ; ; ) {
656                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
657
658                 /*
659                  * Boost the allowable dirty threshold a bit for page
660                  * allocators so they don't get DoS'ed by heavy writers
661                  */
662                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
663
664                 if (global_page_state(NR_UNSTABLE_NFS) +
665                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
666                                 break;
667                 congestion_wait(BLK_RW_ASYNC, HZ/10);
668
669                 /*
670                  * The caller might hold locks which can prevent IO completion
671                  * or progress in the filesystem.  So we cannot just sit here
672                  * waiting for IO to complete.
673                  */
674                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
675                         break;
676         }
677 }
678
679 static void laptop_timer_fn(unsigned long unused);
680
681 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
682
683 /*
684  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
685  */
686 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
687         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
688 {
689         proc_dointvec(table, write, file, buffer, length, ppos);
690         return 0;
691 }
692
693 static void do_laptop_sync(struct work_struct *work)
694 {
695         wakeup_flusher_threads(0);
696         kfree(work);
697 }
698
699 static void laptop_timer_fn(unsigned long unused)
700 {
701         struct work_struct *work;
702
703         work = kmalloc(sizeof(*work), GFP_ATOMIC);
704         if (work) {
705                 INIT_WORK(work, do_laptop_sync);
706                 schedule_work(work);
707         }
708 }
709
710 /*
711  * We've spun up the disk and we're in laptop mode: schedule writeback
712  * of all dirty data a few seconds from now.  If the flush is already scheduled
713  * then push it back - the user is still using the disk.
714  */
715 void laptop_io_completion(void)
716 {
717         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
718 }
719
720 /*
721  * We're in laptop mode and we've just synced. The sync's writes will have
722  * caused another writeback to be scheduled by laptop_io_completion.
723  * Nothing needs to be written back anymore, so we unschedule the writeback.
724  */
725 void laptop_sync_completion(void)
726 {
727         del_timer(&laptop_mode_wb_timer);
728 }
729
730 /*
731  * If ratelimit_pages is too high then we can get into dirty-data overload
732  * if a large number of processes all perform writes at the same time.
733  * If it is too low then SMP machines will call the (expensive)
734  * get_writeback_state too often.
735  *
736  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
737  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
738  * thresholds before writeback cuts in.
739  *
740  * But the limit should not be set too high.  Because it also controls the
741  * amount of memory which the balance_dirty_pages() caller has to write back.
742  * If this is too large then the caller will block on the IO queue all the
743  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
744  * will write six megabyte chunks, max.
745  */
746
747 void writeback_set_ratelimit(void)
748 {
749         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
750         if (ratelimit_pages < 16)
751                 ratelimit_pages = 16;
752         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
753                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
754 }
755
756 static int __cpuinit
757 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
758 {
759         writeback_set_ratelimit();
760         return NOTIFY_DONE;
761 }
762
763 static struct notifier_block __cpuinitdata ratelimit_nb = {
764         .notifier_call  = ratelimit_handler,
765         .next           = NULL,
766 };
767
768 /*
769  * Called early on to tune the page writeback dirty limits.
770  *
771  * We used to scale dirty pages according to how total memory
772  * related to pages that could be allocated for buffers (by
773  * comparing nr_free_buffer_pages() to vm_total_pages.
774  *
775  * However, that was when we used "dirty_ratio" to scale with
776  * all memory, and we don't do that any more. "dirty_ratio"
777  * is now applied to total non-HIGHPAGE memory (by subtracting
778  * totalhigh_pages from vm_total_pages), and as such we can't
779  * get into the old insane situation any more where we had
780  * large amounts of dirty pages compared to a small amount of
781  * non-HIGHMEM memory.
782  *
783  * But we might still want to scale the dirty_ratio by how
784  * much memory the box has..
785  */
786 void __init page_writeback_init(void)
787 {
788         int shift;
789
790         writeback_set_ratelimit();
791         register_cpu_notifier(&ratelimit_nb);
792
793         shift = calc_period_shift();
794         prop_descriptor_init(&vm_completions, shift);
795         prop_descriptor_init(&vm_dirties, shift);
796 }
797
798 /**
799  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
800  * @mapping: address space structure to write
801  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
802  * @writepage: function called for each page
803  * @data: data passed to writepage function
804  *
805  * If a page is already under I/O, write_cache_pages() skips it, even
806  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
807  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
808  * and msync() need to guarantee that all the data which was dirty at the time
809  * the call was made get new I/O started against them.  If wbc->sync_mode is
810  * WB_SYNC_ALL then we were called for data integrity and we must wait for
811  * existing IO to complete.
812  */
813 int write_cache_pages(struct address_space *mapping,
814                       struct writeback_control *wbc, writepage_t writepage,
815                       void *data)
816 {
817         struct backing_dev_info *bdi = mapping->backing_dev_info;
818         int ret = 0;
819         int done = 0;
820         struct pagevec pvec;
821         int nr_pages;
822         pgoff_t uninitialized_var(writeback_index);
823         pgoff_t index;
824         pgoff_t end;            /* Inclusive */
825         pgoff_t done_index;
826         int cycled;
827         int range_whole = 0;
828         long nr_to_write = wbc->nr_to_write;
829
830         if (wbc->nonblocking && bdi_write_congested(bdi)) {
831                 wbc->encountered_congestion = 1;
832                 return 0;
833         }
834
835         pagevec_init(&pvec, 0);
836         if (wbc->range_cyclic) {
837                 writeback_index = mapping->writeback_index; /* prev offset */
838                 index = writeback_index;
839                 if (index == 0)
840                         cycled = 1;
841                 else
842                         cycled = 0;
843                 end = -1;
844         } else {
845                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
846                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
847                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
848                         range_whole = 1;
849                 cycled = 1; /* ignore range_cyclic tests */
850         }
851 retry:
852         done_index = index;
853         while (!done && (index <= end)) {
854                 int i;
855
856                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
857                               PAGECACHE_TAG_DIRTY,
858                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
859                 if (nr_pages == 0)
860                         break;
861
862                 for (i = 0; i < nr_pages; i++) {
863                         struct page *page = pvec.pages[i];
864
865                         /*
866                          * At this point, the page may be truncated or
867                          * invalidated (changing page->mapping to NULL), or
868                          * even swizzled back from swapper_space to tmpfs file
869                          * mapping. However, page->index will not change
870                          * because we have a reference on the page.
871                          */
872                         if (page->index > end) {
873                                 /*
874                                  * can't be range_cyclic (1st pass) because
875                                  * end == -1 in that case.
876                                  */
877                                 done = 1;
878                                 break;
879                         }
880
881                         done_index = page->index + 1;
882
883                         lock_page(page);
884
885                         /*
886                          * Page truncated or invalidated. We can freely skip it
887                          * then, even for data integrity operations: the page
888                          * has disappeared concurrently, so there could be no
889                          * real expectation of this data interity operation
890                          * even if there is now a new, dirty page at the same
891                          * pagecache address.
892                          */
893                         if (unlikely(page->mapping != mapping)) {
894 continue_unlock:
895                                 unlock_page(page);
896                                 continue;
897                         }
898
899                         if (!PageDirty(page)) {
900                                 /* someone wrote it for us */
901                                 goto continue_unlock;
902                         }
903
904                         if (PageWriteback(page)) {
905                                 if (wbc->sync_mode != WB_SYNC_NONE)
906                                         wait_on_page_writeback(page);
907                                 else
908                                         goto continue_unlock;
909                         }
910
911                         BUG_ON(PageWriteback(page));
912                         if (!clear_page_dirty_for_io(page))
913                                 goto continue_unlock;
914
915                         ret = (*writepage)(page, wbc, data);
916                         if (unlikely(ret)) {
917                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
918                                         unlock_page(page);
919                                         ret = 0;
920                                 } else {
921                                         /*
922                                          * done_index is set past this page,
923                                          * so media errors will not choke
924                                          * background writeout for the entire
925                                          * file. This has consequences for
926                                          * range_cyclic semantics (ie. it may
927                                          * not be suitable for data integrity
928                                          * writeout).
929                                          */
930                                         done = 1;
931                                         break;
932                                 }
933                         }
934
935                         if (nr_to_write > 0) {
936                                 nr_to_write--;
937                                 if (nr_to_write == 0 &&
938                                     wbc->sync_mode == WB_SYNC_NONE) {
939                                         /*
940                                          * We stop writing back only if we are
941                                          * not doing integrity sync. In case of
942                                          * integrity sync we have to keep going
943                                          * because someone may be concurrently
944                                          * dirtying pages, and we might have
945                                          * synced a lot of newly appeared dirty
946                                          * pages, but have not synced all of the
947                                          * old dirty pages.
948                                          */
949                                         done = 1;
950                                         break;
951                                 }
952                         }
953
954                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
955                                 wbc->encountered_congestion = 1;
956                                 done = 1;
957                                 break;
958                         }
959                 }
960                 pagevec_release(&pvec);
961                 cond_resched();
962         }
963         if (!cycled && !done) {
964                 /*
965                  * range_cyclic:
966                  * We hit the last page and there is more work to be done: wrap
967                  * back to the start of the file
968                  */
969                 cycled = 1;
970                 index = 0;
971                 end = writeback_index - 1;
972                 goto retry;
973         }
974         if (!wbc->no_nrwrite_index_update) {
975                 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
976                         mapping->writeback_index = done_index;
977                 wbc->nr_to_write = nr_to_write;
978         }
979
980         return ret;
981 }
982 EXPORT_SYMBOL(write_cache_pages);
983
984 /*
985  * Function used by generic_writepages to call the real writepage
986  * function and set the mapping flags on error
987  */
988 static int __writepage(struct page *page, struct writeback_control *wbc,
989                        void *data)
990 {
991         struct address_space *mapping = data;
992         int ret = mapping->a_ops->writepage(page, wbc);
993         mapping_set_error(mapping, ret);
994         return ret;
995 }
996
997 /**
998  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
999  * @mapping: address space structure to write
1000  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1001  *
1002  * This is a library function, which implements the writepages()
1003  * address_space_operation.
1004  */
1005 int generic_writepages(struct address_space *mapping,
1006                        struct writeback_control *wbc)
1007 {
1008         /* deal with chardevs and other special file */
1009         if (!mapping->a_ops->writepage)
1010                 return 0;
1011
1012         return write_cache_pages(mapping, wbc, __writepage, mapping);
1013 }
1014
1015 EXPORT_SYMBOL(generic_writepages);
1016
1017 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1018 {
1019         int ret;
1020
1021         if (wbc->nr_to_write <= 0)
1022                 return 0;
1023         if (mapping->a_ops->writepages)
1024                 ret = mapping->a_ops->writepages(mapping, wbc);
1025         else
1026                 ret = generic_writepages(mapping, wbc);
1027         return ret;
1028 }
1029
1030 /**
1031  * write_one_page - write out a single page and optionally wait on I/O
1032  * @page: the page to write
1033  * @wait: if true, wait on writeout
1034  *
1035  * The page must be locked by the caller and will be unlocked upon return.
1036  *
1037  * write_one_page() returns a negative error code if I/O failed.
1038  */
1039 int write_one_page(struct page *page, int wait)
1040 {
1041         struct address_space *mapping = page->mapping;
1042         int ret = 0;
1043         struct writeback_control wbc = {
1044                 .sync_mode = WB_SYNC_ALL,
1045                 .nr_to_write = 1,
1046         };
1047
1048         BUG_ON(!PageLocked(page));
1049
1050         if (wait)
1051                 wait_on_page_writeback(page);
1052
1053         if (clear_page_dirty_for_io(page)) {
1054                 page_cache_get(page);
1055                 ret = mapping->a_ops->writepage(page, &wbc);
1056                 if (ret == 0 && wait) {
1057                         wait_on_page_writeback(page);
1058                         if (PageError(page))
1059                                 ret = -EIO;
1060                 }
1061                 page_cache_release(page);
1062         } else {
1063                 unlock_page(page);
1064         }
1065         return ret;
1066 }
1067 EXPORT_SYMBOL(write_one_page);
1068
1069 /*
1070  * For address_spaces which do not use buffers nor write back.
1071  */
1072 int __set_page_dirty_no_writeback(struct page *page)
1073 {
1074         if (!PageDirty(page))
1075                 SetPageDirty(page);
1076         return 0;
1077 }
1078
1079 /*
1080  * Helper function for set_page_dirty family.
1081  * NOTE: This relies on being atomic wrt interrupts.
1082  */
1083 void account_page_dirtied(struct page *page, struct address_space *mapping)
1084 {
1085         if (mapping_cap_account_dirty(mapping)) {
1086                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1087                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1088                 task_dirty_inc(current);
1089                 task_io_account_write(PAGE_CACHE_SIZE);
1090         }
1091 }
1092
1093 /*
1094  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1095  * its radix tree.
1096  *
1097  * This is also used when a single buffer is being dirtied: we want to set the
1098  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1099  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1100  *
1101  * Most callers have locked the page, which pins the address_space in memory.
1102  * But zap_pte_range() does not lock the page, however in that case the
1103  * mapping is pinned by the vma's ->vm_file reference.
1104  *
1105  * We take care to handle the case where the page was truncated from the
1106  * mapping by re-checking page_mapping() inside tree_lock.
1107  */
1108 int __set_page_dirty_nobuffers(struct page *page)
1109 {
1110         if (!TestSetPageDirty(page)) {
1111                 struct address_space *mapping = page_mapping(page);
1112                 struct address_space *mapping2;
1113
1114                 if (!mapping)
1115                         return 1;
1116
1117                 spin_lock_irq(&mapping->tree_lock);
1118                 mapping2 = page_mapping(page);
1119                 if (mapping2) { /* Race with truncate? */
1120                         BUG_ON(mapping2 != mapping);
1121                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1122                         account_page_dirtied(page, mapping);
1123                         radix_tree_tag_set(&mapping->page_tree,
1124                                 page_index(page), PAGECACHE_TAG_DIRTY);
1125                 }
1126                 spin_unlock_irq(&mapping->tree_lock);
1127                 if (mapping->host) {
1128                         /* !PageAnon && !swapper_space */
1129                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1130                 }
1131                 return 1;
1132         }
1133         return 0;
1134 }
1135 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1136
1137 /*
1138  * When a writepage implementation decides that it doesn't want to write this
1139  * page for some reason, it should redirty the locked page via
1140  * redirty_page_for_writepage() and it should then unlock the page and return 0
1141  */
1142 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1143 {
1144         wbc->pages_skipped++;
1145         return __set_page_dirty_nobuffers(page);
1146 }
1147 EXPORT_SYMBOL(redirty_page_for_writepage);
1148
1149 /*
1150  * If the mapping doesn't provide a set_page_dirty a_op, then
1151  * just fall through and assume that it wants buffer_heads.
1152  */
1153 int set_page_dirty(struct page *page)
1154 {
1155         struct address_space *mapping = page_mapping(page);
1156
1157         if (likely(mapping)) {
1158                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1159 #ifdef CONFIG_BLOCK
1160                 if (!spd)
1161                         spd = __set_page_dirty_buffers;
1162 #endif
1163                 return (*spd)(page);
1164         }
1165         if (!PageDirty(page)) {
1166                 if (!TestSetPageDirty(page))
1167                         return 1;
1168         }
1169         return 0;
1170 }
1171 EXPORT_SYMBOL(set_page_dirty);
1172
1173 /*
1174  * set_page_dirty() is racy if the caller has no reference against
1175  * page->mapping->host, and if the page is unlocked.  This is because another
1176  * CPU could truncate the page off the mapping and then free the mapping.
1177  *
1178  * Usually, the page _is_ locked, or the caller is a user-space process which
1179  * holds a reference on the inode by having an open file.
1180  *
1181  * In other cases, the page should be locked before running set_page_dirty().
1182  */
1183 int set_page_dirty_lock(struct page *page)
1184 {
1185         int ret;
1186
1187         lock_page_nosync(page);
1188         ret = set_page_dirty(page);
1189         unlock_page(page);
1190         return ret;
1191 }
1192 EXPORT_SYMBOL(set_page_dirty_lock);
1193
1194 /*
1195  * Clear a page's dirty flag, while caring for dirty memory accounting.
1196  * Returns true if the page was previously dirty.
1197  *
1198  * This is for preparing to put the page under writeout.  We leave the page
1199  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1200  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1201  * implementation will run either set_page_writeback() or set_page_dirty(),
1202  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1203  * back into sync.
1204  *
1205  * This incoherency between the page's dirty flag and radix-tree tag is
1206  * unfortunate, but it only exists while the page is locked.
1207  */
1208 int clear_page_dirty_for_io(struct page *page)
1209 {
1210         struct address_space *mapping = page_mapping(page);
1211
1212         BUG_ON(!PageLocked(page));
1213
1214         ClearPageReclaim(page);
1215         if (mapping && mapping_cap_account_dirty(mapping)) {
1216                 /*
1217                  * Yes, Virginia, this is indeed insane.
1218                  *
1219                  * We use this sequence to make sure that
1220                  *  (a) we account for dirty stats properly
1221                  *  (b) we tell the low-level filesystem to
1222                  *      mark the whole page dirty if it was
1223                  *      dirty in a pagetable. Only to then
1224                  *  (c) clean the page again and return 1 to
1225                  *      cause the writeback.
1226                  *
1227                  * This way we avoid all nasty races with the
1228                  * dirty bit in multiple places and clearing
1229                  * them concurrently from different threads.
1230                  *
1231                  * Note! Normally the "set_page_dirty(page)"
1232                  * has no effect on the actual dirty bit - since
1233                  * that will already usually be set. But we
1234                  * need the side effects, and it can help us
1235                  * avoid races.
1236                  *
1237                  * We basically use the page "master dirty bit"
1238                  * as a serialization point for all the different
1239                  * threads doing their things.
1240                  */
1241                 if (page_mkclean(page))
1242                         set_page_dirty(page);
1243                 /*
1244                  * We carefully synchronise fault handlers against
1245                  * installing a dirty pte and marking the page dirty
1246                  * at this point. We do this by having them hold the
1247                  * page lock at some point after installing their
1248                  * pte, but before marking the page dirty.
1249                  * Pages are always locked coming in here, so we get
1250                  * the desired exclusion. See mm/memory.c:do_wp_page()
1251                  * for more comments.
1252                  */
1253                 if (TestClearPageDirty(page)) {
1254                         dec_zone_page_state(page, NR_FILE_DIRTY);
1255                         dec_bdi_stat(mapping->backing_dev_info,
1256                                         BDI_RECLAIMABLE);
1257                         return 1;
1258                 }
1259                 return 0;
1260         }
1261         return TestClearPageDirty(page);
1262 }
1263 EXPORT_SYMBOL(clear_page_dirty_for_io);
1264
1265 int test_clear_page_writeback(struct page *page)
1266 {
1267         struct address_space *mapping = page_mapping(page);
1268         int ret;
1269
1270         if (mapping) {
1271                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1272                 unsigned long flags;
1273
1274                 spin_lock_irqsave(&mapping->tree_lock, flags);
1275                 ret = TestClearPageWriteback(page);
1276                 if (ret) {
1277                         radix_tree_tag_clear(&mapping->page_tree,
1278                                                 page_index(page),
1279                                                 PAGECACHE_TAG_WRITEBACK);
1280                         if (bdi_cap_account_writeback(bdi)) {
1281                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1282                                 __bdi_writeout_inc(bdi);
1283                         }
1284                 }
1285                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1286         } else {
1287                 ret = TestClearPageWriteback(page);
1288         }
1289         if (ret)
1290                 dec_zone_page_state(page, NR_WRITEBACK);
1291         return ret;
1292 }
1293
1294 int test_set_page_writeback(struct page *page)
1295 {
1296         struct address_space *mapping = page_mapping(page);
1297         int ret;
1298
1299         if (mapping) {
1300                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1301                 unsigned long flags;
1302
1303                 spin_lock_irqsave(&mapping->tree_lock, flags);
1304                 ret = TestSetPageWriteback(page);
1305                 if (!ret) {
1306                         radix_tree_tag_set(&mapping->page_tree,
1307                                                 page_index(page),
1308                                                 PAGECACHE_TAG_WRITEBACK);
1309                         if (bdi_cap_account_writeback(bdi))
1310                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1311                 }
1312                 if (!PageDirty(page))
1313                         radix_tree_tag_clear(&mapping->page_tree,
1314                                                 page_index(page),
1315                                                 PAGECACHE_TAG_DIRTY);
1316                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1317         } else {
1318                 ret = TestSetPageWriteback(page);
1319         }
1320         if (!ret)
1321                 inc_zone_page_state(page, NR_WRITEBACK);
1322         return ret;
1323
1324 }
1325 EXPORT_SYMBOL(test_set_page_writeback);
1326
1327 /*
1328  * Return true if any of the pages in the mapping are marked with the
1329  * passed tag.
1330  */
1331 int mapping_tagged(struct address_space *mapping, int tag)
1332 {
1333         int ret;
1334         rcu_read_lock();
1335         ret = radix_tree_tagged(&mapping->page_tree, tag);
1336         rcu_read_unlock();
1337         return ret;
1338 }
1339 EXPORT_SYMBOL(mapping_tagged);