On Tue, Nov 06, 2007 at 02:33:53AM -0800, akpm@linux-foundation.org wrote:
[mmotm.git] / mm / page-writeback.c
blob2bfd1d5d310f01e787e4886c3b2640f19c1b570b
1 /*
2 * mm/page-writeback.c
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
10 * 10Apr2002 Andrew Morton
11 * Initial version
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h>
36 #include <linux/pagevec.h>
39 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
40 * will look to see if it needs to force writeback or throttling.
42 static long ratelimit_pages = 32;
45 * When balance_dirty_pages decides that the caller needs to perform some
46 * non-background writeback, this is how many pages it will attempt to write.
47 * It should be somewhat larger than dirtied pages to ensure that reasonably
48 * large amounts of I/O are submitted.
50 static inline long sync_writeback_pages(unsigned long dirtied)
52 if (dirtied < ratelimit_pages)
53 dirtied = ratelimit_pages;
55 return dirtied + dirtied / 2;
58 /* The following parameters are exported via /proc/sys/vm */
61 * Start background writeback (via writeback threads) at this percentage
63 int dirty_background_ratio = 10;
66 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
67 * dirty_background_ratio * the amount of dirtyable memory
69 unsigned long dirty_background_bytes;
72 * free highmem will not be subtracted from the total free memory
73 * for calculating free ratios if vm_highmem_is_dirtyable is true
75 int vm_highmem_is_dirtyable;
78 * The generator of dirty data starts writeback at this percentage
80 int vm_dirty_ratio = 20;
83 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
84 * vm_dirty_ratio * the amount of dirtyable memory
86 unsigned long vm_dirty_bytes;
89 * The interval between `kupdate'-style writebacks
91 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
94 * The longest time for which data is allowed to remain dirty
96 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
99 * Flag that makes the machine dump writes/reads and block dirtyings.
101 int block_dump;
104 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
105 * a full sync is triggered after this time elapses without any disk activity.
107 int laptop_mode;
109 EXPORT_SYMBOL(laptop_mode);
111 /* End of sysctl-exported parameters */
115 * Scale the writeback cache size proportional to the relative writeout speeds.
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
120 * share.
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
130 static struct prop_descriptor vm_completions;
131 static struct prop_descriptor vm_dirties;
134 * couple the period to the dirty_ratio:
136 * period/2 ~ roundup_pow_of_two(dirty limit)
138 static int calc_period_shift(void)
140 unsigned long dirty_total;
142 if (vm_dirty_bytes)
143 dirty_total = vm_dirty_bytes / PAGE_SIZE;
144 else
145 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146 100;
147 return 2 + ilog2(dirty_total - 1);
151 * update the period when the dirty threshold changes.
153 static void update_completion_period(void)
155 int shift = calc_period_shift();
156 prop_change_shift(&vm_completions, shift);
157 prop_change_shift(&vm_dirties, shift);
160 int dirty_background_ratio_handler(struct ctl_table *table, int write,
161 void __user *buffer, size_t *lenp,
162 loff_t *ppos)
164 int ret;
166 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
167 if (ret == 0 && write)
168 dirty_background_bytes = 0;
169 return ret;
172 int dirty_background_bytes_handler(struct ctl_table *table, int write,
173 void __user *buffer, size_t *lenp,
174 loff_t *ppos)
176 int ret;
178 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
179 if (ret == 0 && write)
180 dirty_background_ratio = 0;
181 return ret;
184 int dirty_ratio_handler(struct ctl_table *table, int write,
185 void __user *buffer, size_t *lenp,
186 loff_t *ppos)
188 int old_ratio = vm_dirty_ratio;
189 int ret;
191 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
192 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
193 update_completion_period();
194 vm_dirty_bytes = 0;
196 return ret;
200 int dirty_bytes_handler(struct ctl_table *table, int write,
201 void __user *buffer, size_t *lenp,
202 loff_t *ppos)
204 unsigned long old_bytes = vm_dirty_bytes;
205 int ret;
207 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
208 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
209 update_completion_period();
210 vm_dirty_ratio = 0;
212 return ret;
216 * Increment the BDI's writeout completion count and the global writeout
217 * completion count. Called from test_clear_page_writeback().
219 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
221 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
222 bdi->max_prop_frac);
225 void bdi_writeout_inc(struct backing_dev_info *bdi)
227 unsigned long flags;
229 local_irq_save(flags);
230 __bdi_writeout_inc(bdi);
231 local_irq_restore(flags);
233 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
235 void task_dirty_inc(struct task_struct *tsk)
237 prop_inc_single(&vm_dirties, &tsk->dirties);
241 * Obtain an accurate fraction of the BDI's portion.
243 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
244 long *numerator, long *denominator)
246 if (bdi_cap_writeback_dirty(bdi)) {
247 prop_fraction_percpu(&vm_completions, &bdi->completions,
248 numerator, denominator);
249 } else {
250 *numerator = 0;
251 *denominator = 1;
256 * Clip the earned share of dirty pages to that which is actually available.
257 * This avoids exceeding the total dirty_limit when the floating averages
258 * fluctuate too quickly.
260 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
261 unsigned long dirty, unsigned long *pbdi_dirty)
263 unsigned long avail_dirty;
265 avail_dirty = global_page_state(NR_FILE_DIRTY) +
266 global_page_state(NR_WRITEBACK) +
267 global_page_state(NR_UNSTABLE_NFS) +
268 global_page_state(NR_WRITEBACK_TEMP);
270 if (avail_dirty < dirty)
271 avail_dirty = dirty - avail_dirty;
272 else
273 avail_dirty = 0;
275 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
276 bdi_stat(bdi, BDI_WRITEBACK);
278 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
281 static inline void task_dirties_fraction(struct task_struct *tsk,
282 long *numerator, long *denominator)
284 prop_fraction_single(&vm_dirties, &tsk->dirties,
285 numerator, denominator);
289 * scale the dirty limit
291 * task specific dirty limit:
293 * dirty -= (dirty/8) * p_{t}
295 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
297 long numerator, denominator;
298 unsigned long dirty = *pdirty;
299 u64 inv = dirty >> 3;
301 task_dirties_fraction(tsk, &numerator, &denominator);
302 inv *= numerator;
303 do_div(inv, denominator);
305 dirty -= inv;
306 if (dirty < *pdirty/2)
307 dirty = *pdirty/2;
309 *pdirty = dirty;
315 static unsigned int bdi_min_ratio;
317 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
319 int ret = 0;
321 spin_lock_bh(&bdi_lock);
322 if (min_ratio > bdi->max_ratio) {
323 ret = -EINVAL;
324 } else {
325 min_ratio -= bdi->min_ratio;
326 if (bdi_min_ratio + min_ratio < 100) {
327 bdi_min_ratio += min_ratio;
328 bdi->min_ratio += min_ratio;
329 } else {
330 ret = -EINVAL;
333 spin_unlock_bh(&bdi_lock);
335 return ret;
338 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
340 int ret = 0;
342 if (max_ratio > 100)
343 return -EINVAL;
345 spin_lock_bh(&bdi_lock);
346 if (bdi->min_ratio > max_ratio) {
347 ret = -EINVAL;
348 } else {
349 bdi->max_ratio = max_ratio;
350 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
352 spin_unlock_bh(&bdi_lock);
354 return ret;
356 EXPORT_SYMBOL(bdi_set_max_ratio);
359 * Work out the current dirty-memory clamping and background writeout
360 * thresholds.
362 * The main aim here is to lower them aggressively if there is a lot of mapped
363 * memory around. To avoid stressing page reclaim with lots of unreclaimable
364 * pages. It is better to clamp down on writers than to start swapping, and
365 * performing lots of scanning.
367 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
369 * We don't permit the clamping level to fall below 5% - that is getting rather
370 * excessive.
372 * We make sure that the background writeout level is below the adjusted
373 * clamping level.
376 static unsigned long highmem_dirtyable_memory(unsigned long total)
378 #ifdef CONFIG_HIGHMEM
379 int node;
380 unsigned long x = 0;
382 for_each_node_state(node, N_HIGH_MEMORY) {
383 struct zone *z =
384 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
386 x += zone_page_state(z, NR_FREE_PAGES) +
387 zone_reclaimable_pages(z);
390 * Make sure that the number of highmem pages is never larger
391 * than the number of the total dirtyable memory. This can only
392 * occur in very strange VM situations but we want to make sure
393 * that this does not occur.
395 return min(x, total);
396 #else
397 return 0;
398 #endif
402 * determine_dirtyable_memory - amount of memory that may be used
404 * Returns the numebr of pages that can currently be freed and used
405 * by the kernel for direct mappings.
407 unsigned long determine_dirtyable_memory(void)
409 unsigned long x;
411 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
413 if (!vm_highmem_is_dirtyable)
414 x -= highmem_dirtyable_memory(x);
416 return x + 1; /* Ensure that we never return 0 */
419 void
420 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
421 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
423 unsigned long background;
424 unsigned long dirty;
425 unsigned long available_memory = determine_dirtyable_memory();
426 struct task_struct *tsk;
428 if (vm_dirty_bytes)
429 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
430 else {
431 int dirty_ratio;
433 dirty_ratio = vm_dirty_ratio;
434 if (dirty_ratio < 5)
435 dirty_ratio = 5;
436 dirty = (dirty_ratio * available_memory) / 100;
439 if (dirty_background_bytes)
440 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
441 else
442 background = (dirty_background_ratio * available_memory) / 100;
444 if (background >= dirty)
445 background = dirty / 2;
446 tsk = current;
447 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
448 background += background / 4;
449 dirty += dirty / 4;
451 *pbackground = background;
452 *pdirty = dirty;
454 if (bdi) {
455 u64 bdi_dirty;
456 long numerator, denominator;
459 * Calculate this BDI's share of the dirty ratio.
461 bdi_writeout_fraction(bdi, &numerator, &denominator);
463 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
464 bdi_dirty *= numerator;
465 do_div(bdi_dirty, denominator);
466 bdi_dirty += (dirty * bdi->min_ratio) / 100;
467 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
468 bdi_dirty = dirty * bdi->max_ratio / 100;
470 *pbdi_dirty = bdi_dirty;
471 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
472 task_dirty_limit(current, pbdi_dirty);
477 * balance_dirty_pages() must be called by processes which are generating dirty
478 * data. It looks at the number of dirty pages in the machine and will force
479 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
480 * If we're over `background_thresh' then the writeback threads are woken to
481 * perform some writeout.
483 static void balance_dirty_pages(struct address_space *mapping,
484 unsigned long write_chunk)
486 long nr_reclaimable, bdi_nr_reclaimable;
487 long nr_writeback, bdi_nr_writeback;
488 unsigned long background_thresh;
489 unsigned long dirty_thresh;
490 unsigned long bdi_thresh;
491 unsigned long pages_written = 0;
492 unsigned long pause = 1;
494 struct backing_dev_info *bdi = mapping->backing_dev_info;
496 for (;;) {
497 struct writeback_control wbc = {
498 .bdi = bdi,
499 .sync_mode = WB_SYNC_NONE,
500 .older_than_this = NULL,
501 .nr_to_write = write_chunk,
502 .range_cyclic = 1,
505 get_dirty_limits(&background_thresh, &dirty_thresh,
506 &bdi_thresh, bdi);
508 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
509 global_page_state(NR_UNSTABLE_NFS);
510 nr_writeback = global_page_state(NR_WRITEBACK);
512 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
513 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
515 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
516 break;
519 * Throttle it only when the background writeback cannot
520 * catch-up. This avoids (excessively) small writeouts
521 * when the bdi limits are ramping up.
523 if (nr_reclaimable + nr_writeback <
524 (background_thresh + dirty_thresh) / 2)
525 break;
527 if (!bdi->dirty_exceeded)
528 bdi->dirty_exceeded = 1;
530 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
531 * Unstable writes are a feature of certain networked
532 * filesystems (i.e. NFS) in which data may have been
533 * written to the server's write cache, but has not yet
534 * been flushed to permanent storage.
535 * Only move pages to writeback if this bdi is over its
536 * threshold otherwise wait until the disk writes catch
537 * up.
539 if (bdi_nr_reclaimable > bdi_thresh) {
540 writeback_inodes_wbc(&wbc);
541 pages_written += write_chunk - wbc.nr_to_write;
542 get_dirty_limits(&background_thresh, &dirty_thresh,
543 &bdi_thresh, bdi);
547 * In order to avoid the stacked BDI deadlock we need
548 * to ensure we accurately count the 'dirty' pages when
549 * the threshold is low.
551 * Otherwise it would be possible to get thresh+n pages
552 * reported dirty, even though there are thresh-m pages
553 * actually dirty; with m+n sitting in the percpu
554 * deltas.
556 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
557 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
558 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
559 } else if (bdi_nr_reclaimable) {
560 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
561 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
564 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
565 break;
566 if (pages_written >= write_chunk)
567 break; /* We've done our duty */
569 __set_current_state(TASK_INTERRUPTIBLE);
570 io_schedule_timeout(pause);
573 * Increase the delay for each loop, up to our previous
574 * default of taking a 100ms nap.
576 pause <<= 1;
577 if (pause > HZ / 10)
578 pause = HZ / 10;
581 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
582 bdi->dirty_exceeded)
583 bdi->dirty_exceeded = 0;
585 if (writeback_in_progress(bdi))
586 return;
589 * In laptop mode, we wait until hitting the higher threshold before
590 * starting background writeout, and then write out all the way down
591 * to the lower threshold. So slow writers cause minimal disk activity.
593 * In normal mode, we start background writeout at the lower
594 * background_thresh, to keep the amount of dirty memory low.
596 if ((laptop_mode && pages_written) ||
597 (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
598 + global_page_state(NR_UNSTABLE_NFS))
599 > background_thresh)))
600 bdi_start_writeback(bdi, NULL, 0);
603 void set_page_dirty_balance(struct page *page, int page_mkwrite)
605 if (set_page_dirty(page) || page_mkwrite) {
606 struct address_space *mapping = page_mapping(page);
608 if (mapping)
609 balance_dirty_pages_ratelimited(mapping);
613 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
616 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
617 * @mapping: address_space which was dirtied
618 * @nr_pages_dirtied: number of pages which the caller has just dirtied
620 * Processes which are dirtying memory should call in here once for each page
621 * which was newly dirtied. The function will periodically check the system's
622 * dirty state and will initiate writeback if needed.
624 * On really big machines, get_writeback_state is expensive, so try to avoid
625 * calling it too often (ratelimiting). But once we're over the dirty memory
626 * limit we decrease the ratelimiting by a lot, to prevent individual processes
627 * from overshooting the limit by (ratelimit_pages) each.
629 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
630 unsigned long nr_pages_dirtied)
632 unsigned long ratelimit;
633 unsigned long *p;
635 ratelimit = ratelimit_pages;
636 if (mapping->backing_dev_info->dirty_exceeded)
637 ratelimit = 8;
640 * Check the rate limiting. Also, we do not want to throttle real-time
641 * tasks in balance_dirty_pages(). Period.
643 preempt_disable();
644 p = &__get_cpu_var(bdp_ratelimits);
645 *p += nr_pages_dirtied;
646 if (unlikely(*p >= ratelimit)) {
647 ratelimit = sync_writeback_pages(*p);
648 *p = 0;
649 preempt_enable();
650 balance_dirty_pages(mapping, ratelimit);
651 return;
653 preempt_enable();
655 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
657 void throttle_vm_writeout(gfp_t gfp_mask)
659 unsigned long background_thresh;
660 unsigned long dirty_thresh;
662 for ( ; ; ) {
663 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
666 * Boost the allowable dirty threshold a bit for page
667 * allocators so they don't get DoS'ed by heavy writers
669 dirty_thresh += dirty_thresh / 10; /* wheeee... */
671 if (global_page_state(NR_UNSTABLE_NFS) +
672 global_page_state(NR_WRITEBACK) <= dirty_thresh)
673 break;
674 congestion_wait(BLK_RW_ASYNC, HZ/10);
677 * The caller might hold locks which can prevent IO completion
678 * or progress in the filesystem. So we cannot just sit here
679 * waiting for IO to complete.
681 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
682 break;
686 static void laptop_timer_fn(unsigned long unused);
688 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
691 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
693 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
694 void __user *buffer, size_t *length, loff_t *ppos)
696 proc_dointvec(table, write, buffer, length, ppos);
697 return 0;
700 static void do_laptop_sync(struct work_struct *work)
702 wakeup_flusher_threads(0);
703 kfree(work);
706 static void laptop_timer_fn(unsigned long unused)
708 struct work_struct *work;
710 work = kmalloc(sizeof(*work), GFP_ATOMIC);
711 if (work) {
712 INIT_WORK(work, do_laptop_sync);
713 schedule_work(work);
718 * We've spun up the disk and we're in laptop mode: schedule writeback
719 * of all dirty data a few seconds from now. If the flush is already scheduled
720 * then push it back - the user is still using the disk.
722 void laptop_io_completion(void)
724 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
728 * We're in laptop mode and we've just synced. The sync's writes will have
729 * caused another writeback to be scheduled by laptop_io_completion.
730 * Nothing needs to be written back anymore, so we unschedule the writeback.
732 void laptop_sync_completion(void)
734 del_timer(&laptop_mode_wb_timer);
738 * If ratelimit_pages is too high then we can get into dirty-data overload
739 * if a large number of processes all perform writes at the same time.
740 * If it is too low then SMP machines will call the (expensive)
741 * get_writeback_state too often.
743 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
744 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
745 * thresholds before writeback cuts in.
747 * But the limit should not be set too high. Because it also controls the
748 * amount of memory which the balance_dirty_pages() caller has to write back.
749 * If this is too large then the caller will block on the IO queue all the
750 * time. So limit it to four megabytes - the balance_dirty_pages() caller
751 * will write six megabyte chunks, max.
754 void writeback_set_ratelimit(void)
756 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
757 if (ratelimit_pages < 16)
758 ratelimit_pages = 16;
759 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
760 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
763 static int __cpuinit
764 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
766 writeback_set_ratelimit();
767 return NOTIFY_DONE;
770 static struct notifier_block __cpuinitdata ratelimit_nb = {
771 .notifier_call = ratelimit_handler,
772 .next = NULL,
776 * Called early on to tune the page writeback dirty limits.
778 * We used to scale dirty pages according to how total memory
779 * related to pages that could be allocated for buffers (by
780 * comparing nr_free_buffer_pages() to vm_total_pages.
782 * However, that was when we used "dirty_ratio" to scale with
783 * all memory, and we don't do that any more. "dirty_ratio"
784 * is now applied to total non-HIGHPAGE memory (by subtracting
785 * totalhigh_pages from vm_total_pages), and as such we can't
786 * get into the old insane situation any more where we had
787 * large amounts of dirty pages compared to a small amount of
788 * non-HIGHMEM memory.
790 * But we might still want to scale the dirty_ratio by how
791 * much memory the box has..
793 void __init page_writeback_init(void)
795 int shift;
797 writeback_set_ratelimit();
798 register_cpu_notifier(&ratelimit_nb);
800 shift = calc_period_shift();
801 prop_descriptor_init(&vm_completions, shift);
802 prop_descriptor_init(&vm_dirties, shift);
806 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
807 * @mapping: address space structure to write
808 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
809 * @writepage: function called for each page
810 * @data: data passed to writepage function
812 * If a page is already under I/O, write_cache_pages() skips it, even
813 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
814 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
815 * and msync() need to guarantee that all the data which was dirty at the time
816 * the call was made get new I/O started against them. If wbc->sync_mode is
817 * WB_SYNC_ALL then we were called for data integrity and we must wait for
818 * existing IO to complete.
820 int write_cache_pages(struct address_space *mapping,
821 struct writeback_control *wbc, writepage_t writepage,
822 void *data)
824 struct backing_dev_info *bdi = mapping->backing_dev_info;
825 int ret = 0;
826 int done = 0;
827 struct pagevec pvec;
828 int nr_pages;
829 pgoff_t uninitialized_var(writeback_index);
830 pgoff_t index;
831 pgoff_t end; /* Inclusive */
832 pgoff_t done_index;
833 int cycled;
834 int range_whole = 0;
835 long nr_to_write = wbc->nr_to_write;
837 if (wbc->nonblocking && bdi_write_congested(bdi)) {
838 wbc->encountered_congestion = 1;
839 return 0;
842 pagevec_init(&pvec, 0);
843 if (wbc->range_cyclic) {
844 writeback_index = mapping->writeback_index; /* prev offset */
845 index = writeback_index;
846 if (index == 0)
847 cycled = 1;
848 else
849 cycled = 0;
850 end = -1;
851 } else {
852 index = wbc->range_start >> PAGE_CACHE_SHIFT;
853 end = wbc->range_end >> PAGE_CACHE_SHIFT;
854 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
855 range_whole = 1;
856 cycled = 1; /* ignore range_cyclic tests */
858 retry:
859 done_index = index;
860 while (!done && (index <= end)) {
861 int i;
863 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
864 PAGECACHE_TAG_DIRTY,
865 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
866 if (nr_pages == 0)
867 break;
869 for (i = 0; i < nr_pages; i++) {
870 struct page *page = pvec.pages[i];
873 * At this point, the page may be truncated or
874 * invalidated (changing page->mapping to NULL), or
875 * even swizzled back from swapper_space to tmpfs file
876 * mapping. However, page->index will not change
877 * because we have a reference on the page.
879 if (page->index > end) {
881 * can't be range_cyclic (1st pass) because
882 * end == -1 in that case.
884 done = 1;
885 break;
888 done_index = page->index + 1;
890 lock_page(page);
893 * Page truncated or invalidated. We can freely skip it
894 * then, even for data integrity operations: the page
895 * has disappeared concurrently, so there could be no
896 * real expectation of this data interity operation
897 * even if there is now a new, dirty page at the same
898 * pagecache address.
900 if (unlikely(page->mapping != mapping)) {
901 continue_unlock:
902 unlock_page(page);
903 continue;
906 if (!PageDirty(page)) {
907 /* someone wrote it for us */
908 goto continue_unlock;
911 if (PageWriteback(page)) {
912 if (wbc->sync_mode != WB_SYNC_NONE)
913 wait_on_page_writeback(page);
914 else
915 goto continue_unlock;
918 BUG_ON(PageWriteback(page));
919 if (!clear_page_dirty_for_io(page))
920 goto continue_unlock;
922 ret = (*writepage)(page, wbc, data);
923 if (unlikely(ret)) {
924 if (ret == AOP_WRITEPAGE_ACTIVATE) {
925 unlock_page(page);
926 ret = 0;
927 } else {
929 * done_index is set past this page,
930 * so media errors will not choke
931 * background writeout for the entire
932 * file. This has consequences for
933 * range_cyclic semantics (ie. it may
934 * not be suitable for data integrity
935 * writeout).
937 done = 1;
938 break;
942 if (nr_to_write > 0) {
943 nr_to_write--;
944 if (nr_to_write == 0 &&
945 wbc->sync_mode == WB_SYNC_NONE) {
947 * We stop writing back only if we are
948 * not doing integrity sync. In case of
949 * integrity sync we have to keep going
950 * because someone may be concurrently
951 * dirtying pages, and we might have
952 * synced a lot of newly appeared dirty
953 * pages, but have not synced all of the
954 * old dirty pages.
956 done = 1;
957 break;
961 if (wbc->nonblocking && bdi_write_congested(bdi)) {
962 wbc->encountered_congestion = 1;
963 done = 1;
964 break;
967 pagevec_release(&pvec);
968 cond_resched();
970 if (!cycled && !done) {
972 * range_cyclic:
973 * We hit the last page and there is more work to be done: wrap
974 * back to the start of the file
976 cycled = 1;
977 index = 0;
978 end = writeback_index - 1;
979 goto retry;
981 if (!wbc->no_nrwrite_index_update) {
982 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
983 mapping->writeback_index = done_index;
984 wbc->nr_to_write = nr_to_write;
987 return ret;
989 EXPORT_SYMBOL(write_cache_pages);
992 * Function used by generic_writepages to call the real writepage
993 * function and set the mapping flags on error
995 static int __writepage(struct page *page, struct writeback_control *wbc,
996 void *data)
998 struct address_space *mapping = data;
999 int ret = mapping->a_ops->writepage(page, wbc);
1000 mapping_set_error(mapping, ret);
1001 return ret;
1005 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1006 * @mapping: address space structure to write
1007 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1009 * This is a library function, which implements the writepages()
1010 * address_space_operation.
1012 int generic_writepages(struct address_space *mapping,
1013 struct writeback_control *wbc)
1015 /* deal with chardevs and other special file */
1016 if (!mapping->a_ops->writepage)
1017 return 0;
1019 return write_cache_pages(mapping, wbc, __writepage, mapping);
1022 EXPORT_SYMBOL(generic_writepages);
1024 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1026 int ret;
1028 if (wbc->nr_to_write <= 0)
1029 return 0;
1030 if (mapping->a_ops->writepages)
1031 ret = mapping->a_ops->writepages(mapping, wbc);
1032 else
1033 ret = generic_writepages(mapping, wbc);
1034 return ret;
1038 * write_one_page - write out a single page and optionally wait on I/O
1039 * @page: the page to write
1040 * @wait: if true, wait on writeout
1042 * The page must be locked by the caller and will be unlocked upon return.
1044 * write_one_page() returns a negative error code if I/O failed.
1046 int write_one_page(struct page *page, int wait)
1048 struct address_space *mapping = page->mapping;
1049 int ret = 0;
1050 struct writeback_control wbc = {
1051 .sync_mode = WB_SYNC_ALL,
1052 .nr_to_write = 1,
1055 BUG_ON(!PageLocked(page));
1057 if (wait)
1058 wait_on_page_writeback(page);
1060 if (clear_page_dirty_for_io(page)) {
1061 page_cache_get(page);
1062 ret = mapping->a_ops->writepage(page, &wbc);
1063 if (ret == 0 && wait) {
1064 wait_on_page_writeback(page);
1065 if (PageError(page))
1066 ret = -EIO;
1068 page_cache_release(page);
1069 } else {
1070 unlock_page(page);
1072 return ret;
1074 EXPORT_SYMBOL(write_one_page);
1077 * For address_spaces which do not use buffers nor write back.
1079 int __set_page_dirty_no_writeback(struct page *page)
1081 if (!PageDirty(page))
1082 SetPageDirty(page);
1083 return 0;
1087 * Helper function for set_page_dirty family.
1088 * NOTE: This relies on being atomic wrt interrupts.
1090 void account_page_dirtied(struct page *page, struct address_space *mapping)
1092 if (mapping_cap_account_dirty(mapping)) {
1093 __inc_zone_page_state(page, NR_FILE_DIRTY);
1094 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1095 task_dirty_inc(current);
1096 task_io_account_write(PAGE_CACHE_SIZE);
1101 * For address_spaces which do not use buffers. Just tag the page as dirty in
1102 * its radix tree.
1104 * This is also used when a single buffer is being dirtied: we want to set the
1105 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1106 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1108 * Most callers have locked the page, which pins the address_space in memory.
1109 * But zap_pte_range() does not lock the page, however in that case the
1110 * mapping is pinned by the vma's ->vm_file reference.
1112 * We take care to handle the case where the page was truncated from the
1113 * mapping by re-checking page_mapping() inside tree_lock.
1115 int __set_page_dirty_nobuffers(struct page *page)
1117 if (!TestSetPageDirty(page)) {
1118 struct address_space *mapping = page_mapping(page);
1119 struct address_space *mapping2;
1121 if (!mapping)
1122 return 1;
1124 spin_lock_irq(&mapping->tree_lock);
1125 mapping2 = page_mapping(page);
1126 if (mapping2) { /* Race with truncate? */
1127 BUG_ON(mapping2 != mapping);
1128 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1129 account_page_dirtied(page, mapping);
1130 radix_tree_tag_set(&mapping->page_tree,
1131 page_index(page), PAGECACHE_TAG_DIRTY);
1133 spin_unlock_irq(&mapping->tree_lock);
1134 if (mapping->host) {
1135 /* !PageAnon && !swapper_space */
1136 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1138 return 1;
1140 return 0;
1142 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1145 * set_page_dirty_notag() -- similar to __set_page_dirty_nobuffers()
1146 * except it doesn't tag the page dirty in the page-cache radix tree.
1147 * This means that the address space using this cannot use the regular
1148 * filemap ->writepages() helpers and must provide its own means of
1149 * tracking and finding non-tagged dirty pages.
1151 * NOTE: furthermore, this version also doesn't handle truncate races.
1153 int set_page_dirty_notag(struct page *page)
1155 struct address_space *mapping = page->mapping;
1157 if (!TestSetPageDirty(page)) {
1158 unsigned long flags;
1159 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1160 local_irq_save(flags);
1161 account_page_dirtied(page, mapping);
1162 local_irq_restore(flags);
1163 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1164 return 1;
1166 return 0;
1168 EXPORT_SYMBOL(set_page_dirty_notag);
1171 * When a writepage implementation decides that it doesn't want to write this
1172 * page for some reason, it should redirty the locked page via
1173 * redirty_page_for_writepage() and it should then unlock the page and return 0
1175 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1177 wbc->pages_skipped++;
1178 return __set_page_dirty_nobuffers(page);
1180 EXPORT_SYMBOL(redirty_page_for_writepage);
1183 * Dirty a page.
1185 * For pages with a mapping this should be done under the page lock
1186 * for the benefit of asynchronous memory errors who prefer a consistent
1187 * dirty state. This rule can be broken in some special cases,
1188 * but should be better not to.
1190 * If the mapping doesn't provide a set_page_dirty a_op, then
1191 * just fall through and assume that it wants buffer_heads.
1193 int set_page_dirty(struct page *page)
1195 struct address_space *mapping = page_mapping(page);
1197 if (likely(mapping)) {
1198 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1199 #ifdef CONFIG_BLOCK
1200 if (!spd)
1201 spd = __set_page_dirty_buffers;
1202 #endif
1203 return (*spd)(page);
1205 if (!PageDirty(page)) {
1206 if (!TestSetPageDirty(page))
1207 return 1;
1209 return 0;
1211 EXPORT_SYMBOL(set_page_dirty);
1214 * set_page_dirty() is racy if the caller has no reference against
1215 * page->mapping->host, and if the page is unlocked. This is because another
1216 * CPU could truncate the page off the mapping and then free the mapping.
1218 * Usually, the page _is_ locked, or the caller is a user-space process which
1219 * holds a reference on the inode by having an open file.
1221 * In other cases, the page should be locked before running set_page_dirty().
1223 int set_page_dirty_lock(struct page *page)
1225 int ret;
1227 lock_page_nosync(page);
1228 ret = set_page_dirty(page);
1229 unlock_page(page);
1230 return ret;
1232 EXPORT_SYMBOL(set_page_dirty_lock);
1235 * Clear a page's dirty flag, while caring for dirty memory accounting.
1236 * Returns true if the page was previously dirty.
1238 * This is for preparing to put the page under writeout. We leave the page
1239 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1240 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1241 * implementation will run either set_page_writeback() or set_page_dirty(),
1242 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1243 * back into sync.
1245 * This incoherency between the page's dirty flag and radix-tree tag is
1246 * unfortunate, but it only exists while the page is locked.
1248 int clear_page_dirty_for_io(struct page *page)
1250 struct address_space *mapping = page_mapping(page);
1252 BUG_ON(!PageLocked(page));
1254 ClearPageReclaim(page);
1255 if (mapping && mapping_cap_account_dirty(mapping)) {
1257 * Yes, Virginia, this is indeed insane.
1259 * We use this sequence to make sure that
1260 * (a) we account for dirty stats properly
1261 * (b) we tell the low-level filesystem to
1262 * mark the whole page dirty if it was
1263 * dirty in a pagetable. Only to then
1264 * (c) clean the page again and return 1 to
1265 * cause the writeback.
1267 * This way we avoid all nasty races with the
1268 * dirty bit in multiple places and clearing
1269 * them concurrently from different threads.
1271 * Note! Normally the "set_page_dirty(page)"
1272 * has no effect on the actual dirty bit - since
1273 * that will already usually be set. But we
1274 * need the side effects, and it can help us
1275 * avoid races.
1277 * We basically use the page "master dirty bit"
1278 * as a serialization point for all the different
1279 * threads doing their things.
1281 if (page_mkclean(page))
1282 set_page_dirty(page);
1284 * We carefully synchronise fault handlers against
1285 * installing a dirty pte and marking the page dirty
1286 * at this point. We do this by having them hold the
1287 * page lock at some point after installing their
1288 * pte, but before marking the page dirty.
1289 * Pages are always locked coming in here, so we get
1290 * the desired exclusion. See mm/memory.c:do_wp_page()
1291 * for more comments.
1293 if (TestClearPageDirty(page)) {
1294 dec_zone_page_state(page, NR_FILE_DIRTY);
1295 dec_bdi_stat(mapping->backing_dev_info,
1296 BDI_RECLAIMABLE);
1297 return 1;
1299 return 0;
1301 return TestClearPageDirty(page);
1303 EXPORT_SYMBOL(clear_page_dirty_for_io);
1305 int test_clear_page_writeback(struct page *page)
1307 struct address_space *mapping = page_mapping(page);
1308 int ret;
1310 if (mapping) {
1311 struct backing_dev_info *bdi = mapping->backing_dev_info;
1312 unsigned long flags;
1314 spin_lock_irqsave(&mapping->tree_lock, flags);
1315 ret = TestClearPageWriteback(page);
1316 if (ret) {
1317 radix_tree_tag_clear(&mapping->page_tree,
1318 page_index(page),
1319 PAGECACHE_TAG_WRITEBACK);
1320 if (bdi_cap_account_writeback(bdi)) {
1321 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1322 __bdi_writeout_inc(bdi);
1325 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1326 } else {
1327 ret = TestClearPageWriteback(page);
1329 if (ret)
1330 dec_zone_page_state(page, NR_WRITEBACK);
1331 return ret;
1334 int test_set_page_writeback(struct page *page)
1336 struct address_space *mapping = page_mapping(page);
1337 int ret;
1339 if (mapping) {
1340 struct backing_dev_info *bdi = mapping->backing_dev_info;
1341 unsigned long flags;
1343 spin_lock_irqsave(&mapping->tree_lock, flags);
1344 ret = TestSetPageWriteback(page);
1345 if (!ret) {
1346 radix_tree_tag_set(&mapping->page_tree,
1347 page_index(page),
1348 PAGECACHE_TAG_WRITEBACK);
1349 if (bdi_cap_account_writeback(bdi))
1350 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1352 if (!PageDirty(page))
1353 radix_tree_tag_clear(&mapping->page_tree,
1354 page_index(page),
1355 PAGECACHE_TAG_DIRTY);
1356 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1357 } else {
1358 ret = TestSetPageWriteback(page);
1360 if (!ret)
1361 inc_zone_page_state(page, NR_WRITEBACK);
1362 return ret;
1365 EXPORT_SYMBOL(test_set_page_writeback);
1368 * Return true if any of the pages in the mapping are marked with the
1369 * passed tag.
1371 int mapping_tagged(struct address_space *mapping, int tag)
1373 int ret;
1374 rcu_read_lock();
1375 ret = radix_tree_tagged(&mapping->page_tree, tag);
1376 rcu_read_unlock();
1377 return ret;
1379 EXPORT_SYMBOL(mapping_tagged);