writeback: balance_dirty_pages(): reduce calls to global_page_state
[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 static inline void task_dirties_fraction(struct task_struct *tsk,
257                 long *numerator, long *denominator)
258 {
259         prop_fraction_single(&vm_dirties, &tsk->dirties,
260                                 numerator, denominator);
261 }
262
263 /*
264  * scale the dirty limit
265  *
266  * task specific dirty limit:
267  *
268  *   dirty -= (dirty/8) * p_{t}
269  */
270 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
271 {
272         long numerator, denominator;
273         unsigned long dirty = *pdirty;
274         u64 inv = dirty >> 3;
275
276         task_dirties_fraction(tsk, &numerator, &denominator);
277         inv *= numerator;
278         do_div(inv, denominator);
279
280         dirty -= inv;
281         if (dirty < *pdirty/2)
282                 dirty = *pdirty/2;
283
284         *pdirty = dirty;
285 }
286
287 /*
288  *
289  */
290 static unsigned int bdi_min_ratio;
291
292 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
293 {
294         int ret = 0;
295
296         spin_lock_bh(&bdi_lock);
297         if (min_ratio > bdi->max_ratio) {
298                 ret = -EINVAL;
299         } else {
300                 min_ratio -= bdi->min_ratio;
301                 if (bdi_min_ratio + min_ratio < 100) {
302                         bdi_min_ratio += min_ratio;
303                         bdi->min_ratio += min_ratio;
304                 } else {
305                         ret = -EINVAL;
306                 }
307         }
308         spin_unlock_bh(&bdi_lock);
309
310         return ret;
311 }
312
313 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
314 {
315         int ret = 0;
316
317         if (max_ratio > 100)
318                 return -EINVAL;
319
320         spin_lock_bh(&bdi_lock);
321         if (bdi->min_ratio > max_ratio) {
322                 ret = -EINVAL;
323         } else {
324                 bdi->max_ratio = max_ratio;
325                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
326         }
327         spin_unlock_bh(&bdi_lock);
328
329         return ret;
330 }
331 EXPORT_SYMBOL(bdi_set_max_ratio);
332
333 /*
334  * Work out the current dirty-memory clamping and background writeout
335  * thresholds.
336  *
337  * The main aim here is to lower them aggressively if there is a lot of mapped
338  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
339  * pages.  It is better to clamp down on writers than to start swapping, and
340  * performing lots of scanning.
341  *
342  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
343  *
344  * We don't permit the clamping level to fall below 5% - that is getting rather
345  * excessive.
346  *
347  * We make sure that the background writeout level is below the adjusted
348  * clamping level.
349  */
350
351 static unsigned long highmem_dirtyable_memory(unsigned long total)
352 {
353 #ifdef CONFIG_HIGHMEM
354         int node;
355         unsigned long x = 0;
356
357         for_each_node_state(node, N_HIGH_MEMORY) {
358                 struct zone *z =
359                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
360
361                 x += zone_page_state(z, NR_FREE_PAGES) +
362                      zone_reclaimable_pages(z);
363         }
364         /*
365          * Make sure that the number of highmem pages is never larger
366          * than the number of the total dirtyable memory. This can only
367          * occur in very strange VM situations but we want to make sure
368          * that this does not occur.
369          */
370         return min(x, total);
371 #else
372         return 0;
373 #endif
374 }
375
376 /**
377  * determine_dirtyable_memory - amount of memory that may be used
378  *
379  * Returns the numebr of pages that can currently be freed and used
380  * by the kernel for direct mappings.
381  */
382 unsigned long determine_dirtyable_memory(void)
383 {
384         unsigned long x;
385
386         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
387
388         if (!vm_highmem_is_dirtyable)
389                 x -= highmem_dirtyable_memory(x);
390
391         return x + 1;   /* Ensure that we never return 0 */
392 }
393
394 void
395 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
396                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
397 {
398         unsigned long background;
399         unsigned long dirty;
400         unsigned long available_memory = determine_dirtyable_memory();
401         struct task_struct *tsk;
402
403         if (vm_dirty_bytes)
404                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
405         else {
406                 int dirty_ratio;
407
408                 dirty_ratio = vm_dirty_ratio;
409                 if (dirty_ratio < 5)
410                         dirty_ratio = 5;
411                 dirty = (dirty_ratio * available_memory) / 100;
412         }
413
414         if (dirty_background_bytes)
415                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
416         else
417                 background = (dirty_background_ratio * available_memory) / 100;
418
419         if (background >= dirty)
420                 background = dirty / 2;
421         tsk = current;
422         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
423                 background += background / 4;
424                 dirty += dirty / 4;
425         }
426         *pbackground = background;
427         *pdirty = dirty;
428
429         if (bdi) {
430                 u64 bdi_dirty;
431                 long numerator, denominator;
432
433                 /*
434                  * Calculate this BDI's share of the dirty ratio.
435                  */
436                 bdi_writeout_fraction(bdi, &numerator, &denominator);
437
438                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
439                 bdi_dirty *= numerator;
440                 do_div(bdi_dirty, denominator);
441                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
442                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
443                         bdi_dirty = dirty * bdi->max_ratio / 100;
444
445                 *pbdi_dirty = bdi_dirty;
446                 task_dirty_limit(current, pbdi_dirty);
447         }
448 }
449
450 /*
451  * balance_dirty_pages() must be called by processes which are generating dirty
452  * data.  It looks at the number of dirty pages in the machine and will force
453  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
454  * If we're over `background_thresh' then the writeback threads are woken to
455  * perform some writeout.
456  */
457 static void balance_dirty_pages(struct address_space *mapping,
458                                 unsigned long write_chunk)
459 {
460         long nr_reclaimable, bdi_nr_reclaimable;
461         long nr_writeback, bdi_nr_writeback;
462         unsigned long background_thresh;
463         unsigned long dirty_thresh;
464         unsigned long bdi_thresh;
465         unsigned long pages_written = 0;
466         unsigned long pause = 1;
467         bool dirty_exceeded = false;
468         struct backing_dev_info *bdi = mapping->backing_dev_info;
469
470         for (;;) {
471                 struct writeback_control wbc = {
472                         .sync_mode      = WB_SYNC_NONE,
473                         .older_than_this = NULL,
474                         .nr_to_write    = write_chunk,
475                         .range_cyclic   = 1,
476                 };
477
478                 get_dirty_limits(&background_thresh, &dirty_thresh,
479                                 &bdi_thresh, bdi);
480
481                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
482                                         global_page_state(NR_UNSTABLE_NFS);
483                 nr_writeback = global_page_state(NR_WRITEBACK);
484
485                 /*
486                  * In order to avoid the stacked BDI deadlock we need
487                  * to ensure we accurately count the 'dirty' pages when
488                  * the threshold is low.
489                  *
490                  * Otherwise it would be possible to get thresh+n pages
491                  * reported dirty, even though there are thresh-m pages
492                  * actually dirty; with m+n sitting in the percpu
493                  * deltas.
494                  */
495                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
496                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
497                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
498                 } else {
499                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
500                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
501                 }
502
503                 /*
504                  * The bdi thresh is somehow "soft" limit derived from the
505                  * global "hard" limit. The former helps to prevent heavy IO
506                  * bdi or process from holding back light ones; The latter is
507                  * the last resort safeguard.
508                  */
509                 dirty_exceeded =
510                         (bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
511                         || (nr_reclaimable + nr_writeback >= dirty_thresh);
512
513                 if (!dirty_exceeded)
514                         break;
515
516                 /*
517                  * Throttle it only when the background writeback cannot
518                  * catch-up. This avoids (excessively) small writeouts
519                  * when the bdi limits are ramping up.
520                  */
521                 if (nr_reclaimable + nr_writeback <
522                                 (background_thresh + dirty_thresh) / 2)
523                         break;
524
525                 if (!bdi->dirty_exceeded)
526                         bdi->dirty_exceeded = 1;
527
528                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
529                  * Unstable writes are a feature of certain networked
530                  * filesystems (i.e. NFS) in which data may have been
531                  * written to the server's write cache, but has not yet
532                  * been flushed to permanent storage.
533                  * Only move pages to writeback if this bdi is over its
534                  * threshold otherwise wait until the disk writes catch
535                  * up.
536                  */
537                 trace_wbc_balance_dirty_start(&wbc, bdi);
538                 if (bdi_nr_reclaimable > bdi_thresh) {
539                         writeback_inodes_wb(&bdi->wb, &wbc);
540                         pages_written += write_chunk - wbc.nr_to_write;
541                         trace_wbc_balance_dirty_written(&wbc, bdi);
542                         if (pages_written >= write_chunk)
543                                 break;          /* We've done our duty */
544                 }
545                 trace_wbc_balance_dirty_wait(&wbc, bdi);
546                 __set_current_state(TASK_INTERRUPTIBLE);
547                 io_schedule_timeout(pause);
548
549                 /*
550                  * Increase the delay for each loop, up to our previous
551                  * default of taking a 100ms nap.
552                  */
553                 pause <<= 1;
554                 if (pause > HZ / 10)
555                         pause = HZ / 10;
556         }
557
558         if (!dirty_exceeded && bdi->dirty_exceeded)
559                 bdi->dirty_exceeded = 0;
560
561         if (writeback_in_progress(bdi))
562                 return;
563
564         /*
565          * In laptop mode, we wait until hitting the higher threshold before
566          * starting background writeout, and then write out all the way down
567          * to the lower threshold.  So slow writers cause minimal disk activity.
568          *
569          * In normal mode, we start background writeout at the lower
570          * background_thresh, to keep the amount of dirty memory low.
571          */
572         if ((laptop_mode && pages_written) ||
573             (!laptop_mode && (nr_reclaimable > background_thresh)))
574                 bdi_start_background_writeback(bdi);
575 }
576
577 void set_page_dirty_balance(struct page *page, int page_mkwrite)
578 {
579         if (set_page_dirty(page) || page_mkwrite) {
580                 struct address_space *mapping = page_mapping(page);
581
582                 if (mapping)
583                         balance_dirty_pages_ratelimited(mapping);
584         }
585 }
586
587 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
588
589 /**
590  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
591  * @mapping: address_space which was dirtied
592  * @nr_pages_dirtied: number of pages which the caller has just dirtied
593  *
594  * Processes which are dirtying memory should call in here once for each page
595  * which was newly dirtied.  The function will periodically check the system's
596  * dirty state and will initiate writeback if needed.
597  *
598  * On really big machines, get_writeback_state is expensive, so try to avoid
599  * calling it too often (ratelimiting).  But once we're over the dirty memory
600  * limit we decrease the ratelimiting by a lot, to prevent individual processes
601  * from overshooting the limit by (ratelimit_pages) each.
602  */
603 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
604                                         unsigned long nr_pages_dirtied)
605 {
606         unsigned long ratelimit;
607         unsigned long *p;
608
609         ratelimit = ratelimit_pages;
610         if (mapping->backing_dev_info->dirty_exceeded)
611                 ratelimit = 8;
612
613         /*
614          * Check the rate limiting. Also, we do not want to throttle real-time
615          * tasks in balance_dirty_pages(). Period.
616          */
617         preempt_disable();
618         p =  &__get_cpu_var(bdp_ratelimits);
619         *p += nr_pages_dirtied;
620         if (unlikely(*p >= ratelimit)) {
621                 ratelimit = sync_writeback_pages(*p);
622                 *p = 0;
623                 preempt_enable();
624                 balance_dirty_pages(mapping, ratelimit);
625                 return;
626         }
627         preempt_enable();
628 }
629 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
630
631 void throttle_vm_writeout(gfp_t gfp_mask)
632 {
633         unsigned long background_thresh;
634         unsigned long dirty_thresh;
635
636         for ( ; ; ) {
637                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
638
639                 /*
640                  * Boost the allowable dirty threshold a bit for page
641                  * allocators so they don't get DoS'ed by heavy writers
642                  */
643                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
644
645                 if (global_page_state(NR_UNSTABLE_NFS) +
646                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
647                                 break;
648                 congestion_wait(BLK_RW_ASYNC, HZ/10);
649
650                 /*
651                  * The caller might hold locks which can prevent IO completion
652                  * or progress in the filesystem.  So we cannot just sit here
653                  * waiting for IO to complete.
654                  */
655                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
656                         break;
657         }
658 }
659
660 /*
661  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
662  */
663 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
664         void __user *buffer, size_t *length, loff_t *ppos)
665 {
666         proc_dointvec(table, write, buffer, length, ppos);
667         bdi_arm_supers_timer();
668         return 0;
669 }
670
671 #ifdef CONFIG_BLOCK
672 void laptop_mode_timer_fn(unsigned long data)
673 {
674         struct request_queue *q = (struct request_queue *)data;
675         int nr_pages = global_page_state(NR_FILE_DIRTY) +
676                 global_page_state(NR_UNSTABLE_NFS);
677
678         /*
679          * We want to write everything out, not just down to the dirty
680          * threshold
681          */
682         if (bdi_has_dirty_io(&q->backing_dev_info))
683                 bdi_start_writeback(&q->backing_dev_info, nr_pages);
684 }
685
686 /*
687  * We've spun up the disk and we're in laptop mode: schedule writeback
688  * of all dirty data a few seconds from now.  If the flush is already scheduled
689  * then push it back - the user is still using the disk.
690  */
691 void laptop_io_completion(struct backing_dev_info *info)
692 {
693         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
694 }
695
696 /*
697  * We're in laptop mode and we've just synced. The sync's writes will have
698  * caused another writeback to be scheduled by laptop_io_completion.
699  * Nothing needs to be written back anymore, so we unschedule the writeback.
700  */
701 void laptop_sync_completion(void)
702 {
703         struct backing_dev_info *bdi;
704
705         rcu_read_lock();
706
707         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
708                 del_timer(&bdi->laptop_mode_wb_timer);
709
710         rcu_read_unlock();
711 }
712 #endif
713
714 /*
715  * If ratelimit_pages is too high then we can get into dirty-data overload
716  * if a large number of processes all perform writes at the same time.
717  * If it is too low then SMP machines will call the (expensive)
718  * get_writeback_state too often.
719  *
720  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
721  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
722  * thresholds before writeback cuts in.
723  *
724  * But the limit should not be set too high.  Because it also controls the
725  * amount of memory which the balance_dirty_pages() caller has to write back.
726  * If this is too large then the caller will block on the IO queue all the
727  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
728  * will write six megabyte chunks, max.
729  */
730
731 void writeback_set_ratelimit(void)
732 {
733         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
734         if (ratelimit_pages < 16)
735                 ratelimit_pages = 16;
736         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
737                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
738 }
739
740 static int __cpuinit
741 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
742 {
743         writeback_set_ratelimit();
744         return NOTIFY_DONE;
745 }
746
747 static struct notifier_block __cpuinitdata ratelimit_nb = {
748         .notifier_call  = ratelimit_handler,
749         .next           = NULL,
750 };
751
752 /*
753  * Called early on to tune the page writeback dirty limits.
754  *
755  * We used to scale dirty pages according to how total memory
756  * related to pages that could be allocated for buffers (by
757  * comparing nr_free_buffer_pages() to vm_total_pages.
758  *
759  * However, that was when we used "dirty_ratio" to scale with
760  * all memory, and we don't do that any more. "dirty_ratio"
761  * is now applied to total non-HIGHPAGE memory (by subtracting
762  * totalhigh_pages from vm_total_pages), and as such we can't
763  * get into the old insane situation any more where we had
764  * large amounts of dirty pages compared to a small amount of
765  * non-HIGHMEM memory.
766  *
767  * But we might still want to scale the dirty_ratio by how
768  * much memory the box has..
769  */
770 void __init page_writeback_init(void)
771 {
772         int shift;
773
774         writeback_set_ratelimit();
775         register_cpu_notifier(&ratelimit_nb);
776
777         shift = calc_period_shift();
778         prop_descriptor_init(&vm_completions, shift);
779         prop_descriptor_init(&vm_dirties, shift);
780 }
781
782 /**
783  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
784  * @mapping: address space structure to write
785  * @start: starting page index
786  * @end: ending page index (inclusive)
787  *
788  * This function scans the page range from @start to @end (inclusive) and tags
789  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
790  * that write_cache_pages (or whoever calls this function) will then use
791  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
792  * used to avoid livelocking of writeback by a process steadily creating new
793  * dirty pages in the file (thus it is important for this function to be quick
794  * so that it can tag pages faster than a dirtying process can create them).
795  */
796 /*
797  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
798  */
799 void tag_pages_for_writeback(struct address_space *mapping,
800                              pgoff_t start, pgoff_t end)
801 {
802 #define WRITEBACK_TAG_BATCH 4096
803         unsigned long tagged;
804
805         do {
806                 spin_lock_irq(&mapping->tree_lock);
807                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
808                                 &start, end, WRITEBACK_TAG_BATCH,
809                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
810                 spin_unlock_irq(&mapping->tree_lock);
811                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
812                 cond_resched();
813         } while (tagged >= WRITEBACK_TAG_BATCH);
814 }
815 EXPORT_SYMBOL(tag_pages_for_writeback);
816
817 /**
818  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
819  * @mapping: address space structure to write
820  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
821  * @writepage: function called for each page
822  * @data: data passed to writepage function
823  *
824  * If a page is already under I/O, write_cache_pages() skips it, even
825  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
826  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
827  * and msync() need to guarantee that all the data which was dirty at the time
828  * the call was made get new I/O started against them.  If wbc->sync_mode is
829  * WB_SYNC_ALL then we were called for data integrity and we must wait for
830  * existing IO to complete.
831  *
832  * To avoid livelocks (when other process dirties new pages), we first tag
833  * pages which should be written back with TOWRITE tag and only then start
834  * writing them. For data-integrity sync we have to be careful so that we do
835  * not miss some pages (e.g., because some other process has cleared TOWRITE
836  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
837  * by the process clearing the DIRTY tag (and submitting the page for IO).
838  */
839 int write_cache_pages(struct address_space *mapping,
840                       struct writeback_control *wbc, writepage_t writepage,
841                       void *data)
842 {
843         int ret = 0;
844         int done = 0;
845         struct pagevec pvec;
846         int nr_pages;
847         pgoff_t uninitialized_var(writeback_index);
848         pgoff_t index;
849         pgoff_t end;            /* Inclusive */
850         pgoff_t done_index;
851         int cycled;
852         int range_whole = 0;
853         int tag;
854
855         pagevec_init(&pvec, 0);
856         if (wbc->range_cyclic) {
857                 writeback_index = mapping->writeback_index; /* prev offset */
858                 index = writeback_index;
859                 if (index == 0)
860                         cycled = 1;
861                 else
862                         cycled = 0;
863                 end = -1;
864         } else {
865                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
866                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
867                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
868                         range_whole = 1;
869                 cycled = 1; /* ignore range_cyclic tests */
870         }
871         if (wbc->sync_mode == WB_SYNC_ALL)
872                 tag = PAGECACHE_TAG_TOWRITE;
873         else
874                 tag = PAGECACHE_TAG_DIRTY;
875 retry:
876         if (wbc->sync_mode == WB_SYNC_ALL)
877                 tag_pages_for_writeback(mapping, index, end);
878         done_index = index;
879         while (!done && (index <= end)) {
880                 int i;
881
882                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
883                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
884                 if (nr_pages == 0)
885                         break;
886
887                 for (i = 0; i < nr_pages; i++) {
888                         struct page *page = pvec.pages[i];
889
890                         /*
891                          * At this point, the page may be truncated or
892                          * invalidated (changing page->mapping to NULL), or
893                          * even swizzled back from swapper_space to tmpfs file
894                          * mapping. However, page->index will not change
895                          * because we have a reference on the page.
896                          */
897                         if (page->index > end) {
898                                 /*
899                                  * can't be range_cyclic (1st pass) because
900                                  * end == -1 in that case.
901                                  */
902                                 done = 1;
903                                 break;
904                         }
905
906                         done_index = page->index + 1;
907
908                         lock_page(page);
909
910                         /*
911                          * Page truncated or invalidated. We can freely skip it
912                          * then, even for data integrity operations: the page
913                          * has disappeared concurrently, so there could be no
914                          * real expectation of this data interity operation
915                          * even if there is now a new, dirty page at the same
916                          * pagecache address.
917                          */
918                         if (unlikely(page->mapping != mapping)) {
919 continue_unlock:
920                                 unlock_page(page);
921                                 continue;
922                         }
923
924                         if (!PageDirty(page)) {
925                                 /* someone wrote it for us */
926                                 goto continue_unlock;
927                         }
928
929                         if (PageWriteback(page)) {
930                                 if (wbc->sync_mode != WB_SYNC_NONE)
931                                         wait_on_page_writeback(page);
932                                 else
933                                         goto continue_unlock;
934                         }
935
936                         BUG_ON(PageWriteback(page));
937                         if (!clear_page_dirty_for_io(page))
938                                 goto continue_unlock;
939
940                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
941                         ret = (*writepage)(page, wbc, data);
942                         if (unlikely(ret)) {
943                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
944                                         unlock_page(page);
945                                         ret = 0;
946                                 } else {
947                                         /*
948                                          * done_index is set past this page,
949                                          * so media errors will not choke
950                                          * background writeout for the entire
951                                          * file. This has consequences for
952                                          * range_cyclic semantics (ie. it may
953                                          * not be suitable for data integrity
954                                          * writeout).
955                                          */
956                                         done = 1;
957                                         break;
958                                 }
959                         }
960
961                         if (wbc->nr_to_write > 0) {
962                                 if (--wbc->nr_to_write == 0 &&
963                                     wbc->sync_mode == WB_SYNC_NONE) {
964                                         /*
965                                          * We stop writing back only if we are
966                                          * not doing integrity sync. In case of
967                                          * integrity sync we have to keep going
968                                          * because someone may be concurrently
969                                          * dirtying pages, and we might have
970                                          * synced a lot of newly appeared dirty
971                                          * pages, but have not synced all of the
972                                          * old dirty pages.
973                                          */
974                                         done = 1;
975                                         break;
976                                 }
977                         }
978                 }
979                 pagevec_release(&pvec);
980                 cond_resched();
981         }
982         if (!cycled && !done) {
983                 /*
984                  * range_cyclic:
985                  * We hit the last page and there is more work to be done: wrap
986                  * back to the start of the file
987                  */
988                 cycled = 1;
989                 index = 0;
990                 end = writeback_index - 1;
991                 goto retry;
992         }
993         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
994                 mapping->writeback_index = done_index;
995
996         return ret;
997 }
998 EXPORT_SYMBOL(write_cache_pages);
999
1000 /*
1001  * Function used by generic_writepages to call the real writepage
1002  * function and set the mapping flags on error
1003  */
1004 static int __writepage(struct page *page, struct writeback_control *wbc,
1005                        void *data)
1006 {
1007         struct address_space *mapping = data;
1008         int ret = mapping->a_ops->writepage(page, wbc);
1009         mapping_set_error(mapping, ret);
1010         return ret;
1011 }
1012
1013 /**
1014  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1015  * @mapping: address space structure to write
1016  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1017  *
1018  * This is a library function, which implements the writepages()
1019  * address_space_operation.
1020  */
1021 int generic_writepages(struct address_space *mapping,
1022                        struct writeback_control *wbc)
1023 {
1024         /* deal with chardevs and other special file */
1025         if (!mapping->a_ops->writepage)
1026                 return 0;
1027
1028         return write_cache_pages(mapping, wbc, __writepage, mapping);
1029 }
1030
1031 EXPORT_SYMBOL(generic_writepages);
1032
1033 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1034 {
1035         int ret;
1036
1037         if (wbc->nr_to_write <= 0)
1038                 return 0;
1039         if (mapping->a_ops->writepages)
1040                 ret = mapping->a_ops->writepages(mapping, wbc);
1041         else
1042                 ret = generic_writepages(mapping, wbc);
1043         return ret;
1044 }
1045
1046 /**
1047  * write_one_page - write out a single page and optionally wait on I/O
1048  * @page: the page to write
1049  * @wait: if true, wait on writeout
1050  *
1051  * The page must be locked by the caller and will be unlocked upon return.
1052  *
1053  * write_one_page() returns a negative error code if I/O failed.
1054  */
1055 int write_one_page(struct page *page, int wait)
1056 {
1057         struct address_space *mapping = page->mapping;
1058         int ret = 0;
1059         struct writeback_control wbc = {
1060                 .sync_mode = WB_SYNC_ALL,
1061                 .nr_to_write = 1,
1062         };
1063
1064         BUG_ON(!PageLocked(page));
1065
1066         if (wait)
1067                 wait_on_page_writeback(page);
1068
1069         if (clear_page_dirty_for_io(page)) {
1070                 page_cache_get(page);
1071                 ret = mapping->a_ops->writepage(page, &wbc);
1072                 if (ret == 0 && wait) {
1073                         wait_on_page_writeback(page);
1074                         if (PageError(page))
1075                                 ret = -EIO;
1076                 }
1077                 page_cache_release(page);
1078         } else {
1079                 unlock_page(page);
1080         }
1081         return ret;
1082 }
1083 EXPORT_SYMBOL(write_one_page);
1084
1085 /*
1086  * For address_spaces which do not use buffers nor write back.
1087  */
1088 int __set_page_dirty_no_writeback(struct page *page)
1089 {
1090         if (!PageDirty(page))
1091                 SetPageDirty(page);
1092         return 0;
1093 }
1094
1095 /*
1096  * Helper function for set_page_dirty family.
1097  * NOTE: This relies on being atomic wrt interrupts.
1098  */
1099 void account_page_dirtied(struct page *page, struct address_space *mapping)
1100 {
1101         if (mapping_cap_account_dirty(mapping)) {
1102                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1103                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1104                 task_dirty_inc(current);
1105                 task_io_account_write(PAGE_CACHE_SIZE);
1106         }
1107 }
1108
1109 /*
1110  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1111  * its radix tree.
1112  *
1113  * This is also used when a single buffer is being dirtied: we want to set the
1114  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1115  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1116  *
1117  * Most callers have locked the page, which pins the address_space in memory.
1118  * But zap_pte_range() does not lock the page, however in that case the
1119  * mapping is pinned by the vma's ->vm_file reference.
1120  *
1121  * We take care to handle the case where the page was truncated from the
1122  * mapping by re-checking page_mapping() inside tree_lock.
1123  */
1124 int __set_page_dirty_nobuffers(struct page *page)
1125 {
1126         if (!TestSetPageDirty(page)) {
1127                 struct address_space *mapping = page_mapping(page);
1128                 struct address_space *mapping2;
1129
1130                 if (!mapping)
1131                         return 1;
1132
1133                 spin_lock_irq(&mapping->tree_lock);
1134                 mapping2 = page_mapping(page);
1135                 if (mapping2) { /* Race with truncate? */
1136                         BUG_ON(mapping2 != mapping);
1137                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1138                         account_page_dirtied(page, mapping);
1139                         radix_tree_tag_set(&mapping->page_tree,
1140                                 page_index(page), PAGECACHE_TAG_DIRTY);
1141                 }
1142                 spin_unlock_irq(&mapping->tree_lock);
1143                 if (mapping->host) {
1144                         /* !PageAnon && !swapper_space */
1145                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1146                 }
1147                 return 1;
1148         }
1149         return 0;
1150 }
1151 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1152
1153 /*
1154  * When a writepage implementation decides that it doesn't want to write this
1155  * page for some reason, it should redirty the locked page via
1156  * redirty_page_for_writepage() and it should then unlock the page and return 0
1157  */
1158 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1159 {
1160         wbc->pages_skipped++;
1161         return __set_page_dirty_nobuffers(page);
1162 }
1163 EXPORT_SYMBOL(redirty_page_for_writepage);
1164
1165 /*
1166  * Dirty a page.
1167  *
1168  * For pages with a mapping this should be done under the page lock
1169  * for the benefit of asynchronous memory errors who prefer a consistent
1170  * dirty state. This rule can be broken in some special cases,
1171  * but should be better not to.
1172  *
1173  * If the mapping doesn't provide a set_page_dirty a_op, then
1174  * just fall through and assume that it wants buffer_heads.
1175  */
1176 int set_page_dirty(struct page *page)
1177 {
1178         struct address_space *mapping = page_mapping(page);
1179
1180         if (likely(mapping)) {
1181                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1182 #ifdef CONFIG_BLOCK
1183                 if (!spd)
1184                         spd = __set_page_dirty_buffers;
1185 #endif
1186                 return (*spd)(page);
1187         }
1188         if (!PageDirty(page)) {
1189                 if (!TestSetPageDirty(page))
1190                         return 1;
1191         }
1192         return 0;
1193 }
1194 EXPORT_SYMBOL(set_page_dirty);
1195
1196 /*
1197  * set_page_dirty() is racy if the caller has no reference against
1198  * page->mapping->host, and if the page is unlocked.  This is because another
1199  * CPU could truncate the page off the mapping and then free the mapping.
1200  *
1201  * Usually, the page _is_ locked, or the caller is a user-space process which
1202  * holds a reference on the inode by having an open file.
1203  *
1204  * In other cases, the page should be locked before running set_page_dirty().
1205  */
1206 int set_page_dirty_lock(struct page *page)
1207 {
1208         int ret;
1209
1210         lock_page_nosync(page);
1211         ret = set_page_dirty(page);
1212         unlock_page(page);
1213         return ret;
1214 }
1215 EXPORT_SYMBOL(set_page_dirty_lock);
1216
1217 /*
1218  * Clear a page's dirty flag, while caring for dirty memory accounting.
1219  * Returns true if the page was previously dirty.
1220  *
1221  * This is for preparing to put the page under writeout.  We leave the page
1222  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1223  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1224  * implementation will run either set_page_writeback() or set_page_dirty(),
1225  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1226  * back into sync.
1227  *
1228  * This incoherency between the page's dirty flag and radix-tree tag is
1229  * unfortunate, but it only exists while the page is locked.
1230  */
1231 int clear_page_dirty_for_io(struct page *page)
1232 {
1233         struct address_space *mapping = page_mapping(page);
1234
1235         BUG_ON(!PageLocked(page));
1236
1237         ClearPageReclaim(page);
1238         if (mapping && mapping_cap_account_dirty(mapping)) {
1239                 /*
1240                  * Yes, Virginia, this is indeed insane.
1241                  *
1242                  * We use this sequence to make sure that
1243                  *  (a) we account for dirty stats properly
1244                  *  (b) we tell the low-level filesystem to
1245                  *      mark the whole page dirty if it was
1246                  *      dirty in a pagetable. Only to then
1247                  *  (c) clean the page again and return 1 to
1248                  *      cause the writeback.
1249                  *
1250                  * This way we avoid all nasty races with the
1251                  * dirty bit in multiple places and clearing
1252                  * them concurrently from different threads.
1253                  *
1254                  * Note! Normally the "set_page_dirty(page)"
1255                  * has no effect on the actual dirty bit - since
1256                  * that will already usually be set. But we
1257                  * need the side effects, and it can help us
1258                  * avoid races.
1259                  *
1260                  * We basically use the page "master dirty bit"
1261                  * as a serialization point for all the different
1262                  * threads doing their things.
1263                  */
1264                 if (page_mkclean(page))
1265                         set_page_dirty(page);
1266                 /*
1267                  * We carefully synchronise fault handlers against
1268                  * installing a dirty pte and marking the page dirty
1269                  * at this point. We do this by having them hold the
1270                  * page lock at some point after installing their
1271                  * pte, but before marking the page dirty.
1272                  * Pages are always locked coming in here, so we get
1273                  * the desired exclusion. See mm/memory.c:do_wp_page()
1274                  * for more comments.
1275                  */
1276                 if (TestClearPageDirty(page)) {
1277                         dec_zone_page_state(page, NR_FILE_DIRTY);
1278                         dec_bdi_stat(mapping->backing_dev_info,
1279                                         BDI_RECLAIMABLE);
1280                         return 1;
1281                 }
1282                 return 0;
1283         }
1284         return TestClearPageDirty(page);
1285 }
1286 EXPORT_SYMBOL(clear_page_dirty_for_io);
1287
1288 int test_clear_page_writeback(struct page *page)
1289 {
1290         struct address_space *mapping = page_mapping(page);
1291         int ret;
1292
1293         if (mapping) {
1294                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1295                 unsigned long flags;
1296
1297                 spin_lock_irqsave(&mapping->tree_lock, flags);
1298                 ret = TestClearPageWriteback(page);
1299                 if (ret) {
1300                         radix_tree_tag_clear(&mapping->page_tree,
1301                                                 page_index(page),
1302                                                 PAGECACHE_TAG_WRITEBACK);
1303                         if (bdi_cap_account_writeback(bdi)) {
1304                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1305                                 __bdi_writeout_inc(bdi);
1306                         }
1307                 }
1308                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1309         } else {
1310                 ret = TestClearPageWriteback(page);
1311         }
1312         if (ret)
1313                 dec_zone_page_state(page, NR_WRITEBACK);
1314         return ret;
1315 }
1316
1317 int test_set_page_writeback(struct page *page)
1318 {
1319         struct address_space *mapping = page_mapping(page);
1320         int ret;
1321
1322         if (mapping) {
1323                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1324                 unsigned long flags;
1325
1326                 spin_lock_irqsave(&mapping->tree_lock, flags);
1327                 ret = TestSetPageWriteback(page);
1328                 if (!ret) {
1329                         radix_tree_tag_set(&mapping->page_tree,
1330                                                 page_index(page),
1331                                                 PAGECACHE_TAG_WRITEBACK);
1332                         if (bdi_cap_account_writeback(bdi))
1333                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1334                 }
1335                 if (!PageDirty(page))
1336                         radix_tree_tag_clear(&mapping->page_tree,
1337                                                 page_index(page),
1338                                                 PAGECACHE_TAG_DIRTY);
1339                 radix_tree_tag_clear(&mapping->page_tree,
1340                                      page_index(page),
1341                                      PAGECACHE_TAG_TOWRITE);
1342                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1343         } else {
1344                 ret = TestSetPageWriteback(page);
1345         }
1346         if (!ret)
1347                 inc_zone_page_state(page, NR_WRITEBACK);
1348         return ret;
1349
1350 }
1351 EXPORT_SYMBOL(test_set_page_writeback);
1352
1353 /*
1354  * Return true if any of the pages in the mapping are marked with the
1355  * passed tag.
1356  */
1357 int mapping_tagged(struct address_space *mapping, int tag)
1358 {
1359         int ret;
1360         rcu_read_lock();
1361         ret = radix_tree_tagged(&mapping->page_tree, tag);
1362         rcu_read_unlock();
1363         return ret;
1364 }
1365 EXPORT_SYMBOL(mapping_tagged);