xref: /openbmc/linux/mm/page-writeback.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
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/task_io_accounting_ops.h>
25 #include <linux/blkdev.h>
26 #include <linux/mpage.h>
27 #include <linux/rmap.h>
28 #include <linux/percpu.h>
29 #include <linux/notifier.h>
30 #include <linux/smp.h>
31 #include <linux/sysctl.h>
32 #include <linux/cpu.h>
33 #include <linux/syscalls.h>
34 #include <linux/buffer_head.h>
35 #include <linux/pagevec.h>
36 
37 /*
38  * The maximum number of pages to writeout in a single bdflush/kupdate
39  * operation.  We do this so we don't hold I_LOCK against an inode for
40  * enormous amounts of time, which would block a userspace task which has
41  * been forced to throttle against that inode.  Also, the code reevaluates
42  * the dirty each time it has written this many pages.
43  */
44 #define MAX_WRITEBACK_PAGES	1024
45 
46 /*
47  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
48  * will look to see if it needs to force writeback or throttling.
49  */
50 static long ratelimit_pages = 32;
51 
52 static int dirty_exceeded __cacheline_aligned_in_smp;	/* Dirty mem may be over limit */
53 
54 /*
55  * When balance_dirty_pages decides that the caller needs to perform some
56  * non-background writeback, this is how many pages it will attempt to write.
57  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
58  * large amounts of I/O are submitted.
59  */
60 static inline long sync_writeback_pages(void)
61 {
62 	return ratelimit_pages + ratelimit_pages / 2;
63 }
64 
65 /* The following parameters are exported via /proc/sys/vm */
66 
67 /*
68  * Start background writeback (via pdflush) at this percentage
69  */
70 int dirty_background_ratio = 5;
71 
72 /*
73  * The generator of dirty data starts writeback at this percentage
74  */
75 int vm_dirty_ratio = 10;
76 
77 /*
78  * The interval between `kupdate'-style writebacks, in jiffies
79  */
80 int dirty_writeback_interval = 5 * HZ;
81 
82 /*
83  * The longest number of jiffies for which data is allowed to remain dirty
84  */
85 int dirty_expire_interval = 30 * HZ;
86 
87 /*
88  * Flag that makes the machine dump writes/reads and block dirtyings.
89  */
90 int block_dump;
91 
92 /*
93  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
94  * a full sync is triggered after this time elapses without any disk activity.
95  */
96 int laptop_mode;
97 
98 EXPORT_SYMBOL(laptop_mode);
99 
100 /* End of sysctl-exported parameters */
101 
102 
103 static void background_writeout(unsigned long _min_pages);
104 
105 /*
106  * Work out the current dirty-memory clamping and background writeout
107  * thresholds.
108  *
109  * The main aim here is to lower them aggressively if there is a lot of mapped
110  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
111  * pages.  It is better to clamp down on writers than to start swapping, and
112  * performing lots of scanning.
113  *
114  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
115  *
116  * We don't permit the clamping level to fall below 5% - that is getting rather
117  * excessive.
118  *
119  * We make sure that the background writeout level is below the adjusted
120  * clamping level.
121  */
122 
123 static unsigned long highmem_dirtyable_memory(unsigned long total)
124 {
125 #ifdef CONFIG_HIGHMEM
126 	int node;
127 	unsigned long x = 0;
128 
129 	for_each_online_node(node) {
130 		struct zone *z =
131 			&NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
132 
133 		x += zone_page_state(z, NR_FREE_PAGES)
134 			+ zone_page_state(z, NR_INACTIVE)
135 			+ zone_page_state(z, NR_ACTIVE);
136 	}
137 	/*
138 	 * Make sure that the number of highmem pages is never larger
139 	 * than the number of the total dirtyable memory. This can only
140 	 * occur in very strange VM situations but we want to make sure
141 	 * that this does not occur.
142 	 */
143 	return min(x, total);
144 #else
145 	return 0;
146 #endif
147 }
148 
149 static unsigned long determine_dirtyable_memory(void)
150 {
151 	unsigned long x;
152 
153 	x = global_page_state(NR_FREE_PAGES)
154 		+ global_page_state(NR_INACTIVE)
155 		+ global_page_state(NR_ACTIVE);
156 	x -= highmem_dirtyable_memory(x);
157 	return x + 1;	/* Ensure that we never return 0 */
158 }
159 
160 static void
161 get_dirty_limits(long *pbackground, long *pdirty,
162 					struct address_space *mapping)
163 {
164 	int background_ratio;		/* Percentages */
165 	int dirty_ratio;
166 	int unmapped_ratio;
167 	long background;
168 	long dirty;
169 	unsigned long available_memory = determine_dirtyable_memory();
170 	struct task_struct *tsk;
171 
172 	unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
173 				global_page_state(NR_ANON_PAGES)) * 100) /
174 					available_memory;
175 
176 	dirty_ratio = vm_dirty_ratio;
177 	if (dirty_ratio > unmapped_ratio / 2)
178 		dirty_ratio = unmapped_ratio / 2;
179 
180 	if (dirty_ratio < 5)
181 		dirty_ratio = 5;
182 
183 	background_ratio = dirty_background_ratio;
184 	if (background_ratio >= dirty_ratio)
185 		background_ratio = dirty_ratio / 2;
186 
187 	background = (background_ratio * available_memory) / 100;
188 	dirty = (dirty_ratio * available_memory) / 100;
189 	tsk = current;
190 	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
191 		background += background / 4;
192 		dirty += dirty / 4;
193 	}
194 	*pbackground = background;
195 	*pdirty = dirty;
196 }
197 
198 /*
199  * balance_dirty_pages() must be called by processes which are generating dirty
200  * data.  It looks at the number of dirty pages in the machine and will force
201  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
202  * If we're over `background_thresh' then pdflush is woken to perform some
203  * writeout.
204  */
205 static void balance_dirty_pages(struct address_space *mapping)
206 {
207 	long nr_reclaimable;
208 	long background_thresh;
209 	long dirty_thresh;
210 	unsigned long pages_written = 0;
211 	unsigned long write_chunk = sync_writeback_pages();
212 
213 	struct backing_dev_info *bdi = mapping->backing_dev_info;
214 
215 	for (;;) {
216 		struct writeback_control wbc = {
217 			.bdi		= bdi,
218 			.sync_mode	= WB_SYNC_NONE,
219 			.older_than_this = NULL,
220 			.nr_to_write	= write_chunk,
221 			.range_cyclic	= 1,
222 		};
223 
224 		get_dirty_limits(&background_thresh, &dirty_thresh, mapping);
225 		nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
226 					global_page_state(NR_UNSTABLE_NFS);
227 		if (nr_reclaimable + global_page_state(NR_WRITEBACK) <=
228 			dirty_thresh)
229 				break;
230 
231 		if (!dirty_exceeded)
232 			dirty_exceeded = 1;
233 
234 		/* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
235 		 * Unstable writes are a feature of certain networked
236 		 * filesystems (i.e. NFS) in which data may have been
237 		 * written to the server's write cache, but has not yet
238 		 * been flushed to permanent storage.
239 		 */
240 		if (nr_reclaimable) {
241 			writeback_inodes(&wbc);
242 			get_dirty_limits(&background_thresh,
243 					 	&dirty_thresh, mapping);
244 			nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
245 					global_page_state(NR_UNSTABLE_NFS);
246 			if (nr_reclaimable +
247 				global_page_state(NR_WRITEBACK)
248 					<= dirty_thresh)
249 						break;
250 			pages_written += write_chunk - wbc.nr_to_write;
251 			if (pages_written >= write_chunk)
252 				break;		/* We've done our duty */
253 		}
254 		congestion_wait(WRITE, HZ/10);
255 	}
256 
257 	if (nr_reclaimable + global_page_state(NR_WRITEBACK)
258 		<= dirty_thresh && dirty_exceeded)
259 			dirty_exceeded = 0;
260 
261 	if (writeback_in_progress(bdi))
262 		return;		/* pdflush is already working this queue */
263 
264 	/*
265 	 * In laptop mode, we wait until hitting the higher threshold before
266 	 * starting background writeout, and then write out all the way down
267 	 * to the lower threshold.  So slow writers cause minimal disk activity.
268 	 *
269 	 * In normal mode, we start background writeout at the lower
270 	 * background_thresh, to keep the amount of dirty memory low.
271 	 */
272 	if ((laptop_mode && pages_written) ||
273 	     (!laptop_mode && (nr_reclaimable > background_thresh)))
274 		pdflush_operation(background_writeout, 0);
275 }
276 
277 void set_page_dirty_balance(struct page *page)
278 {
279 	if (set_page_dirty(page)) {
280 		struct address_space *mapping = page_mapping(page);
281 
282 		if (mapping)
283 			balance_dirty_pages_ratelimited(mapping);
284 	}
285 }
286 
287 /**
288  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
289  * @mapping: address_space which was dirtied
290  * @nr_pages_dirtied: number of pages which the caller has just dirtied
291  *
292  * Processes which are dirtying memory should call in here once for each page
293  * which was newly dirtied.  The function will periodically check the system's
294  * dirty state and will initiate writeback if needed.
295  *
296  * On really big machines, get_writeback_state is expensive, so try to avoid
297  * calling it too often (ratelimiting).  But once we're over the dirty memory
298  * limit we decrease the ratelimiting by a lot, to prevent individual processes
299  * from overshooting the limit by (ratelimit_pages) each.
300  */
301 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
302 					unsigned long nr_pages_dirtied)
303 {
304 	static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
305 	unsigned long ratelimit;
306 	unsigned long *p;
307 
308 	ratelimit = ratelimit_pages;
309 	if (dirty_exceeded)
310 		ratelimit = 8;
311 
312 	/*
313 	 * Check the rate limiting. Also, we do not want to throttle real-time
314 	 * tasks in balance_dirty_pages(). Period.
315 	 */
316 	preempt_disable();
317 	p =  &__get_cpu_var(ratelimits);
318 	*p += nr_pages_dirtied;
319 	if (unlikely(*p >= ratelimit)) {
320 		*p = 0;
321 		preempt_enable();
322 		balance_dirty_pages(mapping);
323 		return;
324 	}
325 	preempt_enable();
326 }
327 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
328 
329 void throttle_vm_writeout(gfp_t gfp_mask)
330 {
331 	long background_thresh;
332 	long dirty_thresh;
333 
334 	if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) {
335 		/*
336 		 * The caller might hold locks which can prevent IO completion
337 		 * or progress in the filesystem.  So we cannot just sit here
338 		 * waiting for IO to complete.
339 		 */
340 		congestion_wait(WRITE, HZ/10);
341 		return;
342 	}
343 
344         for ( ; ; ) {
345 		get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
346 
347                 /*
348                  * Boost the allowable dirty threshold a bit for page
349                  * allocators so they don't get DoS'ed by heavy writers
350                  */
351                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
352 
353                 if (global_page_state(NR_UNSTABLE_NFS) +
354 			global_page_state(NR_WRITEBACK) <= dirty_thresh)
355                         	break;
356                 congestion_wait(WRITE, HZ/10);
357         }
358 }
359 
360 /*
361  * writeback at least _min_pages, and keep writing until the amount of dirty
362  * memory is less than the background threshold, or until we're all clean.
363  */
364 static void background_writeout(unsigned long _min_pages)
365 {
366 	long min_pages = _min_pages;
367 	struct writeback_control wbc = {
368 		.bdi		= NULL,
369 		.sync_mode	= WB_SYNC_NONE,
370 		.older_than_this = NULL,
371 		.nr_to_write	= 0,
372 		.nonblocking	= 1,
373 		.range_cyclic	= 1,
374 	};
375 
376 	for ( ; ; ) {
377 		long background_thresh;
378 		long dirty_thresh;
379 
380 		get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
381 		if (global_page_state(NR_FILE_DIRTY) +
382 			global_page_state(NR_UNSTABLE_NFS) < background_thresh
383 				&& min_pages <= 0)
384 			break;
385 		wbc.encountered_congestion = 0;
386 		wbc.nr_to_write = MAX_WRITEBACK_PAGES;
387 		wbc.pages_skipped = 0;
388 		writeback_inodes(&wbc);
389 		min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
390 		if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
391 			/* Wrote less than expected */
392 			congestion_wait(WRITE, HZ/10);
393 			if (!wbc.encountered_congestion)
394 				break;
395 		}
396 	}
397 }
398 
399 /*
400  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
401  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
402  * -1 if all pdflush threads were busy.
403  */
404 int wakeup_pdflush(long nr_pages)
405 {
406 	if (nr_pages == 0)
407 		nr_pages = global_page_state(NR_FILE_DIRTY) +
408 				global_page_state(NR_UNSTABLE_NFS);
409 	return pdflush_operation(background_writeout, nr_pages);
410 }
411 
412 static void wb_timer_fn(unsigned long unused);
413 static void laptop_timer_fn(unsigned long unused);
414 
415 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
416 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
417 
418 /*
419  * Periodic writeback of "old" data.
420  *
421  * Define "old": the first time one of an inode's pages is dirtied, we mark the
422  * dirtying-time in the inode's address_space.  So this periodic writeback code
423  * just walks the superblock inode list, writing back any inodes which are
424  * older than a specific point in time.
425  *
426  * Try to run once per dirty_writeback_interval.  But if a writeback event
427  * takes longer than a dirty_writeback_interval interval, then leave a
428  * one-second gap.
429  *
430  * older_than_this takes precedence over nr_to_write.  So we'll only write back
431  * all dirty pages if they are all attached to "old" mappings.
432  */
433 static void wb_kupdate(unsigned long arg)
434 {
435 	unsigned long oldest_jif;
436 	unsigned long start_jif;
437 	unsigned long next_jif;
438 	long nr_to_write;
439 	struct writeback_control wbc = {
440 		.bdi		= NULL,
441 		.sync_mode	= WB_SYNC_NONE,
442 		.older_than_this = &oldest_jif,
443 		.nr_to_write	= 0,
444 		.nonblocking	= 1,
445 		.for_kupdate	= 1,
446 		.range_cyclic	= 1,
447 	};
448 
449 	sync_supers();
450 
451 	oldest_jif = jiffies - dirty_expire_interval;
452 	start_jif = jiffies;
453 	next_jif = start_jif + dirty_writeback_interval;
454 	nr_to_write = global_page_state(NR_FILE_DIRTY) +
455 			global_page_state(NR_UNSTABLE_NFS) +
456 			(inodes_stat.nr_inodes - inodes_stat.nr_unused);
457 	while (nr_to_write > 0) {
458 		wbc.encountered_congestion = 0;
459 		wbc.nr_to_write = MAX_WRITEBACK_PAGES;
460 		writeback_inodes(&wbc);
461 		if (wbc.nr_to_write > 0) {
462 			if (wbc.encountered_congestion)
463 				congestion_wait(WRITE, HZ/10);
464 			else
465 				break;	/* All the old data is written */
466 		}
467 		nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
468 	}
469 	if (time_before(next_jif, jiffies + HZ))
470 		next_jif = jiffies + HZ;
471 	if (dirty_writeback_interval)
472 		mod_timer(&wb_timer, next_jif);
473 }
474 
475 /*
476  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
477  */
478 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
479 		struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
480 {
481 	proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
482 	if (dirty_writeback_interval) {
483 		mod_timer(&wb_timer,
484 			jiffies + dirty_writeback_interval);
485 		} else {
486 		del_timer(&wb_timer);
487 	}
488 	return 0;
489 }
490 
491 static void wb_timer_fn(unsigned long unused)
492 {
493 	if (pdflush_operation(wb_kupdate, 0) < 0)
494 		mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
495 }
496 
497 static void laptop_flush(unsigned long unused)
498 {
499 	sys_sync();
500 }
501 
502 static void laptop_timer_fn(unsigned long unused)
503 {
504 	pdflush_operation(laptop_flush, 0);
505 }
506 
507 /*
508  * We've spun up the disk and we're in laptop mode: schedule writeback
509  * of all dirty data a few seconds from now.  If the flush is already scheduled
510  * then push it back - the user is still using the disk.
511  */
512 void laptop_io_completion(void)
513 {
514 	mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
515 }
516 
517 /*
518  * We're in laptop mode and we've just synced. The sync's writes will have
519  * caused another writeback to be scheduled by laptop_io_completion.
520  * Nothing needs to be written back anymore, so we unschedule the writeback.
521  */
522 void laptop_sync_completion(void)
523 {
524 	del_timer(&laptop_mode_wb_timer);
525 }
526 
527 /*
528  * If ratelimit_pages is too high then we can get into dirty-data overload
529  * if a large number of processes all perform writes at the same time.
530  * If it is too low then SMP machines will call the (expensive)
531  * get_writeback_state too often.
532  *
533  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
534  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
535  * thresholds before writeback cuts in.
536  *
537  * But the limit should not be set too high.  Because it also controls the
538  * amount of memory which the balance_dirty_pages() caller has to write back.
539  * If this is too large then the caller will block on the IO queue all the
540  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
541  * will write six megabyte chunks, max.
542  */
543 
544 void writeback_set_ratelimit(void)
545 {
546 	ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
547 	if (ratelimit_pages < 16)
548 		ratelimit_pages = 16;
549 	if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
550 		ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
551 }
552 
553 static int __cpuinit
554 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
555 {
556 	writeback_set_ratelimit();
557 	return NOTIFY_DONE;
558 }
559 
560 static struct notifier_block __cpuinitdata ratelimit_nb = {
561 	.notifier_call	= ratelimit_handler,
562 	.next		= NULL,
563 };
564 
565 /*
566  * Called early on to tune the page writeback dirty limits.
567  *
568  * We used to scale dirty pages according to how total memory
569  * related to pages that could be allocated for buffers (by
570  * comparing nr_free_buffer_pages() to vm_total_pages.
571  *
572  * However, that was when we used "dirty_ratio" to scale with
573  * all memory, and we don't do that any more. "dirty_ratio"
574  * is now applied to total non-HIGHPAGE memory (by subtracting
575  * totalhigh_pages from vm_total_pages), and as such we can't
576  * get into the old insane situation any more where we had
577  * large amounts of dirty pages compared to a small amount of
578  * non-HIGHMEM memory.
579  *
580  * But we might still want to scale the dirty_ratio by how
581  * much memory the box has..
582  */
583 void __init page_writeback_init(void)
584 {
585 	mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
586 	writeback_set_ratelimit();
587 	register_cpu_notifier(&ratelimit_nb);
588 }
589 
590 /**
591  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
592  * @mapping: address space structure to write
593  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
594  * @writepage: function called for each page
595  * @data: data passed to writepage function
596  *
597  * If a page is already under I/O, write_cache_pages() skips it, even
598  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
599  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
600  * and msync() need to guarantee that all the data which was dirty at the time
601  * the call was made get new I/O started against them.  If wbc->sync_mode is
602  * WB_SYNC_ALL then we were called for data integrity and we must wait for
603  * existing IO to complete.
604  */
605 int write_cache_pages(struct address_space *mapping,
606 		      struct writeback_control *wbc, writepage_t writepage,
607 		      void *data)
608 {
609 	struct backing_dev_info *bdi = mapping->backing_dev_info;
610 	int ret = 0;
611 	int done = 0;
612 	struct pagevec pvec;
613 	int nr_pages;
614 	pgoff_t index;
615 	pgoff_t end;		/* Inclusive */
616 	int scanned = 0;
617 	int range_whole = 0;
618 
619 	if (wbc->nonblocking && bdi_write_congested(bdi)) {
620 		wbc->encountered_congestion = 1;
621 		return 0;
622 	}
623 
624 	pagevec_init(&pvec, 0);
625 	if (wbc->range_cyclic) {
626 		index = mapping->writeback_index; /* Start from prev offset */
627 		end = -1;
628 	} else {
629 		index = wbc->range_start >> PAGE_CACHE_SHIFT;
630 		end = wbc->range_end >> PAGE_CACHE_SHIFT;
631 		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
632 			range_whole = 1;
633 		scanned = 1;
634 	}
635 retry:
636 	while (!done && (index <= end) &&
637 	       (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
638 					      PAGECACHE_TAG_DIRTY,
639 					      min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
640 		unsigned i;
641 
642 		scanned = 1;
643 		for (i = 0; i < nr_pages; i++) {
644 			struct page *page = pvec.pages[i];
645 
646 			/*
647 			 * At this point we hold neither mapping->tree_lock nor
648 			 * lock on the page itself: the page may be truncated or
649 			 * invalidated (changing page->mapping to NULL), or even
650 			 * swizzled back from swapper_space to tmpfs file
651 			 * mapping
652 			 */
653 			lock_page(page);
654 
655 			if (unlikely(page->mapping != mapping)) {
656 				unlock_page(page);
657 				continue;
658 			}
659 
660 			if (!wbc->range_cyclic && page->index > end) {
661 				done = 1;
662 				unlock_page(page);
663 				continue;
664 			}
665 
666 			if (wbc->sync_mode != WB_SYNC_NONE)
667 				wait_on_page_writeback(page);
668 
669 			if (PageWriteback(page) ||
670 			    !clear_page_dirty_for_io(page)) {
671 				unlock_page(page);
672 				continue;
673 			}
674 
675 			ret = (*writepage)(page, wbc, data);
676 
677 			if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
678 				unlock_page(page);
679 			if (ret || (--(wbc->nr_to_write) <= 0))
680 				done = 1;
681 			if (wbc->nonblocking && bdi_write_congested(bdi)) {
682 				wbc->encountered_congestion = 1;
683 				done = 1;
684 			}
685 		}
686 		pagevec_release(&pvec);
687 		cond_resched();
688 	}
689 	if (!scanned && !done) {
690 		/*
691 		 * We hit the last page and there is more work to be done: wrap
692 		 * back to the start of the file
693 		 */
694 		scanned = 1;
695 		index = 0;
696 		goto retry;
697 	}
698 	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
699 		mapping->writeback_index = index;
700 	return ret;
701 }
702 EXPORT_SYMBOL(write_cache_pages);
703 
704 /*
705  * Function used by generic_writepages to call the real writepage
706  * function and set the mapping flags on error
707  */
708 static int __writepage(struct page *page, struct writeback_control *wbc,
709 		       void *data)
710 {
711 	struct address_space *mapping = data;
712 	int ret = mapping->a_ops->writepage(page, wbc);
713 	mapping_set_error(mapping, ret);
714 	return ret;
715 }
716 
717 /**
718  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
719  * @mapping: address space structure to write
720  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
721  *
722  * This is a library function, which implements the writepages()
723  * address_space_operation.
724  */
725 int generic_writepages(struct address_space *mapping,
726 		       struct writeback_control *wbc)
727 {
728 	/* deal with chardevs and other special file */
729 	if (!mapping->a_ops->writepage)
730 		return 0;
731 
732 	return write_cache_pages(mapping, wbc, __writepage, mapping);
733 }
734 
735 EXPORT_SYMBOL(generic_writepages);
736 
737 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
738 {
739 	int ret;
740 
741 	if (wbc->nr_to_write <= 0)
742 		return 0;
743 	wbc->for_writepages = 1;
744 	if (mapping->a_ops->writepages)
745 		ret = mapping->a_ops->writepages(mapping, wbc);
746 	else
747 		ret = generic_writepages(mapping, wbc);
748 	wbc->for_writepages = 0;
749 	return ret;
750 }
751 
752 /**
753  * write_one_page - write out a single page and optionally wait on I/O
754  * @page: the page to write
755  * @wait: if true, wait on writeout
756  *
757  * The page must be locked by the caller and will be unlocked upon return.
758  *
759  * write_one_page() returns a negative error code if I/O failed.
760  */
761 int write_one_page(struct page *page, int wait)
762 {
763 	struct address_space *mapping = page->mapping;
764 	int ret = 0;
765 	struct writeback_control wbc = {
766 		.sync_mode = WB_SYNC_ALL,
767 		.nr_to_write = 1,
768 	};
769 
770 	BUG_ON(!PageLocked(page));
771 
772 	if (wait)
773 		wait_on_page_writeback(page);
774 
775 	if (clear_page_dirty_for_io(page)) {
776 		page_cache_get(page);
777 		ret = mapping->a_ops->writepage(page, &wbc);
778 		if (ret == 0 && wait) {
779 			wait_on_page_writeback(page);
780 			if (PageError(page))
781 				ret = -EIO;
782 		}
783 		page_cache_release(page);
784 	} else {
785 		unlock_page(page);
786 	}
787 	return ret;
788 }
789 EXPORT_SYMBOL(write_one_page);
790 
791 /*
792  * For address_spaces which do not use buffers nor write back.
793  */
794 int __set_page_dirty_no_writeback(struct page *page)
795 {
796 	if (!PageDirty(page))
797 		SetPageDirty(page);
798 	return 0;
799 }
800 
801 /*
802  * For address_spaces which do not use buffers.  Just tag the page as dirty in
803  * its radix tree.
804  *
805  * This is also used when a single buffer is being dirtied: we want to set the
806  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
807  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
808  *
809  * Most callers have locked the page, which pins the address_space in memory.
810  * But zap_pte_range() does not lock the page, however in that case the
811  * mapping is pinned by the vma's ->vm_file reference.
812  *
813  * We take care to handle the case where the page was truncated from the
814  * mapping by re-checking page_mapping() insode tree_lock.
815  */
816 int __set_page_dirty_nobuffers(struct page *page)
817 {
818 	if (!TestSetPageDirty(page)) {
819 		struct address_space *mapping = page_mapping(page);
820 		struct address_space *mapping2;
821 
822 		if (!mapping)
823 			return 1;
824 
825 		write_lock_irq(&mapping->tree_lock);
826 		mapping2 = page_mapping(page);
827 		if (mapping2) { /* Race with truncate? */
828 			BUG_ON(mapping2 != mapping);
829 			if (mapping_cap_account_dirty(mapping)) {
830 				__inc_zone_page_state(page, NR_FILE_DIRTY);
831 				task_io_account_write(PAGE_CACHE_SIZE);
832 			}
833 			radix_tree_tag_set(&mapping->page_tree,
834 				page_index(page), PAGECACHE_TAG_DIRTY);
835 		}
836 		write_unlock_irq(&mapping->tree_lock);
837 		if (mapping->host) {
838 			/* !PageAnon && !swapper_space */
839 			__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
840 		}
841 		return 1;
842 	}
843 	return 0;
844 }
845 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
846 
847 /*
848  * When a writepage implementation decides that it doesn't want to write this
849  * page for some reason, it should redirty the locked page via
850  * redirty_page_for_writepage() and it should then unlock the page and return 0
851  */
852 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
853 {
854 	wbc->pages_skipped++;
855 	return __set_page_dirty_nobuffers(page);
856 }
857 EXPORT_SYMBOL(redirty_page_for_writepage);
858 
859 /*
860  * If the mapping doesn't provide a set_page_dirty a_op, then
861  * just fall through and assume that it wants buffer_heads.
862  */
863 int fastcall set_page_dirty(struct page *page)
864 {
865 	struct address_space *mapping = page_mapping(page);
866 
867 	if (likely(mapping)) {
868 		int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
869 #ifdef CONFIG_BLOCK
870 		if (!spd)
871 			spd = __set_page_dirty_buffers;
872 #endif
873 		return (*spd)(page);
874 	}
875 	if (!PageDirty(page)) {
876 		if (!TestSetPageDirty(page))
877 			return 1;
878 	}
879 	return 0;
880 }
881 EXPORT_SYMBOL(set_page_dirty);
882 
883 /*
884  * set_page_dirty() is racy if the caller has no reference against
885  * page->mapping->host, and if the page is unlocked.  This is because another
886  * CPU could truncate the page off the mapping and then free the mapping.
887  *
888  * Usually, the page _is_ locked, or the caller is a user-space process which
889  * holds a reference on the inode by having an open file.
890  *
891  * In other cases, the page should be locked before running set_page_dirty().
892  */
893 int set_page_dirty_lock(struct page *page)
894 {
895 	int ret;
896 
897 	lock_page_nosync(page);
898 	ret = set_page_dirty(page);
899 	unlock_page(page);
900 	return ret;
901 }
902 EXPORT_SYMBOL(set_page_dirty_lock);
903 
904 /*
905  * Clear a page's dirty flag, while caring for dirty memory accounting.
906  * Returns true if the page was previously dirty.
907  *
908  * This is for preparing to put the page under writeout.  We leave the page
909  * tagged as dirty in the radix tree so that a concurrent write-for-sync
910  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
911  * implementation will run either set_page_writeback() or set_page_dirty(),
912  * at which stage we bring the page's dirty flag and radix-tree dirty tag
913  * back into sync.
914  *
915  * This incoherency between the page's dirty flag and radix-tree tag is
916  * unfortunate, but it only exists while the page is locked.
917  */
918 int clear_page_dirty_for_io(struct page *page)
919 {
920 	struct address_space *mapping = page_mapping(page);
921 
922 	if (mapping && mapping_cap_account_dirty(mapping)) {
923 		/*
924 		 * Yes, Virginia, this is indeed insane.
925 		 *
926 		 * We use this sequence to make sure that
927 		 *  (a) we account for dirty stats properly
928 		 *  (b) we tell the low-level filesystem to
929 		 *      mark the whole page dirty if it was
930 		 *      dirty in a pagetable. Only to then
931 		 *  (c) clean the page again and return 1 to
932 		 *      cause the writeback.
933 		 *
934 		 * This way we avoid all nasty races with the
935 		 * dirty bit in multiple places and clearing
936 		 * them concurrently from different threads.
937 		 *
938 		 * Note! Normally the "set_page_dirty(page)"
939 		 * has no effect on the actual dirty bit - since
940 		 * that will already usually be set. But we
941 		 * need the side effects, and it can help us
942 		 * avoid races.
943 		 *
944 		 * We basically use the page "master dirty bit"
945 		 * as a serialization point for all the different
946 		 * threads doing their things.
947 		 *
948 		 * FIXME! We still have a race here: if somebody
949 		 * adds the page back to the page tables in
950 		 * between the "page_mkclean()" and the "TestClearPageDirty()",
951 		 * we might have it mapped without the dirty bit set.
952 		 */
953 		if (page_mkclean(page))
954 			set_page_dirty(page);
955 		if (TestClearPageDirty(page)) {
956 			dec_zone_page_state(page, NR_FILE_DIRTY);
957 			return 1;
958 		}
959 		return 0;
960 	}
961 	return TestClearPageDirty(page);
962 }
963 EXPORT_SYMBOL(clear_page_dirty_for_io);
964 
965 int test_clear_page_writeback(struct page *page)
966 {
967 	struct address_space *mapping = page_mapping(page);
968 	int ret;
969 
970 	if (mapping) {
971 		unsigned long flags;
972 
973 		write_lock_irqsave(&mapping->tree_lock, flags);
974 		ret = TestClearPageWriteback(page);
975 		if (ret)
976 			radix_tree_tag_clear(&mapping->page_tree,
977 						page_index(page),
978 						PAGECACHE_TAG_WRITEBACK);
979 		write_unlock_irqrestore(&mapping->tree_lock, flags);
980 	} else {
981 		ret = TestClearPageWriteback(page);
982 	}
983 	return ret;
984 }
985 
986 int test_set_page_writeback(struct page *page)
987 {
988 	struct address_space *mapping = page_mapping(page);
989 	int ret;
990 
991 	if (mapping) {
992 		unsigned long flags;
993 
994 		write_lock_irqsave(&mapping->tree_lock, flags);
995 		ret = TestSetPageWriteback(page);
996 		if (!ret)
997 			radix_tree_tag_set(&mapping->page_tree,
998 						page_index(page),
999 						PAGECACHE_TAG_WRITEBACK);
1000 		if (!PageDirty(page))
1001 			radix_tree_tag_clear(&mapping->page_tree,
1002 						page_index(page),
1003 						PAGECACHE_TAG_DIRTY);
1004 		write_unlock_irqrestore(&mapping->tree_lock, flags);
1005 	} else {
1006 		ret = TestSetPageWriteback(page);
1007 	}
1008 	return ret;
1009 
1010 }
1011 EXPORT_SYMBOL(test_set_page_writeback);
1012 
1013 /*
1014  * Return true if any of the pages in the mapping are marged with the
1015  * passed tag.
1016  */
1017 int mapping_tagged(struct address_space *mapping, int tag)
1018 {
1019 	unsigned long flags;
1020 	int ret;
1021 
1022 	read_lock_irqsave(&mapping->tree_lock, flags);
1023 	ret = radix_tree_tagged(&mapping->page_tree, tag);
1024 	read_unlock_irqrestore(&mapping->tree_lock, flags);
1025 	return ret;
1026 }
1027 EXPORT_SYMBOL(mapping_tagged);
1028