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/export.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> /* __set_page_dirty_buffers */
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/mm_inline.h>
40#include <trace/events/writeback.h>
41
42#include "internal.h"
43
44/*
45 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
56 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
60#define RATELIMIT_CALC_SHIFT 10
61
62/*
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
68/* The following parameters are exported via /proc/sys/vm */
69
70/*
71 * Start background writeback (via writeback threads) at this percentage
72 */
73int dirty_background_ratio = 10;
74
75/*
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
88 * The generator of dirty data starts writeback at this percentage
89 */
90int vm_dirty_ratio = 20;
91
92/*
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
99 * The interval between `kupdate'-style writebacks
100 */
101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102
103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
105/*
106 * The longest time for which data is allowed to remain dirty
107 */
108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int block_dump;
114
115/*
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
118 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
125unsigned long global_dirty_limit;
126
127/*
128 * Scale the writeback cache size proportional to the relative writeout speeds.
129 *
130 * We do this by keeping a floating proportion between BDIs, based on page
131 * writeback completions [end_page_writeback()]. Those devices that write out
132 * pages fastest will get the larger share, while the slower will get a smaller
133 * share.
134 *
135 * We use page writeout completions because we are interested in getting rid of
136 * dirty pages. Having them written out is the primary goal.
137 *
138 * We introduce a concept of time, a period over which we measure these events,
139 * because demand can/will vary over time. The length of this period itself is
140 * measured in page writeback completions.
141 *
142 */
143static struct fprop_global writeout_completions;
144
145static void writeout_period(unsigned long t);
146/* Timer for aging of writeout_completions */
147static struct timer_list writeout_period_timer =
148        TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
149static unsigned long writeout_period_time = 0;
150
151/*
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
155 */
156#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
157
158/*
159 * In a memory zone, there is a certain amount of pages we consider
160 * available for the page cache, which is essentially the number of
161 * free and reclaimable pages, minus some zone reserves to protect
162 * lowmem and the ability to uphold the zone's watermarks without
163 * requiring writeback.
164 *
165 * This number of dirtyable pages is the base value of which the
166 * user-configurable dirty ratio is the effictive number of pages that
167 * are allowed to be actually dirtied. Per individual zone, or
168 * globally by using the sum of dirtyable pages over all zones.
169 *
170 * Because the user is allowed to specify the dirty limit globally as
171 * absolute number of bytes, calculating the per-zone dirty limit can
172 * require translating the configured limit into a percentage of
173 * global dirtyable memory first.
174 */
175
176/**
177 * zone_dirtyable_memory - number of dirtyable pages in a zone
178 * @zone: the zone
179 *
180 * Returns the zone's number of pages potentially available for dirty
181 * page cache. This is the base value for the per-zone dirty limits.
182 */
183static unsigned long zone_dirtyable_memory(struct zone *zone)
184{
185    unsigned long nr_pages;
186
187    nr_pages = zone_page_state(zone, NR_FREE_PAGES);
188    nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
189
190    nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
191    nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
192
193    return nr_pages;
194}
195
196static unsigned long highmem_dirtyable_memory(unsigned long total)
197{
198#ifdef CONFIG_HIGHMEM
199    int node;
200    unsigned long x = 0;
201
202    for_each_node_state(node, N_HIGH_MEMORY) {
203        struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
204
205        x += zone_dirtyable_memory(z);
206    }
207    /*
208     * Unreclaimable memory (kernel memory or anonymous memory
209     * without swap) can bring down the dirtyable pages below
210     * the zone's dirty balance reserve and the above calculation
211     * will underflow. However we still want to add in nodes
212     * which are below threshold (negative values) to get a more
213     * accurate calculation but make sure that the total never
214     * underflows.
215     */
216    if ((long)x < 0)
217        x = 0;
218
219    /*
220     * Make sure that the number of highmem pages is never larger
221     * than the number of the total dirtyable memory. This can only
222     * occur in very strange VM situations but we want to make sure
223     * that this does not occur.
224     */
225    return min(x, total);
226#else
227    return 0;
228#endif
229}
230
231/**
232 * global_dirtyable_memory - number of globally dirtyable pages
233 *
234 * Returns the global number of pages potentially available for dirty
235 * page cache. This is the base value for the global dirty limits.
236 */
237static unsigned long global_dirtyable_memory(void)
238{
239    unsigned long x;
240
241    x = global_page_state(NR_FREE_PAGES);
242    x -= min(x, dirty_balance_reserve);
243
244    x += global_page_state(NR_INACTIVE_FILE);
245    x += global_page_state(NR_ACTIVE_FILE);
246
247    if (!vm_highmem_is_dirtyable)
248        x -= highmem_dirtyable_memory(x);
249
250    return x + 1; /* Ensure that we never return 0 */
251}
252
253/*
254 * global_dirty_limits - background-writeback and dirty-throttling thresholds
255 *
256 * Calculate the dirty thresholds based on sysctl parameters
257 * - vm.dirty_background_ratio or vm.dirty_background_bytes
258 * - vm.dirty_ratio or vm.dirty_bytes
259 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
260 * real-time tasks.
261 */
262void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
263{
264    unsigned long background;
265    unsigned long dirty;
266    unsigned long uninitialized_var(available_memory);
267    struct task_struct *tsk;
268
269    if (!vm_dirty_bytes || !dirty_background_bytes)
270        available_memory = global_dirtyable_memory();
271
272    if (vm_dirty_bytes)
273        dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
274    else
275        dirty = (vm_dirty_ratio * available_memory) / 100;
276
277    if (dirty_background_bytes)
278        background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
279    else
280        background = (dirty_background_ratio * available_memory) / 100;
281
282    if (background >= dirty)
283        background = dirty / 2;
284    tsk = current;
285    if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
286        background += background / 4;
287        dirty += dirty / 4;
288    }
289    *pbackground = background;
290    *pdirty = dirty;
291    trace_global_dirty_state(background, dirty);
292}
293
294/**
295 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
296 * @zone: the zone
297 *
298 * Returns the maximum number of dirty pages allowed in a zone, based
299 * on the zone's dirtyable memory.
300 */
301static unsigned long zone_dirty_limit(struct zone *zone)
302{
303    unsigned long zone_memory = zone_dirtyable_memory(zone);
304    struct task_struct *tsk = current;
305    unsigned long dirty;
306
307    if (vm_dirty_bytes)
308        dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
309            zone_memory / global_dirtyable_memory();
310    else
311        dirty = vm_dirty_ratio * zone_memory / 100;
312
313    if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
314        dirty += dirty / 4;
315
316    return dirty;
317}
318
319/**
320 * zone_dirty_ok - tells whether a zone is within its dirty limits
321 * @zone: the zone to check
322 *
323 * Returns %true when the dirty pages in @zone are within the zone's
324 * dirty limit, %false if the limit is exceeded.
325 */
326bool zone_dirty_ok(struct zone *zone)
327{
328    unsigned long limit = zone_dirty_limit(zone);
329
330    return zone_page_state(zone, NR_FILE_DIRTY) +
331           zone_page_state(zone, NR_UNSTABLE_NFS) +
332           zone_page_state(zone, NR_WRITEBACK) <= limit;
333}
334
335int dirty_background_ratio_handler(struct ctl_table *table, int write,
336        void __user *buffer, size_t *lenp,
337        loff_t *ppos)
338{
339    int ret;
340
341    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
342    if (ret == 0 && write)
343        dirty_background_bytes = 0;
344    return ret;
345}
346
347int dirty_background_bytes_handler(struct ctl_table *table, int write,
348        void __user *buffer, size_t *lenp,
349        loff_t *ppos)
350{
351    int ret;
352
353    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
354    if (ret == 0 && write)
355        dirty_background_ratio = 0;
356    return ret;
357}
358
359int dirty_ratio_handler(struct ctl_table *table, int write,
360        void __user *buffer, size_t *lenp,
361        loff_t *ppos)
362{
363    int old_ratio = vm_dirty_ratio;
364    int ret;
365
366    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
367    if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
368        writeback_set_ratelimit();
369        vm_dirty_bytes = 0;
370    }
371    return ret;
372}
373
374int dirty_bytes_handler(struct ctl_table *table, int write,
375        void __user *buffer, size_t *lenp,
376        loff_t *ppos)
377{
378    unsigned long old_bytes = vm_dirty_bytes;
379    int ret;
380
381    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
382    if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
383        writeback_set_ratelimit();
384        vm_dirty_ratio = 0;
385    }
386    return ret;
387}
388
389static unsigned long wp_next_time(unsigned long cur_time)
390{
391    cur_time += VM_COMPLETIONS_PERIOD_LEN;
392    /* 0 has a special meaning... */
393    if (!cur_time)
394        return 1;
395    return cur_time;
396}
397
398/*
399 * Increment the BDI's writeout completion count and the global writeout
400 * completion count. Called from test_clear_page_writeback().
401 */
402static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
403{
404    __inc_bdi_stat(bdi, BDI_WRITTEN);
405    __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
406                   bdi->max_prop_frac);
407    /* First event after period switching was turned off? */
408    if (!unlikely(writeout_period_time)) {
409        /*
410         * We can race with other __bdi_writeout_inc calls here but
411         * it does not cause any harm since the resulting time when
412         * timer will fire and what is in writeout_period_time will be
413         * roughly the same.
414         */
415        writeout_period_time = wp_next_time(jiffies);
416        mod_timer(&writeout_period_timer, writeout_period_time);
417    }
418}
419
420void bdi_writeout_inc(struct backing_dev_info *bdi)
421{
422    unsigned long flags;
423
424    local_irq_save(flags);
425    __bdi_writeout_inc(bdi);
426    local_irq_restore(flags);
427}
428EXPORT_SYMBOL_GPL(bdi_writeout_inc);
429
430/*
431 * Obtain an accurate fraction of the BDI's portion.
432 */
433static void bdi_writeout_fraction(struct backing_dev_info *bdi,
434        long *numerator, long *denominator)
435{
436    fprop_fraction_percpu(&writeout_completions, &bdi->completions,
437                numerator, denominator);
438}
439
440/*
441 * On idle system, we can be called long after we scheduled because we use
442 * deferred timers so count with missed periods.
443 */
444static void writeout_period(unsigned long t)
445{
446    int miss_periods = (jiffies - writeout_period_time) /
447                         VM_COMPLETIONS_PERIOD_LEN;
448
449    if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
450        writeout_period_time = wp_next_time(writeout_period_time +
451                miss_periods * VM_COMPLETIONS_PERIOD_LEN);
452        mod_timer(&writeout_period_timer, writeout_period_time);
453    } else {
454        /*
455         * Aging has zeroed all fractions. Stop wasting CPU on period
456         * updates.
457         */
458        writeout_period_time = 0;
459    }
460}
461
462/*
463 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
464 * registered backing devices, which, for obvious reasons, can not
465 * exceed 100%.
466 */
467static unsigned int bdi_min_ratio;
468
469int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
470{
471    int ret = 0;
472
473    spin_lock_bh(&bdi_lock);
474    if (min_ratio > bdi->max_ratio) {
475        ret = -EINVAL;
476    } else {
477        min_ratio -= bdi->min_ratio;
478        if (bdi_min_ratio + min_ratio < 100) {
479            bdi_min_ratio += min_ratio;
480            bdi->min_ratio += min_ratio;
481        } else {
482            ret = -EINVAL;
483        }
484    }
485    spin_unlock_bh(&bdi_lock);
486
487    return ret;
488}
489
490int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
491{
492    int ret = 0;
493
494    if (max_ratio > 100)
495        return -EINVAL;
496
497    spin_lock_bh(&bdi_lock);
498    if (bdi->min_ratio > max_ratio) {
499        ret = -EINVAL;
500    } else {
501        bdi->max_ratio = max_ratio;
502        bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
503    }
504    spin_unlock_bh(&bdi_lock);
505
506    return ret;
507}
508EXPORT_SYMBOL(bdi_set_max_ratio);
509
510static unsigned long dirty_freerun_ceiling(unsigned long thresh,
511                       unsigned long bg_thresh)
512{
513    return (thresh + bg_thresh) / 2;
514}
515
516static unsigned long hard_dirty_limit(unsigned long thresh)
517{
518    return max(thresh, global_dirty_limit);
519}
520
521/**
522 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
523 * @bdi: the backing_dev_info to query
524 * @dirty: global dirty limit in pages
525 *
526 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
527 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
528 *
529 * Note that balance_dirty_pages() will only seriously take it as a hard limit
530 * when sleeping max_pause per page is not enough to keep the dirty pages under
531 * control. For example, when the device is completely stalled due to some error
532 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
533 * In the other normal situations, it acts more gently by throttling the tasks
534 * more (rather than completely block them) when the bdi dirty pages go high.
535 *
536 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
537 * - starving fast devices
538 * - piling up dirty pages (that will take long time to sync) on slow devices
539 *
540 * The bdi's share of dirty limit will be adapting to its throughput and
541 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
542 */
543unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
544{
545    u64 bdi_dirty;
546    long numerator, denominator;
547
548    /*
549     * Calculate this BDI's share of the dirty ratio.
550     */
551    bdi_writeout_fraction(bdi, &numerator, &denominator);
552
553    bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
554    bdi_dirty *= numerator;
555    do_div(bdi_dirty, denominator);
556
557    bdi_dirty += (dirty * bdi->min_ratio) / 100;
558    if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
559        bdi_dirty = dirty * bdi->max_ratio / 100;
560
561    return bdi_dirty;
562}
563
564/*
565 * setpoint - dirty 3
566 * f(dirty) := 1.0 + (----------------)
567 * limit - setpoint
568 *
569 * it's a 3rd order polynomial that subjects to
570 *
571 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
572 * (2) f(setpoint) = 1.0 => the balance point
573 * (3) f(limit) = 0 => the hard limit
574 * (4) df/dx <= 0 => negative feedback control
575 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
576 * => fast response on large errors; small oscillation near setpoint
577 */
578static long long pos_ratio_polynom(unsigned long setpoint,
579                      unsigned long dirty,
580                      unsigned long limit)
581{
582    long long pos_ratio;
583    long x;
584
585    x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
586            limit - setpoint + 1);
587    pos_ratio = x;
588    pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
589    pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
590    pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
591
592    return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
593}
594
595/*
596 * Dirty position control.
597 *
598 * (o) global/bdi setpoints
599 *
600 * We want the dirty pages be balanced around the global/bdi setpoints.
601 * When the number of dirty pages is higher/lower than the setpoint, the
602 * dirty position control ratio (and hence task dirty ratelimit) will be
603 * decreased/increased to bring the dirty pages back to the setpoint.
604 *
605 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
606 *
607 * if (dirty < setpoint) scale up pos_ratio
608 * if (dirty > setpoint) scale down pos_ratio
609 *
610 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
611 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
612 *
613 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
614 *
615 * (o) global control line
616 *
617 * ^ pos_ratio
618 * |
619 * | |<===== global dirty control scope ======>|
620 * 2.0 .............*
621 * | .*
622 * | . *
623 * | . *
624 * | . *
625 * | . *
626 * | . *
627 * 1.0 ................................*
628 * | . . *
629 * | . . *
630 * | . . *
631 * | . . *
632 * | . . *
633 * 0 +------------.------------------.----------------------*------------->
634 * freerun^ setpoint^ limit^ dirty pages
635 *
636 * (o) bdi control line
637 *
638 * ^ pos_ratio
639 * |
640 * | *
641 * | *
642 * | *
643 * | *
644 * | * |<=========== span ============>|
645 * 1.0 .......................*
646 * | . *
647 * | . *
648 * | . *
649 * | . *
650 * | . *
651 * | . *
652 * | . *
653 * | . *
654 * | . *
655 * | . *
656 * | . *
657 * 1/4 ...............................................* * * * * * * * * * * *
658 * | . .
659 * | . .
660 * | . .
661 * 0 +----------------------.-------------------------------.------------->
662 * bdi_setpoint^ x_intercept^
663 *
664 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
665 * be smoothly throttled down to normal if it starts high in situations like
666 * - start writing to a slow SD card and a fast disk at the same time. The SD
667 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
668 * - the bdi dirty thresh drops quickly due to change of JBOD workload
669 */
670static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
671                    unsigned long thresh,
672                    unsigned long bg_thresh,
673                    unsigned long dirty,
674                    unsigned long bdi_thresh,
675                    unsigned long bdi_dirty)
676{
677    unsigned long write_bw = bdi->avg_write_bandwidth;
678    unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
679    unsigned long limit = hard_dirty_limit(thresh);
680    unsigned long x_intercept;
681    unsigned long setpoint; /* dirty pages' target balance point */
682    unsigned long bdi_setpoint;
683    unsigned long span;
684    long long pos_ratio; /* for scaling up/down the rate limit */
685    long x;
686
687    if (unlikely(dirty >= limit))
688        return 0;
689
690    /*
691     * global setpoint
692     *
693     * See comment for pos_ratio_polynom().
694     */
695    setpoint = (freerun + limit) / 2;
696    pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
697
698    /*
699     * The strictlimit feature is a tool preventing mistrusted filesystems
700     * from growing a large number of dirty pages before throttling. For
701     * such filesystems balance_dirty_pages always checks bdi counters
702     * against bdi limits. Even if global "nr_dirty" is under "freerun".
703     * This is especially important for fuse which sets bdi->max_ratio to
704     * 1% by default. Without strictlimit feature, fuse writeback may
705     * consume arbitrary amount of RAM because it is accounted in
706     * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
707     *
708     * Here, in bdi_position_ratio(), we calculate pos_ratio based on
709     * two values: bdi_dirty and bdi_thresh. Let's consider an example:
710     * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
711     * limits are set by default to 10% and 20% (background and throttle).
712     * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
713     * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
714     * about ~6K pages (as the average of background and throttle bdi
715     * limits). The 3rd order polynomial will provide positive feedback if
716     * bdi_dirty is under bdi_setpoint and vice versa.
717     *
718     * Note, that we cannot use global counters in these calculations
719     * because we want to throttle process writing to a strictlimit BDI
720     * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
721     * in the example above).
722     */
723    if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
724        long long bdi_pos_ratio;
725        unsigned long bdi_bg_thresh;
726
727        if (bdi_dirty < 8)
728            return min_t(long long, pos_ratio * 2,
729                     2 << RATELIMIT_CALC_SHIFT);
730
731        if (bdi_dirty >= bdi_thresh)
732            return 0;
733
734        bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
735        bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
736                             bdi_bg_thresh);
737
738        if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
739            return 0;
740
741        bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
742                          bdi_thresh);
743
744        /*
745         * Typically, for strictlimit case, bdi_setpoint << setpoint
746         * and pos_ratio >> bdi_pos_ratio. In the other words global
747         * state ("dirty") is not limiting factor and we have to
748         * make decision based on bdi counters. But there is an
749         * important case when global pos_ratio should get precedence:
750         * global limits are exceeded (e.g. due to activities on other
751         * BDIs) while given strictlimit BDI is below limit.
752         *
753         * "pos_ratio * bdi_pos_ratio" would work for the case above,
754         * but it would look too non-natural for the case of all
755         * activity in the system coming from a single strictlimit BDI
756         * with bdi->max_ratio == 100%.
757         *
758         * Note that min() below somewhat changes the dynamics of the
759         * control system. Normally, pos_ratio value can be well over 3
760         * (when globally we are at freerun and bdi is well below bdi
761         * setpoint). Now the maximum pos_ratio in the same situation
762         * is 2. We might want to tweak this if we observe the control
763         * system is too slow to adapt.
764         */
765        return min(pos_ratio, bdi_pos_ratio);
766    }
767
768    /*
769     * We have computed basic pos_ratio above based on global situation. If
770     * the bdi is over/under its share of dirty pages, we want to scale
771     * pos_ratio further down/up. That is done by the following mechanism.
772     */
773
774    /*
775     * bdi setpoint
776     *
777     * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
778     *
779     * x_intercept - bdi_dirty
780     * := --------------------------
781     * x_intercept - bdi_setpoint
782     *
783     * The main bdi control line is a linear function that subjects to
784     *
785     * (1) f(bdi_setpoint) = 1.0
786     * (2) k = - 1 / (8 * write_bw) (in single bdi case)
787     * or equally: x_intercept = bdi_setpoint + 8 * write_bw
788     *
789     * For single bdi case, the dirty pages are observed to fluctuate
790     * regularly within range
791     * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
792     * for various filesystems, where (2) can yield in a reasonable 12.5%
793     * fluctuation range for pos_ratio.
794     *
795     * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
796     * own size, so move the slope over accordingly and choose a slope that
797     * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
798     */
799    if (unlikely(bdi_thresh > thresh))
800        bdi_thresh = thresh;
801    /*
802     * It's very possible that bdi_thresh is close to 0 not because the
803     * device is slow, but that it has remained inactive for long time.
804     * Honour such devices a reasonable good (hopefully IO efficient)
805     * threshold, so that the occasional writes won't be blocked and active
806     * writes can rampup the threshold quickly.
807     */
808    bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
809    /*
810     * scale global setpoint to bdi's:
811     * bdi_setpoint = setpoint * bdi_thresh / thresh
812     */
813    x = div_u64((u64)bdi_thresh << 16, thresh + 1);
814    bdi_setpoint = setpoint * (u64)x >> 16;
815    /*
816     * Use span=(8*write_bw) in single bdi case as indicated by
817     * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
818     *
819     * bdi_thresh thresh - bdi_thresh
820     * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
821     * thresh thresh
822     */
823    span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
824    x_intercept = bdi_setpoint + span;
825
826    if (bdi_dirty < x_intercept - span / 4) {
827        pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty),
828                    x_intercept - bdi_setpoint + 1);
829    } else
830        pos_ratio /= 4;
831
832    /*
833     * bdi reserve area, safeguard against dirty pool underrun and disk idle
834     * It may push the desired control point of global dirty pages higher
835     * than setpoint.
836     */
837    x_intercept = bdi_thresh / 2;
838    if (bdi_dirty < x_intercept) {
839        if (bdi_dirty > x_intercept / 8)
840            pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
841        else
842            pos_ratio *= 8;
843    }
844
845    return pos_ratio;
846}
847
848static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
849                       unsigned long elapsed,
850                       unsigned long written)
851{
852    const unsigned long period = roundup_pow_of_two(3 * HZ);
853    unsigned long avg = bdi->avg_write_bandwidth;
854    unsigned long old = bdi->write_bandwidth;
855    u64 bw;
856
857    /*
858     * bw = written * HZ / elapsed
859     *
860     * bw * elapsed + write_bandwidth * (period - elapsed)
861     * write_bandwidth = ---------------------------------------------------
862     * period
863     */
864    bw = written - bdi->written_stamp;
865    bw *= HZ;
866    if (unlikely(elapsed > period)) {
867        do_div(bw, elapsed);
868        avg = bw;
869        goto out;
870    }
871    bw += (u64)bdi->write_bandwidth * (period - elapsed);
872    bw >>= ilog2(period);
873
874    /*
875     * one more level of smoothing, for filtering out sudden spikes
876     */
877    if (avg > old && old >= (unsigned long)bw)
878        avg -= (avg - old) >> 3;
879
880    if (avg < old && old <= (unsigned long)bw)
881        avg += (old - avg) >> 3;
882
883out:
884    bdi->write_bandwidth = bw;
885    bdi->avg_write_bandwidth = avg;
886}
887
888/*
889 * The global dirtyable memory and dirty threshold could be suddenly knocked
890 * down by a large amount (eg. on the startup of KVM in a swapless system).
891 * This may throw the system into deep dirty exceeded state and throttle
892 * heavy/light dirtiers alike. To retain good responsiveness, maintain
893 * global_dirty_limit for tracking slowly down to the knocked down dirty
894 * threshold.
895 */
896static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
897{
898    unsigned long limit = global_dirty_limit;
899
900    /*
901     * Follow up in one step.
902     */
903    if (limit < thresh) {
904        limit = thresh;
905        goto update;
906    }
907
908    /*
909     * Follow down slowly. Use the higher one as the target, because thresh
910     * may drop below dirty. This is exactly the reason to introduce
911     * global_dirty_limit which is guaranteed to lie above the dirty pages.
912     */
913    thresh = max(thresh, dirty);
914    if (limit > thresh) {
915        limit -= (limit - thresh) >> 5;
916        goto update;
917    }
918    return;
919update:
920    global_dirty_limit = limit;
921}
922
923static void global_update_bandwidth(unsigned long thresh,
924                    unsigned long dirty,
925                    unsigned long now)
926{
927    static DEFINE_SPINLOCK(dirty_lock);
928    static unsigned long update_time;
929
930    /*
931     * check locklessly first to optimize away locking for the most time
932     */
933    if (time_before(now, update_time + BANDWIDTH_INTERVAL))
934        return;
935
936    spin_lock(&dirty_lock);
937    if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
938        update_dirty_limit(thresh, dirty);
939        update_time = now;
940    }
941    spin_unlock(&dirty_lock);
942}
943
944/*
945 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
946 *
947 * Normal bdi tasks will be curbed at or below it in long term.
948 * Obviously it should be around (write_bw / N) when there are N dd tasks.
949 */
950static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
951                       unsigned long thresh,
952                       unsigned long bg_thresh,
953                       unsigned long dirty,
954                       unsigned long bdi_thresh,
955                       unsigned long bdi_dirty,
956                       unsigned long dirtied,
957                       unsigned long elapsed)
958{
959    unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
960    unsigned long limit = hard_dirty_limit(thresh);
961    unsigned long setpoint = (freerun + limit) / 2;
962    unsigned long write_bw = bdi->avg_write_bandwidth;
963    unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
964    unsigned long dirty_rate;
965    unsigned long task_ratelimit;
966    unsigned long balanced_dirty_ratelimit;
967    unsigned long pos_ratio;
968    unsigned long step;
969    unsigned long x;
970
971    /*
972     * The dirty rate will match the writeout rate in long term, except
973     * when dirty pages are truncated by userspace or re-dirtied by FS.
974     */
975    dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
976
977    pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
978                       bdi_thresh, bdi_dirty);
979    /*
980     * task_ratelimit reflects each dd's dirty rate for the past 200ms.
981     */
982    task_ratelimit = (u64)dirty_ratelimit *
983                    pos_ratio >> RATELIMIT_CALC_SHIFT;
984    task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
985
986    /*
987     * A linear estimation of the "balanced" throttle rate. The theory is,
988     * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
989     * dirty_rate will be measured to be (N * task_ratelimit). So the below
990     * formula will yield the balanced rate limit (write_bw / N).
991     *
992     * Note that the expanded form is not a pure rate feedback:
993     * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
994     * but also takes pos_ratio into account:
995     * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
996     *
997     * (1) is not realistic because pos_ratio also takes part in balancing
998     * the dirty rate. Consider the state
999     * pos_ratio = 0.5 (3)
1000     * rate = 2 * (write_bw / N) (4)
1001     * If (1) is used, it will stuck in that state! Because each dd will
1002     * be throttled at
1003     * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1004     * yielding
1005     * dirty_rate = N * task_ratelimit = write_bw (6)
1006     * put (6) into (1) we get
1007     * rate_(i+1) = rate_(i) (7)
1008     *
1009     * So we end up using (2) to always keep
1010     * rate_(i+1) ~= (write_bw / N) (8)
1011     * regardless of the value of pos_ratio. As long as (8) is satisfied,
1012     * pos_ratio is able to drive itself to 1.0, which is not only where
1013     * the dirty count meet the setpoint, but also where the slope of
1014     * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1015     */
1016    balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1017                       dirty_rate | 1);
1018    /*
1019     * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1020     */
1021    if (unlikely(balanced_dirty_ratelimit > write_bw))
1022        balanced_dirty_ratelimit = write_bw;
1023
1024    /*
1025     * We could safely do this and return immediately:
1026     *
1027     * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
1028     *
1029     * However to get a more stable dirty_ratelimit, the below elaborated
1030     * code makes use of task_ratelimit to filter out singular points and
1031     * limit the step size.
1032     *
1033     * The below code essentially only uses the relative value of
1034     *
1035     * task_ratelimit - dirty_ratelimit
1036     * = (pos_ratio - 1) * dirty_ratelimit
1037     *
1038     * which reflects the direction and size of dirty position error.
1039     */
1040
1041    /*
1042     * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1043     * task_ratelimit is on the same side of dirty_ratelimit, too.
1044     * For example, when
1045     * - dirty_ratelimit > balanced_dirty_ratelimit
1046     * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1047     * lowering dirty_ratelimit will help meet both the position and rate
1048     * control targets. Otherwise, don't update dirty_ratelimit if it will
1049     * only help meet the rate target. After all, what the users ultimately
1050     * feel and care are stable dirty rate and small position error.
1051     *
1052     * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1053     * and filter out the singular points of balanced_dirty_ratelimit. Which
1054     * keeps jumping around randomly and can even leap far away at times
1055     * due to the small 200ms estimation period of dirty_rate (we want to
1056     * keep that period small to reduce time lags).
1057     */
1058    step = 0;
1059
1060    /*
1061     * For strictlimit case, calculations above were based on bdi counters
1062     * and limits (starting from pos_ratio = bdi_position_ratio() and up to
1063     * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1064     * Hence, to calculate "step" properly, we have to use bdi_dirty as
1065     * "dirty" and bdi_setpoint as "setpoint".
1066     *
1067     * We rampup dirty_ratelimit forcibly if bdi_dirty is low because
1068     * it's possible that bdi_thresh is close to zero due to inactivity
1069     * of backing device (see the implementation of bdi_dirty_limit()).
1070     */
1071    if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1072        dirty = bdi_dirty;
1073        if (bdi_dirty < 8)
1074            setpoint = bdi_dirty + 1;
1075        else
1076            setpoint = (bdi_thresh +
1077                    bdi_dirty_limit(bdi, bg_thresh)) / 2;
1078    }
1079
1080    if (dirty < setpoint) {
1081        x = min(bdi->balanced_dirty_ratelimit,
1082             min(balanced_dirty_ratelimit, task_ratelimit));
1083        if (dirty_ratelimit < x)
1084            step = x - dirty_ratelimit;
1085    } else {
1086        x = max(bdi->balanced_dirty_ratelimit,
1087             max(balanced_dirty_ratelimit, task_ratelimit));
1088        if (dirty_ratelimit > x)
1089            step = dirty_ratelimit - x;
1090    }
1091
1092    /*
1093     * Don't pursue 100% rate matching. It's impossible since the balanced
1094     * rate itself is constantly fluctuating. So decrease the track speed
1095     * when it gets close to the target. Helps eliminate pointless tremors.
1096     */
1097    step >>= dirty_ratelimit / (2 * step + 1);
1098    /*
1099     * Limit the tracking speed to avoid overshooting.
1100     */
1101    step = (step + 7) / 8;
1102
1103    if (dirty_ratelimit < balanced_dirty_ratelimit)
1104        dirty_ratelimit += step;
1105    else
1106        dirty_ratelimit -= step;
1107
1108    bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1109    bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1110
1111    trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1112}
1113
1114void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1115                unsigned long thresh,
1116                unsigned long bg_thresh,
1117                unsigned long dirty,
1118                unsigned long bdi_thresh,
1119                unsigned long bdi_dirty,
1120                unsigned long start_time)
1121{
1122    unsigned long now = jiffies;
1123    unsigned long elapsed = now - bdi->bw_time_stamp;
1124    unsigned long dirtied;
1125    unsigned long written;
1126
1127    /*
1128     * rate-limit, only update once every 200ms.
1129     */
1130    if (elapsed < BANDWIDTH_INTERVAL)
1131        return;
1132
1133    dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1134    written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1135
1136    /*
1137     * Skip quiet periods when disk bandwidth is under-utilized.
1138     * (at least 1s idle time between two flusher runs)
1139     */
1140    if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1141        goto snapshot;
1142
1143    if (thresh) {
1144        global_update_bandwidth(thresh, dirty, now);
1145        bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1146                       bdi_thresh, bdi_dirty,
1147                       dirtied, elapsed);
1148    }
1149    bdi_update_write_bandwidth(bdi, elapsed, written);
1150
1151snapshot:
1152    bdi->dirtied_stamp = dirtied;
1153    bdi->written_stamp = written;
1154    bdi->bw_time_stamp = now;
1155}
1156
1157static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1158                 unsigned long thresh,
1159                 unsigned long bg_thresh,
1160                 unsigned long dirty,
1161                 unsigned long bdi_thresh,
1162                 unsigned long bdi_dirty,
1163                 unsigned long start_time)
1164{
1165    if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1166        return;
1167    spin_lock(&bdi->wb.list_lock);
1168    __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1169                   bdi_thresh, bdi_dirty, start_time);
1170    spin_unlock(&bdi->wb.list_lock);
1171}
1172
1173/*
1174 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1175 * will look to see if it needs to start dirty throttling.
1176 *
1177 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1178 * global_page_state() too often. So scale it near-sqrt to the safety margin
1179 * (the number of pages we may dirty without exceeding the dirty limits).
1180 */
1181static unsigned long dirty_poll_interval(unsigned long dirty,
1182                     unsigned long thresh)
1183{
1184    if (thresh > dirty)
1185        return 1UL << (ilog2(thresh - dirty) >> 1);
1186
1187    return 1;
1188}
1189
1190static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1191                   unsigned long bdi_dirty)
1192{
1193    unsigned long bw = bdi->avg_write_bandwidth;
1194    unsigned long t;
1195
1196    /*
1197     * Limit pause time for small memory systems. If sleeping for too long
1198     * time, a small pool of dirty/writeback pages may go empty and disk go
1199     * idle.
1200     *
1201     * 8 serves as the safety ratio.
1202     */
1203    t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1204    t++;
1205
1206    return min_t(unsigned long, t, MAX_PAUSE);
1207}
1208
1209static long bdi_min_pause(struct backing_dev_info *bdi,
1210              long max_pause,
1211              unsigned long task_ratelimit,
1212              unsigned long dirty_ratelimit,
1213              int *nr_dirtied_pause)
1214{
1215    long hi = ilog2(bdi->avg_write_bandwidth);
1216    long lo = ilog2(bdi->dirty_ratelimit);
1217    long t; /* target pause */
1218    long pause; /* estimated next pause */
1219    int pages; /* target nr_dirtied_pause */
1220
1221    /* target for 10ms pause on 1-dd case */
1222    t = max(1, HZ / 100);
1223
1224    /*
1225     * Scale up pause time for concurrent dirtiers in order to reduce CPU
1226     * overheads.
1227     *
1228     * (N * 10ms) on 2^N concurrent tasks.
1229     */
1230    if (hi > lo)
1231        t += (hi - lo) * (10 * HZ) / 1024;
1232
1233    /*
1234     * This is a bit convoluted. We try to base the next nr_dirtied_pause
1235     * on the much more stable dirty_ratelimit. However the next pause time
1236     * will be computed based on task_ratelimit and the two rate limits may
1237     * depart considerably at some time. Especially if task_ratelimit goes
1238     * below dirty_ratelimit/2 and the target pause is max_pause, the next
1239     * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1240     * result task_ratelimit won't be executed faithfully, which could
1241     * eventually bring down dirty_ratelimit.
1242     *
1243     * We apply two rules to fix it up:
1244     * 1) try to estimate the next pause time and if necessary, use a lower
1245     * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1246     * nr_dirtied_pause will be "dancing" with task_ratelimit.
1247     * 2) limit the target pause time to max_pause/2, so that the normal
1248     * small fluctuations of task_ratelimit won't trigger rule (1) and
1249     * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1250     */
1251    t = min(t, 1 + max_pause / 2);
1252    pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1253
1254    /*
1255     * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1256     * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1257     * When the 16 consecutive reads are often interrupted by some dirty
1258     * throttling pause during the async writes, cfq will go into idles
1259     * (deadline is fine). So push nr_dirtied_pause as high as possible
1260     * until reaches DIRTY_POLL_THRESH=32 pages.
1261     */
1262    if (pages < DIRTY_POLL_THRESH) {
1263        t = max_pause;
1264        pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1265        if (pages > DIRTY_POLL_THRESH) {
1266            pages = DIRTY_POLL_THRESH;
1267            t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1268        }
1269    }
1270
1271    pause = HZ * pages / (task_ratelimit + 1);
1272    if (pause > max_pause) {
1273        t = max_pause;
1274        pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1275    }
1276
1277    *nr_dirtied_pause = pages;
1278    /*
1279     * The minimal pause time will normally be half the target pause time.
1280     */
1281    return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1282}
1283
1284static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
1285                    unsigned long dirty_thresh,
1286                    unsigned long background_thresh,
1287                    unsigned long *bdi_dirty,
1288                    unsigned long *bdi_thresh,
1289                    unsigned long *bdi_bg_thresh)
1290{
1291    unsigned long bdi_reclaimable;
1292
1293    /*
1294     * bdi_thresh is not treated as some limiting factor as
1295     * dirty_thresh, due to reasons
1296     * - in JBOD setup, bdi_thresh can fluctuate a lot
1297     * - in a system with HDD and USB key, the USB key may somehow
1298     * go into state (bdi_dirty >> bdi_thresh) either because
1299     * bdi_dirty starts high, or because bdi_thresh drops low.
1300     * In this case we don't want to hard throttle the USB key
1301     * dirtiers for 100 seconds until bdi_dirty drops under
1302     * bdi_thresh. Instead the auxiliary bdi control line in
1303     * bdi_position_ratio() will let the dirtier task progress
1304     * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1305     */
1306    *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1307
1308    if (bdi_bg_thresh)
1309        *bdi_bg_thresh = dirty_thresh ? div_u64((u64)*bdi_thresh *
1310                            background_thresh,
1311                            dirty_thresh) : 0;
1312
1313    /*
1314     * In order to avoid the stacked BDI deadlock we need
1315     * to ensure we accurately count the 'dirty' pages when
1316     * the threshold is low.
1317     *
1318     * Otherwise it would be possible to get thresh+n pages
1319     * reported dirty, even though there are thresh-m pages
1320     * actually dirty; with m+n sitting in the percpu
1321     * deltas.
1322     */
1323    if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
1324        bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1325        *bdi_dirty = bdi_reclaimable +
1326            bdi_stat_sum(bdi, BDI_WRITEBACK);
1327    } else {
1328        bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1329        *bdi_dirty = bdi_reclaimable +
1330            bdi_stat(bdi, BDI_WRITEBACK);
1331    }
1332}
1333
1334/*
1335 * balance_dirty_pages() must be called by processes which are generating dirty
1336 * data. It looks at the number of dirty pages in the machine and will force
1337 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1338 * If we're over `background_thresh' then the writeback threads are woken to
1339 * perform some writeout.
1340 */
1341static void balance_dirty_pages(struct address_space *mapping,
1342                unsigned long pages_dirtied)
1343{
1344    unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1345    unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1346    unsigned long background_thresh;
1347    unsigned long dirty_thresh;
1348    long period;
1349    long pause;
1350    long max_pause;
1351    long min_pause;
1352    int nr_dirtied_pause;
1353    bool dirty_exceeded = false;
1354    unsigned long task_ratelimit;
1355    unsigned long dirty_ratelimit;
1356    unsigned long pos_ratio;
1357    struct backing_dev_info *bdi = mapping->backing_dev_info;
1358    bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1359    unsigned long start_time = jiffies;
1360
1361    for (;;) {
1362        unsigned long now = jiffies;
1363        unsigned long uninitialized_var(bdi_thresh);
1364        unsigned long thresh;
1365        unsigned long uninitialized_var(bdi_dirty);
1366        unsigned long dirty;
1367        unsigned long bg_thresh;
1368
1369        /*
1370         * Unstable writes are a feature of certain networked
1371         * filesystems (i.e. NFS) in which data may have been
1372         * written to the server's write cache, but has not yet
1373         * been flushed to permanent storage.
1374         */
1375        nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1376                    global_page_state(NR_UNSTABLE_NFS);
1377        nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1378
1379        global_dirty_limits(&background_thresh, &dirty_thresh);
1380
1381        if (unlikely(strictlimit)) {
1382            bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1383                     &bdi_dirty, &bdi_thresh, &bg_thresh);
1384
1385            dirty = bdi_dirty;
1386            thresh = bdi_thresh;
1387        } else {
1388            dirty = nr_dirty;
1389            thresh = dirty_thresh;
1390            bg_thresh = background_thresh;
1391        }
1392
1393        /*
1394         * Throttle it only when the background writeback cannot
1395         * catch-up. This avoids (excessively) small writeouts
1396         * when the bdi limits are ramping up in case of !strictlimit.
1397         *
1398         * In strictlimit case make decision based on the bdi counters
1399         * and limits. Small writeouts when the bdi limits are ramping
1400         * up are the price we consciously pay for strictlimit-ing.
1401         */
1402        if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1403            current->dirty_paused_when = now;
1404            current->nr_dirtied = 0;
1405            current->nr_dirtied_pause =
1406                dirty_poll_interval(dirty, thresh);
1407            break;
1408        }
1409
1410        if (unlikely(!writeback_in_progress(bdi)))
1411            bdi_start_background_writeback(bdi);
1412
1413        if (!strictlimit)
1414            bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1415                     &bdi_dirty, &bdi_thresh, NULL);
1416
1417        dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1418                 ((nr_dirty > dirty_thresh) || strictlimit);
1419        if (dirty_exceeded && !bdi->dirty_exceeded)
1420            bdi->dirty_exceeded = 1;
1421
1422        bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1423                     nr_dirty, bdi_thresh, bdi_dirty,
1424                     start_time);
1425
1426        dirty_ratelimit = bdi->dirty_ratelimit;
1427        pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1428                           background_thresh, nr_dirty,
1429                           bdi_thresh, bdi_dirty);
1430        task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1431                            RATELIMIT_CALC_SHIFT;
1432        max_pause = bdi_max_pause(bdi, bdi_dirty);
1433        min_pause = bdi_min_pause(bdi, max_pause,
1434                      task_ratelimit, dirty_ratelimit,
1435                      &nr_dirtied_pause);
1436
1437        if (unlikely(task_ratelimit == 0)) {
1438            period = max_pause;
1439            pause = max_pause;
1440            goto pause;
1441        }
1442        period = HZ * pages_dirtied / task_ratelimit;
1443        pause = period;
1444        if (current->dirty_paused_when)
1445            pause -= now - current->dirty_paused_when;
1446        /*
1447         * For less than 1s think time (ext3/4 may block the dirtier
1448         * for up to 800ms from time to time on 1-HDD; so does xfs,
1449         * however at much less frequency), try to compensate it in
1450         * future periods by updating the virtual time; otherwise just
1451         * do a reset, as it may be a light dirtier.
1452         */
1453        if (pause < min_pause) {
1454            trace_balance_dirty_pages(bdi,
1455                          dirty_thresh,
1456                          background_thresh,
1457                          nr_dirty,
1458                          bdi_thresh,
1459                          bdi_dirty,
1460                          dirty_ratelimit,
1461                          task_ratelimit,
1462                          pages_dirtied,
1463                          period,
1464                          min(pause, 0L),
1465                          start_time);
1466            if (pause < -HZ) {
1467                current->dirty_paused_when = now;
1468                current->nr_dirtied = 0;
1469            } else if (period) {
1470                current->dirty_paused_when += period;
1471                current->nr_dirtied = 0;
1472            } else if (current->nr_dirtied_pause <= pages_dirtied)
1473                current->nr_dirtied_pause += pages_dirtied;
1474            break;
1475        }
1476        if (unlikely(pause > max_pause)) {
1477            /* for occasional dropped task_ratelimit */
1478            now += min(pause - max_pause, max_pause);
1479            pause = max_pause;
1480        }
1481
1482pause:
1483        trace_balance_dirty_pages(bdi,
1484                      dirty_thresh,
1485                      background_thresh,
1486                      nr_dirty,
1487                      bdi_thresh,
1488                      bdi_dirty,
1489                      dirty_ratelimit,
1490                      task_ratelimit,
1491                      pages_dirtied,
1492                      period,
1493                      pause,
1494                      start_time);
1495        __set_current_state(TASK_KILLABLE);
1496        io_schedule_timeout(pause);
1497
1498        current->dirty_paused_when = now + pause;
1499        current->nr_dirtied = 0;
1500        current->nr_dirtied_pause = nr_dirtied_pause;
1501
1502        /*
1503         * This is typically equal to (nr_dirty < dirty_thresh) and can
1504         * also keep "1000+ dd on a slow USB stick" under control.
1505         */
1506        if (task_ratelimit)
1507            break;
1508
1509        /*
1510         * In the case of an unresponding NFS server and the NFS dirty
1511         * pages exceeds dirty_thresh, give the other good bdi's a pipe
1512         * to go through, so that tasks on them still remain responsive.
1513         *
1514         * In theory 1 page is enough to keep the comsumer-producer
1515         * pipe going: the flusher cleans 1 page => the task dirties 1
1516         * more page. However bdi_dirty has accounting errors. So use
1517         * the larger and more IO friendly bdi_stat_error.
1518         */
1519        if (bdi_dirty <= bdi_stat_error(bdi))
1520            break;
1521
1522        if (fatal_signal_pending(current))
1523            break;
1524    }
1525
1526    if (!dirty_exceeded && bdi->dirty_exceeded)
1527        bdi->dirty_exceeded = 0;
1528
1529    if (writeback_in_progress(bdi))
1530        return;
1531
1532    /*
1533     * In laptop mode, we wait until hitting the higher threshold before
1534     * starting background writeout, and then write out all the way down
1535     * to the lower threshold. So slow writers cause minimal disk activity.
1536     *
1537     * In normal mode, we start background writeout at the lower
1538     * background_thresh, to keep the amount of dirty memory low.
1539     */
1540    if (laptop_mode)
1541        return;
1542
1543    if (nr_reclaimable > background_thresh)
1544        bdi_start_background_writeback(bdi);
1545}
1546
1547void set_page_dirty_balance(struct page *page)
1548{
1549    if (set_page_dirty(page)) {
1550        struct address_space *mapping = page_mapping(page);
1551
1552        if (mapping)
1553            balance_dirty_pages_ratelimited(mapping);
1554    }
1555}
1556
1557static DEFINE_PER_CPU(int, bdp_ratelimits);
1558
1559/*
1560 * Normal tasks are throttled by
1561 * loop {
1562 * dirty tsk->nr_dirtied_pause pages;
1563 * take a snap in balance_dirty_pages();
1564 * }
1565 * However there is a worst case. If every task exit immediately when dirtied
1566 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1567 * called to throttle the page dirties. The solution is to save the not yet
1568 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1569 * randomly into the running tasks. This works well for the above worst case,
1570 * as the new task will pick up and accumulate the old task's leaked dirty
1571 * count and eventually get throttled.
1572 */
1573DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1574
1575/**
1576 * balance_dirty_pages_ratelimited - balance dirty memory state
1577 * @mapping: address_space which was dirtied
1578 *
1579 * Processes which are dirtying memory should call in here once for each page
1580 * which was newly dirtied. The function will periodically check the system's
1581 * dirty state and will initiate writeback if needed.
1582 *
1583 * On really big machines, get_writeback_state is expensive, so try to avoid
1584 * calling it too often (ratelimiting). But once we're over the dirty memory
1585 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1586 * from overshooting the limit by (ratelimit_pages) each.
1587 */
1588void balance_dirty_pages_ratelimited(struct address_space *mapping)
1589{
1590    struct backing_dev_info *bdi = mapping->backing_dev_info;
1591    int ratelimit;
1592    int *p;
1593
1594    if (!bdi_cap_account_dirty(bdi))
1595        return;
1596
1597    ratelimit = current->nr_dirtied_pause;
1598    if (bdi->dirty_exceeded)
1599        ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1600
1601    preempt_disable();
1602    /*
1603     * This prevents one CPU to accumulate too many dirtied pages without
1604     * calling into balance_dirty_pages(), which can happen when there are
1605     * 1000+ tasks, all of them start dirtying pages at exactly the same
1606     * time, hence all honoured too large initial task->nr_dirtied_pause.
1607     */
1608    p = this_cpu_ptr(&bdp_ratelimits);
1609    if (unlikely(current->nr_dirtied >= ratelimit))
1610        *p = 0;
1611    else if (unlikely(*p >= ratelimit_pages)) {
1612        *p = 0;
1613        ratelimit = 0;
1614    }
1615    /*
1616     * Pick up the dirtied pages by the exited tasks. This avoids lots of
1617     * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1618     * the dirty throttling and livelock other long-run dirtiers.
1619     */
1620    p = this_cpu_ptr(&dirty_throttle_leaks);
1621    if (*p > 0 && current->nr_dirtied < ratelimit) {
1622        unsigned long nr_pages_dirtied;
1623        nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1624        *p -= nr_pages_dirtied;
1625        current->nr_dirtied += nr_pages_dirtied;
1626    }
1627    preempt_enable();
1628
1629    if (unlikely(current->nr_dirtied >= ratelimit))
1630        balance_dirty_pages(mapping, current->nr_dirtied);
1631}
1632EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1633
1634void throttle_vm_writeout(gfp_t gfp_mask)
1635{
1636    unsigned long background_thresh;
1637    unsigned long dirty_thresh;
1638
1639        for ( ; ; ) {
1640        global_dirty_limits(&background_thresh, &dirty_thresh);
1641        dirty_thresh = hard_dirty_limit(dirty_thresh);
1642
1643                /*
1644                 * Boost the allowable dirty threshold a bit for page
1645                 * allocators so they don't get DoS'ed by heavy writers
1646                 */
1647                dirty_thresh += dirty_thresh / 10; /* wheeee... */
1648
1649                if (global_page_state(NR_UNSTABLE_NFS) +
1650            global_page_state(NR_WRITEBACK) <= dirty_thresh)
1651                            break;
1652                congestion_wait(BLK_RW_ASYNC, HZ/10);
1653
1654        /*
1655         * The caller might hold locks which can prevent IO completion
1656         * or progress in the filesystem. So we cannot just sit here
1657         * waiting for IO to complete.
1658         */
1659        if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1660            break;
1661        }
1662}
1663
1664/*
1665 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1666 */
1667int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1668    void __user *buffer, size_t *length, loff_t *ppos)
1669{
1670    proc_dointvec(table, write, buffer, length, ppos);
1671    return 0;
1672}
1673
1674#ifdef CONFIG_BLOCK
1675void laptop_mode_timer_fn(unsigned long data)
1676{
1677    struct request_queue *q = (struct request_queue *)data;
1678    int nr_pages = global_page_state(NR_FILE_DIRTY) +
1679        global_page_state(NR_UNSTABLE_NFS);
1680
1681    /*
1682     * We want to write everything out, not just down to the dirty
1683     * threshold
1684     */
1685    if (bdi_has_dirty_io(&q->backing_dev_info))
1686        bdi_start_writeback(&q->backing_dev_info, nr_pages,
1687                    WB_REASON_LAPTOP_TIMER);
1688}
1689
1690/*
1691 * We've spun up the disk and we're in laptop mode: schedule writeback
1692 * of all dirty data a few seconds from now. If the flush is already scheduled
1693 * then push it back - the user is still using the disk.
1694 */
1695void laptop_io_completion(struct backing_dev_info *info)
1696{
1697    mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1698}
1699
1700/*
1701 * We're in laptop mode and we've just synced. The sync's writes will have
1702 * caused another writeback to be scheduled by laptop_io_completion.
1703 * Nothing needs to be written back anymore, so we unschedule the writeback.
1704 */
1705void laptop_sync_completion(void)
1706{
1707    struct backing_dev_info *bdi;
1708
1709    rcu_read_lock();
1710
1711    list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1712        del_timer(&bdi->laptop_mode_wb_timer);
1713
1714    rcu_read_unlock();
1715}
1716#endif
1717
1718/*
1719 * If ratelimit_pages is too high then we can get into dirty-data overload
1720 * if a large number of processes all perform writes at the same time.
1721 * If it is too low then SMP machines will call the (expensive)
1722 * get_writeback_state too often.
1723 *
1724 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1725 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1726 * thresholds.
1727 */
1728
1729void writeback_set_ratelimit(void)
1730{
1731    unsigned long background_thresh;
1732    unsigned long dirty_thresh;
1733    global_dirty_limits(&background_thresh, &dirty_thresh);
1734    global_dirty_limit = dirty_thresh;
1735    ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1736    if (ratelimit_pages < 16)
1737        ratelimit_pages = 16;
1738}
1739
1740static int
1741ratelimit_handler(struct notifier_block *self, unsigned long action,
1742          void *hcpu)
1743{
1744
1745    switch (action & ~CPU_TASKS_FROZEN) {
1746    case CPU_ONLINE:
1747    case CPU_DEAD:
1748        writeback_set_ratelimit();
1749        return NOTIFY_OK;
1750    default:
1751        return NOTIFY_DONE;
1752    }
1753}
1754
1755static struct notifier_block ratelimit_nb = {
1756    .notifier_call = ratelimit_handler,
1757    .next = NULL,
1758};
1759
1760/*
1761 * Called early on to tune the page writeback dirty limits.
1762 *
1763 * We used to scale dirty pages according to how total memory
1764 * related to pages that could be allocated for buffers (by
1765 * comparing nr_free_buffer_pages() to vm_total_pages.
1766 *
1767 * However, that was when we used "dirty_ratio" to scale with
1768 * all memory, and we don't do that any more. "dirty_ratio"
1769 * is now applied to total non-HIGHPAGE memory (by subtracting
1770 * totalhigh_pages from vm_total_pages), and as such we can't
1771 * get into the old insane situation any more where we had
1772 * large amounts of dirty pages compared to a small amount of
1773 * non-HIGHMEM memory.
1774 *
1775 * But we might still want to scale the dirty_ratio by how
1776 * much memory the box has..
1777 */
1778void __init page_writeback_init(void)
1779{
1780    writeback_set_ratelimit();
1781    register_cpu_notifier(&ratelimit_nb);
1782
1783    fprop_global_init(&writeout_completions);
1784}
1785
1786/**
1787 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1788 * @mapping: address space structure to write
1789 * @start: starting page index
1790 * @end: ending page index (inclusive)
1791 *
1792 * This function scans the page range from @start to @end (inclusive) and tags
1793 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1794 * that write_cache_pages (or whoever calls this function) will then use
1795 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1796 * used to avoid livelocking of writeback by a process steadily creating new
1797 * dirty pages in the file (thus it is important for this function to be quick
1798 * so that it can tag pages faster than a dirtying process can create them).
1799 */
1800/*
1801 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1802 */
1803void tag_pages_for_writeback(struct address_space *mapping,
1804                 pgoff_t start, pgoff_t end)
1805{
1806#define WRITEBACK_TAG_BATCH 4096
1807    unsigned long tagged;
1808
1809    do {
1810        spin_lock_irq(&mapping->tree_lock);
1811        tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1812                &start, end, WRITEBACK_TAG_BATCH,
1813                PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1814        spin_unlock_irq(&mapping->tree_lock);
1815        WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1816        cond_resched();
1817        /* We check 'start' to handle wrapping when end == ~0UL */
1818    } while (tagged >= WRITEBACK_TAG_BATCH && start);
1819}
1820EXPORT_SYMBOL(tag_pages_for_writeback);
1821
1822/**
1823 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1824 * @mapping: address space structure to write
1825 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1826 * @writepage: function called for each page
1827 * @data: data passed to writepage function
1828 *
1829 * If a page is already under I/O, write_cache_pages() skips it, even
1830 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1831 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1832 * and msync() need to guarantee that all the data which was dirty at the time
1833 * the call was made get new I/O started against them. If wbc->sync_mode is
1834 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1835 * existing IO to complete.
1836 *
1837 * To avoid livelocks (when other process dirties new pages), we first tag
1838 * pages which should be written back with TOWRITE tag and only then start
1839 * writing them. For data-integrity sync we have to be careful so that we do
1840 * not miss some pages (e.g., because some other process has cleared TOWRITE
1841 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1842 * by the process clearing the DIRTY tag (and submitting the page for IO).
1843 */
1844int write_cache_pages(struct address_space *mapping,
1845              struct writeback_control *wbc, writepage_t writepage,
1846              void *data)
1847{
1848    int ret = 0;
1849    int done = 0;
1850    struct pagevec pvec;
1851    int nr_pages;
1852    pgoff_t uninitialized_var(writeback_index);
1853    pgoff_t index;
1854    pgoff_t end; /* Inclusive */
1855    pgoff_t done_index;
1856    int cycled;
1857    int range_whole = 0;
1858    int tag;
1859
1860    pagevec_init(&pvec, 0);
1861    if (wbc->range_cyclic) {
1862        writeback_index = mapping->writeback_index; /* prev offset */
1863        index = writeback_index;
1864        if (index == 0)
1865            cycled = 1;
1866        else
1867            cycled = 0;
1868        end = -1;
1869    } else {
1870        index = wbc->range_start >> PAGE_CACHE_SHIFT;
1871        end = wbc->range_end >> PAGE_CACHE_SHIFT;
1872        if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1873            range_whole = 1;
1874        cycled = 1; /* ignore range_cyclic tests */
1875    }
1876    if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1877        tag = PAGECACHE_TAG_TOWRITE;
1878    else
1879        tag = PAGECACHE_TAG_DIRTY;
1880retry:
1881    if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1882        tag_pages_for_writeback(mapping, index, end);
1883    done_index = index;
1884    while (!done && (index <= end)) {
1885        int i;
1886
1887        nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1888                  min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1889        if (nr_pages == 0)
1890            break;
1891
1892        for (i = 0; i < nr_pages; i++) {
1893            struct page *page = pvec.pages[i];
1894
1895            /*
1896             * At this point, the page may be truncated or
1897             * invalidated (changing page->mapping to NULL), or
1898             * even swizzled back from swapper_space to tmpfs file
1899             * mapping. However, page->index will not change
1900             * because we have a reference on the page.
1901             */
1902            if (page->index > end) {
1903                /*
1904                 * can't be range_cyclic (1st pass) because
1905                 * end == -1 in that case.
1906                 */
1907                done = 1;
1908                break;
1909            }
1910
1911            done_index = page->index;
1912
1913            lock_page(page);
1914
1915            /*
1916             * Page truncated or invalidated. We can freely skip it
1917             * then, even for data integrity operations: the page
1918             * has disappeared concurrently, so there could be no
1919             * real expectation of this data interity operation
1920             * even if there is now a new, dirty page at the same
1921             * pagecache address.
1922             */
1923            if (unlikely(page->mapping != mapping)) {
1924continue_unlock:
1925                unlock_page(page);
1926                continue;
1927            }
1928
1929            if (!PageDirty(page)) {
1930                /* someone wrote it for us */
1931                goto continue_unlock;
1932            }
1933
1934            if (PageWriteback(page)) {
1935                if (wbc->sync_mode != WB_SYNC_NONE)
1936                    wait_on_page_writeback(page);
1937                else
1938                    goto continue_unlock;
1939            }
1940
1941            BUG_ON(PageWriteback(page));
1942            if (!clear_page_dirty_for_io(page))
1943                goto continue_unlock;
1944
1945            trace_wbc_writepage(wbc, mapping->backing_dev_info);
1946            ret = (*writepage)(page, wbc, data);
1947            if (unlikely(ret)) {
1948                if (ret == AOP_WRITEPAGE_ACTIVATE) {
1949                    unlock_page(page);
1950                    ret = 0;
1951                } else {
1952                    /*
1953                     * done_index is set past this page,
1954                     * so media errors will not choke
1955                     * background writeout for the entire
1956                     * file. This has consequences for
1957                     * range_cyclic semantics (ie. it may
1958                     * not be suitable for data integrity
1959                     * writeout).
1960                     */
1961                    done_index = page->index + 1;
1962                    done = 1;
1963                    break;
1964                }
1965            }
1966
1967            /*
1968             * We stop writing back only if we are not doing
1969             * integrity sync. In case of integrity sync we have to
1970             * keep going until we have written all the pages
1971             * we tagged for writeback prior to entering this loop.
1972             */
1973            if (--wbc->nr_to_write <= 0 &&
1974                wbc->sync_mode == WB_SYNC_NONE) {
1975                done = 1;
1976                break;
1977            }
1978        }
1979        pagevec_release(&pvec);
1980        cond_resched();
1981    }
1982    if (!cycled && !done) {
1983        /*
1984         * range_cyclic:
1985         * We hit the last page and there is more work to be done: wrap
1986         * back to the start of the file
1987         */
1988        cycled = 1;
1989        index = 0;
1990        end = writeback_index - 1;
1991        goto retry;
1992    }
1993    if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1994        mapping->writeback_index = done_index;
1995
1996    return ret;
1997}
1998EXPORT_SYMBOL(write_cache_pages);
1999
2000/*
2001 * Function used by generic_writepages to call the real writepage
2002 * function and set the mapping flags on error
2003 */
2004static int __writepage(struct page *page, struct writeback_control *wbc,
2005               void *data)
2006{
2007    struct address_space *mapping = data;
2008    int ret = mapping->a_ops->writepage(page, wbc);
2009    mapping_set_error(mapping, ret);
2010    return ret;
2011}
2012
2013/**
2014 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2015 * @mapping: address space structure to write
2016 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2017 *
2018 * This is a library function, which implements the writepages()
2019 * address_space_operation.
2020 */
2021int generic_writepages(struct address_space *mapping,
2022               struct writeback_control *wbc)
2023{
2024    struct blk_plug plug;
2025    int ret;
2026
2027    /* deal with chardevs and other special file */
2028    if (!mapping->a_ops->writepage)
2029        return 0;
2030
2031    blk_start_plug(&plug);
2032    ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2033    blk_finish_plug(&plug);
2034    return ret;
2035}
2036
2037EXPORT_SYMBOL(generic_writepages);
2038
2039int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2040{
2041    int ret;
2042
2043    if (wbc->nr_to_write <= 0)
2044        return 0;
2045    if (mapping->a_ops->writepages)
2046        ret = mapping->a_ops->writepages(mapping, wbc);
2047    else
2048        ret = generic_writepages(mapping, wbc);
2049    return ret;
2050}
2051
2052/**
2053 * write_one_page - write out a single page and optionally wait on I/O
2054 * @page: the page to write
2055 * @wait: if true, wait on writeout
2056 *
2057 * The page must be locked by the caller and will be unlocked upon return.
2058 *
2059 * write_one_page() returns a negative error code if I/O failed.
2060 */
2061int write_one_page(struct page *page, int wait)
2062{
2063    struct address_space *mapping = page->mapping;
2064    int ret = 0;
2065    struct writeback_control wbc = {
2066        .sync_mode = WB_SYNC_ALL,
2067        .nr_to_write = 1,
2068    };
2069
2070    BUG_ON(!PageLocked(page));
2071
2072    if (wait)
2073        wait_on_page_writeback(page);
2074
2075    if (clear_page_dirty_for_io(page)) {
2076        page_cache_get(page);
2077        ret = mapping->a_ops->writepage(page, &wbc);
2078        if (ret == 0 && wait) {
2079            wait_on_page_writeback(page);
2080            if (PageError(page))
2081                ret = -EIO;
2082        }
2083        page_cache_release(page);
2084    } else {
2085        unlock_page(page);
2086    }
2087    return ret;
2088}
2089EXPORT_SYMBOL(write_one_page);
2090
2091/*
2092 * For address_spaces which do not use buffers nor write back.
2093 */
2094int __set_page_dirty_no_writeback(struct page *page)
2095{
2096    if (!PageDirty(page))
2097        return !TestSetPageDirty(page);
2098    return 0;
2099}
2100
2101/*
2102 * Helper function for set_page_dirty family.
2103 * NOTE: This relies on being atomic wrt interrupts.
2104 */
2105void account_page_dirtied(struct page *page, struct address_space *mapping)
2106{
2107    trace_writeback_dirty_page(page, mapping);
2108
2109    if (mapping_cap_account_dirty(mapping)) {
2110        __inc_zone_page_state(page, NR_FILE_DIRTY);
2111        __inc_zone_page_state(page, NR_DIRTIED);
2112        __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
2113        __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2114        task_io_account_write(PAGE_CACHE_SIZE);
2115        current->nr_dirtied++;
2116        this_cpu_inc(bdp_ratelimits);
2117    }
2118}
2119EXPORT_SYMBOL(account_page_dirtied);
2120
2121/*
2122 * Helper function for set_page_writeback family.
2123 *
2124 * The caller must hold mem_cgroup_begin/end_update_page_stat() lock
2125 * while calling this function.
2126 * See test_set_page_writeback for example.
2127 *
2128 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2129 * wrt interrupts.
2130 */
2131void account_page_writeback(struct page *page)
2132{
2133    mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2134    inc_zone_page_state(page, NR_WRITEBACK);
2135}
2136EXPORT_SYMBOL(account_page_writeback);
2137
2138/*
2139 * For address_spaces which do not use buffers. Just tag the page as dirty in
2140 * its radix tree.
2141 *
2142 * This is also used when a single buffer is being dirtied: we want to set the
2143 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2144 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2145 *
2146 * Most callers have locked the page, which pins the address_space in memory.
2147 * But zap_pte_range() does not lock the page, however in that case the
2148 * mapping is pinned by the vma's ->vm_file reference.
2149 *
2150 * We take care to handle the case where the page was truncated from the
2151 * mapping by re-checking page_mapping() inside tree_lock.
2152 */
2153int __set_page_dirty_nobuffers(struct page *page)
2154{
2155    if (!TestSetPageDirty(page)) {
2156        struct address_space *mapping = page_mapping(page);
2157        struct address_space *mapping2;
2158        unsigned long flags;
2159
2160        if (!mapping)
2161            return 1;
2162
2163        spin_lock_irqsave(&mapping->tree_lock, flags);
2164        mapping2 = page_mapping(page);
2165        if (mapping2) { /* Race with truncate? */
2166            BUG_ON(mapping2 != mapping);
2167            WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2168            account_page_dirtied(page, mapping);
2169            radix_tree_tag_set(&mapping->page_tree,
2170                page_index(page), PAGECACHE_TAG_DIRTY);
2171        }
2172        spin_unlock_irqrestore(&mapping->tree_lock, flags);
2173        if (mapping->host) {
2174            /* !PageAnon && !swapper_space */
2175            __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2176        }
2177        return 1;
2178    }
2179    return 0;
2180}
2181EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2182
2183/*
2184 * Call this whenever redirtying a page, to de-account the dirty counters
2185 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2186 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2187 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2188 * control.
2189 */
2190void account_page_redirty(struct page *page)
2191{
2192    struct address_space *mapping = page->mapping;
2193    if (mapping && mapping_cap_account_dirty(mapping)) {
2194        current->nr_dirtied--;
2195        dec_zone_page_state(page, NR_DIRTIED);
2196        dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2197    }
2198}
2199EXPORT_SYMBOL(account_page_redirty);
2200
2201/*
2202 * When a writepage implementation decides that it doesn't want to write this
2203 * page for some reason, it should redirty the locked page via
2204 * redirty_page_for_writepage() and it should then unlock the page and return 0
2205 */
2206int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2207{
2208    wbc->pages_skipped++;
2209    account_page_redirty(page);
2210    return __set_page_dirty_nobuffers(page);
2211}
2212EXPORT_SYMBOL(redirty_page_for_writepage);
2213
2214/*
2215 * Dirty a page.
2216 *
2217 * For pages with a mapping this should be done under the page lock
2218 * for the benefit of asynchronous memory errors who prefer a consistent
2219 * dirty state. This rule can be broken in some special cases,
2220 * but should be better not to.
2221 *
2222 * If the mapping doesn't provide a set_page_dirty a_op, then
2223 * just fall through and assume that it wants buffer_heads.
2224 */
2225int set_page_dirty(struct page *page)
2226{
2227    struct address_space *mapping = page_mapping(page);
2228
2229    if (likely(mapping)) {
2230        int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2231        /*
2232         * readahead/lru_deactivate_page could remain
2233         * PG_readahead/PG_reclaim due to race with end_page_writeback
2234         * About readahead, if the page is written, the flags would be
2235         * reset. So no problem.
2236         * About lru_deactivate_page, if the page is redirty, the flag
2237         * will be reset. So no problem. but if the page is used by readahead
2238         * it will confuse readahead and make it restart the size rampup
2239         * process. But it's a trivial problem.
2240         */
2241        ClearPageReclaim(page);
2242#ifdef CONFIG_BLOCK
2243        if (!spd)
2244            spd = __set_page_dirty_buffers;
2245#endif
2246        return (*spd)(page);
2247    }
2248    if (!PageDirty(page)) {
2249        if (!TestSetPageDirty(page))
2250            return 1;
2251    }
2252    return 0;
2253}
2254EXPORT_SYMBOL(set_page_dirty);
2255
2256/*
2257 * set_page_dirty() is racy if the caller has no reference against
2258 * page->mapping->host, and if the page is unlocked. This is because another
2259 * CPU could truncate the page off the mapping and then free the mapping.
2260 *
2261 * Usually, the page _is_ locked, or the caller is a user-space process which
2262 * holds a reference on the inode by having an open file.
2263 *
2264 * In other cases, the page should be locked before running set_page_dirty().
2265 */
2266int set_page_dirty_lock(struct page *page)
2267{
2268    int ret;
2269
2270    lock_page(page);
2271    ret = set_page_dirty(page);
2272    unlock_page(page);
2273    return ret;
2274}
2275EXPORT_SYMBOL(set_page_dirty_lock);
2276
2277/*
2278 * Clear a page's dirty flag, while caring for dirty memory accounting.
2279 * Returns true if the page was previously dirty.
2280 *
2281 * This is for preparing to put the page under writeout. We leave the page
2282 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2283 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2284 * implementation will run either set_page_writeback() or set_page_dirty(),
2285 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2286 * back into sync.
2287 *
2288 * This incoherency between the page's dirty flag and radix-tree tag is
2289 * unfortunate, but it only exists while the page is locked.
2290 */
2291int clear_page_dirty_for_io(struct page *page)
2292{
2293    struct address_space *mapping = page_mapping(page);
2294
2295    BUG_ON(!PageLocked(page));
2296
2297    if (mapping && mapping_cap_account_dirty(mapping)) {
2298        /*
2299         * Yes, Virginia, this is indeed insane.
2300         *
2301         * We use this sequence to make sure that
2302         * (a) we account for dirty stats properly
2303         * (b) we tell the low-level filesystem to
2304         * mark the whole page dirty if it was
2305         * dirty in a pagetable. Only to then
2306         * (c) clean the page again and return 1 to
2307         * cause the writeback.
2308         *
2309         * This way we avoid all nasty races with the
2310         * dirty bit in multiple places and clearing
2311         * them concurrently from different threads.
2312         *
2313         * Note! Normally the "set_page_dirty(page)"
2314         * has no effect on the actual dirty bit - since
2315         * that will already usually be set. But we
2316         * need the side effects, and it can help us
2317         * avoid races.
2318         *
2319         * We basically use the page "master dirty bit"
2320         * as a serialization point for all the different
2321         * threads doing their things.
2322         */
2323        if (page_mkclean(page))
2324            set_page_dirty(page);
2325        /*
2326         * We carefully synchronise fault handlers against
2327         * installing a dirty pte and marking the page dirty
2328         * at this point. We do this by having them hold the
2329         * page lock at some point after installing their
2330         * pte, but before marking the page dirty.
2331         * Pages are always locked coming in here, so we get
2332         * the desired exclusion. See mm/memory.c:do_wp_page()
2333         * for more comments.
2334         */
2335        if (TestClearPageDirty(page)) {
2336            dec_zone_page_state(page, NR_FILE_DIRTY);
2337            dec_bdi_stat(mapping->backing_dev_info,
2338                    BDI_RECLAIMABLE);
2339            return 1;
2340        }
2341        return 0;
2342    }
2343    return TestClearPageDirty(page);
2344}
2345EXPORT_SYMBOL(clear_page_dirty_for_io);
2346
2347int test_clear_page_writeback(struct page *page)
2348{
2349    struct address_space *mapping = page_mapping(page);
2350    int ret;
2351    bool locked;
2352    unsigned long memcg_flags;
2353
2354    mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2355    if (mapping) {
2356        struct backing_dev_info *bdi = mapping->backing_dev_info;
2357        unsigned long flags;
2358
2359        spin_lock_irqsave(&mapping->tree_lock, flags);
2360        ret = TestClearPageWriteback(page);
2361        if (ret) {
2362            radix_tree_tag_clear(&mapping->page_tree,
2363                        page_index(page),
2364                        PAGECACHE_TAG_WRITEBACK);
2365            if (bdi_cap_account_writeback(bdi)) {
2366                __dec_bdi_stat(bdi, BDI_WRITEBACK);
2367                __bdi_writeout_inc(bdi);
2368            }
2369        }
2370        spin_unlock_irqrestore(&mapping->tree_lock, flags);
2371    } else {
2372        ret = TestClearPageWriteback(page);
2373    }
2374    if (ret) {
2375        mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2376        dec_zone_page_state(page, NR_WRITEBACK);
2377        inc_zone_page_state(page, NR_WRITTEN);
2378    }
2379    mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2380    return ret;
2381}
2382
2383int __test_set_page_writeback(struct page *page, bool keep_write)
2384{
2385    struct address_space *mapping = page_mapping(page);
2386    int ret;
2387    bool locked;
2388    unsigned long memcg_flags;
2389
2390    mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2391    if (mapping) {
2392        struct backing_dev_info *bdi = mapping->backing_dev_info;
2393        unsigned long flags;
2394
2395        spin_lock_irqsave(&mapping->tree_lock, flags);
2396        ret = TestSetPageWriteback(page);
2397        if (!ret) {
2398            radix_tree_tag_set(&mapping->page_tree,
2399                        page_index(page),
2400                        PAGECACHE_TAG_WRITEBACK);
2401            if (bdi_cap_account_writeback(bdi))
2402                __inc_bdi_stat(bdi, BDI_WRITEBACK);
2403        }
2404        if (!PageDirty(page))
2405            radix_tree_tag_clear(&mapping->page_tree,
2406                        page_index(page),
2407                        PAGECACHE_TAG_DIRTY);
2408        if (!keep_write)
2409            radix_tree_tag_clear(&mapping->page_tree,
2410                        page_index(page),
2411                        PAGECACHE_TAG_TOWRITE);
2412        spin_unlock_irqrestore(&mapping->tree_lock, flags);
2413    } else {
2414        ret = TestSetPageWriteback(page);
2415    }
2416    if (!ret)
2417        account_page_writeback(page);
2418    mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2419    return ret;
2420
2421}
2422EXPORT_SYMBOL(__test_set_page_writeback);
2423
2424/*
2425 * Return true if any of the pages in the mapping are marked with the
2426 * passed tag.
2427 */
2428int mapping_tagged(struct address_space *mapping, int tag)
2429{
2430    return radix_tree_tagged(&mapping->page_tree, tag);
2431}
2432EXPORT_SYMBOL(mapping_tagged);
2433
2434/**
2435 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2436 * @page: The page to wait on.
2437 *
2438 * This function determines if the given page is related to a backing device
2439 * that requires page contents to be held stable during writeback. If so, then
2440 * it will wait for any pending writeback to complete.
2441 */
2442void wait_for_stable_page(struct page *page)
2443{
2444    struct address_space *mapping = page_mapping(page);
2445    struct backing_dev_info *bdi = mapping->backing_dev_info;
2446
2447    if (!bdi_cap_stable_pages_required(bdi))
2448        return;
2449
2450    wait_on_page_writeback(page);
2451}
2452EXPORT_SYMBOL_GPL(wait_for_stable_page);
2453

Archive Download this file



interactive