[PATCH] call mm/page-writeback.c:set_ratelimit() when new pages are hot-added
[linux-2.6.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c.
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains functions related to writing back dirty pages at the
7  * address_space level.
8  *
9  * 10Apr2002    akpm@zip.com.au
10  *              Initial version
11  */
12
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/spinlock.h>
16 #include <linux/fs.h>
17 #include <linux/mm.h>
18 #include <linux/swap.h>
19 #include <linux/slab.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/init.h>
23 #include <linux/backing-dev.h>
24 #include <linux/blkdev.h>
25 #include <linux/mpage.h>
26 #include <linux/rmap.h>
27 #include <linux/percpu.h>
28 #include <linux/notifier.h>
29 #include <linux/smp.h>
30 #include <linux/sysctl.h>
31 #include <linux/cpu.h>
32 #include <linux/syscalls.h>
33
34 /*
35  * The maximum number of pages to writeout in a single bdflush/kupdate
36  * operation.  We do this so we don't hold I_LOCK against an inode for
37  * enormous amounts of time, which would block a userspace task which has
38  * been forced to throttle against that inode.  Also, the code reevaluates
39  * the dirty each time it has written this many pages.
40  */
41 #define MAX_WRITEBACK_PAGES     1024
42
43 /*
44  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
45  * will look to see if it needs to force writeback or throttling.
46  */
47 static long ratelimit_pages = 32;
48
49 static int dirty_exceeded __cacheline_aligned_in_smp;   /* Dirty mem may be over limit */
50
51 /*
52  * When balance_dirty_pages decides that the caller needs to perform some
53  * non-background writeback, this is how many pages it will attempt to write.
54  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
55  * large amounts of I/O are submitted.
56  */
57 static inline long sync_writeback_pages(void)
58 {
59         return ratelimit_pages + ratelimit_pages / 2;
60 }
61
62 /* The following parameters are exported via /proc/sys/vm */
63
64 /*
65  * Start background writeback (via pdflush) at this percentage
66  */
67 int dirty_background_ratio = 10;
68
69 /*
70  * The generator of dirty data starts writeback at this percentage
71  */
72 int vm_dirty_ratio = 40;
73
74 /*
75  * The interval between `kupdate'-style writebacks, in jiffies
76  */
77 int dirty_writeback_interval = 5 * HZ;
78
79 /*
80  * The longest number of jiffies for which data is allowed to remain dirty
81  */
82 int dirty_expire_interval = 30 * HZ;
83
84 /*
85  * Flag that makes the machine dump writes/reads and block dirtyings.
86  */
87 int block_dump;
88
89 /*
90  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
91  * a full sync is triggered after this time elapses without any disk activity.
92  */
93 int laptop_mode;
94
95 EXPORT_SYMBOL(laptop_mode);
96
97 /* End of sysctl-exported parameters */
98
99
100 static void background_writeout(unsigned long _min_pages);
101
102 /*
103  * Work out the current dirty-memory clamping and background writeout
104  * thresholds.
105  *
106  * The main aim here is to lower them aggressively if there is a lot of mapped
107  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
108  * pages.  It is better to clamp down on writers than to start swapping, and
109  * performing lots of scanning.
110  *
111  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
112  *
113  * We don't permit the clamping level to fall below 5% - that is getting rather
114  * excessive.
115  *
116  * We make sure that the background writeout level is below the adjusted
117  * clamping level.
118  */
119 static void
120 get_dirty_limits(long *pbackground, long *pdirty,
121                                         struct address_space *mapping)
122 {
123         int background_ratio;           /* Percentages */
124         int dirty_ratio;
125         int unmapped_ratio;
126         long background;
127         long dirty;
128         unsigned long available_memory = vm_total_pages;
129         struct task_struct *tsk;
130
131 #ifdef CONFIG_HIGHMEM
132         /*
133          * If this mapping can only allocate from low memory,
134          * we exclude high memory from our count.
135          */
136         if (mapping && !(mapping_gfp_mask(mapping) & __GFP_HIGHMEM))
137                 available_memory -= totalhigh_pages;
138 #endif
139
140
141         unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
142                                 global_page_state(NR_ANON_PAGES)) * 100) /
143                                         vm_total_pages;
144
145         dirty_ratio = vm_dirty_ratio;
146         if (dirty_ratio > unmapped_ratio / 2)
147                 dirty_ratio = unmapped_ratio / 2;
148
149         if (dirty_ratio < 5)
150                 dirty_ratio = 5;
151
152         background_ratio = dirty_background_ratio;
153         if (background_ratio >= dirty_ratio)
154                 background_ratio = dirty_ratio / 2;
155
156         background = (background_ratio * available_memory) / 100;
157         dirty = (dirty_ratio * available_memory) / 100;
158         tsk = current;
159         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
160                 background += background / 4;
161                 dirty += dirty / 4;
162         }
163         *pbackground = background;
164         *pdirty = dirty;
165 }
166
167 /*
168  * balance_dirty_pages() must be called by processes which are generating dirty
169  * data.  It looks at the number of dirty pages in the machine and will force
170  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
171  * If we're over `background_thresh' then pdflush is woken to perform some
172  * writeout.
173  */
174 static void balance_dirty_pages(struct address_space *mapping)
175 {
176         long nr_reclaimable;
177         long background_thresh;
178         long dirty_thresh;
179         unsigned long pages_written = 0;
180         unsigned long write_chunk = sync_writeback_pages();
181
182         struct backing_dev_info *bdi = mapping->backing_dev_info;
183
184         for (;;) {
185                 struct writeback_control wbc = {
186                         .bdi            = bdi,
187                         .sync_mode      = WB_SYNC_NONE,
188                         .older_than_this = NULL,
189                         .nr_to_write    = write_chunk,
190                         .range_cyclic   = 1,
191                 };
192
193                 get_dirty_limits(&background_thresh, &dirty_thresh, mapping);
194                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
195                                         global_page_state(NR_UNSTABLE_NFS);
196                 if (nr_reclaimable + global_page_state(NR_WRITEBACK) <=
197                         dirty_thresh)
198                                 break;
199
200                 if (!dirty_exceeded)
201                         dirty_exceeded = 1;
202
203                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
204                  * Unstable writes are a feature of certain networked
205                  * filesystems (i.e. NFS) in which data may have been
206                  * written to the server's write cache, but has not yet
207                  * been flushed to permanent storage.
208                  */
209                 if (nr_reclaimable) {
210                         writeback_inodes(&wbc);
211                         get_dirty_limits(&background_thresh,
212                                                 &dirty_thresh, mapping);
213                         nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
214                                         global_page_state(NR_UNSTABLE_NFS);
215                         if (nr_reclaimable +
216                                 global_page_state(NR_WRITEBACK)
217                                         <= dirty_thresh)
218                                                 break;
219                         pages_written += write_chunk - wbc.nr_to_write;
220                         if (pages_written >= write_chunk)
221                                 break;          /* We've done our duty */
222                 }
223                 blk_congestion_wait(WRITE, HZ/10);
224         }
225
226         if (nr_reclaimable + global_page_state(NR_WRITEBACK)
227                 <= dirty_thresh && dirty_exceeded)
228                         dirty_exceeded = 0;
229
230         if (writeback_in_progress(bdi))
231                 return;         /* pdflush is already working this queue */
232
233         /*
234          * In laptop mode, we wait until hitting the higher threshold before
235          * starting background writeout, and then write out all the way down
236          * to the lower threshold.  So slow writers cause minimal disk activity.
237          *
238          * In normal mode, we start background writeout at the lower
239          * background_thresh, to keep the amount of dirty memory low.
240          */
241         if ((laptop_mode && pages_written) ||
242              (!laptop_mode && (nr_reclaimable > background_thresh)))
243                 pdflush_operation(background_writeout, 0);
244 }
245
246 void set_page_dirty_balance(struct page *page)
247 {
248         if (set_page_dirty(page)) {
249                 struct address_space *mapping = page_mapping(page);
250
251                 if (mapping)
252                         balance_dirty_pages_ratelimited(mapping);
253         }
254 }
255
256 /**
257  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
258  * @mapping: address_space which was dirtied
259  * @nr_pages_dirtied: number of pages which the caller has just dirtied
260  *
261  * Processes which are dirtying memory should call in here once for each page
262  * which was newly dirtied.  The function will periodically check the system's
263  * dirty state and will initiate writeback if needed.
264  *
265  * On really big machines, get_writeback_state is expensive, so try to avoid
266  * calling it too often (ratelimiting).  But once we're over the dirty memory
267  * limit we decrease the ratelimiting by a lot, to prevent individual processes
268  * from overshooting the limit by (ratelimit_pages) each.
269  */
270 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
271                                         unsigned long nr_pages_dirtied)
272 {
273         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
274         unsigned long ratelimit;
275         unsigned long *p;
276
277         ratelimit = ratelimit_pages;
278         if (dirty_exceeded)
279                 ratelimit = 8;
280
281         /*
282          * Check the rate limiting. Also, we do not want to throttle real-time
283          * tasks in balance_dirty_pages(). Period.
284          */
285         preempt_disable();
286         p =  &__get_cpu_var(ratelimits);
287         *p += nr_pages_dirtied;
288         if (unlikely(*p >= ratelimit)) {
289                 *p = 0;
290                 preempt_enable();
291                 balance_dirty_pages(mapping);
292                 return;
293         }
294         preempt_enable();
295 }
296 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
297
298 void throttle_vm_writeout(void)
299 {
300         long background_thresh;
301         long dirty_thresh;
302
303         for ( ; ; ) {
304                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
305
306                 /*
307                  * Boost the allowable dirty threshold a bit for page
308                  * allocators so they don't get DoS'ed by heavy writers
309                  */
310                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
311
312                 if (global_page_state(NR_UNSTABLE_NFS) +
313                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
314                                 break;
315                 blk_congestion_wait(WRITE, HZ/10);
316         }
317 }
318
319
320 /*
321  * writeback at least _min_pages, and keep writing until the amount of dirty
322  * memory is less than the background threshold, or until we're all clean.
323  */
324 static void background_writeout(unsigned long _min_pages)
325 {
326         long min_pages = _min_pages;
327         struct writeback_control wbc = {
328                 .bdi            = NULL,
329                 .sync_mode      = WB_SYNC_NONE,
330                 .older_than_this = NULL,
331                 .nr_to_write    = 0,
332                 .nonblocking    = 1,
333                 .range_cyclic   = 1,
334         };
335
336         for ( ; ; ) {
337                 long background_thresh;
338                 long dirty_thresh;
339
340                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
341                 if (global_page_state(NR_FILE_DIRTY) +
342                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
343                                 && min_pages <= 0)
344                         break;
345                 wbc.encountered_congestion = 0;
346                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
347                 wbc.pages_skipped = 0;
348                 writeback_inodes(&wbc);
349                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
350                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
351                         /* Wrote less than expected */
352                         blk_congestion_wait(WRITE, HZ/10);
353                         if (!wbc.encountered_congestion)
354                                 break;
355                 }
356         }
357 }
358
359 /*
360  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
361  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
362  * -1 if all pdflush threads were busy.
363  */
364 int wakeup_pdflush(long nr_pages)
365 {
366         if (nr_pages == 0)
367                 nr_pages = global_page_state(NR_FILE_DIRTY) +
368                                 global_page_state(NR_UNSTABLE_NFS);
369         return pdflush_operation(background_writeout, nr_pages);
370 }
371
372 static void wb_timer_fn(unsigned long unused);
373 static void laptop_timer_fn(unsigned long unused);
374
375 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
376 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
377
378 /*
379  * Periodic writeback of "old" data.
380  *
381  * Define "old": the first time one of an inode's pages is dirtied, we mark the
382  * dirtying-time in the inode's address_space.  So this periodic writeback code
383  * just walks the superblock inode list, writing back any inodes which are
384  * older than a specific point in time.
385  *
386  * Try to run once per dirty_writeback_interval.  But if a writeback event
387  * takes longer than a dirty_writeback_interval interval, then leave a
388  * one-second gap.
389  *
390  * older_than_this takes precedence over nr_to_write.  So we'll only write back
391  * all dirty pages if they are all attached to "old" mappings.
392  */
393 static void wb_kupdate(unsigned long arg)
394 {
395         unsigned long oldest_jif;
396         unsigned long start_jif;
397         unsigned long next_jif;
398         long nr_to_write;
399         struct writeback_control wbc = {
400                 .bdi            = NULL,
401                 .sync_mode      = WB_SYNC_NONE,
402                 .older_than_this = &oldest_jif,
403                 .nr_to_write    = 0,
404                 .nonblocking    = 1,
405                 .for_kupdate    = 1,
406                 .range_cyclic   = 1,
407         };
408
409         sync_supers();
410
411         oldest_jif = jiffies - dirty_expire_interval;
412         start_jif = jiffies;
413         next_jif = start_jif + dirty_writeback_interval;
414         nr_to_write = global_page_state(NR_FILE_DIRTY) +
415                         global_page_state(NR_UNSTABLE_NFS) +
416                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
417         while (nr_to_write > 0) {
418                 wbc.encountered_congestion = 0;
419                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
420                 writeback_inodes(&wbc);
421                 if (wbc.nr_to_write > 0) {
422                         if (wbc.encountered_congestion)
423                                 blk_congestion_wait(WRITE, HZ/10);
424                         else
425                                 break;  /* All the old data is written */
426                 }
427                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
428         }
429         if (time_before(next_jif, jiffies + HZ))
430                 next_jif = jiffies + HZ;
431         if (dirty_writeback_interval)
432                 mod_timer(&wb_timer, next_jif);
433 }
434
435 /*
436  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
437  */
438 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
439                 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
440 {
441         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
442         if (dirty_writeback_interval) {
443                 mod_timer(&wb_timer,
444                         jiffies + dirty_writeback_interval);
445                 } else {
446                 del_timer(&wb_timer);
447         }
448         return 0;
449 }
450
451 static void wb_timer_fn(unsigned long unused)
452 {
453         if (pdflush_operation(wb_kupdate, 0) < 0)
454                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
455 }
456
457 static void laptop_flush(unsigned long unused)
458 {
459         sys_sync();
460 }
461
462 static void laptop_timer_fn(unsigned long unused)
463 {
464         pdflush_operation(laptop_flush, 0);
465 }
466
467 /*
468  * We've spun up the disk and we're in laptop mode: schedule writeback
469  * of all dirty data a few seconds from now.  If the flush is already scheduled
470  * then push it back - the user is still using the disk.
471  */
472 void laptop_io_completion(void)
473 {
474         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
475 }
476
477 /*
478  * We're in laptop mode and we've just synced. The sync's writes will have
479  * caused another writeback to be scheduled by laptop_io_completion.
480  * Nothing needs to be written back anymore, so we unschedule the writeback.
481  */
482 void laptop_sync_completion(void)
483 {
484         del_timer(&laptop_mode_wb_timer);
485 }
486
487 /*
488  * If ratelimit_pages is too high then we can get into dirty-data overload
489  * if a large number of processes all perform writes at the same time.
490  * If it is too low then SMP machines will call the (expensive)
491  * get_writeback_state too often.
492  *
493  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
494  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
495  * thresholds before writeback cuts in.
496  *
497  * But the limit should not be set too high.  Because it also controls the
498  * amount of memory which the balance_dirty_pages() caller has to write back.
499  * If this is too large then the caller will block on the IO queue all the
500  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
501  * will write six megabyte chunks, max.
502  */
503
504 void writeback_set_ratelimit(void)
505 {
506         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
507         if (ratelimit_pages < 16)
508                 ratelimit_pages = 16;
509         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
510                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
511 }
512
513 static int __cpuinit
514 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
515 {
516         writeback_set_ratelimit();
517         return 0;
518 }
519
520 static struct notifier_block __cpuinitdata ratelimit_nb = {
521         .notifier_call  = ratelimit_handler,
522         .next           = NULL,
523 };
524
525 /*
526  * If the machine has a large highmem:lowmem ratio then scale back the default
527  * dirty memory thresholds: allowing too much dirty highmem pins an excessive
528  * number of buffer_heads.
529  */
530 void __init page_writeback_init(void)
531 {
532         long buffer_pages = nr_free_buffer_pages();
533         long correction;
534
535         correction = (100 * 4 * buffer_pages) / vm_total_pages;
536
537         if (correction < 100) {
538                 dirty_background_ratio *= correction;
539                 dirty_background_ratio /= 100;
540                 vm_dirty_ratio *= correction;
541                 vm_dirty_ratio /= 100;
542
543                 if (dirty_background_ratio <= 0)
544                         dirty_background_ratio = 1;
545                 if (vm_dirty_ratio <= 0)
546                         vm_dirty_ratio = 1;
547         }
548         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
549         writeback_set_ratelimit();
550         register_cpu_notifier(&ratelimit_nb);
551 }
552
553 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
554 {
555         int ret;
556
557         if (wbc->nr_to_write <= 0)
558                 return 0;
559         wbc->for_writepages = 1;
560         if (mapping->a_ops->writepages)
561                 ret = mapping->a_ops->writepages(mapping, wbc);
562         else
563                 ret = generic_writepages(mapping, wbc);
564         wbc->for_writepages = 0;
565         return ret;
566 }
567
568 /**
569  * write_one_page - write out a single page and optionally wait on I/O
570  *
571  * @page: the page to write
572  * @wait: if true, wait on writeout
573  *
574  * The page must be locked by the caller and will be unlocked upon return.
575  *
576  * write_one_page() returns a negative error code if I/O failed.
577  */
578 int write_one_page(struct page *page, int wait)
579 {
580         struct address_space *mapping = page->mapping;
581         int ret = 0;
582         struct writeback_control wbc = {
583                 .sync_mode = WB_SYNC_ALL,
584                 .nr_to_write = 1,
585         };
586
587         BUG_ON(!PageLocked(page));
588
589         if (wait)
590                 wait_on_page_writeback(page);
591
592         if (clear_page_dirty_for_io(page)) {
593                 page_cache_get(page);
594                 ret = mapping->a_ops->writepage(page, &wbc);
595                 if (ret == 0 && wait) {
596                         wait_on_page_writeback(page);
597                         if (PageError(page))
598                                 ret = -EIO;
599                 }
600                 page_cache_release(page);
601         } else {
602                 unlock_page(page);
603         }
604         return ret;
605 }
606 EXPORT_SYMBOL(write_one_page);
607
608 /*
609  * For address_spaces which do not use buffers.  Just tag the page as dirty in
610  * its radix tree.
611  *
612  * This is also used when a single buffer is being dirtied: we want to set the
613  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
614  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
615  *
616  * Most callers have locked the page, which pins the address_space in memory.
617  * But zap_pte_range() does not lock the page, however in that case the
618  * mapping is pinned by the vma's ->vm_file reference.
619  *
620  * We take care to handle the case where the page was truncated from the
621  * mapping by re-checking page_mapping() insode tree_lock.
622  */
623 int __set_page_dirty_nobuffers(struct page *page)
624 {
625         if (!TestSetPageDirty(page)) {
626                 struct address_space *mapping = page_mapping(page);
627                 struct address_space *mapping2;
628
629                 if (mapping) {
630                         write_lock_irq(&mapping->tree_lock);
631                         mapping2 = page_mapping(page);
632                         if (mapping2) { /* Race with truncate? */
633                                 BUG_ON(mapping2 != mapping);
634                                 if (mapping_cap_account_dirty(mapping))
635                                         __inc_zone_page_state(page,
636                                                                 NR_FILE_DIRTY);
637                                 radix_tree_tag_set(&mapping->page_tree,
638                                         page_index(page), PAGECACHE_TAG_DIRTY);
639                         }
640                         write_unlock_irq(&mapping->tree_lock);
641                         if (mapping->host) {
642                                 /* !PageAnon && !swapper_space */
643                                 __mark_inode_dirty(mapping->host,
644                                                         I_DIRTY_PAGES);
645                         }
646                 }
647                 return 1;
648         }
649         return 0;
650 }
651 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
652
653 /*
654  * When a writepage implementation decides that it doesn't want to write this
655  * page for some reason, it should redirty the locked page via
656  * redirty_page_for_writepage() and it should then unlock the page and return 0
657  */
658 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
659 {
660         wbc->pages_skipped++;
661         return __set_page_dirty_nobuffers(page);
662 }
663 EXPORT_SYMBOL(redirty_page_for_writepage);
664
665 /*
666  * If the mapping doesn't provide a set_page_dirty a_op, then
667  * just fall through and assume that it wants buffer_heads.
668  */
669 int fastcall set_page_dirty(struct page *page)
670 {
671         struct address_space *mapping = page_mapping(page);
672
673         if (likely(mapping)) {
674                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
675                 if (spd)
676                         return (*spd)(page);
677                 return __set_page_dirty_buffers(page);
678         }
679         if (!PageDirty(page)) {
680                 if (!TestSetPageDirty(page))
681                         return 1;
682         }
683         return 0;
684 }
685 EXPORT_SYMBOL(set_page_dirty);
686
687 /*
688  * set_page_dirty() is racy if the caller has no reference against
689  * page->mapping->host, and if the page is unlocked.  This is because another
690  * CPU could truncate the page off the mapping and then free the mapping.
691  *
692  * Usually, the page _is_ locked, or the caller is a user-space process which
693  * holds a reference on the inode by having an open file.
694  *
695  * In other cases, the page should be locked before running set_page_dirty().
696  */
697 int set_page_dirty_lock(struct page *page)
698 {
699         int ret;
700
701         lock_page_nosync(page);
702         ret = set_page_dirty(page);
703         unlock_page(page);
704         return ret;
705 }
706 EXPORT_SYMBOL(set_page_dirty_lock);
707
708 /*
709  * Clear a page's dirty flag, while caring for dirty memory accounting. 
710  * Returns true if the page was previously dirty.
711  */
712 int test_clear_page_dirty(struct page *page)
713 {
714         struct address_space *mapping = page_mapping(page);
715         unsigned long flags;
716
717         if (mapping) {
718                 write_lock_irqsave(&mapping->tree_lock, flags);
719                 if (TestClearPageDirty(page)) {
720                         radix_tree_tag_clear(&mapping->page_tree,
721                                                 page_index(page),
722                                                 PAGECACHE_TAG_DIRTY);
723                         write_unlock_irqrestore(&mapping->tree_lock, flags);
724                         /*
725                          * We can continue to use `mapping' here because the
726                          * page is locked, which pins the address_space
727                          */
728                         if (mapping_cap_account_dirty(mapping)) {
729                                 page_mkclean(page);
730                                 dec_zone_page_state(page, NR_FILE_DIRTY);
731                         }
732                         return 1;
733                 }
734                 write_unlock_irqrestore(&mapping->tree_lock, flags);
735                 return 0;
736         }
737         return TestClearPageDirty(page);
738 }
739 EXPORT_SYMBOL(test_clear_page_dirty);
740
741 /*
742  * Clear a page's dirty flag, while caring for dirty memory accounting.
743  * Returns true if the page was previously dirty.
744  *
745  * This is for preparing to put the page under writeout.  We leave the page
746  * tagged as dirty in the radix tree so that a concurrent write-for-sync
747  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
748  * implementation will run either set_page_writeback() or set_page_dirty(),
749  * at which stage we bring the page's dirty flag and radix-tree dirty tag
750  * back into sync.
751  *
752  * This incoherency between the page's dirty flag and radix-tree tag is
753  * unfortunate, but it only exists while the page is locked.
754  */
755 int clear_page_dirty_for_io(struct page *page)
756 {
757         struct address_space *mapping = page_mapping(page);
758
759         if (mapping) {
760                 if (TestClearPageDirty(page)) {
761                         if (mapping_cap_account_dirty(mapping)) {
762                                 page_mkclean(page);
763                                 dec_zone_page_state(page, NR_FILE_DIRTY);
764                         }
765                         return 1;
766                 }
767                 return 0;
768         }
769         return TestClearPageDirty(page);
770 }
771 EXPORT_SYMBOL(clear_page_dirty_for_io);
772
773 int test_clear_page_writeback(struct page *page)
774 {
775         struct address_space *mapping = page_mapping(page);
776         int ret;
777
778         if (mapping) {
779                 unsigned long flags;
780
781                 write_lock_irqsave(&mapping->tree_lock, flags);
782                 ret = TestClearPageWriteback(page);
783                 if (ret)
784                         radix_tree_tag_clear(&mapping->page_tree,
785                                                 page_index(page),
786                                                 PAGECACHE_TAG_WRITEBACK);
787                 write_unlock_irqrestore(&mapping->tree_lock, flags);
788         } else {
789                 ret = TestClearPageWriteback(page);
790         }
791         return ret;
792 }
793
794 int test_set_page_writeback(struct page *page)
795 {
796         struct address_space *mapping = page_mapping(page);
797         int ret;
798
799         if (mapping) {
800                 unsigned long flags;
801
802                 write_lock_irqsave(&mapping->tree_lock, flags);
803                 ret = TestSetPageWriteback(page);
804                 if (!ret)
805                         radix_tree_tag_set(&mapping->page_tree,
806                                                 page_index(page),
807                                                 PAGECACHE_TAG_WRITEBACK);
808                 if (!PageDirty(page))
809                         radix_tree_tag_clear(&mapping->page_tree,
810                                                 page_index(page),
811                                                 PAGECACHE_TAG_DIRTY);
812                 write_unlock_irqrestore(&mapping->tree_lock, flags);
813         } else {
814                 ret = TestSetPageWriteback(page);
815         }
816         return ret;
817
818 }
819 EXPORT_SYMBOL(test_set_page_writeback);
820
821 /*
822  * Wakes up tasks that are being throttled due to writeback congestion
823  */
824 void writeback_congestion_end(void)
825 {
826         blk_congestion_end(WRITE);
827 }
828 EXPORT_SYMBOL(writeback_congestion_end);
829
830 /*
831  * Return true if any of the pages in the mapping are marged with the
832  * passed tag.
833  */
834 int mapping_tagged(struct address_space *mapping, int tag)
835 {
836         unsigned long flags;
837         int ret;
838
839         read_lock_irqsave(&mapping->tree_lock, flags);
840         ret = radix_tree_tagged(&mapping->page_tree, tag);
841         read_unlock_irqrestore(&mapping->tree_lock, flags);
842         return ret;
843 }
844 EXPORT_SYMBOL(mapping_tagged);