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 * real-time tasks.
408 */
409void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
410{
411    unsigned long background;
412    unsigned long dirty;
413    unsigned long uninitialized_var(available_memory);
414    struct task_struct *tsk;
415
416    if (!vm_dirty_bytes || !dirty_background_bytes)
417        available_memory = determine_dirtyable_memory();
418
419    if (vm_dirty_bytes)
420        dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
421    else
422        dirty = (vm_dirty_ratio * available_memory) / 100;
423
424    if (dirty_background_bytes)
425        background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
426    else
427        background = (dirty_background_ratio * available_memory) / 100;
428
429    if (background >= dirty)
430        background = dirty / 2;
431    tsk = current;
432    if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
433        background += background / 4;
434        dirty += dirty / 4;
435    }
436    *pbackground = background;
437    *pdirty = dirty;
438}
439
440/*
441 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
442 *
443 * Allocate high/low dirty limits to fast/slow devices, in order to prevent
444 * - starving fast devices
445 * - piling up dirty pages (that will take long time to sync) on slow devices
446 *
447 * The bdi's share of dirty limit will be adapting to its throughput and
448 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
449 */
450unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
451{
452    u64 bdi_dirty;
453    long numerator, denominator;
454
455    /*
456     * Calculate this BDI's share of the dirty ratio.
457     */
458    bdi_writeout_fraction(bdi, &numerator, &denominator);
459
460    bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
461    bdi_dirty *= numerator;
462    do_div(bdi_dirty, denominator);
463
464    bdi_dirty += (dirty * bdi->min_ratio) / 100;
465    if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
466        bdi_dirty = dirty * bdi->max_ratio / 100;
467
468    return bdi_dirty;
469}
470
471/*
472 * balance_dirty_pages() must be called by processes which are generating dirty
473 * data. It looks at the number of dirty pages in the machine and will force
474 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
475 * If we're over `background_thresh' then the writeback threads are woken to
476 * perform some writeout.
477 */
478static void balance_dirty_pages(struct address_space *mapping,
479                unsigned long write_chunk)
480{
481    long nr_reclaimable, bdi_nr_reclaimable;
482    long nr_writeback, bdi_nr_writeback;
483    unsigned long background_thresh;
484    unsigned long dirty_thresh;
485    unsigned long bdi_thresh;
486    unsigned long pages_written = 0;
487    unsigned long pause = 1;
488    bool dirty_exceeded = false;
489    struct backing_dev_info *bdi = mapping->backing_dev_info;
490
491    for (;;) {
492        struct writeback_control wbc = {
493            .sync_mode = WB_SYNC_NONE,
494            .older_than_this = NULL,
495            .nr_to_write = write_chunk,
496            .range_cyclic = 1,
497        };
498
499        nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
500                    global_page_state(NR_UNSTABLE_NFS);
501        nr_writeback = global_page_state(NR_WRITEBACK);
502
503        global_dirty_limits(&background_thresh, &dirty_thresh);
504
505        /*
506         * Throttle it only when the background writeback cannot
507         * catch-up. This avoids (excessively) small writeouts
508         * when the bdi limits are ramping up.
509         */
510        if (nr_reclaimable + nr_writeback <=
511                (background_thresh + dirty_thresh) / 2)
512            break;
513
514        bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
515        bdi_thresh = task_dirty_limit(current, bdi_thresh);
516
517        /*
518         * In order to avoid the stacked BDI deadlock we need
519         * to ensure we accurately count the 'dirty' pages when
520         * the threshold is low.
521         *
522         * Otherwise it would be possible to get thresh+n pages
523         * reported dirty, even though there are thresh-m pages
524         * actually dirty; with m+n sitting in the percpu
525         * deltas.
526         */
527        if (bdi_thresh < 2*bdi_stat_error(bdi)) {
528            bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
529            bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
530        } else {
531            bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
532            bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
533        }
534
535        /*
536         * The bdi thresh is somehow "soft" limit derived from the
537         * global "hard" limit. The former helps to prevent heavy IO
538         * bdi or process from holding back light ones; The latter is
539         * the last resort safeguard.
540         */
541        dirty_exceeded =
542            (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh)
543            || (nr_reclaimable + nr_writeback > dirty_thresh);
544
545        if (!dirty_exceeded)
546            break;
547
548        if (!bdi->dirty_exceeded)
549            bdi->dirty_exceeded = 1;
550
551        /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
552         * Unstable writes are a feature of certain networked
553         * filesystems (i.e. NFS) in which data may have been
554         * written to the server's write cache, but has not yet
555         * been flushed to permanent storage.
556         * Only move pages to writeback if this bdi is over its
557         * threshold otherwise wait until the disk writes catch
558         * up.
559         */
560        trace_wbc_balance_dirty_start(&wbc, bdi);
561        if (bdi_nr_reclaimable > bdi_thresh) {
562            writeback_inodes_wb(&bdi->wb, &wbc);
563            pages_written += write_chunk - wbc.nr_to_write;
564            trace_wbc_balance_dirty_written(&wbc, bdi);
565            if (pages_written >= write_chunk)
566                break; /* We've done our duty */
567        }
568        trace_wbc_balance_dirty_wait(&wbc, bdi);
569        __set_current_state(TASK_UNINTERRUPTIBLE);
570        io_schedule_timeout(pause);
571
572        /*
573         * Increase the delay for each loop, up to our previous
574         * default of taking a 100ms nap.
575         */
576        pause <<= 1;
577        if (pause > HZ / 10)
578            pause = HZ / 10;
579    }
580
581    if (!dirty_exceeded && bdi->dirty_exceeded)
582        bdi->dirty_exceeded = 0;
583
584    if (writeback_in_progress(bdi))
585        return;
586
587    /*
588     * In laptop mode, we wait until hitting the higher threshold before
589     * starting background writeout, and then write out all the way down
590     * to the lower threshold. So slow writers cause minimal disk activity.
591     *
592     * In normal mode, we start background writeout at the lower
593     * background_thresh, to keep the amount of dirty memory low.
594     */
595    if ((laptop_mode && pages_written) ||
596        (!laptop_mode && (nr_reclaimable > background_thresh)))
597        bdi_start_background_writeback(bdi);
598}
599
600void set_page_dirty_balance(struct page *page, int page_mkwrite)
601{
602    if (set_page_dirty(page) || page_mkwrite) {
603        struct address_space *mapping = page_mapping(page);
604
605        if (mapping)
606            balance_dirty_pages_ratelimited(mapping);
607    }
608}
609
610static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
611
612/**
613 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
614 * @mapping: address_space which was dirtied
615 * @nr_pages_dirtied: number of pages which the caller has just dirtied
616 *
617 * Processes which are dirtying memory should call in here once for each page
618 * which was newly dirtied. The function will periodically check the system's
619 * dirty state and will initiate writeback if needed.
620 *
621 * On really big machines, get_writeback_state is expensive, so try to avoid
622 * calling it too often (ratelimiting). But once we're over the dirty memory
623 * limit we decrease the ratelimiting by a lot, to prevent individual processes
624 * from overshooting the limit by (ratelimit_pages) each.
625 */
626void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
627                    unsigned long nr_pages_dirtied)
628{
629    unsigned long ratelimit;
630    unsigned long *p;
631
632    ratelimit = ratelimit_pages;
633    if (mapping->backing_dev_info->dirty_exceeded)
634        ratelimit = 8;
635
636    /*
637     * Check the rate limiting. Also, we do not want to throttle real-time
638     * tasks in balance_dirty_pages(). Period.
639     */
640    preempt_disable();
641    p = &__get_cpu_var(bdp_ratelimits);
642    *p += nr_pages_dirtied;
643    if (unlikely(*p >= ratelimit)) {
644        ratelimit = sync_writeback_pages(*p);
645        *p = 0;
646        preempt_enable();
647        balance_dirty_pages(mapping, ratelimit);
648        return;
649    }
650    preempt_enable();
651}
652EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
653
654void throttle_vm_writeout(gfp_t gfp_mask)
655{
656    unsigned long background_thresh;
657    unsigned long dirty_thresh;
658
659        for ( ; ; ) {
660        global_dirty_limits(&background_thresh, &dirty_thresh);
661
662                /*
663                 * Boost the allowable dirty threshold a bit for page
664                 * allocators so they don't get DoS'ed by heavy writers
665                 */
666                dirty_thresh += dirty_thresh / 10; /* wheeee... */
667
668                if (global_page_state(NR_UNSTABLE_NFS) +
669            global_page_state(NR_WRITEBACK) <= dirty_thresh)
670                            break;
671                congestion_wait(BLK_RW_ASYNC, HZ/10);
672
673        /*
674         * The caller might hold locks which can prevent IO completion
675         * or progress in the filesystem. So we cannot just sit here
676         * waiting for IO to complete.
677         */
678        if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
679            break;
680        }
681}
682
683/*
684 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
685 */
686int dirty_writeback_centisecs_handler(ctl_table *table, int write,
687    void __user *buffer, size_t *length, loff_t *ppos)
688{
689    proc_dointvec(table, write, buffer, length, ppos);
690    bdi_arm_supers_timer();
691    return 0;
692}
693
694#ifdef CONFIG_BLOCK
695void laptop_mode_timer_fn(unsigned long data)
696{
697    struct request_queue *q = (struct request_queue *)data;
698    int nr_pages = global_page_state(NR_FILE_DIRTY) +
699        global_page_state(NR_UNSTABLE_NFS);
700
701    /*
702     * We want to write everything out, not just down to the dirty
703     * threshold
704     */
705    if (bdi_has_dirty_io(&q->backing_dev_info))
706        bdi_start_writeback(&q->backing_dev_info, nr_pages);
707}
708
709/*
710 * We've spun up the disk and we're in laptop mode: schedule writeback
711 * of all dirty data a few seconds from now. If the flush is already scheduled
712 * then push it back - the user is still using the disk.
713 */
714void laptop_io_completion(struct backing_dev_info *info)
715{
716    mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
717}
718
719/*
720 * We're in laptop mode and we've just synced. The sync's writes will have
721 * caused another writeback to be scheduled by laptop_io_completion.
722 * Nothing needs to be written back anymore, so we unschedule the writeback.
723 */
724void laptop_sync_completion(void)
725{
726    struct backing_dev_info *bdi;
727
728    rcu_read_lock();
729
730    list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
731        del_timer(&bdi->laptop_mode_wb_timer);
732
733    rcu_read_unlock();
734}
735#endif
736
737/*
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.
742 *
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.
746 *
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.
752 */
753
754void writeback_set_ratelimit(void)
755{
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;
761}
762
763static int __cpuinit
764ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
765{
766    writeback_set_ratelimit();
767    return NOTIFY_DONE;
768}
769
770static struct notifier_block __cpuinitdata ratelimit_nb = {
771    .notifier_call = ratelimit_handler,
772    .next = NULL,
773};
774
775/*
776 * Called early on to tune the page writeback dirty limits.
777 *
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.
781 *
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.
789 *
790 * But we might still want to scale the dirty_ratio by how
791 * much memory the box has..
792 */
793void __init page_writeback_init(void)
794{
795    int shift;
796
797    writeback_set_ratelimit();
798    register_cpu_notifier(&ratelimit_nb);
799
800    shift = calc_period_shift();
801    prop_descriptor_init(&vm_completions, shift);
802    prop_descriptor_init(&vm_dirties, shift);
803}
804
805/**
806 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
807 * @mapping: address space structure to write
808 * @start: starting page index
809 * @end: ending page index (inclusive)
810 *
811 * This function scans the page range from @start to @end (inclusive) and tags
812 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
813 * that write_cache_pages (or whoever calls this function) will then use
814 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
815 * used to avoid livelocking of writeback by a process steadily creating new
816 * dirty pages in the file (thus it is important for this function to be quick
817 * so that it can tag pages faster than a dirtying process can create them).
818 */
819/*
820 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
821 */
822void tag_pages_for_writeback(struct address_space *mapping,
823                 pgoff_t start, pgoff_t end)
824{
825#define WRITEBACK_TAG_BATCH 4096
826    unsigned long tagged;
827
828    do {
829        spin_lock_irq(&mapping->tree_lock);
830        tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
831                &start, end, WRITEBACK_TAG_BATCH,
832                PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
833        spin_unlock_irq(&mapping->tree_lock);
834        WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
835        cond_resched();
836        /* We check 'start' to handle wrapping when end == ~0UL */
837    } while (tagged >= WRITEBACK_TAG_BATCH && start);
838}
839EXPORT_SYMBOL(tag_pages_for_writeback);
840
841/**
842 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
843 * @mapping: address space structure to write
844 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
845 * @writepage: function called for each page
846 * @data: data passed to writepage function
847 *
848 * If a page is already under I/O, write_cache_pages() skips it, even
849 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
850 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
851 * and msync() need to guarantee that all the data which was dirty at the time
852 * the call was made get new I/O started against them. If wbc->sync_mode is
853 * WB_SYNC_ALL then we were called for data integrity and we must wait for
854 * existing IO to complete.
855 *
856 * To avoid livelocks (when other process dirties new pages), we first tag
857 * pages which should be written back with TOWRITE tag and only then start
858 * writing them. For data-integrity sync we have to be careful so that we do
859 * not miss some pages (e.g., because some other process has cleared TOWRITE
860 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
861 * by the process clearing the DIRTY tag (and submitting the page for IO).
862 */
863int write_cache_pages(struct address_space *mapping,
864              struct writeback_control *wbc, writepage_t writepage,
865              void *data)
866{
867    int ret = 0;
868    int done = 0;
869    struct pagevec pvec;
870    int nr_pages;
871    pgoff_t uninitialized_var(writeback_index);
872    pgoff_t index;
873    pgoff_t end; /* Inclusive */
874    pgoff_t done_index;
875    int cycled;
876    int range_whole = 0;
877    int tag;
878
879    pagevec_init(&pvec, 0);
880    if (wbc->range_cyclic) {
881        writeback_index = mapping->writeback_index; /* prev offset */
882        index = writeback_index;
883        if (index == 0)
884            cycled = 1;
885        else
886            cycled = 0;
887        end = -1;
888    } else {
889        index = wbc->range_start >> PAGE_CACHE_SHIFT;
890        end = wbc->range_end >> PAGE_CACHE_SHIFT;
891        if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
892            range_whole = 1;
893        cycled = 1; /* ignore range_cyclic tests */
894    }
895    if (wbc->sync_mode == WB_SYNC_ALL)
896        tag = PAGECACHE_TAG_TOWRITE;
897    else
898        tag = PAGECACHE_TAG_DIRTY;
899retry:
900    if (wbc->sync_mode == WB_SYNC_ALL)
901        tag_pages_for_writeback(mapping, index, end);
902    done_index = index;
903    while (!done && (index <= end)) {
904        int i;
905
906        nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
907                  min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
908        if (nr_pages == 0)
909            break;
910
911        for (i = 0; i < nr_pages; i++) {
912            struct page *page = pvec.pages[i];
913
914            /*
915             * At this point, the page may be truncated or
916             * invalidated (changing page->mapping to NULL), or
917             * even swizzled back from swapper_space to tmpfs file
918             * mapping. However, page->index will not change
919             * because we have a reference on the page.
920             */
921            if (page->index > end) {
922                /*
923                 * can't be range_cyclic (1st pass) because
924                 * end == -1 in that case.
925                 */
926                done = 1;
927                break;
928            }
929
930            done_index = page->index;
931
932            lock_page(page);
933
934            /*
935             * Page truncated or invalidated. We can freely skip it
936             * then, even for data integrity operations: the page
937             * has disappeared concurrently, so there could be no
938             * real expectation of this data interity operation
939             * even if there is now a new, dirty page at the same
940             * pagecache address.
941             */
942            if (unlikely(page->mapping != mapping)) {
943continue_unlock:
944                unlock_page(page);
945                continue;
946            }
947
948            if (!PageDirty(page)) {
949                /* someone wrote it for us */
950                goto continue_unlock;
951            }
952
953            if (PageWriteback(page)) {
954                if (wbc->sync_mode != WB_SYNC_NONE)
955                    wait_on_page_writeback(page);
956                else
957                    goto continue_unlock;
958            }
959
960            BUG_ON(PageWriteback(page));
961            if (!clear_page_dirty_for_io(page))
962                goto continue_unlock;
963
964            trace_wbc_writepage(wbc, mapping->backing_dev_info);
965            ret = (*writepage)(page, wbc, data);
966            if (unlikely(ret)) {
967                if (ret == AOP_WRITEPAGE_ACTIVATE) {
968                    unlock_page(page);
969                    ret = 0;
970                } else {
971                    /*
972                     * done_index is set past this page,
973                     * so media errors will not choke
974                     * background writeout for the entire
975                     * file. This has consequences for
976                     * range_cyclic semantics (ie. it may
977                     * not be suitable for data integrity
978                     * writeout).
979                     */
980                    done_index = page->index + 1;
981                    done = 1;
982                    break;
983                }
984            }
985
986            /*
987             * We stop writing back only if we are not doing
988             * integrity sync. In case of integrity sync we have to
989             * keep going until we have written all the pages
990             * we tagged for writeback prior to entering this loop.
991             */
992            if (--wbc->nr_to_write <= 0 &&
993                wbc->sync_mode == WB_SYNC_NONE) {
994                done = 1;
995                break;
996            }
997        }
998        pagevec_release(&pvec);
999        cond_resched();
1000    }
1001    if (!cycled && !done) {
1002        /*
1003         * range_cyclic:
1004         * We hit the last page and there is more work to be done: wrap
1005         * back to the start of the file
1006         */
1007        cycled = 1;
1008        index = 0;
1009        end = writeback_index - 1;
1010        goto retry;
1011    }
1012    if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1013        mapping->writeback_index = done_index;
1014
1015    return ret;
1016}
1017EXPORT_SYMBOL(write_cache_pages);
1018
1019/*
1020 * Function used by generic_writepages to call the real writepage
1021 * function and set the mapping flags on error
1022 */
1023static int __writepage(struct page *page, struct writeback_control *wbc,
1024               void *data)
1025{
1026    struct address_space *mapping = data;
1027    int ret = mapping->a_ops->writepage(page, wbc);
1028    mapping_set_error(mapping, ret);
1029    return ret;
1030}
1031
1032/**
1033 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1034 * @mapping: address space structure to write
1035 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1036 *
1037 * This is a library function, which implements the writepages()
1038 * address_space_operation.
1039 */
1040int generic_writepages(struct address_space *mapping,
1041               struct writeback_control *wbc)
1042{
1043    struct blk_plug plug;
1044    int ret;
1045
1046    /* deal with chardevs and other special file */
1047    if (!mapping->a_ops->writepage)
1048        return 0;
1049
1050    blk_start_plug(&plug);
1051    ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1052    blk_finish_plug(&plug);
1053    return ret;
1054}
1055
1056EXPORT_SYMBOL(generic_writepages);
1057
1058int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1059{
1060    int ret;
1061
1062    if (wbc->nr_to_write <= 0)
1063        return 0;
1064    if (mapping->a_ops->writepages)
1065        ret = mapping->a_ops->writepages(mapping, wbc);
1066    else
1067        ret = generic_writepages(mapping, wbc);
1068    return ret;
1069}
1070
1071/**
1072 * write_one_page - write out a single page and optionally wait on I/O
1073 * @page: the page to write
1074 * @wait: if true, wait on writeout
1075 *
1076 * The page must be locked by the caller and will be unlocked upon return.
1077 *
1078 * write_one_page() returns a negative error code if I/O failed.
1079 */
1080int write_one_page(struct page *page, int wait)
1081{
1082    struct address_space *mapping = page->mapping;
1083    int ret = 0;
1084    struct writeback_control wbc = {
1085        .sync_mode = WB_SYNC_ALL,
1086        .nr_to_write = 1,
1087    };
1088
1089    BUG_ON(!PageLocked(page));
1090
1091    if (wait)
1092        wait_on_page_writeback(page);
1093
1094    if (clear_page_dirty_for_io(page)) {
1095        page_cache_get(page);
1096        ret = mapping->a_ops->writepage(page, &wbc);
1097        if (ret == 0 && wait) {
1098            wait_on_page_writeback(page);
1099            if (PageError(page))
1100                ret = -EIO;
1101        }
1102        page_cache_release(page);
1103    } else {
1104        unlock_page(page);
1105    }
1106    return ret;
1107}
1108EXPORT_SYMBOL(write_one_page);
1109
1110/*
1111 * For address_spaces which do not use buffers nor write back.
1112 */
1113int __set_page_dirty_no_writeback(struct page *page)
1114{
1115    if (!PageDirty(page))
1116        return !TestSetPageDirty(page);
1117    return 0;
1118}
1119
1120/*
1121 * Helper function for set_page_dirty family.
1122 * NOTE: This relies on being atomic wrt interrupts.
1123 */
1124void account_page_dirtied(struct page *page, struct address_space *mapping)
1125{
1126    if (mapping_cap_account_dirty(mapping)) {
1127        __inc_zone_page_state(page, NR_FILE_DIRTY);
1128        __inc_zone_page_state(page, NR_DIRTIED);
1129        __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1130        task_dirty_inc(current);
1131        task_io_account_write(PAGE_CACHE_SIZE);
1132    }
1133}
1134EXPORT_SYMBOL(account_page_dirtied);
1135
1136/*
1137 * Helper function for set_page_writeback family.
1138 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1139 * wrt interrupts.
1140 */
1141void account_page_writeback(struct page *page)
1142{
1143    inc_zone_page_state(page, NR_WRITEBACK);
1144    inc_zone_page_state(page, NR_WRITTEN);
1145}
1146EXPORT_SYMBOL(account_page_writeback);
1147
1148/*
1149 * For address_spaces which do not use buffers. Just tag the page as dirty in
1150 * its radix tree.
1151 *
1152 * This is also used when a single buffer is being dirtied: we want to set the
1153 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1154 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1155 *
1156 * Most callers have locked the page, which pins the address_space in memory.
1157 * But zap_pte_range() does not lock the page, however in that case the
1158 * mapping is pinned by the vma's ->vm_file reference.
1159 *
1160 * We take care to handle the case where the page was truncated from the
1161 * mapping by re-checking page_mapping() inside tree_lock.
1162 */
1163int __set_page_dirty_nobuffers(struct page *page)
1164{
1165    if (!TestSetPageDirty(page)) {
1166        struct address_space *mapping = page_mapping(page);
1167        struct address_space *mapping2;
1168
1169        if (!mapping)
1170            return 1;
1171
1172        spin_lock_irq(&mapping->tree_lock);
1173        mapping2 = page_mapping(page);
1174        if (mapping2) { /* Race with truncate? */
1175            BUG_ON(mapping2 != mapping);
1176            WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1177            account_page_dirtied(page, mapping);
1178            radix_tree_tag_set(&mapping->page_tree,
1179                page_index(page), PAGECACHE_TAG_DIRTY);
1180        }
1181        spin_unlock_irq(&mapping->tree_lock);
1182        if (mapping->host) {
1183            /* !PageAnon && !swapper_space */
1184            __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1185        }
1186        return 1;
1187    }
1188    return 0;
1189}
1190EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1191
1192/*
1193 * When a writepage implementation decides that it doesn't want to write this
1194 * page for some reason, it should redirty the locked page via
1195 * redirty_page_for_writepage() and it should then unlock the page and return 0
1196 */
1197int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1198{
1199    wbc->pages_skipped++;
1200    return __set_page_dirty_nobuffers(page);
1201}
1202EXPORT_SYMBOL(redirty_page_for_writepage);
1203
1204/*
1205 * Dirty a page.
1206 *
1207 * For pages with a mapping this should be done under the page lock
1208 * for the benefit of asynchronous memory errors who prefer a consistent
1209 * dirty state. This rule can be broken in some special cases,
1210 * but should be better not to.
1211 *
1212 * If the mapping doesn't provide a set_page_dirty a_op, then
1213 * just fall through and assume that it wants buffer_heads.
1214 */
1215int set_page_dirty(struct page *page)
1216{
1217    struct address_space *mapping = page_mapping(page);
1218
1219    if (likely(mapping)) {
1220        int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1221        /*
1222         * readahead/lru_deactivate_page could remain
1223         * PG_readahead/PG_reclaim due to race with end_page_writeback
1224         * About readahead, if the page is written, the flags would be
1225         * reset. So no problem.
1226         * About lru_deactivate_page, if the page is redirty, the flag
1227         * will be reset. So no problem. but if the page is used by readahead
1228         * it will confuse readahead and make it restart the size rampup
1229         * process. But it's a trivial problem.
1230         */
1231        ClearPageReclaim(page);
1232#ifdef CONFIG_BLOCK
1233        if (!spd)
1234            spd = __set_page_dirty_buffers;
1235#endif
1236        return (*spd)(page);
1237    }
1238    if (!PageDirty(page)) {
1239        if (!TestSetPageDirty(page))
1240            return 1;
1241    }
1242    return 0;
1243}
1244EXPORT_SYMBOL(set_page_dirty);
1245
1246/*
1247 * set_page_dirty() is racy if the caller has no reference against
1248 * page->mapping->host, and if the page is unlocked. This is because another
1249 * CPU could truncate the page off the mapping and then free the mapping.
1250 *
1251 * Usually, the page _is_ locked, or the caller is a user-space process which
1252 * holds a reference on the inode by having an open file.
1253 *
1254 * In other cases, the page should be locked before running set_page_dirty().
1255 */
1256int set_page_dirty_lock(struct page *page)
1257{
1258    int ret;
1259
1260    lock_page(page);
1261    ret = set_page_dirty(page);
1262    unlock_page(page);
1263    return ret;
1264}
1265EXPORT_SYMBOL(set_page_dirty_lock);
1266
1267/*
1268 * Clear a page's dirty flag, while caring for dirty memory accounting.
1269 * Returns true if the page was previously dirty.
1270 *
1271 * This is for preparing to put the page under writeout. We leave the page
1272 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1273 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1274 * implementation will run either set_page_writeback() or set_page_dirty(),
1275 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1276 * back into sync.
1277 *
1278 * This incoherency between the page's dirty flag and radix-tree tag is
1279 * unfortunate, but it only exists while the page is locked.
1280 */
1281int clear_page_dirty_for_io(struct page *page)
1282{
1283    struct address_space *mapping = page_mapping(page);
1284
1285    BUG_ON(!PageLocked(page));
1286
1287    if (mapping && mapping_cap_account_dirty(mapping)) {
1288        /*
1289         * Yes, Virginia, this is indeed insane.
1290         *
1291         * We use this sequence to make sure that
1292         * (a) we account for dirty stats properly
1293         * (b) we tell the low-level filesystem to
1294         * mark the whole page dirty if it was
1295         * dirty in a pagetable. Only to then
1296         * (c) clean the page again and return 1 to
1297         * cause the writeback.
1298         *
1299         * This way we avoid all nasty races with the
1300         * dirty bit in multiple places and clearing
1301         * them concurrently from different threads.
1302         *
1303         * Note! Normally the "set_page_dirty(page)"
1304         * has no effect on the actual dirty bit - since
1305         * that will already usually be set. But we
1306         * need the side effects, and it can help us
1307         * avoid races.
1308         *
1309         * We basically use the page "master dirty bit"
1310         * as a serialization point for all the different
1311         * threads doing their things.
1312         */
1313        if (page_mkclean(page))
1314            set_page_dirty(page);
1315        /*
1316         * We carefully synchronise fault handlers against
1317         * installing a dirty pte and marking the page dirty
1318         * at this point. We do this by having them hold the
1319         * page lock at some point after installing their
1320         * pte, but before marking the page dirty.
1321         * Pages are always locked coming in here, so we get
1322         * the desired exclusion. See mm/memory.c:do_wp_page()
1323         * for more comments.
1324         */
1325        if (TestClearPageDirty(page)) {
1326            dec_zone_page_state(page, NR_FILE_DIRTY);
1327            dec_bdi_stat(mapping->backing_dev_info,
1328                    BDI_RECLAIMABLE);
1329            return 1;
1330        }
1331        return 0;
1332    }
1333    return TestClearPageDirty(page);
1334}
1335EXPORT_SYMBOL(clear_page_dirty_for_io);
1336
1337int test_clear_page_writeback(struct page *page)
1338{
1339    struct address_space *mapping = page_mapping(page);
1340    int ret;
1341
1342    if (mapping) {
1343        struct backing_dev_info *bdi = mapping->backing_dev_info;
1344        unsigned long flags;
1345
1346        spin_lock_irqsave(&mapping->tree_lock, flags);
1347        ret = TestClearPageWriteback(page);
1348        if (ret) {
1349            radix_tree_tag_clear(&mapping->page_tree,
1350                        page_index(page),
1351                        PAGECACHE_TAG_WRITEBACK);
1352            if (bdi_cap_account_writeback(bdi)) {
1353                __dec_bdi_stat(bdi, BDI_WRITEBACK);
1354                __bdi_writeout_inc(bdi);
1355            }
1356        }
1357        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1358    } else {
1359        ret = TestClearPageWriteback(page);
1360    }
1361    if (ret)
1362        dec_zone_page_state(page, NR_WRITEBACK);
1363    return ret;
1364}
1365
1366int test_set_page_writeback(struct page *page)
1367{
1368    struct address_space *mapping = page_mapping(page);
1369    int ret;
1370
1371    if (mapping) {
1372        struct backing_dev_info *bdi = mapping->backing_dev_info;
1373        unsigned long flags;
1374
1375        spin_lock_irqsave(&mapping->tree_lock, flags);
1376        ret = TestSetPageWriteback(page);
1377        if (!ret) {
1378            radix_tree_tag_set(&mapping->page_tree,
1379                        page_index(page),
1380                        PAGECACHE_TAG_WRITEBACK);
1381            if (bdi_cap_account_writeback(bdi))
1382                __inc_bdi_stat(bdi, BDI_WRITEBACK);
1383        }
1384        if (!PageDirty(page))
1385            radix_tree_tag_clear(&mapping->page_tree,
1386                        page_index(page),
1387                        PAGECACHE_TAG_DIRTY);
1388        radix_tree_tag_clear(&mapping->page_tree,
1389                     page_index(page),
1390                     PAGECACHE_TAG_TOWRITE);
1391        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1392    } else {
1393        ret = TestSetPageWriteback(page);
1394    }
1395    if (!ret)
1396        account_page_writeback(page);
1397    return ret;
1398
1399}
1400EXPORT_SYMBOL(test_set_page_writeback);
1401
1402/*
1403 * Return true if any of the pages in the mapping are marked with the
1404 * passed tag.
1405 */
1406int mapping_tagged(struct address_space *mapping, int tag)
1407{
1408    int ret;
1409    rcu_read_lock();
1410    ret = radix_tree_tagged(&mapping->page_tree, tag);
1411    rcu_read_unlock();
1412    return ret;
1413}
1414EXPORT_SYMBOL(mapping_tagged);
1415

Archive Download this file



interactive