writeback: Add tracing to write_cache_pages
[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 #include <trace/events/writeback.h>
38
39 /*
40  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
41  * will look to see if it needs to force writeback or throttling.
42  */
43 static long ratelimit_pages = 32;
44
45 /*
46  * When balance_dirty_pages decides that the caller needs to perform some
47  * non-background writeback, this is how many pages it will attempt to write.
48  * It should be somewhat larger than dirtied pages to ensure that reasonably
49  * large amounts of I/O are submitted.
50  */
51 static inline long sync_writeback_pages(unsigned long dirtied)
52 {
53         if (dirtied < ratelimit_pages)
54                 dirtied = ratelimit_pages;
55
56         return dirtied + dirtied / 2;
57 }
58
59 /* The following parameters are exported via /proc/sys/vm */
60
61 /*
62  * Start background writeback (via writeback threads) at this percentage
63  */
64 int dirty_background_ratio = 10;
65
66 /*
67  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
68  * dirty_background_ratio * the amount of dirtyable memory
69  */
70 unsigned long dirty_background_bytes;
71
72 /*
73  * free highmem will not be subtracted from the total free memory
74  * for calculating free ratios if vm_highmem_is_dirtyable is true
75  */
76 int vm_highmem_is_dirtyable;
77
78 /*
79  * The generator of dirty data starts writeback at this percentage
80  */
81 int vm_dirty_ratio = 20;
82
83 /*
84  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
85  * vm_dirty_ratio * the amount of dirtyable memory
86  */
87 unsigned long vm_dirty_bytes;
88
89 /*
90  * The interval between `kupdate'-style writebacks
91  */
92 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
93
94 /*
95  * The longest time for which data is allowed to remain dirty
96  */
97 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
98
99 /*
100  * Flag that makes the machine dump writes/reads and block dirtyings.
101  */
102 int block_dump;
103
104 /*
105  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
106  * a full sync is triggered after this time elapses without any disk activity.
107  */
108 int laptop_mode;
109
110 EXPORT_SYMBOL(laptop_mode);
111
112 /* End of sysctl-exported parameters */
113
114
115 /*
116  * Scale the writeback cache size proportional to the relative writeout speeds.
117  *
118  * We do this by keeping a floating proportion between BDIs, based on page
119  * writeback completions [end_page_writeback()]. Those devices that write out
120  * pages fastest will get the larger share, while the slower will get a smaller
121  * share.
122  *
123  * We use page writeout completions because we are interested in getting rid of
124  * dirty pages. Having them written out is the primary goal.
125  *
126  * We introduce a concept of time, a period over which we measure these events,
127  * because demand can/will vary over time. The length of this period itself is
128  * measured in page writeback completions.
129  *
130  */
131 static struct prop_descriptor vm_completions;
132 static struct prop_descriptor vm_dirties;
133
134 /*
135  * couple the period to the dirty_ratio:
136  *
137  *   period/2 ~ roundup_pow_of_two(dirty limit)
138  */
139 static int calc_period_shift(void)
140 {
141         unsigned long dirty_total;
142
143         if (vm_dirty_bytes)
144                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
145         else
146                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
147                                 100;
148         return 2 + ilog2(dirty_total - 1);
149 }
150
151 /*
152  * update the period when the dirty threshold changes.
153  */
154 static void update_completion_period(void)
155 {
156         int shift = calc_period_shift();
157         prop_change_shift(&vm_completions, shift);
158         prop_change_shift(&vm_dirties, shift);
159 }
160
161 int dirty_background_ratio_handler(struct ctl_table *table, int write,
162                 void __user *buffer, size_t *lenp,
163                 loff_t *ppos)
164 {
165         int ret;
166
167         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
168         if (ret == 0 && write)
169                 dirty_background_bytes = 0;
170         return ret;
171 }
172
173 int dirty_background_bytes_handler(struct ctl_table *table, int write,
174                 void __user *buffer, size_t *lenp,
175                 loff_t *ppos)
176 {
177         int ret;
178
179         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
180         if (ret == 0 && write)
181                 dirty_background_ratio = 0;
182         return ret;
183 }
184
185 int dirty_ratio_handler(struct ctl_table *table, int write,
186                 void __user *buffer, size_t *lenp,
187                 loff_t *ppos)
188 {
189         int old_ratio = vm_dirty_ratio;
190         int ret;
191
192         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
193         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
194                 update_completion_period();
195                 vm_dirty_bytes = 0;
196         }
197         return ret;
198 }
199
200
201 int dirty_bytes_handler(struct ctl_table *table, int write,
202                 void __user *buffer, size_t *lenp,
203                 loff_t *ppos)
204 {
205         unsigned long old_bytes = vm_dirty_bytes;
206         int ret;
207
208         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
209         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
210                 update_completion_period();
211                 vm_dirty_ratio = 0;
212         }
213         return ret;
214 }
215
216 /*
217  * Increment the BDI's writeout completion count and the global writeout
218  * completion count. Called from test_clear_page_writeback().
219  */
220 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
221 {
222         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
223                               bdi->max_prop_frac);
224 }
225
226 void bdi_writeout_inc(struct backing_dev_info *bdi)
227 {
228         unsigned long flags;
229
230         local_irq_save(flags);
231         __bdi_writeout_inc(bdi);
232         local_irq_restore(flags);
233 }
234 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
235
236 void task_dirty_inc(struct task_struct *tsk)
237 {
238         prop_inc_single(&vm_dirties, &tsk->dirties);
239 }
240
241 /*
242  * Obtain an accurate fraction of the BDI's portion.
243  */
244 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
245                 long *numerator, long *denominator)
246 {
247         if (bdi_cap_writeback_dirty(bdi)) {
248                 prop_fraction_percpu(&vm_completions, &bdi->completions,
249                                 numerator, denominator);
250         } else {
251                 *numerator = 0;
252                 *denominator = 1;
253         }
254 }
255
256 /*
257  * Clip the earned share of dirty pages to that which is actually available.
258  * This avoids exceeding the total dirty_limit when the floating averages
259  * fluctuate too quickly.
260  */
261 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
262                 unsigned long dirty, unsigned long *pbdi_dirty)
263 {
264         unsigned long avail_dirty;
265
266         avail_dirty = global_page_state(NR_FILE_DIRTY) +
267                  global_page_state(NR_WRITEBACK) +
268                  global_page_state(NR_UNSTABLE_NFS) +
269                  global_page_state(NR_WRITEBACK_TEMP);
270
271         if (avail_dirty < dirty)
272                 avail_dirty = dirty - avail_dirty;
273         else
274                 avail_dirty = 0;
275
276         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
277                 bdi_stat(bdi, BDI_WRITEBACK);
278
279         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
280 }
281
282 static inline void task_dirties_fraction(struct task_struct *tsk,
283                 long *numerator, long *denominator)
284 {
285         prop_fraction_single(&vm_dirties, &tsk->dirties,
286                                 numerator, denominator);
287 }
288
289 /*
290  * scale the dirty limit
291  *
292  * task specific dirty limit:
293  *
294  *   dirty -= (dirty/8) * p_{t}
295  */
296 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
297 {
298         long numerator, denominator;
299         unsigned long dirty = *pdirty;
300         u64 inv = dirty >> 3;
301
302         task_dirties_fraction(tsk, &numerator, &denominator);
303         inv *= numerator;
304         do_div(inv, denominator);
305
306         dirty -= inv;
307         if (dirty < *pdirty/2)
308                 dirty = *pdirty/2;
309
310         *pdirty = dirty;
311 }
312
313 /*
314  *
315  */
316 static unsigned int bdi_min_ratio;
317
318 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
319 {
320         int ret = 0;
321
322         spin_lock_bh(&bdi_lock);
323         if (min_ratio > bdi->max_ratio) {
324                 ret = -EINVAL;
325         } else {
326                 min_ratio -= bdi->min_ratio;
327                 if (bdi_min_ratio + min_ratio < 100) {
328                         bdi_min_ratio += min_ratio;
329                         bdi->min_ratio += min_ratio;
330                 } else {
331                         ret = -EINVAL;
332                 }
333         }
334         spin_unlock_bh(&bdi_lock);
335
336         return ret;
337 }
338
339 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
340 {
341         int ret = 0;
342
343         if (max_ratio > 100)
344                 return -EINVAL;
345
346         spin_lock_bh(&bdi_lock);
347         if (bdi->min_ratio > max_ratio) {
348                 ret = -EINVAL;
349         } else {
350                 bdi->max_ratio = max_ratio;
351                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
352         }
353         spin_unlock_bh(&bdi_lock);
354
355         return ret;
356 }
357 EXPORT_SYMBOL(bdi_set_max_ratio);
358
359 /*
360  * Work out the current dirty-memory clamping and background writeout
361  * thresholds.
362  *
363  * The main aim here is to lower them aggressively if there is a lot of mapped
364  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
365  * pages.  It is better to clamp down on writers than to start swapping, and
366  * performing lots of scanning.
367  *
368  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
369  *
370  * We don't permit the clamping level to fall below 5% - that is getting rather
371  * excessive.
372  *
373  * We make sure that the background writeout level is below the adjusted
374  * clamping level.
375  */
376
377 static unsigned long highmem_dirtyable_memory(unsigned long total)
378 {
379 #ifdef CONFIG_HIGHMEM
380         int node;
381         unsigned long x = 0;
382
383         for_each_node_state(node, N_HIGH_MEMORY) {
384                 struct zone *z =
385                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
386
387                 x += zone_page_state(z, NR_FREE_PAGES) +
388                      zone_reclaimable_pages(z);
389         }
390         /*
391          * Make sure that the number of highmem pages is never larger
392          * than the number of the total dirtyable memory. This can only
393          * occur in very strange VM situations but we want to make sure
394          * that this does not occur.
395          */
396         return min(x, total);
397 #else
398         return 0;
399 #endif
400 }
401
402 /**
403  * determine_dirtyable_memory - amount of memory that may be used
404  *
405  * Returns the numebr of pages that can currently be freed and used
406  * by the kernel for direct mappings.
407  */
408 unsigned long determine_dirtyable_memory(void)
409 {
410         unsigned long x;
411
412         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
413
414         if (!vm_highmem_is_dirtyable)
415                 x -= highmem_dirtyable_memory(x);
416
417         return x + 1;   /* Ensure that we never return 0 */
418 }
419
420 void
421 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
422                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
423 {
424         unsigned long background;
425         unsigned long dirty;
426         unsigned long available_memory = determine_dirtyable_memory();
427         struct task_struct *tsk;
428
429         if (vm_dirty_bytes)
430                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
431         else {
432                 int dirty_ratio;
433
434                 dirty_ratio = vm_dirty_ratio;
435                 if (dirty_ratio < 5)
436                         dirty_ratio = 5;
437                 dirty = (dirty_ratio * available_memory) / 100;
438         }
439
440         if (dirty_background_bytes)
441                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
442         else
443                 background = (dirty_background_ratio * available_memory) / 100;
444
445         if (background >= dirty)
446                 background = dirty / 2;
447         tsk = current;
448         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
449                 background += background / 4;
450                 dirty += dirty / 4;
451         }
452         *pbackground = background;
453         *pdirty = dirty;
454
455         if (bdi) {
456                 u64 bdi_dirty;
457                 long numerator, denominator;
458
459                 /*
460                  * Calculate this BDI's share of the dirty ratio.
461                  */
462                 bdi_writeout_fraction(bdi, &numerator, &denominator);
463
464                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
465                 bdi_dirty *= numerator;
466                 do_div(bdi_dirty, denominator);
467                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
468                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
469                         bdi_dirty = dirty * bdi->max_ratio / 100;
470
471                 *pbdi_dirty = bdi_dirty;
472                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
473                 task_dirty_limit(current, pbdi_dirty);
474         }
475 }
476
477 /*
478  * balance_dirty_pages() must be called by processes which are generating dirty
479  * data.  It looks at the number of dirty pages in the machine and will force
480  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
481  * If we're over `background_thresh' then the writeback threads are woken to
482  * perform some writeout.
483  */
484 static void balance_dirty_pages(struct address_space *mapping,
485                                 unsigned long write_chunk)
486 {
487         long nr_reclaimable, bdi_nr_reclaimable;
488         long nr_writeback, bdi_nr_writeback;
489         unsigned long background_thresh;
490         unsigned long dirty_thresh;
491         unsigned long bdi_thresh;
492         unsigned long pages_written = 0;
493         unsigned long pause = 1;
494
495         struct backing_dev_info *bdi = mapping->backing_dev_info;
496
497         for (;;) {
498                 struct writeback_control wbc = {
499                         .sync_mode      = WB_SYNC_NONE,
500                         .older_than_this = NULL,
501                         .nr_to_write    = write_chunk,
502                         .range_cyclic   = 1,
503                 };
504
505                 get_dirty_limits(&background_thresh, &dirty_thresh,
506                                 &bdi_thresh, bdi);
507
508                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
509                                         global_page_state(NR_UNSTABLE_NFS);
510                 nr_writeback = global_page_state(NR_WRITEBACK);
511
512                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
513                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
514
515                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
516                         break;
517
518                 /*
519                  * Throttle it only when the background writeback cannot
520                  * catch-up. This avoids (excessively) small writeouts
521                  * when the bdi limits are ramping up.
522                  */
523                 if (nr_reclaimable + nr_writeback <
524                                 (background_thresh + dirty_thresh) / 2)
525                         break;
526
527                 if (!bdi->dirty_exceeded)
528                         bdi->dirty_exceeded = 1;
529
530                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
531                  * Unstable writes are a feature of certain networked
532                  * filesystems (i.e. NFS) in which data may have been
533                  * written to the server's write cache, but has not yet
534                  * been flushed to permanent storage.
535                  * Only move pages to writeback if this bdi is over its
536                  * threshold otherwise wait until the disk writes catch
537                  * up.
538                  */
539                 trace_wbc_balance_dirty_start(&wbc, bdi);
540                 if (bdi_nr_reclaimable > bdi_thresh) {
541                         writeback_inodes_wb(&bdi->wb, &wbc);
542                         pages_written += write_chunk - wbc.nr_to_write;
543                         get_dirty_limits(&background_thresh, &dirty_thresh,
544                                        &bdi_thresh, bdi);
545                         trace_wbc_balance_dirty_written(&wbc, bdi);
546                 }
547
548                 /*
549                  * In order to avoid the stacked BDI deadlock we need
550                  * to ensure we accurately count the 'dirty' pages when
551                  * the threshold is low.
552                  *
553                  * Otherwise it would be possible to get thresh+n pages
554                  * reported dirty, even though there are thresh-m pages
555                  * actually dirty; with m+n sitting in the percpu
556                  * deltas.
557                  */
558                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
559                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
560                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
561                 } else if (bdi_nr_reclaimable) {
562                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
563                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
564                 }
565
566                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
567                         break;
568                 if (pages_written >= write_chunk)
569                         break;          /* We've done our duty */
570
571                 trace_wbc_balance_dirty_wait(&wbc, bdi);
572                 __set_current_state(TASK_INTERRUPTIBLE);
573                 io_schedule_timeout(pause);
574
575                 /*
576                  * Increase the delay for each loop, up to our previous
577                  * default of taking a 100ms nap.
578                  */
579                 pause <<= 1;
580                 if (pause > HZ / 10)
581                         pause = HZ / 10;
582         }
583
584         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
585                         bdi->dirty_exceeded)
586                 bdi->dirty_exceeded = 0;
587
588         if (writeback_in_progress(bdi))
589                 return;
590
591         /*
592          * In laptop mode, we wait until hitting the higher threshold before
593          * starting background writeout, and then write out all the way down
594          * to the lower threshold.  So slow writers cause minimal disk activity.
595          *
596          * In normal mode, we start background writeout at the lower
597          * background_thresh, to keep the amount of dirty memory low.
598          */
599         if ((laptop_mode && pages_written) ||
600             (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
601                                + global_page_state(NR_UNSTABLE_NFS))
602                                           > background_thresh)))
603                 bdi_start_background_writeback(bdi);
604 }
605
606 void set_page_dirty_balance(struct page *page, int page_mkwrite)
607 {
608         if (set_page_dirty(page) || page_mkwrite) {
609                 struct address_space *mapping = page_mapping(page);
610
611                 if (mapping)
612                         balance_dirty_pages_ratelimited(mapping);
613         }
614 }
615
616 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
617
618 /**
619  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
620  * @mapping: address_space which was dirtied
621  * @nr_pages_dirtied: number of pages which the caller has just dirtied
622  *
623  * Processes which are dirtying memory should call in here once for each page
624  * which was newly dirtied.  The function will periodically check the system's
625  * dirty state and will initiate writeback if needed.
626  *
627  * On really big machines, get_writeback_state is expensive, so try to avoid
628  * calling it too often (ratelimiting).  But once we're over the dirty memory
629  * limit we decrease the ratelimiting by a lot, to prevent individual processes
630  * from overshooting the limit by (ratelimit_pages) each.
631  */
632 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
633                                         unsigned long nr_pages_dirtied)
634 {
635         unsigned long ratelimit;
636         unsigned long *p;
637
638         ratelimit = ratelimit_pages;
639         if (mapping->backing_dev_info->dirty_exceeded)
640                 ratelimit = 8;
641
642         /*
643          * Check the rate limiting. Also, we do not want to throttle real-time
644          * tasks in balance_dirty_pages(). Period.
645          */
646         preempt_disable();
647         p =  &__get_cpu_var(bdp_ratelimits);
648         *p += nr_pages_dirtied;
649         if (unlikely(*p >= ratelimit)) {
650                 ratelimit = sync_writeback_pages(*p);
651                 *p = 0;
652                 preempt_enable();
653                 balance_dirty_pages(mapping, ratelimit);
654                 return;
655         }
656         preempt_enable();
657 }
658 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
659
660 void throttle_vm_writeout(gfp_t gfp_mask)
661 {
662         unsigned long background_thresh;
663         unsigned long dirty_thresh;
664
665         for ( ; ; ) {
666                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
667
668                 /*
669                  * Boost the allowable dirty threshold a bit for page
670                  * allocators so they don't get DoS'ed by heavy writers
671                  */
672                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
673
674                 if (global_page_state(NR_UNSTABLE_NFS) +
675                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
676                                 break;
677                 congestion_wait(BLK_RW_ASYNC, HZ/10);
678
679                 /*
680                  * The caller might hold locks which can prevent IO completion
681                  * or progress in the filesystem.  So we cannot just sit here
682                  * waiting for IO to complete.
683                  */
684                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
685                         break;
686         }
687 }
688
689 /*
690  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
691  */
692 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
693         void __user *buffer, size_t *length, loff_t *ppos)
694 {
695         proc_dointvec(table, write, buffer, length, ppos);
696         bdi_arm_supers_timer();
697         return 0;
698 }
699
700 #ifdef CONFIG_BLOCK
701 void laptop_mode_timer_fn(unsigned long data)
702 {
703         struct request_queue *q = (struct request_queue *)data;
704         int nr_pages = global_page_state(NR_FILE_DIRTY) +
705                 global_page_state(NR_UNSTABLE_NFS);
706
707         /*
708          * We want to write everything out, not just down to the dirty
709          * threshold
710          */
711         if (bdi_has_dirty_io(&q->backing_dev_info))
712                 bdi_start_writeback(&q->backing_dev_info, nr_pages);
713 }
714
715 /*
716  * We've spun up the disk and we're in laptop mode: schedule writeback
717  * of all dirty data a few seconds from now.  If the flush is already scheduled
718  * then push it back - the user is still using the disk.
719  */
720 void laptop_io_completion(struct backing_dev_info *info)
721 {
722         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
723 }
724
725 /*
726  * We're in laptop mode and we've just synced. The sync's writes will have
727  * caused another writeback to be scheduled by laptop_io_completion.
728  * Nothing needs to be written back anymore, so we unschedule the writeback.
729  */
730 void laptop_sync_completion(void)
731 {
732         struct backing_dev_info *bdi;
733
734         rcu_read_lock();
735
736         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
737                 del_timer(&bdi->laptop_mode_wb_timer);
738
739         rcu_read_unlock();
740 }
741 #endif
742
743 /*
744  * If ratelimit_pages is too high then we can get into dirty-data overload
745  * if a large number of processes all perform writes at the same time.
746  * If it is too low then SMP machines will call the (expensive)
747  * get_writeback_state too often.
748  *
749  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
750  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
751  * thresholds before writeback cuts in.
752  *
753  * But the limit should not be set too high.  Because it also controls the
754  * amount of memory which the balance_dirty_pages() caller has to write back.
755  * If this is too large then the caller will block on the IO queue all the
756  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
757  * will write six megabyte chunks, max.
758  */
759
760 void writeback_set_ratelimit(void)
761 {
762         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
763         if (ratelimit_pages < 16)
764                 ratelimit_pages = 16;
765         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
766                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
767 }
768
769 static int __cpuinit
770 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
771 {
772         writeback_set_ratelimit();
773         return NOTIFY_DONE;
774 }
775
776 static struct notifier_block __cpuinitdata ratelimit_nb = {
777         .notifier_call  = ratelimit_handler,
778         .next           = NULL,
779 };
780
781 /*
782  * Called early on to tune the page writeback dirty limits.
783  *
784  * We used to scale dirty pages according to how total memory
785  * related to pages that could be allocated for buffers (by
786  * comparing nr_free_buffer_pages() to vm_total_pages.
787  *
788  * However, that was when we used "dirty_ratio" to scale with
789  * all memory, and we don't do that any more. "dirty_ratio"
790  * is now applied to total non-HIGHPAGE memory (by subtracting
791  * totalhigh_pages from vm_total_pages), and as such we can't
792  * get into the old insane situation any more where we had
793  * large amounts of dirty pages compared to a small amount of
794  * non-HIGHMEM memory.
795  *
796  * But we might still want to scale the dirty_ratio by how
797  * much memory the box has..
798  */
799 void __init page_writeback_init(void)
800 {
801         int shift;
802
803         writeback_set_ratelimit();
804         register_cpu_notifier(&ratelimit_nb);
805
806         shift = calc_period_shift();
807         prop_descriptor_init(&vm_completions, shift);
808         prop_descriptor_init(&vm_dirties, shift);
809 }
810
811 /**
812  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
813  * @mapping: address space structure to write
814  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
815  * @writepage: function called for each page
816  * @data: data passed to writepage function
817  *
818  * If a page is already under I/O, write_cache_pages() skips it, even
819  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
820  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
821  * and msync() need to guarantee that all the data which was dirty at the time
822  * the call was made get new I/O started against them.  If wbc->sync_mode is
823  * WB_SYNC_ALL then we were called for data integrity and we must wait for
824  * existing IO to complete.
825  */
826 int write_cache_pages(struct address_space *mapping,
827                       struct writeback_control *wbc, writepage_t writepage,
828                       void *data)
829 {
830         int ret = 0;
831         int done = 0;
832         struct pagevec pvec;
833         int nr_pages;
834         pgoff_t uninitialized_var(writeback_index);
835         pgoff_t index;
836         pgoff_t end;            /* Inclusive */
837         pgoff_t done_index;
838         int cycled;
839         int range_whole = 0;
840
841         pagevec_init(&pvec, 0);
842         if (wbc->range_cyclic) {
843                 writeback_index = mapping->writeback_index; /* prev offset */
844                 index = writeback_index;
845                 if (index == 0)
846                         cycled = 1;
847                 else
848                         cycled = 0;
849                 end = -1;
850         } else {
851                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
852                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
853                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
854                         range_whole = 1;
855                 cycled = 1; /* ignore range_cyclic tests */
856
857                 /*
858                  * If this is a data integrity sync, cap the writeback to the
859                  * current end of file. Any extension to the file that occurs
860                  * after this is a new write and we don't need to write those
861                  * pages out to fulfil our data integrity requirements. If we
862                  * try to write them out, we can get stuck in this scan until
863                  * the concurrent writer stops adding dirty pages and extending
864                  * EOF.
865                  */
866                 if (wbc->sync_mode == WB_SYNC_ALL &&
867                     wbc->range_end == LLONG_MAX) {
868                         end = i_size_read(mapping->host) >> PAGE_CACHE_SHIFT;
869                 }
870         }
871
872 retry:
873         done_index = index;
874         while (!done && (index <= end)) {
875                 int i;
876
877                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
878                               PAGECACHE_TAG_DIRTY,
879                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
880                 if (nr_pages == 0)
881                         break;
882
883                 for (i = 0; i < nr_pages; i++) {
884                         struct page *page = pvec.pages[i];
885
886                         /*
887                          * At this point, the page may be truncated or
888                          * invalidated (changing page->mapping to NULL), or
889                          * even swizzled back from swapper_space to tmpfs file
890                          * mapping. However, page->index will not change
891                          * because we have a reference on the page.
892                          */
893                         if (page->index > end) {
894                                 /*
895                                  * can't be range_cyclic (1st pass) because
896                                  * end == -1 in that case.
897                                  */
898                                 done = 1;
899                                 break;
900                         }
901
902                         done_index = page->index + 1;
903
904                         lock_page(page);
905
906                         /*
907                          * Page truncated or invalidated. We can freely skip it
908                          * then, even for data integrity operations: the page
909                          * has disappeared concurrently, so there could be no
910                          * real expectation of this data interity operation
911                          * even if there is now a new, dirty page at the same
912                          * pagecache address.
913                          */
914                         if (unlikely(page->mapping != mapping)) {
915 continue_unlock:
916                                 unlock_page(page);
917                                 continue;
918                         }
919
920                         if (!PageDirty(page)) {
921                                 /* someone wrote it for us */
922                                 goto continue_unlock;
923                         }
924
925                         if (PageWriteback(page)) {
926                                 if (wbc->sync_mode != WB_SYNC_NONE)
927                                         wait_on_page_writeback(page);
928                                 else
929                                         goto continue_unlock;
930                         }
931
932                         BUG_ON(PageWriteback(page));
933                         if (!clear_page_dirty_for_io(page))
934                                 goto continue_unlock;
935
936                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
937                         ret = (*writepage)(page, wbc, data);
938                         if (unlikely(ret)) {
939                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
940                                         unlock_page(page);
941                                         ret = 0;
942                                 } else {
943                                         /*
944                                          * done_index is set past this page,
945                                          * so media errors will not choke
946                                          * background writeout for the entire
947                                          * file. This has consequences for
948                                          * range_cyclic semantics (ie. it may
949                                          * not be suitable for data integrity
950                                          * writeout).
951                                          */
952                                         done = 1;
953                                         break;
954                                 }
955                         }
956
957                         if (wbc->nr_to_write > 0) {
958                                 if (--wbc->nr_to_write == 0 &&
959                                     wbc->sync_mode == WB_SYNC_NONE) {
960                                         /*
961                                          * We stop writing back only if we are
962                                          * not doing integrity sync. In case of
963                                          * integrity sync we have to keep going
964                                          * because someone may be concurrently
965                                          * dirtying pages, and we might have
966                                          * synced a lot of newly appeared dirty
967                                          * pages, but have not synced all of the
968                                          * old dirty pages.
969                                          */
970                                         done = 1;
971                                         break;
972                                 }
973                         }
974                 }
975                 pagevec_release(&pvec);
976                 cond_resched();
977         }
978         if (!cycled && !done) {
979                 /*
980                  * range_cyclic:
981                  * We hit the last page and there is more work to be done: wrap
982                  * back to the start of the file
983                  */
984                 cycled = 1;
985                 index = 0;
986                 end = writeback_index - 1;
987                 goto retry;
988         }
989         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
990                 mapping->writeback_index = done_index;
991
992         return ret;
993 }
994 EXPORT_SYMBOL(write_cache_pages);
995
996 /*
997  * Function used by generic_writepages to call the real writepage
998  * function and set the mapping flags on error
999  */
1000 static int __writepage(struct page *page, struct writeback_control *wbc,
1001                        void *data)
1002 {
1003         struct address_space *mapping = data;
1004         int ret = mapping->a_ops->writepage(page, wbc);
1005         mapping_set_error(mapping, ret);
1006         return ret;
1007 }
1008
1009 /**
1010  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1011  * @mapping: address space structure to write
1012  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1013  *
1014  * This is a library function, which implements the writepages()
1015  * address_space_operation.
1016  */
1017 int generic_writepages(struct address_space *mapping,
1018                        struct writeback_control *wbc)
1019 {
1020         /* deal with chardevs and other special file */
1021         if (!mapping->a_ops->writepage)
1022                 return 0;
1023
1024         return write_cache_pages(mapping, wbc, __writepage, mapping);
1025 }
1026
1027 EXPORT_SYMBOL(generic_writepages);
1028
1029 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1030 {
1031         int ret;
1032
1033         if (wbc->nr_to_write <= 0)
1034                 return 0;
1035         if (mapping->a_ops->writepages)
1036                 ret = mapping->a_ops->writepages(mapping, wbc);
1037         else
1038                 ret = generic_writepages(mapping, wbc);
1039         return ret;
1040 }
1041
1042 /**
1043  * write_one_page - write out a single page and optionally wait on I/O
1044  * @page: the page to write
1045  * @wait: if true, wait on writeout
1046  *
1047  * The page must be locked by the caller and will be unlocked upon return.
1048  *
1049  * write_one_page() returns a negative error code if I/O failed.
1050  */
1051 int write_one_page(struct page *page, int wait)
1052 {
1053         struct address_space *mapping = page->mapping;
1054         int ret = 0;
1055         struct writeback_control wbc = {
1056                 .sync_mode = WB_SYNC_ALL,
1057                 .nr_to_write = 1,
1058         };
1059
1060         BUG_ON(!PageLocked(page));
1061
1062         if (wait)
1063                 wait_on_page_writeback(page);
1064
1065         if (clear_page_dirty_for_io(page)) {
1066                 page_cache_get(page);
1067                 ret = mapping->a_ops->writepage(page, &wbc);
1068                 if (ret == 0 && wait) {
1069                         wait_on_page_writeback(page);
1070                         if (PageError(page))
1071                                 ret = -EIO;
1072                 }
1073                 page_cache_release(page);
1074         } else {
1075                 unlock_page(page);
1076         }
1077         return ret;
1078 }
1079 EXPORT_SYMBOL(write_one_page);
1080
1081 /*
1082  * For address_spaces which do not use buffers nor write back.
1083  */
1084 int __set_page_dirty_no_writeback(struct page *page)
1085 {
1086         if (!PageDirty(page))
1087                 SetPageDirty(page);
1088         return 0;
1089 }
1090
1091 /*
1092  * Helper function for set_page_dirty family.
1093  * NOTE: This relies on being atomic wrt interrupts.
1094  */
1095 void account_page_dirtied(struct page *page, struct address_space *mapping)
1096 {
1097         if (mapping_cap_account_dirty(mapping)) {
1098                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1099                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1100                 task_dirty_inc(current);
1101                 task_io_account_write(PAGE_CACHE_SIZE);
1102         }
1103 }
1104
1105 /*
1106  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1107  * its radix tree.
1108  *
1109  * This is also used when a single buffer is being dirtied: we want to set the
1110  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1111  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1112  *
1113  * Most callers have locked the page, which pins the address_space in memory.
1114  * But zap_pte_range() does not lock the page, however in that case the
1115  * mapping is pinned by the vma's ->vm_file reference.
1116  *
1117  * We take care to handle the case where the page was truncated from the
1118  * mapping by re-checking page_mapping() inside tree_lock.
1119  */
1120 int __set_page_dirty_nobuffers(struct page *page)
1121 {
1122         if (!TestSetPageDirty(page)) {
1123                 struct address_space *mapping = page_mapping(page);
1124                 struct address_space *mapping2;
1125
1126                 if (!mapping)
1127                         return 1;
1128
1129                 spin_lock_irq(&mapping->tree_lock);
1130                 mapping2 = page_mapping(page);
1131                 if (mapping2) { /* Race with truncate? */
1132                         BUG_ON(mapping2 != mapping);
1133                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1134                         account_page_dirtied(page, mapping);
1135                         radix_tree_tag_set(&mapping->page_tree,
1136                                 page_index(page), PAGECACHE_TAG_DIRTY);
1137                 }
1138                 spin_unlock_irq(&mapping->tree_lock);
1139                 if (mapping->host) {
1140                         /* !PageAnon && !swapper_space */
1141                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1142                 }
1143                 return 1;
1144         }
1145         return 0;
1146 }
1147 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1148
1149 /*
1150  * When a writepage implementation decides that it doesn't want to write this
1151  * page for some reason, it should redirty the locked page via
1152  * redirty_page_for_writepage() and it should then unlock the page and return 0
1153  */
1154 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1155 {
1156         wbc->pages_skipped++;
1157         return __set_page_dirty_nobuffers(page);
1158 }
1159 EXPORT_SYMBOL(redirty_page_for_writepage);
1160
1161 /*
1162  * Dirty a page.
1163  *
1164  * For pages with a mapping this should be done under the page lock
1165  * for the benefit of asynchronous memory errors who prefer a consistent
1166  * dirty state. This rule can be broken in some special cases,
1167  * but should be better not to.
1168  *
1169  * If the mapping doesn't provide a set_page_dirty a_op, then
1170  * just fall through and assume that it wants buffer_heads.
1171  */
1172 int set_page_dirty(struct page *page)
1173 {
1174         struct address_space *mapping = page_mapping(page);
1175
1176         if (likely(mapping)) {
1177                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1178 #ifdef CONFIG_BLOCK
1179                 if (!spd)
1180                         spd = __set_page_dirty_buffers;
1181 #endif
1182                 return (*spd)(page);
1183         }
1184         if (!PageDirty(page)) {
1185                 if (!TestSetPageDirty(page))
1186                         return 1;
1187         }
1188         return 0;
1189 }
1190 EXPORT_SYMBOL(set_page_dirty);
1191
1192 /*
1193  * set_page_dirty() is racy if the caller has no reference against
1194  * page->mapping->host, and if the page is unlocked.  This is because another
1195  * CPU could truncate the page off the mapping and then free the mapping.
1196  *
1197  * Usually, the page _is_ locked, or the caller is a user-space process which
1198  * holds a reference on the inode by having an open file.
1199  *
1200  * In other cases, the page should be locked before running set_page_dirty().
1201  */
1202 int set_page_dirty_lock(struct page *page)
1203 {
1204         int ret;
1205
1206         lock_page_nosync(page);
1207         ret = set_page_dirty(page);
1208         unlock_page(page);
1209         return ret;
1210 }
1211 EXPORT_SYMBOL(set_page_dirty_lock);
1212
1213 /*
1214  * Clear a page's dirty flag, while caring for dirty memory accounting.
1215  * Returns true if the page was previously dirty.
1216  *
1217  * This is for preparing to put the page under writeout.  We leave the page
1218  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1219  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1220  * implementation will run either set_page_writeback() or set_page_dirty(),
1221  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1222  * back into sync.
1223  *
1224  * This incoherency between the page's dirty flag and radix-tree tag is
1225  * unfortunate, but it only exists while the page is locked.
1226  */
1227 int clear_page_dirty_for_io(struct page *page)
1228 {
1229         struct address_space *mapping = page_mapping(page);
1230
1231         BUG_ON(!PageLocked(page));
1232
1233         ClearPageReclaim(page);
1234         if (mapping && mapping_cap_account_dirty(mapping)) {
1235                 /*
1236                  * Yes, Virginia, this is indeed insane.
1237                  *
1238                  * We use this sequence to make sure that
1239                  *  (a) we account for dirty stats properly
1240                  *  (b) we tell the low-level filesystem to
1241                  *      mark the whole page dirty if it was
1242                  *      dirty in a pagetable. Only to then
1243                  *  (c) clean the page again and return 1 to
1244                  *      cause the writeback.
1245                  *
1246                  * This way we avoid all nasty races with the
1247                  * dirty bit in multiple places and clearing
1248                  * them concurrently from different threads.
1249                  *
1250                  * Note! Normally the "set_page_dirty(page)"
1251                  * has no effect on the actual dirty bit - since
1252                  * that will already usually be set. But we
1253                  * need the side effects, and it can help us
1254                  * avoid races.
1255                  *
1256                  * We basically use the page "master dirty bit"
1257                  * as a serialization point for all the different
1258                  * threads doing their things.
1259                  */
1260                 if (page_mkclean(page))
1261                         set_page_dirty(page);
1262                 /*
1263                  * We carefully synchronise fault handlers against
1264                  * installing a dirty pte and marking the page dirty
1265                  * at this point. We do this by having them hold the
1266                  * page lock at some point after installing their
1267                  * pte, but before marking the page dirty.
1268                  * Pages are always locked coming in here, so we get
1269                  * the desired exclusion. See mm/memory.c:do_wp_page()
1270                  * for more comments.
1271                  */
1272                 if (TestClearPageDirty(page)) {
1273                         dec_zone_page_state(page, NR_FILE_DIRTY);
1274                         dec_bdi_stat(mapping->backing_dev_info,
1275                                         BDI_RECLAIMABLE);
1276                         return 1;
1277                 }
1278                 return 0;
1279         }
1280         return TestClearPageDirty(page);
1281 }
1282 EXPORT_SYMBOL(clear_page_dirty_for_io);
1283
1284 int test_clear_page_writeback(struct page *page)
1285 {
1286         struct address_space *mapping = page_mapping(page);
1287         int ret;
1288
1289         if (mapping) {
1290                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1291                 unsigned long flags;
1292
1293                 spin_lock_irqsave(&mapping->tree_lock, flags);
1294                 ret = TestClearPageWriteback(page);
1295                 if (ret) {
1296                         radix_tree_tag_clear(&mapping->page_tree,
1297                                                 page_index(page),
1298                                                 PAGECACHE_TAG_WRITEBACK);
1299                         if (bdi_cap_account_writeback(bdi)) {
1300                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1301                                 __bdi_writeout_inc(bdi);
1302                         }
1303                 }
1304                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1305         } else {
1306                 ret = TestClearPageWriteback(page);
1307         }
1308         if (ret)
1309                 dec_zone_page_state(page, NR_WRITEBACK);
1310         return ret;
1311 }
1312
1313 int test_set_page_writeback(struct page *page)
1314 {
1315         struct address_space *mapping = page_mapping(page);
1316         int ret;
1317
1318         if (mapping) {
1319                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1320                 unsigned long flags;
1321
1322                 spin_lock_irqsave(&mapping->tree_lock, flags);
1323                 ret = TestSetPageWriteback(page);
1324                 if (!ret) {
1325                         radix_tree_tag_set(&mapping->page_tree,
1326                                                 page_index(page),
1327                                                 PAGECACHE_TAG_WRITEBACK);
1328                         if (bdi_cap_account_writeback(bdi))
1329                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1330                 }
1331                 if (!PageDirty(page))
1332                         radix_tree_tag_clear(&mapping->page_tree,
1333                                                 page_index(page),
1334                                                 PAGECACHE_TAG_DIRTY);
1335                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1336         } else {
1337                 ret = TestSetPageWriteback(page);
1338         }
1339         if (!ret)
1340                 inc_zone_page_state(page, NR_WRITEBACK);
1341         return ret;
1342
1343 }
1344 EXPORT_SYMBOL(test_set_page_writeback);
1345
1346 /*
1347  * Return true if any of the pages in the mapping are marked with the
1348  * passed tag.
1349  */
1350 int mapping_tagged(struct address_space *mapping, int tag)
1351 {
1352         int ret;
1353         rcu_read_lock();
1354         ret = radix_tree_tagged(&mapping->page_tree, tag);
1355         rcu_read_unlock();
1356         return ret;
1357 }
1358 EXPORT_SYMBOL(mapping_tagged);