Root/mm/page-writeback.c

1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
25#include <linux/task_io_accounting_ops.h>
26#include <linux/blkdev.h>
27#include <linux/mpage.h>
28#include <linux/rmap.h>
29#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h>
36#include <linux/pagevec.h>
37#include <trace/events/writeback.h>
38
39/*
40 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
41 * will look to see if it needs to force writeback or throttling.
42 */
43static long ratelimit_pages = 32;
44
45/*
46 * When balance_dirty_pages decides that the caller needs to perform some
47 * non-background writeback, this is how many pages it will attempt to write.
48 * It should be somewhat larger than dirtied pages to ensure that reasonably
49 * large amounts of I/O are submitted.
50 */
51static inline long sync_writeback_pages(unsigned long dirtied)
52{
53    if (dirtied < ratelimit_pages)
54        dirtied = ratelimit_pages;
55
56    return dirtied + dirtied / 2;
57}
58
59/* The following parameters are exported via /proc/sys/vm */
60
61/*
62 * Start background writeback (via writeback threads) at this percentage
63 */
64int dirty_background_ratio = 10;
65
66/*
67 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
68 * dirty_background_ratio * the amount of dirtyable memory
69 */
70unsigned long dirty_background_bytes;
71
72/*
73 * free highmem will not be subtracted from the total free memory
74 * for calculating free ratios if vm_highmem_is_dirtyable is true
75 */
76int vm_highmem_is_dirtyable;
77
78/*
79 * The generator of dirty data starts writeback at this percentage
80 */
81int vm_dirty_ratio = 20;
82
83/*
84 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
85 * vm_dirty_ratio * the amount of dirtyable memory
86 */
87unsigned long vm_dirty_bytes;
88
89/*
90 * The interval between `kupdate'-style writebacks
91 */
92unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
93
94/*
95 * The longest time for which data is allowed to remain dirty
96 */
97unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
98
99/*
100 * Flag that makes the machine dump writes/reads and block dirtyings.
101 */
102int block_dump;
103
104/*
105 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
106 * a full sync is triggered after this time elapses without any disk activity.
107 */
108int laptop_mode;
109
110EXPORT_SYMBOL(laptop_mode);
111
112/* End of sysctl-exported parameters */
113
114
115/*
116 * Scale the writeback cache size proportional to the relative writeout speeds.
117 *
118 * We do this by keeping a floating proportion between BDIs, based on page
119 * writeback completions [end_page_writeback()]. Those devices that write out
120 * pages fastest will get the larger share, while the slower will get a smaller
121 * share.
122 *
123 * We use page writeout completions because we are interested in getting rid of
124 * dirty pages. Having them written out is the primary goal.
125 *
126 * We introduce a concept of time, a period over which we measure these events,
127 * because demand can/will vary over time. The length of this period itself is
128 * measured in page writeback completions.
129 *
130 */
131static struct prop_descriptor vm_completions;
132static struct prop_descriptor vm_dirties;
133
134/*
135 * couple the period to the dirty_ratio:
136 *
137 * period/2 ~ roundup_pow_of_two(dirty limit)
138 */
139static int calc_period_shift(void)
140{
141    unsigned long dirty_total;
142
143    if (vm_dirty_bytes)
144        dirty_total = vm_dirty_bytes / PAGE_SIZE;
145    else
146        dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
147                100;
148    return 2 + ilog2(dirty_total - 1);
149}
150
151/*
152 * update the period when the dirty threshold changes.
153 */
154static void update_completion_period(void)
155{
156    int shift = calc_period_shift();
157    prop_change_shift(&vm_completions, shift);
158    prop_change_shift(&vm_dirties, shift);
159}
160
161int dirty_background_ratio_handler(struct ctl_table *table, int write,
162        void __user *buffer, size_t *lenp,
163        loff_t *ppos)
164{
165    int ret;
166
167    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
168    if (ret == 0 && write)
169        dirty_background_bytes = 0;
170    return ret;
171}
172
173int dirty_background_bytes_handler(struct ctl_table *table, int write,
174        void __user *buffer, size_t *lenp,
175        loff_t *ppos)
176{
177    int ret;
178
179    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
180    if (ret == 0 && write)
181        dirty_background_ratio = 0;
182    return ret;
183}
184
185int dirty_ratio_handler(struct ctl_table *table, int write,
186        void __user *buffer, size_t *lenp,
187        loff_t *ppos)
188{
189    int old_ratio = vm_dirty_ratio;
190    int ret;
191
192    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
193    if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
194        update_completion_period();
195        vm_dirty_bytes = 0;
196    }
197    return ret;
198}
199
200
201int dirty_bytes_handler(struct ctl_table *table, int write,
202        void __user *buffer, size_t *lenp,
203        loff_t *ppos)
204{
205    unsigned long old_bytes = vm_dirty_bytes;
206    int ret;
207
208    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
209    if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
210        update_completion_period();
211        vm_dirty_ratio = 0;
212    }
213    return ret;
214}
215
216/*
217 * Increment the BDI's writeout completion count and the global writeout
218 * completion count. Called from test_clear_page_writeback().
219 */
220static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
221{
222    __prop_inc_percpu_max(&vm_completions, &bdi->completions,
223                  bdi->max_prop_frac);
224}
225
226void bdi_writeout_inc(struct backing_dev_info *bdi)
227{
228    unsigned long flags;
229
230    local_irq_save(flags);
231    __bdi_writeout_inc(bdi);
232    local_irq_restore(flags);
233}
234EXPORT_SYMBOL_GPL(bdi_writeout_inc);
235
236void task_dirty_inc(struct task_struct *tsk)
237{
238    prop_inc_single(&vm_dirties, &tsk->dirties);
239}
240
241/*
242 * Obtain an accurate fraction of the BDI's portion.
243 */
244static void bdi_writeout_fraction(struct backing_dev_info *bdi,
245        long *numerator, long *denominator)
246{
247    if (bdi_cap_writeback_dirty(bdi)) {
248        prop_fraction_percpu(&vm_completions, &bdi->completions,
249                numerator, denominator);
250    } else {
251        *numerator = 0;
252        *denominator = 1;
253    }
254}
255
256static inline void task_dirties_fraction(struct task_struct *tsk,
257        long *numerator, long *denominator)
258{
259    prop_fraction_single(&vm_dirties, &tsk->dirties,
260                numerator, denominator);
261}
262
263/*
264 * task_dirty_limit - scale down dirty throttling threshold for one task
265 *
266 * task specific dirty limit:
267 *
268 * dirty -= (dirty/8) * p_{t}
269 *
270 * To protect light/slow dirtying tasks from heavier/fast ones, we start
271 * throttling individual tasks before reaching the bdi dirty limit.
272 * Relatively low thresholds will be allocated to heavy dirtiers. So when
273 * dirty pages grow large, heavy dirtiers will be throttled first, which will
274 * effectively curb the growth of dirty pages. Light dirtiers with high enough
275 * dirty threshold may never get throttled.
276 */
277static unsigned long task_dirty_limit(struct task_struct *tsk,
278                       unsigned long bdi_dirty)
279{
280    long numerator, denominator;
281    unsigned long dirty = bdi_dirty;
282    u64 inv = dirty >> 3;
283
284    task_dirties_fraction(tsk, &numerator, &denominator);
285    inv *= numerator;
286    do_div(inv, denominator);
287
288    dirty -= inv;
289
290    return max(dirty, bdi_dirty/2);
291}
292
293/*
294 *
295 */
296static unsigned int bdi_min_ratio;
297
298int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
299{
300    int ret = 0;
301
302    spin_lock_bh(&bdi_lock);
303    if (min_ratio > bdi->max_ratio) {
304        ret = -EINVAL;
305    } else {
306        min_ratio -= bdi->min_ratio;
307        if (bdi_min_ratio + min_ratio < 100) {
308            bdi_min_ratio += min_ratio;
309            bdi->min_ratio += min_ratio;
310        } else {
311            ret = -EINVAL;
312        }
313    }
314    spin_unlock_bh(&bdi_lock);
315
316    return ret;
317}
318
319int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
320{
321    int ret = 0;
322
323    if (max_ratio > 100)
324        return -EINVAL;
325
326    spin_lock_bh(&bdi_lock);
327    if (bdi->min_ratio > max_ratio) {
328        ret = -EINVAL;
329    } else {
330        bdi->max_ratio = max_ratio;
331        bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
332    }
333    spin_unlock_bh(&bdi_lock);
334
335    return ret;
336}
337EXPORT_SYMBOL(bdi_set_max_ratio);
338
339/*
340 * Work out the current dirty-memory clamping and background writeout
341 * thresholds.
342 *
343 * The main aim here is to lower them aggressively if there is a lot of mapped
344 * memory around. To avoid stressing page reclaim with lots of unreclaimable
345 * pages. It is better to clamp down on writers than to start swapping, and
346 * performing lots of scanning.
347 *
348 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
349 *
350 * We don't permit the clamping level to fall below 5% - that is getting rather
351 * excessive.
352 *
353 * We make sure that the background writeout level is below the adjusted
354 * clamping level.
355 */
356
357static unsigned long highmem_dirtyable_memory(unsigned long total)
358{
359#ifdef CONFIG_HIGHMEM
360    int node;
361    unsigned long x = 0;
362
363    for_each_node_state(node, N_HIGH_MEMORY) {
364        struct zone *z =
365            &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
366
367        x += zone_page_state(z, NR_FREE_PAGES) +
368             zone_reclaimable_pages(z);
369    }
370    /*
371     * Make sure that the number of highmem pages is never larger
372     * than the number of the total dirtyable memory. This can only
373     * occur in very strange VM situations but we want to make sure
374     * that this does not occur.
375     */
376    return min(x, total);
377#else
378    return 0;
379#endif
380}
381
382/**
383 * determine_dirtyable_memory - amount of memory that may be used
384 *
385 * Returns the numebr of pages that can currently be freed and used
386 * by the kernel for direct mappings.
387 */
388unsigned long determine_dirtyable_memory(void)
389{
390    unsigned long x;
391
392    x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
393
394    if (!vm_highmem_is_dirtyable)
395        x -= highmem_dirtyable_memory(x);
396
397    return x + 1; /* Ensure that we never return 0 */
398}
399
400/*
401 * global_dirty_limits - background-writeback and dirty-throttling thresholds
402 *
403 * Calculate the dirty thresholds based on sysctl parameters
404 * - vm.dirty_background_ratio or vm.dirty_background_bytes
405 * - vm.dirty_ratio or vm.dirty_bytes
406 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
407 * runtime tasks.
408 */
409void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
410{
411    unsigned long background;
412    unsigned long dirty;
413    unsigned long available_memory = determine_dirtyable_memory();
414    struct task_struct *tsk;
415
416    if (vm_dirty_bytes)
417        dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
418    else {
419        int dirty_ratio;
420
421        dirty_ratio = vm_dirty_ratio;
422        if (dirty_ratio < 5)
423            dirty_ratio = 5;
424        dirty = (dirty_ratio * available_memory) / 100;
425    }
426
427    if (dirty_background_bytes)
428        background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
429    else
430        background = (dirty_background_ratio * available_memory) / 100;
431
432    if (background >= dirty)
433        background = dirty / 2;
434    tsk = current;
435    if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
436        background += background / 4;
437        dirty += dirty / 4;
438    }
439    *pbackground = background;
440    *pdirty = dirty;
441}
442
443/*
444 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
445 *
446 * Allocate high/low dirty limits to fast/slow devices, in order to prevent
447 * - starving fast devices
448 * - piling up dirty pages (that will take long time to sync) on slow devices
449 *
450 * The bdi's share of dirty limit will be adapting to its throughput and
451 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
452 */
453unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
454{
455    u64 bdi_dirty;
456    long numerator, denominator;
457
458    /*
459     * Calculate this BDI's share of the dirty ratio.
460     */
461    bdi_writeout_fraction(bdi, &numerator, &denominator);
462
463    bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
464    bdi_dirty *= numerator;
465    do_div(bdi_dirty, denominator);
466
467    bdi_dirty += (dirty * bdi->min_ratio) / 100;
468    if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
469        bdi_dirty = dirty * bdi->max_ratio / 100;
470
471    return bdi_dirty;
472}
473
474/*
475 * balance_dirty_pages() must be called by processes which are generating dirty
476 * data. It looks at the number of dirty pages in the machine and will force
477 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
478 * If we're over `background_thresh' then the writeback threads are woken to
479 * perform some writeout.
480 */
481static void balance_dirty_pages(struct address_space *mapping,
482                unsigned long write_chunk)
483{
484    long nr_reclaimable, bdi_nr_reclaimable;
485    long nr_writeback, bdi_nr_writeback;
486    unsigned long background_thresh;
487    unsigned long dirty_thresh;
488    unsigned long bdi_thresh;
489    unsigned long pages_written = 0;
490    unsigned long pause = 1;
491    bool dirty_exceeded = false;
492    struct backing_dev_info *bdi = mapping->backing_dev_info;
493
494    for (;;) {
495        struct writeback_control wbc = {
496            .sync_mode = WB_SYNC_NONE,
497            .older_than_this = NULL,
498            .nr_to_write = write_chunk,
499            .range_cyclic = 1,
500        };
501
502        nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
503                    global_page_state(NR_UNSTABLE_NFS);
504        nr_writeback = global_page_state(NR_WRITEBACK);
505
506        global_dirty_limits(&background_thresh, &dirty_thresh);
507
508        /*
509         * Throttle it only when the background writeback cannot
510         * catch-up. This avoids (excessively) small writeouts
511         * when the bdi limits are ramping up.
512         */
513        if (nr_reclaimable + nr_writeback <
514                (background_thresh + dirty_thresh) / 2)
515            break;
516
517        bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
518        bdi_thresh = task_dirty_limit(current, bdi_thresh);
519
520        /*
521         * In order to avoid the stacked BDI deadlock we need
522         * to ensure we accurately count the 'dirty' pages when
523         * the threshold is low.
524         *
525         * Otherwise it would be possible to get thresh+n pages
526         * reported dirty, even though there are thresh-m pages
527         * actually dirty; with m+n sitting in the percpu
528         * deltas.
529         */
530        if (bdi_thresh < 2*bdi_stat_error(bdi)) {
531            bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
532            bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
533        } else {
534            bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
535            bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
536        }
537
538        /*
539         * The bdi thresh is somehow "soft" limit derived from the
540         * global "hard" limit. The former helps to prevent heavy IO
541         * bdi or process from holding back light ones; The latter is
542         * the last resort safeguard.
543         */
544        dirty_exceeded =
545            (bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
546            || (nr_reclaimable + nr_writeback >= dirty_thresh);
547
548        if (!dirty_exceeded)
549            break;
550
551        if (!bdi->dirty_exceeded)
552            bdi->dirty_exceeded = 1;
553
554        /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
555         * Unstable writes are a feature of certain networked
556         * filesystems (i.e. NFS) in which data may have been
557         * written to the server's write cache, but has not yet
558         * been flushed to permanent storage.
559         * Only move pages to writeback if this bdi is over its
560         * threshold otherwise wait until the disk writes catch
561         * up.
562         */
563        trace_wbc_balance_dirty_start(&wbc, bdi);
564        if (bdi_nr_reclaimable > bdi_thresh) {
565            writeback_inodes_wb(&bdi->wb, &wbc);
566            pages_written += write_chunk - wbc.nr_to_write;
567            trace_wbc_balance_dirty_written(&wbc, bdi);
568            if (pages_written >= write_chunk)
569                break; /* We've done our duty */
570        }
571        trace_wbc_balance_dirty_wait(&wbc, bdi);
572        __set_current_state(TASK_INTERRUPTIBLE);
573        io_schedule_timeout(pause);
574
575        /*
576         * Increase the delay for each loop, up to our previous
577         * default of taking a 100ms nap.
578         */
579        pause <<= 1;
580        if (pause > HZ / 10)
581            pause = HZ / 10;
582    }
583
584    if (!dirty_exceeded && bdi->dirty_exceeded)
585        bdi->dirty_exceeded = 0;
586
587    if (writeback_in_progress(bdi))
588        return;
589
590    /*
591     * In laptop mode, we wait until hitting the higher threshold before
592     * starting background writeout, and then write out all the way down
593     * to the lower threshold. So slow writers cause minimal disk activity.
594     *
595     * In normal mode, we start background writeout at the lower
596     * background_thresh, to keep the amount of dirty memory low.
597     */
598    if ((laptop_mode && pages_written) ||
599        (!laptop_mode && (nr_reclaimable > background_thresh)))
600        bdi_start_background_writeback(bdi);
601}
602
603void set_page_dirty_balance(struct page *page, int page_mkwrite)
604{
605    if (set_page_dirty(page) || page_mkwrite) {
606        struct address_space *mapping = page_mapping(page);
607
608        if (mapping)
609            balance_dirty_pages_ratelimited(mapping);
610    }
611}
612
613static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
614
615/**
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
619 *
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.
623 *
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.
628 */
629void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
630                    unsigned long nr_pages_dirtied)
631{
632    unsigned long ratelimit;
633    unsigned long *p;
634
635    ratelimit = ratelimit_pages;
636    if (mapping->backing_dev_info->dirty_exceeded)
637        ratelimit = 8;
638
639    /*
640     * Check the rate limiting. Also, we do not want to throttle real-time
641     * tasks in balance_dirty_pages(). Period.
642     */
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;
652    }
653    preempt_enable();
654}
655EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
656
657void throttle_vm_writeout(gfp_t gfp_mask)
658{
659    unsigned long background_thresh;
660    unsigned long dirty_thresh;
661
662        for ( ; ; ) {
663        global_dirty_limits(&background_thresh, &dirty_thresh);
664
665                /*
666                 * Boost the allowable dirty threshold a bit for page
667                 * allocators so they don't get DoS'ed by heavy writers
668                 */
669                dirty_thresh += dirty_thresh / 10; /* wheeee... */
670
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);
675
676        /*
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.
680         */
681        if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
682            break;
683        }
684}
685
686/*
687 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
688 */
689int dirty_writeback_centisecs_handler(ctl_table *table, int write,
690    void __user *buffer, size_t *length, loff_t *ppos)
691{
692    proc_dointvec(table, write, buffer, length, ppos);
693    bdi_arm_supers_timer();
694    return 0;
695}
696
697#ifdef CONFIG_BLOCK
698void laptop_mode_timer_fn(unsigned long data)
699{
700    struct request_queue *q = (struct request_queue *)data;
701    int nr_pages = global_page_state(NR_FILE_DIRTY) +
702        global_page_state(NR_UNSTABLE_NFS);
703
704    /*
705     * We want to write everything out, not just down to the dirty
706     * threshold
707     */
708    if (bdi_has_dirty_io(&q->backing_dev_info))
709        bdi_start_writeback(&q->backing_dev_info, nr_pages);
710}
711
712/*
713 * We've spun up the disk and we're in laptop mode: schedule writeback
714 * of all dirty data a few seconds from now. If the flush is already scheduled
715 * then push it back - the user is still using the disk.
716 */
717void laptop_io_completion(struct backing_dev_info *info)
718{
719    mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
720}
721
722/*
723 * We're in laptop mode and we've just synced. The sync's writes will have
724 * caused another writeback to be scheduled by laptop_io_completion.
725 * Nothing needs to be written back anymore, so we unschedule the writeback.
726 */
727void laptop_sync_completion(void)
728{
729    struct backing_dev_info *bdi;
730
731    rcu_read_lock();
732
733    list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
734        del_timer(&bdi->laptop_mode_wb_timer);
735
736    rcu_read_unlock();
737}
738#endif
739
740/*
741 * If ratelimit_pages is too high then we can get into dirty-data overload
742 * if a large number of processes all perform writes at the same time.
743 * If it is too low then SMP machines will call the (expensive)
744 * get_writeback_state too often.
745 *
746 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
747 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
748 * thresholds before writeback cuts in.
749 *
750 * But the limit should not be set too high. Because it also controls the
751 * amount of memory which the balance_dirty_pages() caller has to write back.
752 * If this is too large then the caller will block on the IO queue all the
753 * time. So limit it to four megabytes - the balance_dirty_pages() caller
754 * will write six megabyte chunks, max.
755 */
756
757void writeback_set_ratelimit(void)
758{
759    ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
760    if (ratelimit_pages < 16)
761        ratelimit_pages = 16;
762    if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
763        ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
764}
765
766static int __cpuinit
767ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
768{
769    writeback_set_ratelimit();
770    return NOTIFY_DONE;
771}
772
773static struct notifier_block __cpuinitdata ratelimit_nb = {
774    .notifier_call = ratelimit_handler,
775    .next = NULL,
776};
777
778/*
779 * Called early on to tune the page writeback dirty limits.
780 *
781 * We used to scale dirty pages according to how total memory
782 * related to pages that could be allocated for buffers (by
783 * comparing nr_free_buffer_pages() to vm_total_pages.
784 *
785 * However, that was when we used "dirty_ratio" to scale with
786 * all memory, and we don't do that any more. "dirty_ratio"
787 * is now applied to total non-HIGHPAGE memory (by subtracting
788 * totalhigh_pages from vm_total_pages), and as such we can't
789 * get into the old insane situation any more where we had
790 * large amounts of dirty pages compared to a small amount of
791 * non-HIGHMEM memory.
792 *
793 * But we might still want to scale the dirty_ratio by how
794 * much memory the box has..
795 */
796void __init page_writeback_init(void)
797{
798    int shift;
799
800    writeback_set_ratelimit();
801    register_cpu_notifier(&ratelimit_nb);
802
803    shift = calc_period_shift();
804    prop_descriptor_init(&vm_completions, shift);
805    prop_descriptor_init(&vm_dirties, shift);
806}
807
808/**
809 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
810 * @mapping: address space structure to write
811 * @start: starting page index
812 * @end: ending page index (inclusive)
813 *
814 * This function scans the page range from @start to @end (inclusive) and tags
815 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
816 * that write_cache_pages (or whoever calls this function) will then use
817 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
818 * used to avoid livelocking of writeback by a process steadily creating new
819 * dirty pages in the file (thus it is important for this function to be quick
820 * so that it can tag pages faster than a dirtying process can create them).
821 */
822/*
823 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
824 */
825void tag_pages_for_writeback(struct address_space *mapping,
826                 pgoff_t start, pgoff_t end)
827{
828#define WRITEBACK_TAG_BATCH 4096
829    unsigned long tagged;
830
831    do {
832        spin_lock_irq(&mapping->tree_lock);
833        tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
834                &start, end, WRITEBACK_TAG_BATCH,
835                PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
836        spin_unlock_irq(&mapping->tree_lock);
837        WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
838        cond_resched();
839        /* We check 'start' to handle wrapping when end == ~0UL */
840    } while (tagged >= WRITEBACK_TAG_BATCH && start);
841}
842EXPORT_SYMBOL(tag_pages_for_writeback);
843
844/**
845 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
846 * @mapping: address space structure to write
847 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
848 * @writepage: function called for each page
849 * @data: data passed to writepage function
850 *
851 * If a page is already under I/O, write_cache_pages() skips it, even
852 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
853 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
854 * and msync() need to guarantee that all the data which was dirty at the time
855 * the call was made get new I/O started against them. If wbc->sync_mode is
856 * WB_SYNC_ALL then we were called for data integrity and we must wait for
857 * existing IO to complete.
858 *
859 * To avoid livelocks (when other process dirties new pages), we first tag
860 * pages which should be written back with TOWRITE tag and only then start
861 * writing them. For data-integrity sync we have to be careful so that we do
862 * not miss some pages (e.g., because some other process has cleared TOWRITE
863 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
864 * by the process clearing the DIRTY tag (and submitting the page for IO).
865 */
866int write_cache_pages(struct address_space *mapping,
867              struct writeback_control *wbc, writepage_t writepage,
868              void *data)
869{
870    int ret = 0;
871    int done = 0;
872    struct pagevec pvec;
873    int nr_pages;
874    pgoff_t uninitialized_var(writeback_index);
875    pgoff_t index;
876    pgoff_t end; /* Inclusive */
877    pgoff_t done_index;
878    int cycled;
879    int range_whole = 0;
880    int tag;
881
882    pagevec_init(&pvec, 0);
883    if (wbc->range_cyclic) {
884        writeback_index = mapping->writeback_index; /* prev offset */
885        index = writeback_index;
886        if (index == 0)
887            cycled = 1;
888        else
889            cycled = 0;
890        end = -1;
891    } else {
892        index = wbc->range_start >> PAGE_CACHE_SHIFT;
893        end = wbc->range_end >> PAGE_CACHE_SHIFT;
894        if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
895            range_whole = 1;
896        cycled = 1; /* ignore range_cyclic tests */
897    }
898    if (wbc->sync_mode == WB_SYNC_ALL)
899        tag = PAGECACHE_TAG_TOWRITE;
900    else
901        tag = PAGECACHE_TAG_DIRTY;
902retry:
903    if (wbc->sync_mode == WB_SYNC_ALL)
904        tag_pages_for_writeback(mapping, index, end);
905    done_index = index;
906    while (!done && (index <= end)) {
907        int i;
908
909        nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
910                  min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
911        if (nr_pages == 0)
912            break;
913
914        for (i = 0; i < nr_pages; i++) {
915            struct page *page = pvec.pages[i];
916
917            /*
918             * At this point, the page may be truncated or
919             * invalidated (changing page->mapping to NULL), or
920             * even swizzled back from swapper_space to tmpfs file
921             * mapping. However, page->index will not change
922             * because we have a reference on the page.
923             */
924            if (page->index > end) {
925                /*
926                 * can't be range_cyclic (1st pass) because
927                 * end == -1 in that case.
928                 */
929                done = 1;
930                break;
931            }
932
933            done_index = page->index + 1;
934
935            lock_page(page);
936
937            /*
938             * Page truncated or invalidated. We can freely skip it
939             * then, even for data integrity operations: the page
940             * has disappeared concurrently, so there could be no
941             * real expectation of this data interity operation
942             * even if there is now a new, dirty page at the same
943             * pagecache address.
944             */
945            if (unlikely(page->mapping != mapping)) {
946continue_unlock:
947                unlock_page(page);
948                continue;
949            }
950
951            if (!PageDirty(page)) {
952                /* someone wrote it for us */
953                goto continue_unlock;
954            }
955
956            if (PageWriteback(page)) {
957                if (wbc->sync_mode != WB_SYNC_NONE)
958                    wait_on_page_writeback(page);
959                else
960                    goto continue_unlock;
961            }
962
963            BUG_ON(PageWriteback(page));
964            if (!clear_page_dirty_for_io(page))
965                goto continue_unlock;
966
967            trace_wbc_writepage(wbc, mapping->backing_dev_info);
968            ret = (*writepage)(page, wbc, data);
969            if (unlikely(ret)) {
970                if (ret == AOP_WRITEPAGE_ACTIVATE) {
971                    unlock_page(page);
972                    ret = 0;
973                } else {
974                    /*
975                     * done_index is set past this page,
976                     * so media errors will not choke
977                     * background writeout for the entire
978                     * file. This has consequences for
979                     * range_cyclic semantics (ie. it may
980                     * not be suitable for data integrity
981                     * writeout).
982                     */
983                    done = 1;
984                    break;
985                }
986            }
987
988            /*
989             * We stop writing back only if we are not doing
990             * integrity sync. In case of integrity sync we have to
991             * keep going until we have written all the pages
992             * we tagged for writeback prior to entering this loop.
993             */
994            if (--wbc->nr_to_write <= 0 &&
995                wbc->sync_mode == WB_SYNC_NONE) {
996                done = 1;
997                break;
998            }
999        }
1000        pagevec_release(&pvec);
1001        cond_resched();
1002    }
1003    if (!cycled && !done) {
1004        /*
1005         * range_cyclic:
1006         * We hit the last page and there is more work to be done: wrap
1007         * back to the start of the file
1008         */
1009        cycled = 1;
1010        index = 0;
1011        end = writeback_index - 1;
1012        goto retry;
1013    }
1014    if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1015        mapping->writeback_index = done_index;
1016
1017    return ret;
1018}
1019EXPORT_SYMBOL(write_cache_pages);
1020
1021/*
1022 * Function used by generic_writepages to call the real writepage
1023 * function and set the mapping flags on error
1024 */
1025static int __writepage(struct page *page, struct writeback_control *wbc,
1026               void *data)
1027{
1028    struct address_space *mapping = data;
1029    int ret = mapping->a_ops->writepage(page, wbc);
1030    mapping_set_error(mapping, ret);
1031    return ret;
1032}
1033
1034/**
1035 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1036 * @mapping: address space structure to write
1037 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1038 *
1039 * This is a library function, which implements the writepages()
1040 * address_space_operation.
1041 */
1042int generic_writepages(struct address_space *mapping,
1043               struct writeback_control *wbc)
1044{
1045    /* deal with chardevs and other special file */
1046    if (!mapping->a_ops->writepage)
1047        return 0;
1048
1049    return write_cache_pages(mapping, wbc, __writepage, mapping);
1050}
1051
1052EXPORT_SYMBOL(generic_writepages);
1053
1054int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1055{
1056    int ret;
1057
1058    if (wbc->nr_to_write <= 0)
1059        return 0;
1060    if (mapping->a_ops->writepages)
1061        ret = mapping->a_ops->writepages(mapping, wbc);
1062    else
1063        ret = generic_writepages(mapping, wbc);
1064    return ret;
1065}
1066
1067/**
1068 * write_one_page - write out a single page and optionally wait on I/O
1069 * @page: the page to write
1070 * @wait: if true, wait on writeout
1071 *
1072 * The page must be locked by the caller and will be unlocked upon return.
1073 *
1074 * write_one_page() returns a negative error code if I/O failed.
1075 */
1076int write_one_page(struct page *page, int wait)
1077{
1078    struct address_space *mapping = page->mapping;
1079    int ret = 0;
1080    struct writeback_control wbc = {
1081        .sync_mode = WB_SYNC_ALL,
1082        .nr_to_write = 1,
1083    };
1084
1085    BUG_ON(!PageLocked(page));
1086
1087    if (wait)
1088        wait_on_page_writeback(page);
1089
1090    if (clear_page_dirty_for_io(page)) {
1091        page_cache_get(page);
1092        ret = mapping->a_ops->writepage(page, &wbc);
1093        if (ret == 0 && wait) {
1094            wait_on_page_writeback(page);
1095            if (PageError(page))
1096                ret = -EIO;
1097        }
1098        page_cache_release(page);
1099    } else {
1100        unlock_page(page);
1101    }
1102    return ret;
1103}
1104EXPORT_SYMBOL(write_one_page);
1105
1106/*
1107 * For address_spaces which do not use buffers nor write back.
1108 */
1109int __set_page_dirty_no_writeback(struct page *page)
1110{
1111    if (!PageDirty(page))
1112        SetPageDirty(page);
1113    return 0;
1114}
1115
1116/*
1117 * Helper function for set_page_dirty family.
1118 * NOTE: This relies on being atomic wrt interrupts.
1119 */
1120void account_page_dirtied(struct page *page, struct address_space *mapping)
1121{
1122    if (mapping_cap_account_dirty(mapping)) {
1123        __inc_zone_page_state(page, NR_FILE_DIRTY);
1124        __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1125        task_dirty_inc(current);
1126        task_io_account_write(PAGE_CACHE_SIZE);
1127    }
1128}
1129EXPORT_SYMBOL(account_page_dirtied);
1130
1131/*
1132 * For address_spaces which do not use buffers. Just tag the page as dirty in
1133 * its radix tree.
1134 *
1135 * This is also used when a single buffer is being dirtied: we want to set the
1136 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1137 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1138 *
1139 * Most callers have locked the page, which pins the address_space in memory.
1140 * But zap_pte_range() does not lock the page, however in that case the
1141 * mapping is pinned by the vma's ->vm_file reference.
1142 *
1143 * We take care to handle the case where the page was truncated from the
1144 * mapping by re-checking page_mapping() inside tree_lock.
1145 */
1146int __set_page_dirty_nobuffers(struct page *page)
1147{
1148    if (!TestSetPageDirty(page)) {
1149        struct address_space *mapping = page_mapping(page);
1150        struct address_space *mapping2;
1151
1152        if (!mapping)
1153            return 1;
1154
1155        spin_lock_irq(&mapping->tree_lock);
1156        mapping2 = page_mapping(page);
1157        if (mapping2) { /* Race with truncate? */
1158            BUG_ON(mapping2 != mapping);
1159            WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1160            account_page_dirtied(page, mapping);
1161            radix_tree_tag_set(&mapping->page_tree,
1162                page_index(page), PAGECACHE_TAG_DIRTY);
1163        }
1164        spin_unlock_irq(&mapping->tree_lock);
1165        if (mapping->host) {
1166            /* !PageAnon && !swapper_space */
1167            __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1168        }
1169        return 1;
1170    }
1171    return 0;
1172}
1173EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1174
1175/*
1176 * When a writepage implementation decides that it doesn't want to write this
1177 * page for some reason, it should redirty the locked page via
1178 * redirty_page_for_writepage() and it should then unlock the page and return 0
1179 */
1180int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1181{
1182    wbc->pages_skipped++;
1183    return __set_page_dirty_nobuffers(page);
1184}
1185EXPORT_SYMBOL(redirty_page_for_writepage);
1186
1187/*
1188 * Dirty a page.
1189 *
1190 * For pages with a mapping this should be done under the page lock
1191 * for the benefit of asynchronous memory errors who prefer a consistent
1192 * dirty state. This rule can be broken in some special cases,
1193 * but should be better not to.
1194 *
1195 * If the mapping doesn't provide a set_page_dirty a_op, then
1196 * just fall through and assume that it wants buffer_heads.
1197 */
1198int set_page_dirty(struct page *page)
1199{
1200    struct address_space *mapping = page_mapping(page);
1201
1202    if (likely(mapping)) {
1203        int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1204#ifdef CONFIG_BLOCK
1205        if (!spd)
1206            spd = __set_page_dirty_buffers;
1207#endif
1208        return (*spd)(page);
1209    }
1210    if (!PageDirty(page)) {
1211        if (!TestSetPageDirty(page))
1212            return 1;
1213    }
1214    return 0;
1215}
1216EXPORT_SYMBOL(set_page_dirty);
1217
1218/*
1219 * set_page_dirty() is racy if the caller has no reference against
1220 * page->mapping->host, and if the page is unlocked. This is because another
1221 * CPU could truncate the page off the mapping and then free the mapping.
1222 *
1223 * Usually, the page _is_ locked, or the caller is a user-space process which
1224 * holds a reference on the inode by having an open file.
1225 *
1226 * In other cases, the page should be locked before running set_page_dirty().
1227 */
1228int set_page_dirty_lock(struct page *page)
1229{
1230    int ret;
1231
1232    lock_page_nosync(page);
1233    ret = set_page_dirty(page);
1234    unlock_page(page);
1235    return ret;
1236}
1237EXPORT_SYMBOL(set_page_dirty_lock);
1238
1239/*
1240 * Clear a page's dirty flag, while caring for dirty memory accounting.
1241 * Returns true if the page was previously dirty.
1242 *
1243 * This is for preparing to put the page under writeout. We leave the page
1244 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1245 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1246 * implementation will run either set_page_writeback() or set_page_dirty(),
1247 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1248 * back into sync.
1249 *
1250 * This incoherency between the page's dirty flag and radix-tree tag is
1251 * unfortunate, but it only exists while the page is locked.
1252 */
1253int clear_page_dirty_for_io(struct page *page)
1254{
1255    struct address_space *mapping = page_mapping(page);
1256
1257    BUG_ON(!PageLocked(page));
1258
1259    ClearPageReclaim(page);
1260    if (mapping && mapping_cap_account_dirty(mapping)) {
1261        /*
1262         * Yes, Virginia, this is indeed insane.
1263         *
1264         * We use this sequence to make sure that
1265         * (a) we account for dirty stats properly
1266         * (b) we tell the low-level filesystem to
1267         * mark the whole page dirty if it was
1268         * dirty in a pagetable. Only to then
1269         * (c) clean the page again and return 1 to
1270         * cause the writeback.
1271         *
1272         * This way we avoid all nasty races with the
1273         * dirty bit in multiple places and clearing
1274         * them concurrently from different threads.
1275         *
1276         * Note! Normally the "set_page_dirty(page)"
1277         * has no effect on the actual dirty bit - since
1278         * that will already usually be set. But we
1279         * need the side effects, and it can help us
1280         * avoid races.
1281         *
1282         * We basically use the page "master dirty bit"
1283         * as a serialization point for all the different
1284         * threads doing their things.
1285         */
1286        if (page_mkclean(page))
1287            set_page_dirty(page);
1288        /*
1289         * We carefully synchronise fault handlers against
1290         * installing a dirty pte and marking the page dirty
1291         * at this point. We do this by having them hold the
1292         * page lock at some point after installing their
1293         * pte, but before marking the page dirty.
1294         * Pages are always locked coming in here, so we get
1295         * the desired exclusion. See mm/memory.c:do_wp_page()
1296         * for more comments.
1297         */
1298        if (TestClearPageDirty(page)) {
1299            dec_zone_page_state(page, NR_FILE_DIRTY);
1300            dec_bdi_stat(mapping->backing_dev_info,
1301                    BDI_RECLAIMABLE);
1302            return 1;
1303        }
1304        return 0;
1305    }
1306    return TestClearPageDirty(page);
1307}
1308EXPORT_SYMBOL(clear_page_dirty_for_io);
1309
1310int test_clear_page_writeback(struct page *page)
1311{
1312    struct address_space *mapping = page_mapping(page);
1313    int ret;
1314
1315    if (mapping) {
1316        struct backing_dev_info *bdi = mapping->backing_dev_info;
1317        unsigned long flags;
1318
1319        spin_lock_irqsave(&mapping->tree_lock, flags);
1320        ret = TestClearPageWriteback(page);
1321        if (ret) {
1322            radix_tree_tag_clear(&mapping->page_tree,
1323                        page_index(page),
1324                        PAGECACHE_TAG_WRITEBACK);
1325            if (bdi_cap_account_writeback(bdi)) {
1326                __dec_bdi_stat(bdi, BDI_WRITEBACK);
1327                __bdi_writeout_inc(bdi);
1328            }
1329        }
1330        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1331    } else {
1332        ret = TestClearPageWriteback(page);
1333    }
1334    if (ret)
1335        dec_zone_page_state(page, NR_WRITEBACK);
1336    return ret;
1337}
1338
1339int test_set_page_writeback(struct page *page)
1340{
1341    struct address_space *mapping = page_mapping(page);
1342    int ret;
1343
1344    if (mapping) {
1345        struct backing_dev_info *bdi = mapping->backing_dev_info;
1346        unsigned long flags;
1347
1348        spin_lock_irqsave(&mapping->tree_lock, flags);
1349        ret = TestSetPageWriteback(page);
1350        if (!ret) {
1351            radix_tree_tag_set(&mapping->page_tree,
1352                        page_index(page),
1353                        PAGECACHE_TAG_WRITEBACK);
1354            if (bdi_cap_account_writeback(bdi))
1355                __inc_bdi_stat(bdi, BDI_WRITEBACK);
1356        }
1357        if (!PageDirty(page))
1358            radix_tree_tag_clear(&mapping->page_tree,
1359                        page_index(page),
1360                        PAGECACHE_TAG_DIRTY);
1361        radix_tree_tag_clear(&mapping->page_tree,
1362                     page_index(page),
1363                     PAGECACHE_TAG_TOWRITE);
1364        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1365    } else {
1366        ret = TestSetPageWriteback(page);
1367    }
1368    if (!ret)
1369        inc_zone_page_state(page, NR_WRITEBACK);
1370    return ret;
1371
1372}
1373EXPORT_SYMBOL(test_set_page_writeback);
1374
1375/*
1376 * Return true if any of the pages in the mapping are marked with the
1377 * passed tag.
1378 */
1379int mapping_tagged(struct address_space *mapping, int tag)
1380{
1381    int ret;
1382    rcu_read_lock();
1383    ret = radix_tree_tagged(&mapping->page_tree, tag);
1384    rcu_read_unlock();
1385    return ret;
1386}
1387EXPORT_SYMBOL(mapping_tagged);
1388

Archive Download this file



interactive