[PATCH] writeback: fix range handling
[linux-3.10.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/percpu.h>
27 #include <linux/notifier.h>
28 #include <linux/smp.h>
29 #include <linux/sysctl.h>
30 #include <linux/cpu.h>
31 #include <linux/syscalls.h>
32
33 /*
34  * The maximum number of pages to writeout in a single bdflush/kupdate
35  * operation.  We do this so we don't hold I_LOCK against an inode for
36  * enormous amounts of time, which would block a userspace task which has
37  * been forced to throttle against that inode.  Also, the code reevaluates
38  * the dirty each time it has written this many pages.
39  */
40 #define MAX_WRITEBACK_PAGES     1024
41
42 /*
43  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
44  * will look to see if it needs to force writeback or throttling.
45  */
46 static long ratelimit_pages = 32;
47
48 static long total_pages;        /* The total number of pages in the machine. */
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 struct writeback_state
103 {
104         unsigned long nr_dirty;
105         unsigned long nr_unstable;
106         unsigned long nr_mapped;
107         unsigned long nr_writeback;
108 };
109
110 static void get_writeback_state(struct writeback_state *wbs)
111 {
112         wbs->nr_dirty = read_page_state(nr_dirty);
113         wbs->nr_unstable = read_page_state(nr_unstable);
114         wbs->nr_mapped = read_page_state(nr_mapped);
115         wbs->nr_writeback = read_page_state(nr_writeback);
116 }
117
118 /*
119  * Work out the current dirty-memory clamping and background writeout
120  * thresholds.
121  *
122  * The main aim here is to lower them aggressively if there is a lot of mapped
123  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
124  * pages.  It is better to clamp down on writers than to start swapping, and
125  * performing lots of scanning.
126  *
127  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
128  *
129  * We don't permit the clamping level to fall below 5% - that is getting rather
130  * excessive.
131  *
132  * We make sure that the background writeout level is below the adjusted
133  * clamping level.
134  */
135 static void
136 get_dirty_limits(struct writeback_state *wbs, long *pbackground, long *pdirty,
137                 struct address_space *mapping)
138 {
139         int background_ratio;           /* Percentages */
140         int dirty_ratio;
141         int unmapped_ratio;
142         long background;
143         long dirty;
144         unsigned long available_memory = total_pages;
145         struct task_struct *tsk;
146
147         get_writeback_state(wbs);
148
149 #ifdef CONFIG_HIGHMEM
150         /*
151          * If this mapping can only allocate from low memory,
152          * we exclude high memory from our count.
153          */
154         if (mapping && !(mapping_gfp_mask(mapping) & __GFP_HIGHMEM))
155                 available_memory -= totalhigh_pages;
156 #endif
157
158
159         unmapped_ratio = 100 - (wbs->nr_mapped * 100) / total_pages;
160
161         dirty_ratio = vm_dirty_ratio;
162         if (dirty_ratio > unmapped_ratio / 2)
163                 dirty_ratio = unmapped_ratio / 2;
164
165         if (dirty_ratio < 5)
166                 dirty_ratio = 5;
167
168         background_ratio = dirty_background_ratio;
169         if (background_ratio >= dirty_ratio)
170                 background_ratio = dirty_ratio / 2;
171
172         background = (background_ratio * available_memory) / 100;
173         dirty = (dirty_ratio * available_memory) / 100;
174         tsk = current;
175         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
176                 background += background / 4;
177                 dirty += dirty / 4;
178         }
179         *pbackground = background;
180         *pdirty = dirty;
181 }
182
183 /*
184  * balance_dirty_pages() must be called by processes which are generating dirty
185  * data.  It looks at the number of dirty pages in the machine and will force
186  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
187  * If we're over `background_thresh' then pdflush is woken to perform some
188  * writeout.
189  */
190 static void balance_dirty_pages(struct address_space *mapping)
191 {
192         struct writeback_state wbs;
193         long nr_reclaimable;
194         long background_thresh;
195         long dirty_thresh;
196         unsigned long pages_written = 0;
197         unsigned long write_chunk = sync_writeback_pages();
198
199         struct backing_dev_info *bdi = mapping->backing_dev_info;
200
201         for (;;) {
202                 struct writeback_control wbc = {
203                         .bdi            = bdi,
204                         .sync_mode      = WB_SYNC_NONE,
205                         .older_than_this = NULL,
206                         .nr_to_write    = write_chunk,
207                         .range_cyclic   = 1,
208                 };
209
210                 get_dirty_limits(&wbs, &background_thresh,
211                                         &dirty_thresh, mapping);
212                 nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
213                 if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
214                         break;
215
216                 if (!dirty_exceeded)
217                         dirty_exceeded = 1;
218
219                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
220                  * Unstable writes are a feature of certain networked
221                  * filesystems (i.e. NFS) in which data may have been
222                  * written to the server's write cache, but has not yet
223                  * been flushed to permanent storage.
224                  */
225                 if (nr_reclaimable) {
226                         writeback_inodes(&wbc);
227                         get_dirty_limits(&wbs, &background_thresh,
228                                         &dirty_thresh, mapping);
229                         nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
230                         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
231                                 break;
232                         pages_written += write_chunk - wbc.nr_to_write;
233                         if (pages_written >= write_chunk)
234                                 break;          /* We've done our duty */
235                 }
236                 blk_congestion_wait(WRITE, HZ/10);
237         }
238
239         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh && dirty_exceeded)
240                 dirty_exceeded = 0;
241
242         if (writeback_in_progress(bdi))
243                 return;         /* pdflush is already working this queue */
244
245         /*
246          * In laptop mode, we wait until hitting the higher threshold before
247          * starting background writeout, and then write out all the way down
248          * to the lower threshold.  So slow writers cause minimal disk activity.
249          *
250          * In normal mode, we start background writeout at the lower
251          * background_thresh, to keep the amount of dirty memory low.
252          */
253         if ((laptop_mode && pages_written) ||
254              (!laptop_mode && (nr_reclaimable > background_thresh)))
255                 pdflush_operation(background_writeout, 0);
256 }
257
258 /**
259  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
260  * @mapping: address_space which was dirtied
261  * @nr_pages_dirtied: number of pages which the caller has just dirtied
262  *
263  * Processes which are dirtying memory should call in here once for each page
264  * which was newly dirtied.  The function will periodically check the system's
265  * dirty state and will initiate writeback if needed.
266  *
267  * On really big machines, get_writeback_state is expensive, so try to avoid
268  * calling it too often (ratelimiting).  But once we're over the dirty memory
269  * limit we decrease the ratelimiting by a lot, to prevent individual processes
270  * from overshooting the limit by (ratelimit_pages) each.
271  */
272 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
273                                         unsigned long nr_pages_dirtied)
274 {
275         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
276         unsigned long ratelimit;
277         unsigned long *p;
278
279         ratelimit = ratelimit_pages;
280         if (dirty_exceeded)
281                 ratelimit = 8;
282
283         /*
284          * Check the rate limiting. Also, we do not want to throttle real-time
285          * tasks in balance_dirty_pages(). Period.
286          */
287         preempt_disable();
288         p =  &__get_cpu_var(ratelimits);
289         *p += nr_pages_dirtied;
290         if (unlikely(*p >= ratelimit)) {
291                 *p = 0;
292                 preempt_enable();
293                 balance_dirty_pages(mapping);
294                 return;
295         }
296         preempt_enable();
297 }
298 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
299
300 void throttle_vm_writeout(void)
301 {
302         struct writeback_state wbs;
303         long background_thresh;
304         long dirty_thresh;
305
306         for ( ; ; ) {
307                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
308
309                 /*
310                  * Boost the allowable dirty threshold a bit for page
311                  * allocators so they don't get DoS'ed by heavy writers
312                  */
313                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
314
315                 if (wbs.nr_unstable + wbs.nr_writeback <= dirty_thresh)
316                         break;
317                 blk_congestion_wait(WRITE, HZ/10);
318         }
319 }
320
321
322 /*
323  * writeback at least _min_pages, and keep writing until the amount of dirty
324  * memory is less than the background threshold, or until we're all clean.
325  */
326 static void background_writeout(unsigned long _min_pages)
327 {
328         long min_pages = _min_pages;
329         struct writeback_control wbc = {
330                 .bdi            = NULL,
331                 .sync_mode      = WB_SYNC_NONE,
332                 .older_than_this = NULL,
333                 .nr_to_write    = 0,
334                 .nonblocking    = 1,
335                 .range_cyclic   = 1,
336         };
337
338         for ( ; ; ) {
339                 struct writeback_state wbs;
340                 long background_thresh;
341                 long dirty_thresh;
342
343                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
344                 if (wbs.nr_dirty + wbs.nr_unstable < background_thresh
345                                 && min_pages <= 0)
346                         break;
347                 wbc.encountered_congestion = 0;
348                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
349                 wbc.pages_skipped = 0;
350                 writeback_inodes(&wbc);
351                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
352                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
353                         /* Wrote less than expected */
354                         blk_congestion_wait(WRITE, HZ/10);
355                         if (!wbc.encountered_congestion)
356                                 break;
357                 }
358         }
359 }
360
361 /*
362  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
363  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
364  * -1 if all pdflush threads were busy.
365  */
366 int wakeup_pdflush(long nr_pages)
367 {
368         if (nr_pages == 0) {
369                 struct writeback_state wbs;
370
371                 get_writeback_state(&wbs);
372                 nr_pages = wbs.nr_dirty + wbs.nr_unstable;
373         }
374         return pdflush_operation(background_writeout, nr_pages);
375 }
376
377 static void wb_timer_fn(unsigned long unused);
378 static void laptop_timer_fn(unsigned long unused);
379
380 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
381 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
382
383 /*
384  * Periodic writeback of "old" data.
385  *
386  * Define "old": the first time one of an inode's pages is dirtied, we mark the
387  * dirtying-time in the inode's address_space.  So this periodic writeback code
388  * just walks the superblock inode list, writing back any inodes which are
389  * older than a specific point in time.
390  *
391  * Try to run once per dirty_writeback_interval.  But if a writeback event
392  * takes longer than a dirty_writeback_interval interval, then leave a
393  * one-second gap.
394  *
395  * older_than_this takes precedence over nr_to_write.  So we'll only write back
396  * all dirty pages if they are all attached to "old" mappings.
397  */
398 static void wb_kupdate(unsigned long arg)
399 {
400         unsigned long oldest_jif;
401         unsigned long start_jif;
402         unsigned long next_jif;
403         long nr_to_write;
404         struct writeback_state wbs;
405         struct writeback_control wbc = {
406                 .bdi            = NULL,
407                 .sync_mode      = WB_SYNC_NONE,
408                 .older_than_this = &oldest_jif,
409                 .nr_to_write    = 0,
410                 .nonblocking    = 1,
411                 .for_kupdate    = 1,
412                 .range_cyclic   = 1,
413         };
414
415         sync_supers();
416
417         get_writeback_state(&wbs);
418         oldest_jif = jiffies - dirty_expire_interval;
419         start_jif = jiffies;
420         next_jif = start_jif + dirty_writeback_interval;
421         nr_to_write = wbs.nr_dirty + wbs.nr_unstable +
422                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
423         while (nr_to_write > 0) {
424                 wbc.encountered_congestion = 0;
425                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
426                 writeback_inodes(&wbc);
427                 if (wbc.nr_to_write > 0) {
428                         if (wbc.encountered_congestion)
429                                 blk_congestion_wait(WRITE, HZ/10);
430                         else
431                                 break;  /* All the old data is written */
432                 }
433                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
434         }
435         if (time_before(next_jif, jiffies + HZ))
436                 next_jif = jiffies + HZ;
437         if (dirty_writeback_interval)
438                 mod_timer(&wb_timer, next_jif);
439 }
440
441 /*
442  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
443  */
444 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
445                 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
446 {
447         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
448         if (dirty_writeback_interval) {
449                 mod_timer(&wb_timer,
450                         jiffies + dirty_writeback_interval);
451                 } else {
452                 del_timer(&wb_timer);
453         }
454         return 0;
455 }
456
457 static void wb_timer_fn(unsigned long unused)
458 {
459         if (pdflush_operation(wb_kupdate, 0) < 0)
460                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
461 }
462
463 static void laptop_flush(unsigned long unused)
464 {
465         sys_sync();
466 }
467
468 static void laptop_timer_fn(unsigned long unused)
469 {
470         pdflush_operation(laptop_flush, 0);
471 }
472
473 /*
474  * We've spun up the disk and we're in laptop mode: schedule writeback
475  * of all dirty data a few seconds from now.  If the flush is already scheduled
476  * then push it back - the user is still using the disk.
477  */
478 void laptop_io_completion(void)
479 {
480         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
481 }
482
483 /*
484  * We're in laptop mode and we've just synced. The sync's writes will have
485  * caused another writeback to be scheduled by laptop_io_completion.
486  * Nothing needs to be written back anymore, so we unschedule the writeback.
487  */
488 void laptop_sync_completion(void)
489 {
490         del_timer(&laptop_mode_wb_timer);
491 }
492
493 /*
494  * If ratelimit_pages is too high then we can get into dirty-data overload
495  * if a large number of processes all perform writes at the same time.
496  * If it is too low then SMP machines will call the (expensive)
497  * get_writeback_state too often.
498  *
499  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
500  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
501  * thresholds before writeback cuts in.
502  *
503  * But the limit should not be set too high.  Because it also controls the
504  * amount of memory which the balance_dirty_pages() caller has to write back.
505  * If this is too large then the caller will block on the IO queue all the
506  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
507  * will write six megabyte chunks, max.
508  */
509
510 static void set_ratelimit(void)
511 {
512         ratelimit_pages = total_pages / (num_online_cpus() * 32);
513         if (ratelimit_pages < 16)
514                 ratelimit_pages = 16;
515         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
516                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
517 }
518
519 static int
520 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
521 {
522         set_ratelimit();
523         return 0;
524 }
525
526 static struct notifier_block ratelimit_nb = {
527         .notifier_call  = ratelimit_handler,
528         .next           = NULL,
529 };
530
531 /*
532  * If the machine has a large highmem:lowmem ratio then scale back the default
533  * dirty memory thresholds: allowing too much dirty highmem pins an excessive
534  * number of buffer_heads.
535  */
536 void __init page_writeback_init(void)
537 {
538         long buffer_pages = nr_free_buffer_pages();
539         long correction;
540
541         total_pages = nr_free_pagecache_pages();
542
543         correction = (100 * 4 * buffer_pages) / total_pages;
544
545         if (correction < 100) {
546                 dirty_background_ratio *= correction;
547                 dirty_background_ratio /= 100;
548                 vm_dirty_ratio *= correction;
549                 vm_dirty_ratio /= 100;
550
551                 if (dirty_background_ratio <= 0)
552                         dirty_background_ratio = 1;
553                 if (vm_dirty_ratio <= 0)
554                         vm_dirty_ratio = 1;
555         }
556         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
557         set_ratelimit();
558         register_cpu_notifier(&ratelimit_nb);
559 }
560
561 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
562 {
563         int ret;
564
565         if (wbc->nr_to_write <= 0)
566                 return 0;
567         wbc->for_writepages = 1;
568         if (mapping->a_ops->writepages)
569                 ret =  mapping->a_ops->writepages(mapping, wbc);
570         else
571                 ret = generic_writepages(mapping, wbc);
572         wbc->for_writepages = 0;
573         return ret;
574 }
575
576 /**
577  * write_one_page - write out a single page and optionally wait on I/O
578  *
579  * @page: the page to write
580  * @wait: if true, wait on writeout
581  *
582  * The page must be locked by the caller and will be unlocked upon return.
583  *
584  * write_one_page() returns a negative error code if I/O failed.
585  */
586 int write_one_page(struct page *page, int wait)
587 {
588         struct address_space *mapping = page->mapping;
589         int ret = 0;
590         struct writeback_control wbc = {
591                 .sync_mode = WB_SYNC_ALL,
592                 .nr_to_write = 1,
593         };
594
595         BUG_ON(!PageLocked(page));
596
597         if (wait)
598                 wait_on_page_writeback(page);
599
600         if (clear_page_dirty_for_io(page)) {
601                 page_cache_get(page);
602                 ret = mapping->a_ops->writepage(page, &wbc);
603                 if (ret == 0 && wait) {
604                         wait_on_page_writeback(page);
605                         if (PageError(page))
606                                 ret = -EIO;
607                 }
608                 page_cache_release(page);
609         } else {
610                 unlock_page(page);
611         }
612         return ret;
613 }
614 EXPORT_SYMBOL(write_one_page);
615
616 /*
617  * For address_spaces which do not use buffers.  Just tag the page as dirty in
618  * its radix tree.
619  *
620  * This is also used when a single buffer is being dirtied: we want to set the
621  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
622  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
623  *
624  * Most callers have locked the page, which pins the address_space in memory.
625  * But zap_pte_range() does not lock the page, however in that case the
626  * mapping is pinned by the vma's ->vm_file reference.
627  *
628  * We take care to handle the case where the page was truncated from the
629  * mapping by re-checking page_mapping() insode tree_lock.
630  */
631 int __set_page_dirty_nobuffers(struct page *page)
632 {
633         if (!TestSetPageDirty(page)) {
634                 struct address_space *mapping = page_mapping(page);
635                 struct address_space *mapping2;
636
637                 if (mapping) {
638                         write_lock_irq(&mapping->tree_lock);
639                         mapping2 = page_mapping(page);
640                         if (mapping2) { /* Race with truncate? */
641                                 BUG_ON(mapping2 != mapping);
642                                 if (mapping_cap_account_dirty(mapping))
643                                         inc_page_state(nr_dirty);
644                                 radix_tree_tag_set(&mapping->page_tree,
645                                         page_index(page), PAGECACHE_TAG_DIRTY);
646                         }
647                         write_unlock_irq(&mapping->tree_lock);
648                         if (mapping->host) {
649                                 /* !PageAnon && !swapper_space */
650                                 __mark_inode_dirty(mapping->host,
651                                                         I_DIRTY_PAGES);
652                         }
653                 }
654                 return 1;
655         }
656         return 0;
657 }
658 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
659
660 /*
661  * When a writepage implementation decides that it doesn't want to write this
662  * page for some reason, it should redirty the locked page via
663  * redirty_page_for_writepage() and it should then unlock the page and return 0
664  */
665 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
666 {
667         wbc->pages_skipped++;
668         return __set_page_dirty_nobuffers(page);
669 }
670 EXPORT_SYMBOL(redirty_page_for_writepage);
671
672 /*
673  * If the mapping doesn't provide a set_page_dirty a_op, then
674  * just fall through and assume that it wants buffer_heads.
675  */
676 int fastcall set_page_dirty(struct page *page)
677 {
678         struct address_space *mapping = page_mapping(page);
679
680         if (likely(mapping)) {
681                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
682                 if (spd)
683                         return (*spd)(page);
684                 return __set_page_dirty_buffers(page);
685         }
686         if (!PageDirty(page)) {
687                 if (!TestSetPageDirty(page))
688                         return 1;
689         }
690         return 0;
691 }
692 EXPORT_SYMBOL(set_page_dirty);
693
694 /*
695  * set_page_dirty() is racy if the caller has no reference against
696  * page->mapping->host, and if the page is unlocked.  This is because another
697  * CPU could truncate the page off the mapping and then free the mapping.
698  *
699  * Usually, the page _is_ locked, or the caller is a user-space process which
700  * holds a reference on the inode by having an open file.
701  *
702  * In other cases, the page should be locked before running set_page_dirty().
703  */
704 int set_page_dirty_lock(struct page *page)
705 {
706         int ret;
707
708         lock_page(page);
709         ret = set_page_dirty(page);
710         unlock_page(page);
711         return ret;
712 }
713 EXPORT_SYMBOL(set_page_dirty_lock);
714
715 /*
716  * Clear a page's dirty flag, while caring for dirty memory accounting. 
717  * Returns true if the page was previously dirty.
718  */
719 int test_clear_page_dirty(struct page *page)
720 {
721         struct address_space *mapping = page_mapping(page);
722         unsigned long flags;
723
724         if (mapping) {
725                 write_lock_irqsave(&mapping->tree_lock, flags);
726                 if (TestClearPageDirty(page)) {
727                         radix_tree_tag_clear(&mapping->page_tree,
728                                                 page_index(page),
729                                                 PAGECACHE_TAG_DIRTY);
730                         write_unlock_irqrestore(&mapping->tree_lock, flags);
731                         if (mapping_cap_account_dirty(mapping))
732                                 dec_page_state(nr_dirty);
733                         return 1;
734                 }
735                 write_unlock_irqrestore(&mapping->tree_lock, flags);
736                 return 0;
737         }
738         return TestClearPageDirty(page);
739 }
740 EXPORT_SYMBOL(test_clear_page_dirty);
741
742 /*
743  * Clear a page's dirty flag, while caring for dirty memory accounting.
744  * Returns true if the page was previously dirty.
745  *
746  * This is for preparing to put the page under writeout.  We leave the page
747  * tagged as dirty in the radix tree so that a concurrent write-for-sync
748  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
749  * implementation will run either set_page_writeback() or set_page_dirty(),
750  * at which stage we bring the page's dirty flag and radix-tree dirty tag
751  * back into sync.
752  *
753  * This incoherency between the page's dirty flag and radix-tree tag is
754  * unfortunate, but it only exists while the page is locked.
755  */
756 int clear_page_dirty_for_io(struct page *page)
757 {
758         struct address_space *mapping = page_mapping(page);
759
760         if (mapping) {
761                 if (TestClearPageDirty(page)) {
762                         if (mapping_cap_account_dirty(mapping))
763                                 dec_page_state(nr_dirty);
764                         return 1;
765                 }
766                 return 0;
767         }
768         return TestClearPageDirty(page);
769 }
770 EXPORT_SYMBOL(clear_page_dirty_for_io);
771
772 int test_clear_page_writeback(struct page *page)
773 {
774         struct address_space *mapping = page_mapping(page);
775         int ret;
776
777         if (mapping) {
778                 unsigned long flags;
779
780                 write_lock_irqsave(&mapping->tree_lock, flags);
781                 ret = TestClearPageWriteback(page);
782                 if (ret)
783                         radix_tree_tag_clear(&mapping->page_tree,
784                                                 page_index(page),
785                                                 PAGECACHE_TAG_WRITEBACK);
786                 write_unlock_irqrestore(&mapping->tree_lock, flags);
787         } else {
788                 ret = TestClearPageWriteback(page);
789         }
790         return ret;
791 }
792
793 int test_set_page_writeback(struct page *page)
794 {
795         struct address_space *mapping = page_mapping(page);
796         int ret;
797
798         if (mapping) {
799                 unsigned long flags;
800
801                 write_lock_irqsave(&mapping->tree_lock, flags);
802                 ret = TestSetPageWriteback(page);
803                 if (!ret)
804                         radix_tree_tag_set(&mapping->page_tree,
805                                                 page_index(page),
806                                                 PAGECACHE_TAG_WRITEBACK);
807                 if (!PageDirty(page))
808                         radix_tree_tag_clear(&mapping->page_tree,
809                                                 page_index(page),
810                                                 PAGECACHE_TAG_DIRTY);
811                 write_unlock_irqrestore(&mapping->tree_lock, flags);
812         } else {
813                 ret = TestSetPageWriteback(page);
814         }
815         return ret;
816
817 }
818 EXPORT_SYMBOL(test_set_page_writeback);
819
820 /*
821  * Return true if any of the pages in the mapping are marged with the
822  * passed tag.
823  */
824 int mapping_tagged(struct address_space *mapping, int tag)
825 {
826         unsigned long flags;
827         int ret;
828
829         read_lock_irqsave(&mapping->tree_lock, flags);
830         ret = radix_tree_tagged(&mapping->page_tree, tag);
831         read_unlock_irqrestore(&mapping->tree_lock, flags);
832         return ret;
833 }
834 EXPORT_SYMBOL(mapping_tagged);