Loading...
1/*
2 * linux/mm/vmstat.c
3 *
4 * Manages VM statistics
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * zoned VM statistics
8 * Copyright (C) 2006 Silicon Graphics, Inc.,
9 * Christoph Lameter <christoph@lameter.com>
10 * Copyright (C) 2008-2014 Christoph Lameter
11 */
12#include <linux/fs.h>
13#include <linux/mm.h>
14#include <linux/err.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/cpu.h>
18#include <linux/cpumask.h>
19#include <linux/vmstat.h>
20#include <linux/proc_fs.h>
21#include <linux/seq_file.h>
22#include <linux/debugfs.h>
23#include <linux/sched.h>
24#include <linux/math64.h>
25#include <linux/writeback.h>
26#include <linux/compaction.h>
27#include <linux/mm_inline.h>
28#include <linux/page_ext.h>
29#include <linux/page_owner.h>
30
31#include "internal.h"
32
33#define NUMA_STATS_THRESHOLD (U16_MAX - 2)
34
35#ifdef CONFIG_NUMA
36int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
37
38/* zero numa counters within a zone */
39static void zero_zone_numa_counters(struct zone *zone)
40{
41 int item, cpu;
42
43 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
44 atomic_long_set(&zone->vm_numa_stat[item], 0);
45 for_each_online_cpu(cpu)
46 per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
47 = 0;
48 }
49}
50
51/* zero numa counters of all the populated zones */
52static void zero_zones_numa_counters(void)
53{
54 struct zone *zone;
55
56 for_each_populated_zone(zone)
57 zero_zone_numa_counters(zone);
58}
59
60/* zero global numa counters */
61static void zero_global_numa_counters(void)
62{
63 int item;
64
65 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
66 atomic_long_set(&vm_numa_stat[item], 0);
67}
68
69static void invalid_numa_statistics(void)
70{
71 zero_zones_numa_counters();
72 zero_global_numa_counters();
73}
74
75static DEFINE_MUTEX(vm_numa_stat_lock);
76
77int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
78 void __user *buffer, size_t *length, loff_t *ppos)
79{
80 int ret, oldval;
81
82 mutex_lock(&vm_numa_stat_lock);
83 if (write)
84 oldval = sysctl_vm_numa_stat;
85 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
86 if (ret || !write)
87 goto out;
88
89 if (oldval == sysctl_vm_numa_stat)
90 goto out;
91 else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
92 static_branch_enable(&vm_numa_stat_key);
93 pr_info("enable numa statistics\n");
94 } else {
95 static_branch_disable(&vm_numa_stat_key);
96 invalid_numa_statistics();
97 pr_info("disable numa statistics, and clear numa counters\n");
98 }
99
100out:
101 mutex_unlock(&vm_numa_stat_lock);
102 return ret;
103}
104#endif
105
106#ifdef CONFIG_VM_EVENT_COUNTERS
107DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
108EXPORT_PER_CPU_SYMBOL(vm_event_states);
109
110static void sum_vm_events(unsigned long *ret)
111{
112 int cpu;
113 int i;
114
115 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
116
117 for_each_online_cpu(cpu) {
118 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
119
120 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
121 ret[i] += this->event[i];
122 }
123}
124
125/*
126 * Accumulate the vm event counters across all CPUs.
127 * The result is unavoidably approximate - it can change
128 * during and after execution of this function.
129*/
130void all_vm_events(unsigned long *ret)
131{
132 get_online_cpus();
133 sum_vm_events(ret);
134 put_online_cpus();
135}
136EXPORT_SYMBOL_GPL(all_vm_events);
137
138/*
139 * Fold the foreign cpu events into our own.
140 *
141 * This is adding to the events on one processor
142 * but keeps the global counts constant.
143 */
144void vm_events_fold_cpu(int cpu)
145{
146 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
147 int i;
148
149 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
150 count_vm_events(i, fold_state->event[i]);
151 fold_state->event[i] = 0;
152 }
153}
154
155#endif /* CONFIG_VM_EVENT_COUNTERS */
156
157/*
158 * Manage combined zone based / global counters
159 *
160 * vm_stat contains the global counters
161 */
162atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
163atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
165EXPORT_SYMBOL(vm_zone_stat);
166EXPORT_SYMBOL(vm_numa_stat);
167EXPORT_SYMBOL(vm_node_stat);
168
169#ifdef CONFIG_SMP
170
171int calculate_pressure_threshold(struct zone *zone)
172{
173 int threshold;
174 int watermark_distance;
175
176 /*
177 * As vmstats are not up to date, there is drift between the estimated
178 * and real values. For high thresholds and a high number of CPUs, it
179 * is possible for the min watermark to be breached while the estimated
180 * value looks fine. The pressure threshold is a reduced value such
181 * that even the maximum amount of drift will not accidentally breach
182 * the min watermark
183 */
184 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
185 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
186
187 /*
188 * Maximum threshold is 125
189 */
190 threshold = min(125, threshold);
191
192 return threshold;
193}
194
195int calculate_normal_threshold(struct zone *zone)
196{
197 int threshold;
198 int mem; /* memory in 128 MB units */
199
200 /*
201 * The threshold scales with the number of processors and the amount
202 * of memory per zone. More memory means that we can defer updates for
203 * longer, more processors could lead to more contention.
204 * fls() is used to have a cheap way of logarithmic scaling.
205 *
206 * Some sample thresholds:
207 *
208 * Threshold Processors (fls) Zonesize fls(mem+1)
209 * ------------------------------------------------------------------
210 * 8 1 1 0.9-1 GB 4
211 * 16 2 2 0.9-1 GB 4
212 * 20 2 2 1-2 GB 5
213 * 24 2 2 2-4 GB 6
214 * 28 2 2 4-8 GB 7
215 * 32 2 2 8-16 GB 8
216 * 4 2 2 <128M 1
217 * 30 4 3 2-4 GB 5
218 * 48 4 3 8-16 GB 8
219 * 32 8 4 1-2 GB 4
220 * 32 8 4 0.9-1GB 4
221 * 10 16 5 <128M 1
222 * 40 16 5 900M 4
223 * 70 64 7 2-4 GB 5
224 * 84 64 7 4-8 GB 6
225 * 108 512 9 4-8 GB 6
226 * 125 1024 10 8-16 GB 8
227 * 125 1024 10 16-32 GB 9
228 */
229
230 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
231
232 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
233
234 /*
235 * Maximum threshold is 125
236 */
237 threshold = min(125, threshold);
238
239 return threshold;
240}
241
242/*
243 * Refresh the thresholds for each zone.
244 */
245void refresh_zone_stat_thresholds(void)
246{
247 struct pglist_data *pgdat;
248 struct zone *zone;
249 int cpu;
250 int threshold;
251
252 /* Zero current pgdat thresholds */
253 for_each_online_pgdat(pgdat) {
254 for_each_online_cpu(cpu) {
255 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
256 }
257 }
258
259 for_each_populated_zone(zone) {
260 struct pglist_data *pgdat = zone->zone_pgdat;
261 unsigned long max_drift, tolerate_drift;
262
263 threshold = calculate_normal_threshold(zone);
264
265 for_each_online_cpu(cpu) {
266 int pgdat_threshold;
267
268 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
269 = threshold;
270
271 /* Base nodestat threshold on the largest populated zone. */
272 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
273 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
274 = max(threshold, pgdat_threshold);
275 }
276
277 /*
278 * Only set percpu_drift_mark if there is a danger that
279 * NR_FREE_PAGES reports the low watermark is ok when in fact
280 * the min watermark could be breached by an allocation
281 */
282 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
283 max_drift = num_online_cpus() * threshold;
284 if (max_drift > tolerate_drift)
285 zone->percpu_drift_mark = high_wmark_pages(zone) +
286 max_drift;
287 }
288}
289
290void set_pgdat_percpu_threshold(pg_data_t *pgdat,
291 int (*calculate_pressure)(struct zone *))
292{
293 struct zone *zone;
294 int cpu;
295 int threshold;
296 int i;
297
298 for (i = 0; i < pgdat->nr_zones; i++) {
299 zone = &pgdat->node_zones[i];
300 if (!zone->percpu_drift_mark)
301 continue;
302
303 threshold = (*calculate_pressure)(zone);
304 for_each_online_cpu(cpu)
305 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
306 = threshold;
307 }
308}
309
310/*
311 * For use when we know that interrupts are disabled,
312 * or when we know that preemption is disabled and that
313 * particular counter cannot be updated from interrupt context.
314 */
315void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
316 long delta)
317{
318 struct per_cpu_pageset __percpu *pcp = zone->pageset;
319 s8 __percpu *p = pcp->vm_stat_diff + item;
320 long x;
321 long t;
322
323 x = delta + __this_cpu_read(*p);
324
325 t = __this_cpu_read(pcp->stat_threshold);
326
327 if (unlikely(x > t || x < -t)) {
328 zone_page_state_add(x, zone, item);
329 x = 0;
330 }
331 __this_cpu_write(*p, x);
332}
333EXPORT_SYMBOL(__mod_zone_page_state);
334
335void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
336 long delta)
337{
338 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
339 s8 __percpu *p = pcp->vm_node_stat_diff + item;
340 long x;
341 long t;
342
343 x = delta + __this_cpu_read(*p);
344
345 t = __this_cpu_read(pcp->stat_threshold);
346
347 if (unlikely(x > t || x < -t)) {
348 node_page_state_add(x, pgdat, item);
349 x = 0;
350 }
351 __this_cpu_write(*p, x);
352}
353EXPORT_SYMBOL(__mod_node_page_state);
354
355/*
356 * Optimized increment and decrement functions.
357 *
358 * These are only for a single page and therefore can take a struct page *
359 * argument instead of struct zone *. This allows the inclusion of the code
360 * generated for page_zone(page) into the optimized functions.
361 *
362 * No overflow check is necessary and therefore the differential can be
363 * incremented or decremented in place which may allow the compilers to
364 * generate better code.
365 * The increment or decrement is known and therefore one boundary check can
366 * be omitted.
367 *
368 * NOTE: These functions are very performance sensitive. Change only
369 * with care.
370 *
371 * Some processors have inc/dec instructions that are atomic vs an interrupt.
372 * However, the code must first determine the differential location in a zone
373 * based on the processor number and then inc/dec the counter. There is no
374 * guarantee without disabling preemption that the processor will not change
375 * in between and therefore the atomicity vs. interrupt cannot be exploited
376 * in a useful way here.
377 */
378void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
379{
380 struct per_cpu_pageset __percpu *pcp = zone->pageset;
381 s8 __percpu *p = pcp->vm_stat_diff + item;
382 s8 v, t;
383
384 v = __this_cpu_inc_return(*p);
385 t = __this_cpu_read(pcp->stat_threshold);
386 if (unlikely(v > t)) {
387 s8 overstep = t >> 1;
388
389 zone_page_state_add(v + overstep, zone, item);
390 __this_cpu_write(*p, -overstep);
391 }
392}
393
394void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
395{
396 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
397 s8 __percpu *p = pcp->vm_node_stat_diff + item;
398 s8 v, t;
399
400 v = __this_cpu_inc_return(*p);
401 t = __this_cpu_read(pcp->stat_threshold);
402 if (unlikely(v > t)) {
403 s8 overstep = t >> 1;
404
405 node_page_state_add(v + overstep, pgdat, item);
406 __this_cpu_write(*p, -overstep);
407 }
408}
409
410void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
411{
412 __inc_zone_state(page_zone(page), item);
413}
414EXPORT_SYMBOL(__inc_zone_page_state);
415
416void __inc_node_page_state(struct page *page, enum node_stat_item item)
417{
418 __inc_node_state(page_pgdat(page), item);
419}
420EXPORT_SYMBOL(__inc_node_page_state);
421
422void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
423{
424 struct per_cpu_pageset __percpu *pcp = zone->pageset;
425 s8 __percpu *p = pcp->vm_stat_diff + item;
426 s8 v, t;
427
428 v = __this_cpu_dec_return(*p);
429 t = __this_cpu_read(pcp->stat_threshold);
430 if (unlikely(v < - t)) {
431 s8 overstep = t >> 1;
432
433 zone_page_state_add(v - overstep, zone, item);
434 __this_cpu_write(*p, overstep);
435 }
436}
437
438void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
439{
440 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
441 s8 __percpu *p = pcp->vm_node_stat_diff + item;
442 s8 v, t;
443
444 v = __this_cpu_dec_return(*p);
445 t = __this_cpu_read(pcp->stat_threshold);
446 if (unlikely(v < - t)) {
447 s8 overstep = t >> 1;
448
449 node_page_state_add(v - overstep, pgdat, item);
450 __this_cpu_write(*p, overstep);
451 }
452}
453
454void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
455{
456 __dec_zone_state(page_zone(page), item);
457}
458EXPORT_SYMBOL(__dec_zone_page_state);
459
460void __dec_node_page_state(struct page *page, enum node_stat_item item)
461{
462 __dec_node_state(page_pgdat(page), item);
463}
464EXPORT_SYMBOL(__dec_node_page_state);
465
466#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
467/*
468 * If we have cmpxchg_local support then we do not need to incur the overhead
469 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
470 *
471 * mod_state() modifies the zone counter state through atomic per cpu
472 * operations.
473 *
474 * Overstep mode specifies how overstep should handled:
475 * 0 No overstepping
476 * 1 Overstepping half of threshold
477 * -1 Overstepping minus half of threshold
478*/
479static inline void mod_zone_state(struct zone *zone,
480 enum zone_stat_item item, long delta, int overstep_mode)
481{
482 struct per_cpu_pageset __percpu *pcp = zone->pageset;
483 s8 __percpu *p = pcp->vm_stat_diff + item;
484 long o, n, t, z;
485
486 do {
487 z = 0; /* overflow to zone counters */
488
489 /*
490 * The fetching of the stat_threshold is racy. We may apply
491 * a counter threshold to the wrong the cpu if we get
492 * rescheduled while executing here. However, the next
493 * counter update will apply the threshold again and
494 * therefore bring the counter under the threshold again.
495 *
496 * Most of the time the thresholds are the same anyways
497 * for all cpus in a zone.
498 */
499 t = this_cpu_read(pcp->stat_threshold);
500
501 o = this_cpu_read(*p);
502 n = delta + o;
503
504 if (n > t || n < -t) {
505 int os = overstep_mode * (t >> 1) ;
506
507 /* Overflow must be added to zone counters */
508 z = n + os;
509 n = -os;
510 }
511 } while (this_cpu_cmpxchg(*p, o, n) != o);
512
513 if (z)
514 zone_page_state_add(z, zone, item);
515}
516
517void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
518 long delta)
519{
520 mod_zone_state(zone, item, delta, 0);
521}
522EXPORT_SYMBOL(mod_zone_page_state);
523
524void inc_zone_page_state(struct page *page, enum zone_stat_item item)
525{
526 mod_zone_state(page_zone(page), item, 1, 1);
527}
528EXPORT_SYMBOL(inc_zone_page_state);
529
530void dec_zone_page_state(struct page *page, enum zone_stat_item item)
531{
532 mod_zone_state(page_zone(page), item, -1, -1);
533}
534EXPORT_SYMBOL(dec_zone_page_state);
535
536static inline void mod_node_state(struct pglist_data *pgdat,
537 enum node_stat_item item, int delta, int overstep_mode)
538{
539 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
540 s8 __percpu *p = pcp->vm_node_stat_diff + item;
541 long o, n, t, z;
542
543 do {
544 z = 0; /* overflow to node counters */
545
546 /*
547 * The fetching of the stat_threshold is racy. We may apply
548 * a counter threshold to the wrong the cpu if we get
549 * rescheduled while executing here. However, the next
550 * counter update will apply the threshold again and
551 * therefore bring the counter under the threshold again.
552 *
553 * Most of the time the thresholds are the same anyways
554 * for all cpus in a node.
555 */
556 t = this_cpu_read(pcp->stat_threshold);
557
558 o = this_cpu_read(*p);
559 n = delta + o;
560
561 if (n > t || n < -t) {
562 int os = overstep_mode * (t >> 1) ;
563
564 /* Overflow must be added to node counters */
565 z = n + os;
566 n = -os;
567 }
568 } while (this_cpu_cmpxchg(*p, o, n) != o);
569
570 if (z)
571 node_page_state_add(z, pgdat, item);
572}
573
574void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
575 long delta)
576{
577 mod_node_state(pgdat, item, delta, 0);
578}
579EXPORT_SYMBOL(mod_node_page_state);
580
581void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
582{
583 mod_node_state(pgdat, item, 1, 1);
584}
585
586void inc_node_page_state(struct page *page, enum node_stat_item item)
587{
588 mod_node_state(page_pgdat(page), item, 1, 1);
589}
590EXPORT_SYMBOL(inc_node_page_state);
591
592void dec_node_page_state(struct page *page, enum node_stat_item item)
593{
594 mod_node_state(page_pgdat(page), item, -1, -1);
595}
596EXPORT_SYMBOL(dec_node_page_state);
597#else
598/*
599 * Use interrupt disable to serialize counter updates
600 */
601void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
602 long delta)
603{
604 unsigned long flags;
605
606 local_irq_save(flags);
607 __mod_zone_page_state(zone, item, delta);
608 local_irq_restore(flags);
609}
610EXPORT_SYMBOL(mod_zone_page_state);
611
612void inc_zone_page_state(struct page *page, enum zone_stat_item item)
613{
614 unsigned long flags;
615 struct zone *zone;
616
617 zone = page_zone(page);
618 local_irq_save(flags);
619 __inc_zone_state(zone, item);
620 local_irq_restore(flags);
621}
622EXPORT_SYMBOL(inc_zone_page_state);
623
624void dec_zone_page_state(struct page *page, enum zone_stat_item item)
625{
626 unsigned long flags;
627
628 local_irq_save(flags);
629 __dec_zone_page_state(page, item);
630 local_irq_restore(flags);
631}
632EXPORT_SYMBOL(dec_zone_page_state);
633
634void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
635{
636 unsigned long flags;
637
638 local_irq_save(flags);
639 __inc_node_state(pgdat, item);
640 local_irq_restore(flags);
641}
642EXPORT_SYMBOL(inc_node_state);
643
644void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
645 long delta)
646{
647 unsigned long flags;
648
649 local_irq_save(flags);
650 __mod_node_page_state(pgdat, item, delta);
651 local_irq_restore(flags);
652}
653EXPORT_SYMBOL(mod_node_page_state);
654
655void inc_node_page_state(struct page *page, enum node_stat_item item)
656{
657 unsigned long flags;
658 struct pglist_data *pgdat;
659
660 pgdat = page_pgdat(page);
661 local_irq_save(flags);
662 __inc_node_state(pgdat, item);
663 local_irq_restore(flags);
664}
665EXPORT_SYMBOL(inc_node_page_state);
666
667void dec_node_page_state(struct page *page, enum node_stat_item item)
668{
669 unsigned long flags;
670
671 local_irq_save(flags);
672 __dec_node_page_state(page, item);
673 local_irq_restore(flags);
674}
675EXPORT_SYMBOL(dec_node_page_state);
676#endif
677
678/*
679 * Fold a differential into the global counters.
680 * Returns the number of counters updated.
681 */
682#ifdef CONFIG_NUMA
683static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
684{
685 int i;
686 int changes = 0;
687
688 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
689 if (zone_diff[i]) {
690 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
691 changes++;
692 }
693
694 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
695 if (numa_diff[i]) {
696 atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
697 changes++;
698 }
699
700 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
701 if (node_diff[i]) {
702 atomic_long_add(node_diff[i], &vm_node_stat[i]);
703 changes++;
704 }
705 return changes;
706}
707#else
708static int fold_diff(int *zone_diff, int *node_diff)
709{
710 int i;
711 int changes = 0;
712
713 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
714 if (zone_diff[i]) {
715 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
716 changes++;
717 }
718
719 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
720 if (node_diff[i]) {
721 atomic_long_add(node_diff[i], &vm_node_stat[i]);
722 changes++;
723 }
724 return changes;
725}
726#endif /* CONFIG_NUMA */
727
728/*
729 * Update the zone counters for the current cpu.
730 *
731 * Note that refresh_cpu_vm_stats strives to only access
732 * node local memory. The per cpu pagesets on remote zones are placed
733 * in the memory local to the processor using that pageset. So the
734 * loop over all zones will access a series of cachelines local to
735 * the processor.
736 *
737 * The call to zone_page_state_add updates the cachelines with the
738 * statistics in the remote zone struct as well as the global cachelines
739 * with the global counters. These could cause remote node cache line
740 * bouncing and will have to be only done when necessary.
741 *
742 * The function returns the number of global counters updated.
743 */
744static int refresh_cpu_vm_stats(bool do_pagesets)
745{
746 struct pglist_data *pgdat;
747 struct zone *zone;
748 int i;
749 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
750#ifdef CONFIG_NUMA
751 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
752#endif
753 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
754 int changes = 0;
755
756 for_each_populated_zone(zone) {
757 struct per_cpu_pageset __percpu *p = zone->pageset;
758
759 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
760 int v;
761
762 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
763 if (v) {
764
765 atomic_long_add(v, &zone->vm_stat[i]);
766 global_zone_diff[i] += v;
767#ifdef CONFIG_NUMA
768 /* 3 seconds idle till flush */
769 __this_cpu_write(p->expire, 3);
770#endif
771 }
772 }
773#ifdef CONFIG_NUMA
774 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
775 int v;
776
777 v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
778 if (v) {
779
780 atomic_long_add(v, &zone->vm_numa_stat[i]);
781 global_numa_diff[i] += v;
782 __this_cpu_write(p->expire, 3);
783 }
784 }
785
786 if (do_pagesets) {
787 cond_resched();
788 /*
789 * Deal with draining the remote pageset of this
790 * processor
791 *
792 * Check if there are pages remaining in this pageset
793 * if not then there is nothing to expire.
794 */
795 if (!__this_cpu_read(p->expire) ||
796 !__this_cpu_read(p->pcp.count))
797 continue;
798
799 /*
800 * We never drain zones local to this processor.
801 */
802 if (zone_to_nid(zone) == numa_node_id()) {
803 __this_cpu_write(p->expire, 0);
804 continue;
805 }
806
807 if (__this_cpu_dec_return(p->expire))
808 continue;
809
810 if (__this_cpu_read(p->pcp.count)) {
811 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
812 changes++;
813 }
814 }
815#endif
816 }
817
818 for_each_online_pgdat(pgdat) {
819 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
820
821 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
822 int v;
823
824 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
825 if (v) {
826 atomic_long_add(v, &pgdat->vm_stat[i]);
827 global_node_diff[i] += v;
828 }
829 }
830 }
831
832#ifdef CONFIG_NUMA
833 changes += fold_diff(global_zone_diff, global_numa_diff,
834 global_node_diff);
835#else
836 changes += fold_diff(global_zone_diff, global_node_diff);
837#endif
838 return changes;
839}
840
841/*
842 * Fold the data for an offline cpu into the global array.
843 * There cannot be any access by the offline cpu and therefore
844 * synchronization is simplified.
845 */
846void cpu_vm_stats_fold(int cpu)
847{
848 struct pglist_data *pgdat;
849 struct zone *zone;
850 int i;
851 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
852#ifdef CONFIG_NUMA
853 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
854#endif
855 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
856
857 for_each_populated_zone(zone) {
858 struct per_cpu_pageset *p;
859
860 p = per_cpu_ptr(zone->pageset, cpu);
861
862 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
863 if (p->vm_stat_diff[i]) {
864 int v;
865
866 v = p->vm_stat_diff[i];
867 p->vm_stat_diff[i] = 0;
868 atomic_long_add(v, &zone->vm_stat[i]);
869 global_zone_diff[i] += v;
870 }
871
872#ifdef CONFIG_NUMA
873 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
874 if (p->vm_numa_stat_diff[i]) {
875 int v;
876
877 v = p->vm_numa_stat_diff[i];
878 p->vm_numa_stat_diff[i] = 0;
879 atomic_long_add(v, &zone->vm_numa_stat[i]);
880 global_numa_diff[i] += v;
881 }
882#endif
883 }
884
885 for_each_online_pgdat(pgdat) {
886 struct per_cpu_nodestat *p;
887
888 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
889
890 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
891 if (p->vm_node_stat_diff[i]) {
892 int v;
893
894 v = p->vm_node_stat_diff[i];
895 p->vm_node_stat_diff[i] = 0;
896 atomic_long_add(v, &pgdat->vm_stat[i]);
897 global_node_diff[i] += v;
898 }
899 }
900
901#ifdef CONFIG_NUMA
902 fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
903#else
904 fold_diff(global_zone_diff, global_node_diff);
905#endif
906}
907
908/*
909 * this is only called if !populated_zone(zone), which implies no other users of
910 * pset->vm_stat_diff[] exsist.
911 */
912void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
913{
914 int i;
915
916 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
917 if (pset->vm_stat_diff[i]) {
918 int v = pset->vm_stat_diff[i];
919 pset->vm_stat_diff[i] = 0;
920 atomic_long_add(v, &zone->vm_stat[i]);
921 atomic_long_add(v, &vm_zone_stat[i]);
922 }
923
924#ifdef CONFIG_NUMA
925 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
926 if (pset->vm_numa_stat_diff[i]) {
927 int v = pset->vm_numa_stat_diff[i];
928
929 pset->vm_numa_stat_diff[i] = 0;
930 atomic_long_add(v, &zone->vm_numa_stat[i]);
931 atomic_long_add(v, &vm_numa_stat[i]);
932 }
933#endif
934}
935#endif
936
937#ifdef CONFIG_NUMA
938void __inc_numa_state(struct zone *zone,
939 enum numa_stat_item item)
940{
941 struct per_cpu_pageset __percpu *pcp = zone->pageset;
942 u16 __percpu *p = pcp->vm_numa_stat_diff + item;
943 u16 v;
944
945 v = __this_cpu_inc_return(*p);
946
947 if (unlikely(v > NUMA_STATS_THRESHOLD)) {
948 zone_numa_state_add(v, zone, item);
949 __this_cpu_write(*p, 0);
950 }
951}
952
953/*
954 * Determine the per node value of a stat item. This function
955 * is called frequently in a NUMA machine, so try to be as
956 * frugal as possible.
957 */
958unsigned long sum_zone_node_page_state(int node,
959 enum zone_stat_item item)
960{
961 struct zone *zones = NODE_DATA(node)->node_zones;
962 int i;
963 unsigned long count = 0;
964
965 for (i = 0; i < MAX_NR_ZONES; i++)
966 count += zone_page_state(zones + i, item);
967
968 return count;
969}
970
971/*
972 * Determine the per node value of a numa stat item. To avoid deviation,
973 * the per cpu stat number in vm_numa_stat_diff[] is also included.
974 */
975unsigned long sum_zone_numa_state(int node,
976 enum numa_stat_item item)
977{
978 struct zone *zones = NODE_DATA(node)->node_zones;
979 int i;
980 unsigned long count = 0;
981
982 for (i = 0; i < MAX_NR_ZONES; i++)
983 count += zone_numa_state_snapshot(zones + i, item);
984
985 return count;
986}
987
988/*
989 * Determine the per node value of a stat item.
990 */
991unsigned long node_page_state(struct pglist_data *pgdat,
992 enum node_stat_item item)
993{
994 long x = atomic_long_read(&pgdat->vm_stat[item]);
995#ifdef CONFIG_SMP
996 if (x < 0)
997 x = 0;
998#endif
999 return x;
1000}
1001#endif
1002
1003#ifdef CONFIG_COMPACTION
1004
1005struct contig_page_info {
1006 unsigned long free_pages;
1007 unsigned long free_blocks_total;
1008 unsigned long free_blocks_suitable;
1009};
1010
1011/*
1012 * Calculate the number of free pages in a zone, how many contiguous
1013 * pages are free and how many are large enough to satisfy an allocation of
1014 * the target size. Note that this function makes no attempt to estimate
1015 * how many suitable free blocks there *might* be if MOVABLE pages were
1016 * migrated. Calculating that is possible, but expensive and can be
1017 * figured out from userspace
1018 */
1019static void fill_contig_page_info(struct zone *zone,
1020 unsigned int suitable_order,
1021 struct contig_page_info *info)
1022{
1023 unsigned int order;
1024
1025 info->free_pages = 0;
1026 info->free_blocks_total = 0;
1027 info->free_blocks_suitable = 0;
1028
1029 for (order = 0; order < MAX_ORDER; order++) {
1030 unsigned long blocks;
1031
1032 /* Count number of free blocks */
1033 blocks = zone->free_area[order].nr_free;
1034 info->free_blocks_total += blocks;
1035
1036 /* Count free base pages */
1037 info->free_pages += blocks << order;
1038
1039 /* Count the suitable free blocks */
1040 if (order >= suitable_order)
1041 info->free_blocks_suitable += blocks <<
1042 (order - suitable_order);
1043 }
1044}
1045
1046/*
1047 * A fragmentation index only makes sense if an allocation of a requested
1048 * size would fail. If that is true, the fragmentation index indicates
1049 * whether external fragmentation or a lack of memory was the problem.
1050 * The value can be used to determine if page reclaim or compaction
1051 * should be used
1052 */
1053static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1054{
1055 unsigned long requested = 1UL << order;
1056
1057 if (WARN_ON_ONCE(order >= MAX_ORDER))
1058 return 0;
1059
1060 if (!info->free_blocks_total)
1061 return 0;
1062
1063 /* Fragmentation index only makes sense when a request would fail */
1064 if (info->free_blocks_suitable)
1065 return -1000;
1066
1067 /*
1068 * Index is between 0 and 1 so return within 3 decimal places
1069 *
1070 * 0 => allocation would fail due to lack of memory
1071 * 1 => allocation would fail due to fragmentation
1072 */
1073 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1074}
1075
1076/* Same as __fragmentation index but allocs contig_page_info on stack */
1077int fragmentation_index(struct zone *zone, unsigned int order)
1078{
1079 struct contig_page_info info;
1080
1081 fill_contig_page_info(zone, order, &info);
1082 return __fragmentation_index(order, &info);
1083}
1084#endif
1085
1086#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
1087#ifdef CONFIG_ZONE_DMA
1088#define TEXT_FOR_DMA(xx) xx "_dma",
1089#else
1090#define TEXT_FOR_DMA(xx)
1091#endif
1092
1093#ifdef CONFIG_ZONE_DMA32
1094#define TEXT_FOR_DMA32(xx) xx "_dma32",
1095#else
1096#define TEXT_FOR_DMA32(xx)
1097#endif
1098
1099#ifdef CONFIG_HIGHMEM
1100#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1101#else
1102#define TEXT_FOR_HIGHMEM(xx)
1103#endif
1104
1105#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1106 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1107
1108const char * const vmstat_text[] = {
1109 /* enum zone_stat_item countes */
1110 "nr_free_pages",
1111 "nr_zone_inactive_anon",
1112 "nr_zone_active_anon",
1113 "nr_zone_inactive_file",
1114 "nr_zone_active_file",
1115 "nr_zone_unevictable",
1116 "nr_zone_write_pending",
1117 "nr_mlock",
1118 "nr_page_table_pages",
1119 "nr_kernel_stack",
1120 "nr_bounce",
1121#if IS_ENABLED(CONFIG_ZSMALLOC)
1122 "nr_zspages",
1123#endif
1124 "nr_free_cma",
1125
1126 /* enum numa_stat_item counters */
1127#ifdef CONFIG_NUMA
1128 "numa_hit",
1129 "numa_miss",
1130 "numa_foreign",
1131 "numa_interleave",
1132 "numa_local",
1133 "numa_other",
1134#endif
1135
1136 /* Node-based counters */
1137 "nr_inactive_anon",
1138 "nr_active_anon",
1139 "nr_inactive_file",
1140 "nr_active_file",
1141 "nr_unevictable",
1142 "nr_slab_reclaimable",
1143 "nr_slab_unreclaimable",
1144 "nr_isolated_anon",
1145 "nr_isolated_file",
1146 "workingset_refault",
1147 "workingset_activate",
1148 "workingset_nodereclaim",
1149 "nr_anon_pages",
1150 "nr_mapped",
1151 "nr_file_pages",
1152 "nr_dirty",
1153 "nr_writeback",
1154 "nr_writeback_temp",
1155 "nr_shmem",
1156 "nr_shmem_hugepages",
1157 "nr_shmem_pmdmapped",
1158 "nr_anon_transparent_hugepages",
1159 "nr_unstable",
1160 "nr_vmscan_write",
1161 "nr_vmscan_immediate_reclaim",
1162 "nr_dirtied",
1163 "nr_written",
1164 "", /* nr_indirectly_reclaimable */
1165
1166 /* enum writeback_stat_item counters */
1167 "nr_dirty_threshold",
1168 "nr_dirty_background_threshold",
1169
1170#ifdef CONFIG_VM_EVENT_COUNTERS
1171 /* enum vm_event_item counters */
1172 "pgpgin",
1173 "pgpgout",
1174 "pswpin",
1175 "pswpout",
1176
1177 TEXTS_FOR_ZONES("pgalloc")
1178 TEXTS_FOR_ZONES("allocstall")
1179 TEXTS_FOR_ZONES("pgskip")
1180
1181 "pgfree",
1182 "pgactivate",
1183 "pgdeactivate",
1184 "pglazyfree",
1185
1186 "pgfault",
1187 "pgmajfault",
1188 "pglazyfreed",
1189
1190 "pgrefill",
1191 "pgsteal_kswapd",
1192 "pgsteal_direct",
1193 "pgscan_kswapd",
1194 "pgscan_direct",
1195 "pgscan_direct_throttle",
1196
1197#ifdef CONFIG_NUMA
1198 "zone_reclaim_failed",
1199#endif
1200 "pginodesteal",
1201 "slabs_scanned",
1202 "kswapd_inodesteal",
1203 "kswapd_low_wmark_hit_quickly",
1204 "kswapd_high_wmark_hit_quickly",
1205 "pageoutrun",
1206
1207 "pgrotated",
1208
1209 "drop_pagecache",
1210 "drop_slab",
1211 "oom_kill",
1212
1213#ifdef CONFIG_NUMA_BALANCING
1214 "numa_pte_updates",
1215 "numa_huge_pte_updates",
1216 "numa_hint_faults",
1217 "numa_hint_faults_local",
1218 "numa_pages_migrated",
1219#endif
1220#ifdef CONFIG_MIGRATION
1221 "pgmigrate_success",
1222 "pgmigrate_fail",
1223#endif
1224#ifdef CONFIG_COMPACTION
1225 "compact_migrate_scanned",
1226 "compact_free_scanned",
1227 "compact_isolated",
1228 "compact_stall",
1229 "compact_fail",
1230 "compact_success",
1231 "compact_daemon_wake",
1232 "compact_daemon_migrate_scanned",
1233 "compact_daemon_free_scanned",
1234#endif
1235
1236#ifdef CONFIG_HUGETLB_PAGE
1237 "htlb_buddy_alloc_success",
1238 "htlb_buddy_alloc_fail",
1239#endif
1240 "unevictable_pgs_culled",
1241 "unevictable_pgs_scanned",
1242 "unevictable_pgs_rescued",
1243 "unevictable_pgs_mlocked",
1244 "unevictable_pgs_munlocked",
1245 "unevictable_pgs_cleared",
1246 "unevictable_pgs_stranded",
1247
1248#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1249 "thp_fault_alloc",
1250 "thp_fault_fallback",
1251 "thp_collapse_alloc",
1252 "thp_collapse_alloc_failed",
1253 "thp_file_alloc",
1254 "thp_file_mapped",
1255 "thp_split_page",
1256 "thp_split_page_failed",
1257 "thp_deferred_split_page",
1258 "thp_split_pmd",
1259#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1260 "thp_split_pud",
1261#endif
1262 "thp_zero_page_alloc",
1263 "thp_zero_page_alloc_failed",
1264 "thp_swpout",
1265 "thp_swpout_fallback",
1266#endif
1267#ifdef CONFIG_MEMORY_BALLOON
1268 "balloon_inflate",
1269 "balloon_deflate",
1270#ifdef CONFIG_BALLOON_COMPACTION
1271 "balloon_migrate",
1272#endif
1273#endif /* CONFIG_MEMORY_BALLOON */
1274#ifdef CONFIG_DEBUG_TLBFLUSH
1275#ifdef CONFIG_SMP
1276 "nr_tlb_remote_flush",
1277 "nr_tlb_remote_flush_received",
1278#endif /* CONFIG_SMP */
1279 "nr_tlb_local_flush_all",
1280 "nr_tlb_local_flush_one",
1281#endif /* CONFIG_DEBUG_TLBFLUSH */
1282
1283#ifdef CONFIG_DEBUG_VM_VMACACHE
1284 "vmacache_find_calls",
1285 "vmacache_find_hits",
1286 "vmacache_full_flushes",
1287#endif
1288#ifdef CONFIG_SWAP
1289 "swap_ra",
1290 "swap_ra_hit",
1291#endif
1292#endif /* CONFIG_VM_EVENTS_COUNTERS */
1293};
1294#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
1295
1296#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1297 defined(CONFIG_PROC_FS)
1298static void *frag_start(struct seq_file *m, loff_t *pos)
1299{
1300 pg_data_t *pgdat;
1301 loff_t node = *pos;
1302
1303 for (pgdat = first_online_pgdat();
1304 pgdat && node;
1305 pgdat = next_online_pgdat(pgdat))
1306 --node;
1307
1308 return pgdat;
1309}
1310
1311static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1312{
1313 pg_data_t *pgdat = (pg_data_t *)arg;
1314
1315 (*pos)++;
1316 return next_online_pgdat(pgdat);
1317}
1318
1319static void frag_stop(struct seq_file *m, void *arg)
1320{
1321}
1322
1323/*
1324 * Walk zones in a node and print using a callback.
1325 * If @assert_populated is true, only use callback for zones that are populated.
1326 */
1327static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1328 bool assert_populated, bool nolock,
1329 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1330{
1331 struct zone *zone;
1332 struct zone *node_zones = pgdat->node_zones;
1333 unsigned long flags;
1334
1335 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1336 if (assert_populated && !populated_zone(zone))
1337 continue;
1338
1339 if (!nolock)
1340 spin_lock_irqsave(&zone->lock, flags);
1341 print(m, pgdat, zone);
1342 if (!nolock)
1343 spin_unlock_irqrestore(&zone->lock, flags);
1344 }
1345}
1346#endif
1347
1348#ifdef CONFIG_PROC_FS
1349static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1350 struct zone *zone)
1351{
1352 int order;
1353
1354 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1355 for (order = 0; order < MAX_ORDER; ++order)
1356 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1357 seq_putc(m, '\n');
1358}
1359
1360/*
1361 * This walks the free areas for each zone.
1362 */
1363static int frag_show(struct seq_file *m, void *arg)
1364{
1365 pg_data_t *pgdat = (pg_data_t *)arg;
1366 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1367 return 0;
1368}
1369
1370static void pagetypeinfo_showfree_print(struct seq_file *m,
1371 pg_data_t *pgdat, struct zone *zone)
1372{
1373 int order, mtype;
1374
1375 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1376 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1377 pgdat->node_id,
1378 zone->name,
1379 migratetype_names[mtype]);
1380 for (order = 0; order < MAX_ORDER; ++order) {
1381 unsigned long freecount = 0;
1382 struct free_area *area;
1383 struct list_head *curr;
1384
1385 area = &(zone->free_area[order]);
1386
1387 list_for_each(curr, &area->free_list[mtype])
1388 freecount++;
1389 seq_printf(m, "%6lu ", freecount);
1390 }
1391 seq_putc(m, '\n');
1392 }
1393}
1394
1395/* Print out the free pages at each order for each migatetype */
1396static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1397{
1398 int order;
1399 pg_data_t *pgdat = (pg_data_t *)arg;
1400
1401 /* Print header */
1402 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1403 for (order = 0; order < MAX_ORDER; ++order)
1404 seq_printf(m, "%6d ", order);
1405 seq_putc(m, '\n');
1406
1407 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1408
1409 return 0;
1410}
1411
1412static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1413 pg_data_t *pgdat, struct zone *zone)
1414{
1415 int mtype;
1416 unsigned long pfn;
1417 unsigned long start_pfn = zone->zone_start_pfn;
1418 unsigned long end_pfn = zone_end_pfn(zone);
1419 unsigned long count[MIGRATE_TYPES] = { 0, };
1420
1421 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1422 struct page *page;
1423
1424 page = pfn_to_online_page(pfn);
1425 if (!page)
1426 continue;
1427
1428 /* Watch for unexpected holes punched in the memmap */
1429 if (!memmap_valid_within(pfn, page, zone))
1430 continue;
1431
1432 if (page_zone(page) != zone)
1433 continue;
1434
1435 mtype = get_pageblock_migratetype(page);
1436
1437 if (mtype < MIGRATE_TYPES)
1438 count[mtype]++;
1439 }
1440
1441 /* Print counts */
1442 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1443 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1444 seq_printf(m, "%12lu ", count[mtype]);
1445 seq_putc(m, '\n');
1446}
1447
1448/* Print out the number of pageblocks for each migratetype */
1449static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1450{
1451 int mtype;
1452 pg_data_t *pgdat = (pg_data_t *)arg;
1453
1454 seq_printf(m, "\n%-23s", "Number of blocks type ");
1455 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1456 seq_printf(m, "%12s ", migratetype_names[mtype]);
1457 seq_putc(m, '\n');
1458 walk_zones_in_node(m, pgdat, true, false,
1459 pagetypeinfo_showblockcount_print);
1460
1461 return 0;
1462}
1463
1464/*
1465 * Print out the number of pageblocks for each migratetype that contain pages
1466 * of other types. This gives an indication of how well fallbacks are being
1467 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1468 * to determine what is going on
1469 */
1470static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1471{
1472#ifdef CONFIG_PAGE_OWNER
1473 int mtype;
1474
1475 if (!static_branch_unlikely(&page_owner_inited))
1476 return;
1477
1478 drain_all_pages(NULL);
1479
1480 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1481 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1482 seq_printf(m, "%12s ", migratetype_names[mtype]);
1483 seq_putc(m, '\n');
1484
1485 walk_zones_in_node(m, pgdat, true, true,
1486 pagetypeinfo_showmixedcount_print);
1487#endif /* CONFIG_PAGE_OWNER */
1488}
1489
1490/*
1491 * This prints out statistics in relation to grouping pages by mobility.
1492 * It is expensive to collect so do not constantly read the file.
1493 */
1494static int pagetypeinfo_show(struct seq_file *m, void *arg)
1495{
1496 pg_data_t *pgdat = (pg_data_t *)arg;
1497
1498 /* check memoryless node */
1499 if (!node_state(pgdat->node_id, N_MEMORY))
1500 return 0;
1501
1502 seq_printf(m, "Page block order: %d\n", pageblock_order);
1503 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1504 seq_putc(m, '\n');
1505 pagetypeinfo_showfree(m, pgdat);
1506 pagetypeinfo_showblockcount(m, pgdat);
1507 pagetypeinfo_showmixedcount(m, pgdat);
1508
1509 return 0;
1510}
1511
1512static const struct seq_operations fragmentation_op = {
1513 .start = frag_start,
1514 .next = frag_next,
1515 .stop = frag_stop,
1516 .show = frag_show,
1517};
1518
1519static int fragmentation_open(struct inode *inode, struct file *file)
1520{
1521 return seq_open(file, &fragmentation_op);
1522}
1523
1524static const struct file_operations buddyinfo_file_operations = {
1525 .open = fragmentation_open,
1526 .read = seq_read,
1527 .llseek = seq_lseek,
1528 .release = seq_release,
1529};
1530
1531static const struct seq_operations pagetypeinfo_op = {
1532 .start = frag_start,
1533 .next = frag_next,
1534 .stop = frag_stop,
1535 .show = pagetypeinfo_show,
1536};
1537
1538static int pagetypeinfo_open(struct inode *inode, struct file *file)
1539{
1540 return seq_open(file, &pagetypeinfo_op);
1541}
1542
1543static const struct file_operations pagetypeinfo_file_operations = {
1544 .open = pagetypeinfo_open,
1545 .read = seq_read,
1546 .llseek = seq_lseek,
1547 .release = seq_release,
1548};
1549
1550static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1551{
1552 int zid;
1553
1554 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1555 struct zone *compare = &pgdat->node_zones[zid];
1556
1557 if (populated_zone(compare))
1558 return zone == compare;
1559 }
1560
1561 return false;
1562}
1563
1564static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1565 struct zone *zone)
1566{
1567 int i;
1568 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1569 if (is_zone_first_populated(pgdat, zone)) {
1570 seq_printf(m, "\n per-node stats");
1571 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1572 seq_printf(m, "\n %-12s %lu",
1573 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS +
1574 NR_VM_NUMA_STAT_ITEMS],
1575 node_page_state(pgdat, i));
1576 }
1577 }
1578 seq_printf(m,
1579 "\n pages free %lu"
1580 "\n min %lu"
1581 "\n low %lu"
1582 "\n high %lu"
1583 "\n spanned %lu"
1584 "\n present %lu"
1585 "\n managed %lu",
1586 zone_page_state(zone, NR_FREE_PAGES),
1587 min_wmark_pages(zone),
1588 low_wmark_pages(zone),
1589 high_wmark_pages(zone),
1590 zone->spanned_pages,
1591 zone->present_pages,
1592 zone->managed_pages);
1593
1594 seq_printf(m,
1595 "\n protection: (%ld",
1596 zone->lowmem_reserve[0]);
1597 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1598 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1599 seq_putc(m, ')');
1600
1601 /* If unpopulated, no other information is useful */
1602 if (!populated_zone(zone)) {
1603 seq_putc(m, '\n');
1604 return;
1605 }
1606
1607 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1608 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1609 zone_page_state(zone, i));
1610
1611#ifdef CONFIG_NUMA
1612 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1613 seq_printf(m, "\n %-12s %lu",
1614 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
1615 zone_numa_state_snapshot(zone, i));
1616#endif
1617
1618 seq_printf(m, "\n pagesets");
1619 for_each_online_cpu(i) {
1620 struct per_cpu_pageset *pageset;
1621
1622 pageset = per_cpu_ptr(zone->pageset, i);
1623 seq_printf(m,
1624 "\n cpu: %i"
1625 "\n count: %i"
1626 "\n high: %i"
1627 "\n batch: %i",
1628 i,
1629 pageset->pcp.count,
1630 pageset->pcp.high,
1631 pageset->pcp.batch);
1632#ifdef CONFIG_SMP
1633 seq_printf(m, "\n vm stats threshold: %d",
1634 pageset->stat_threshold);
1635#endif
1636 }
1637 seq_printf(m,
1638 "\n node_unreclaimable: %u"
1639 "\n start_pfn: %lu",
1640 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1641 zone->zone_start_pfn);
1642 seq_putc(m, '\n');
1643}
1644
1645/*
1646 * Output information about zones in @pgdat. All zones are printed regardless
1647 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1648 * set of all zones and userspace would not be aware of such zones if they are
1649 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1650 */
1651static int zoneinfo_show(struct seq_file *m, void *arg)
1652{
1653 pg_data_t *pgdat = (pg_data_t *)arg;
1654 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1655 return 0;
1656}
1657
1658static const struct seq_operations zoneinfo_op = {
1659 .start = frag_start, /* iterate over all zones. The same as in
1660 * fragmentation. */
1661 .next = frag_next,
1662 .stop = frag_stop,
1663 .show = zoneinfo_show,
1664};
1665
1666static int zoneinfo_open(struct inode *inode, struct file *file)
1667{
1668 return seq_open(file, &zoneinfo_op);
1669}
1670
1671static const struct file_operations zoneinfo_file_operations = {
1672 .open = zoneinfo_open,
1673 .read = seq_read,
1674 .llseek = seq_lseek,
1675 .release = seq_release,
1676};
1677
1678enum writeback_stat_item {
1679 NR_DIRTY_THRESHOLD,
1680 NR_DIRTY_BG_THRESHOLD,
1681 NR_VM_WRITEBACK_STAT_ITEMS,
1682};
1683
1684static void *vmstat_start(struct seq_file *m, loff_t *pos)
1685{
1686 unsigned long *v;
1687 int i, stat_items_size;
1688
1689 if (*pos >= ARRAY_SIZE(vmstat_text))
1690 return NULL;
1691 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1692 NR_VM_NUMA_STAT_ITEMS * sizeof(unsigned long) +
1693 NR_VM_NODE_STAT_ITEMS * sizeof(unsigned long) +
1694 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1695
1696#ifdef CONFIG_VM_EVENT_COUNTERS
1697 stat_items_size += sizeof(struct vm_event_state);
1698#endif
1699
1700 v = kmalloc(stat_items_size, GFP_KERNEL);
1701 m->private = v;
1702 if (!v)
1703 return ERR_PTR(-ENOMEM);
1704 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1705 v[i] = global_zone_page_state(i);
1706 v += NR_VM_ZONE_STAT_ITEMS;
1707
1708#ifdef CONFIG_NUMA
1709 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1710 v[i] = global_numa_state(i);
1711 v += NR_VM_NUMA_STAT_ITEMS;
1712#endif
1713
1714 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1715 v[i] = global_node_page_state(i);
1716 v += NR_VM_NODE_STAT_ITEMS;
1717
1718 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1719 v + NR_DIRTY_THRESHOLD);
1720 v += NR_VM_WRITEBACK_STAT_ITEMS;
1721
1722#ifdef CONFIG_VM_EVENT_COUNTERS
1723 all_vm_events(v);
1724 v[PGPGIN] /= 2; /* sectors -> kbytes */
1725 v[PGPGOUT] /= 2;
1726#endif
1727 return (unsigned long *)m->private + *pos;
1728}
1729
1730static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1731{
1732 (*pos)++;
1733 if (*pos >= ARRAY_SIZE(vmstat_text))
1734 return NULL;
1735 return (unsigned long *)m->private + *pos;
1736}
1737
1738static int vmstat_show(struct seq_file *m, void *arg)
1739{
1740 unsigned long *l = arg;
1741 unsigned long off = l - (unsigned long *)m->private;
1742
1743 /* Skip hidden vmstat items. */
1744 if (*vmstat_text[off] == '\0')
1745 return 0;
1746
1747 seq_puts(m, vmstat_text[off]);
1748 seq_put_decimal_ull(m, " ", *l);
1749 seq_putc(m, '\n');
1750 return 0;
1751}
1752
1753static void vmstat_stop(struct seq_file *m, void *arg)
1754{
1755 kfree(m->private);
1756 m->private = NULL;
1757}
1758
1759static const struct seq_operations vmstat_op = {
1760 .start = vmstat_start,
1761 .next = vmstat_next,
1762 .stop = vmstat_stop,
1763 .show = vmstat_show,
1764};
1765
1766static int vmstat_open(struct inode *inode, struct file *file)
1767{
1768 return seq_open(file, &vmstat_op);
1769}
1770
1771static const struct file_operations vmstat_file_operations = {
1772 .open = vmstat_open,
1773 .read = seq_read,
1774 .llseek = seq_lseek,
1775 .release = seq_release,
1776};
1777#endif /* CONFIG_PROC_FS */
1778
1779#ifdef CONFIG_SMP
1780static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1781int sysctl_stat_interval __read_mostly = HZ;
1782
1783#ifdef CONFIG_PROC_FS
1784static void refresh_vm_stats(struct work_struct *work)
1785{
1786 refresh_cpu_vm_stats(true);
1787}
1788
1789int vmstat_refresh(struct ctl_table *table, int write,
1790 void __user *buffer, size_t *lenp, loff_t *ppos)
1791{
1792 long val;
1793 int err;
1794 int i;
1795
1796 /*
1797 * The regular update, every sysctl_stat_interval, may come later
1798 * than expected: leaving a significant amount in per_cpu buckets.
1799 * This is particularly misleading when checking a quantity of HUGE
1800 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1801 * which can equally be echo'ed to or cat'ted from (by root),
1802 * can be used to update the stats just before reading them.
1803 *
1804 * Oh, and since global_zone_page_state() etc. are so careful to hide
1805 * transiently negative values, report an error here if any of
1806 * the stats is negative, so we know to go looking for imbalance.
1807 */
1808 err = schedule_on_each_cpu(refresh_vm_stats);
1809 if (err)
1810 return err;
1811 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1812 val = atomic_long_read(&vm_zone_stat[i]);
1813 if (val < 0) {
1814 pr_warn("%s: %s %ld\n",
1815 __func__, vmstat_text[i], val);
1816 err = -EINVAL;
1817 }
1818 }
1819#ifdef CONFIG_NUMA
1820 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1821 val = atomic_long_read(&vm_numa_stat[i]);
1822 if (val < 0) {
1823 pr_warn("%s: %s %ld\n",
1824 __func__, vmstat_text[i + NR_VM_ZONE_STAT_ITEMS], val);
1825 err = -EINVAL;
1826 }
1827 }
1828#endif
1829 if (err)
1830 return err;
1831 if (write)
1832 *ppos += *lenp;
1833 else
1834 *lenp = 0;
1835 return 0;
1836}
1837#endif /* CONFIG_PROC_FS */
1838
1839static void vmstat_update(struct work_struct *w)
1840{
1841 if (refresh_cpu_vm_stats(true)) {
1842 /*
1843 * Counters were updated so we expect more updates
1844 * to occur in the future. Keep on running the
1845 * update worker thread.
1846 */
1847 preempt_disable();
1848 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1849 this_cpu_ptr(&vmstat_work),
1850 round_jiffies_relative(sysctl_stat_interval));
1851 preempt_enable();
1852 }
1853}
1854
1855/*
1856 * Switch off vmstat processing and then fold all the remaining differentials
1857 * until the diffs stay at zero. The function is used by NOHZ and can only be
1858 * invoked when tick processing is not active.
1859 */
1860/*
1861 * Check if the diffs for a certain cpu indicate that
1862 * an update is needed.
1863 */
1864static bool need_update(int cpu)
1865{
1866 struct zone *zone;
1867
1868 for_each_populated_zone(zone) {
1869 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1870
1871 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1872#ifdef CONFIG_NUMA
1873 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1874#endif
1875
1876 /*
1877 * The fast way of checking if there are any vmstat diffs.
1878 * This works because the diffs are byte sized items.
1879 */
1880 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1881 return true;
1882#ifdef CONFIG_NUMA
1883 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS))
1884 return true;
1885#endif
1886 }
1887 return false;
1888}
1889
1890/*
1891 * Switch off vmstat processing and then fold all the remaining differentials
1892 * until the diffs stay at zero. The function is used by NOHZ and can only be
1893 * invoked when tick processing is not active.
1894 */
1895void quiet_vmstat(void)
1896{
1897 if (system_state != SYSTEM_RUNNING)
1898 return;
1899
1900 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1901 return;
1902
1903 if (!need_update(smp_processor_id()))
1904 return;
1905
1906 /*
1907 * Just refresh counters and do not care about the pending delayed
1908 * vmstat_update. It doesn't fire that often to matter and canceling
1909 * it would be too expensive from this path.
1910 * vmstat_shepherd will take care about that for us.
1911 */
1912 refresh_cpu_vm_stats(false);
1913}
1914
1915/*
1916 * Shepherd worker thread that checks the
1917 * differentials of processors that have their worker
1918 * threads for vm statistics updates disabled because of
1919 * inactivity.
1920 */
1921static void vmstat_shepherd(struct work_struct *w);
1922
1923static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1924
1925static void vmstat_shepherd(struct work_struct *w)
1926{
1927 int cpu;
1928
1929 get_online_cpus();
1930 /* Check processors whose vmstat worker threads have been disabled */
1931 for_each_online_cpu(cpu) {
1932 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1933
1934 if (!delayed_work_pending(dw) && need_update(cpu))
1935 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1936 }
1937 put_online_cpus();
1938
1939 schedule_delayed_work(&shepherd,
1940 round_jiffies_relative(sysctl_stat_interval));
1941}
1942
1943static void __init start_shepherd_timer(void)
1944{
1945 int cpu;
1946
1947 for_each_possible_cpu(cpu)
1948 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1949 vmstat_update);
1950
1951 schedule_delayed_work(&shepherd,
1952 round_jiffies_relative(sysctl_stat_interval));
1953}
1954
1955static void __init init_cpu_node_state(void)
1956{
1957 int node;
1958
1959 for_each_online_node(node) {
1960 if (cpumask_weight(cpumask_of_node(node)) > 0)
1961 node_set_state(node, N_CPU);
1962 }
1963}
1964
1965static int vmstat_cpu_online(unsigned int cpu)
1966{
1967 refresh_zone_stat_thresholds();
1968 node_set_state(cpu_to_node(cpu), N_CPU);
1969 return 0;
1970}
1971
1972static int vmstat_cpu_down_prep(unsigned int cpu)
1973{
1974 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1975 return 0;
1976}
1977
1978static int vmstat_cpu_dead(unsigned int cpu)
1979{
1980 const struct cpumask *node_cpus;
1981 int node;
1982
1983 node = cpu_to_node(cpu);
1984
1985 refresh_zone_stat_thresholds();
1986 node_cpus = cpumask_of_node(node);
1987 if (cpumask_weight(node_cpus) > 0)
1988 return 0;
1989
1990 node_clear_state(node, N_CPU);
1991 return 0;
1992}
1993
1994#endif
1995
1996struct workqueue_struct *mm_percpu_wq;
1997
1998void __init init_mm_internals(void)
1999{
2000 int ret __maybe_unused;
2001
2002 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2003
2004#ifdef CONFIG_SMP
2005 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2006 NULL, vmstat_cpu_dead);
2007 if (ret < 0)
2008 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2009
2010 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2011 vmstat_cpu_online,
2012 vmstat_cpu_down_prep);
2013 if (ret < 0)
2014 pr_err("vmstat: failed to register 'online' hotplug state\n");
2015
2016 get_online_cpus();
2017 init_cpu_node_state();
2018 put_online_cpus();
2019
2020 start_shepherd_timer();
2021#endif
2022#ifdef CONFIG_PROC_FS
2023 proc_create("buddyinfo", 0444, NULL, &buddyinfo_file_operations);
2024 proc_create("pagetypeinfo", 0444, NULL, &pagetypeinfo_file_operations);
2025 proc_create("vmstat", 0444, NULL, &vmstat_file_operations);
2026 proc_create("zoneinfo", 0444, NULL, &zoneinfo_file_operations);
2027#endif
2028}
2029
2030#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2031
2032/*
2033 * Return an index indicating how much of the available free memory is
2034 * unusable for an allocation of the requested size.
2035 */
2036static int unusable_free_index(unsigned int order,
2037 struct contig_page_info *info)
2038{
2039 /* No free memory is interpreted as all free memory is unusable */
2040 if (info->free_pages == 0)
2041 return 1000;
2042
2043 /*
2044 * Index should be a value between 0 and 1. Return a value to 3
2045 * decimal places.
2046 *
2047 * 0 => no fragmentation
2048 * 1 => high fragmentation
2049 */
2050 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2051
2052}
2053
2054static void unusable_show_print(struct seq_file *m,
2055 pg_data_t *pgdat, struct zone *zone)
2056{
2057 unsigned int order;
2058 int index;
2059 struct contig_page_info info;
2060
2061 seq_printf(m, "Node %d, zone %8s ",
2062 pgdat->node_id,
2063 zone->name);
2064 for (order = 0; order < MAX_ORDER; ++order) {
2065 fill_contig_page_info(zone, order, &info);
2066 index = unusable_free_index(order, &info);
2067 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2068 }
2069
2070 seq_putc(m, '\n');
2071}
2072
2073/*
2074 * Display unusable free space index
2075 *
2076 * The unusable free space index measures how much of the available free
2077 * memory cannot be used to satisfy an allocation of a given size and is a
2078 * value between 0 and 1. The higher the value, the more of free memory is
2079 * unusable and by implication, the worse the external fragmentation is. This
2080 * can be expressed as a percentage by multiplying by 100.
2081 */
2082static int unusable_show(struct seq_file *m, void *arg)
2083{
2084 pg_data_t *pgdat = (pg_data_t *)arg;
2085
2086 /* check memoryless node */
2087 if (!node_state(pgdat->node_id, N_MEMORY))
2088 return 0;
2089
2090 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2091
2092 return 0;
2093}
2094
2095static const struct seq_operations unusable_op = {
2096 .start = frag_start,
2097 .next = frag_next,
2098 .stop = frag_stop,
2099 .show = unusable_show,
2100};
2101
2102static int unusable_open(struct inode *inode, struct file *file)
2103{
2104 return seq_open(file, &unusable_op);
2105}
2106
2107static const struct file_operations unusable_file_ops = {
2108 .open = unusable_open,
2109 .read = seq_read,
2110 .llseek = seq_lseek,
2111 .release = seq_release,
2112};
2113
2114static void extfrag_show_print(struct seq_file *m,
2115 pg_data_t *pgdat, struct zone *zone)
2116{
2117 unsigned int order;
2118 int index;
2119
2120 /* Alloc on stack as interrupts are disabled for zone walk */
2121 struct contig_page_info info;
2122
2123 seq_printf(m, "Node %d, zone %8s ",
2124 pgdat->node_id,
2125 zone->name);
2126 for (order = 0; order < MAX_ORDER; ++order) {
2127 fill_contig_page_info(zone, order, &info);
2128 index = __fragmentation_index(order, &info);
2129 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2130 }
2131
2132 seq_putc(m, '\n');
2133}
2134
2135/*
2136 * Display fragmentation index for orders that allocations would fail for
2137 */
2138static int extfrag_show(struct seq_file *m, void *arg)
2139{
2140 pg_data_t *pgdat = (pg_data_t *)arg;
2141
2142 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2143
2144 return 0;
2145}
2146
2147static const struct seq_operations extfrag_op = {
2148 .start = frag_start,
2149 .next = frag_next,
2150 .stop = frag_stop,
2151 .show = extfrag_show,
2152};
2153
2154static int extfrag_open(struct inode *inode, struct file *file)
2155{
2156 return seq_open(file, &extfrag_op);
2157}
2158
2159static const struct file_operations extfrag_file_ops = {
2160 .open = extfrag_open,
2161 .read = seq_read,
2162 .llseek = seq_lseek,
2163 .release = seq_release,
2164};
2165
2166static int __init extfrag_debug_init(void)
2167{
2168 struct dentry *extfrag_debug_root;
2169
2170 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2171 if (!extfrag_debug_root)
2172 return -ENOMEM;
2173
2174 if (!debugfs_create_file("unusable_index", 0444,
2175 extfrag_debug_root, NULL, &unusable_file_ops))
2176 goto fail;
2177
2178 if (!debugfs_create_file("extfrag_index", 0444,
2179 extfrag_debug_root, NULL, &extfrag_file_ops))
2180 goto fail;
2181
2182 return 0;
2183fail:
2184 debugfs_remove_recursive(extfrag_debug_root);
2185 return -ENOMEM;
2186}
2187
2188module_init(extfrag_debug_init);
2189#endif
1/*
2 * linux/mm/vmstat.c
3 *
4 * Manages VM statistics
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * zoned VM statistics
8 * Copyright (C) 2006 Silicon Graphics, Inc.,
9 * Christoph Lameter <christoph@lameter.com>
10 * Copyright (C) 2008-2014 Christoph Lameter
11 */
12#include <linux/fs.h>
13#include <linux/mm.h>
14#include <linux/err.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/cpu.h>
18#include <linux/cpumask.h>
19#include <linux/vmstat.h>
20#include <linux/proc_fs.h>
21#include <linux/seq_file.h>
22#include <linux/debugfs.h>
23#include <linux/sched.h>
24#include <linux/math64.h>
25#include <linux/writeback.h>
26#include <linux/compaction.h>
27#include <linux/mm_inline.h>
28#include <linux/page_ext.h>
29#include <linux/page_owner.h>
30
31#include "internal.h"
32
33#ifdef CONFIG_VM_EVENT_COUNTERS
34DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
35EXPORT_PER_CPU_SYMBOL(vm_event_states);
36
37static void sum_vm_events(unsigned long *ret)
38{
39 int cpu;
40 int i;
41
42 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
43
44 for_each_online_cpu(cpu) {
45 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
46
47 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
48 ret[i] += this->event[i];
49 }
50}
51
52/*
53 * Accumulate the vm event counters across all CPUs.
54 * The result is unavoidably approximate - it can change
55 * during and after execution of this function.
56*/
57void all_vm_events(unsigned long *ret)
58{
59 get_online_cpus();
60 sum_vm_events(ret);
61 put_online_cpus();
62}
63EXPORT_SYMBOL_GPL(all_vm_events);
64
65/*
66 * Fold the foreign cpu events into our own.
67 *
68 * This is adding to the events on one processor
69 * but keeps the global counts constant.
70 */
71void vm_events_fold_cpu(int cpu)
72{
73 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
74 int i;
75
76 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
77 count_vm_events(i, fold_state->event[i]);
78 fold_state->event[i] = 0;
79 }
80}
81
82#endif /* CONFIG_VM_EVENT_COUNTERS */
83
84/*
85 * Manage combined zone based / global counters
86 *
87 * vm_stat contains the global counters
88 */
89atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
90EXPORT_SYMBOL(vm_stat);
91
92#ifdef CONFIG_SMP
93
94int calculate_pressure_threshold(struct zone *zone)
95{
96 int threshold;
97 int watermark_distance;
98
99 /*
100 * As vmstats are not up to date, there is drift between the estimated
101 * and real values. For high thresholds and a high number of CPUs, it
102 * is possible for the min watermark to be breached while the estimated
103 * value looks fine. The pressure threshold is a reduced value such
104 * that even the maximum amount of drift will not accidentally breach
105 * the min watermark
106 */
107 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
108 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
109
110 /*
111 * Maximum threshold is 125
112 */
113 threshold = min(125, threshold);
114
115 return threshold;
116}
117
118int calculate_normal_threshold(struct zone *zone)
119{
120 int threshold;
121 int mem; /* memory in 128 MB units */
122
123 /*
124 * The threshold scales with the number of processors and the amount
125 * of memory per zone. More memory means that we can defer updates for
126 * longer, more processors could lead to more contention.
127 * fls() is used to have a cheap way of logarithmic scaling.
128 *
129 * Some sample thresholds:
130 *
131 * Threshold Processors (fls) Zonesize fls(mem+1)
132 * ------------------------------------------------------------------
133 * 8 1 1 0.9-1 GB 4
134 * 16 2 2 0.9-1 GB 4
135 * 20 2 2 1-2 GB 5
136 * 24 2 2 2-4 GB 6
137 * 28 2 2 4-8 GB 7
138 * 32 2 2 8-16 GB 8
139 * 4 2 2 <128M 1
140 * 30 4 3 2-4 GB 5
141 * 48 4 3 8-16 GB 8
142 * 32 8 4 1-2 GB 4
143 * 32 8 4 0.9-1GB 4
144 * 10 16 5 <128M 1
145 * 40 16 5 900M 4
146 * 70 64 7 2-4 GB 5
147 * 84 64 7 4-8 GB 6
148 * 108 512 9 4-8 GB 6
149 * 125 1024 10 8-16 GB 8
150 * 125 1024 10 16-32 GB 9
151 */
152
153 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
154
155 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
156
157 /*
158 * Maximum threshold is 125
159 */
160 threshold = min(125, threshold);
161
162 return threshold;
163}
164
165/*
166 * Refresh the thresholds for each zone.
167 */
168void refresh_zone_stat_thresholds(void)
169{
170 struct zone *zone;
171 int cpu;
172 int threshold;
173
174 for_each_populated_zone(zone) {
175 unsigned long max_drift, tolerate_drift;
176
177 threshold = calculate_normal_threshold(zone);
178
179 for_each_online_cpu(cpu)
180 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
181 = threshold;
182
183 /*
184 * Only set percpu_drift_mark if there is a danger that
185 * NR_FREE_PAGES reports the low watermark is ok when in fact
186 * the min watermark could be breached by an allocation
187 */
188 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
189 max_drift = num_online_cpus() * threshold;
190 if (max_drift > tolerate_drift)
191 zone->percpu_drift_mark = high_wmark_pages(zone) +
192 max_drift;
193 }
194}
195
196void set_pgdat_percpu_threshold(pg_data_t *pgdat,
197 int (*calculate_pressure)(struct zone *))
198{
199 struct zone *zone;
200 int cpu;
201 int threshold;
202 int i;
203
204 for (i = 0; i < pgdat->nr_zones; i++) {
205 zone = &pgdat->node_zones[i];
206 if (!zone->percpu_drift_mark)
207 continue;
208
209 threshold = (*calculate_pressure)(zone);
210 for_each_online_cpu(cpu)
211 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
212 = threshold;
213 }
214}
215
216/*
217 * For use when we know that interrupts are disabled,
218 * or when we know that preemption is disabled and that
219 * particular counter cannot be updated from interrupt context.
220 */
221void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
222 long delta)
223{
224 struct per_cpu_pageset __percpu *pcp = zone->pageset;
225 s8 __percpu *p = pcp->vm_stat_diff + item;
226 long x;
227 long t;
228
229 x = delta + __this_cpu_read(*p);
230
231 t = __this_cpu_read(pcp->stat_threshold);
232
233 if (unlikely(x > t || x < -t)) {
234 zone_page_state_add(x, zone, item);
235 x = 0;
236 }
237 __this_cpu_write(*p, x);
238}
239EXPORT_SYMBOL(__mod_zone_page_state);
240
241/*
242 * Optimized increment and decrement functions.
243 *
244 * These are only for a single page and therefore can take a struct page *
245 * argument instead of struct zone *. This allows the inclusion of the code
246 * generated for page_zone(page) into the optimized functions.
247 *
248 * No overflow check is necessary and therefore the differential can be
249 * incremented or decremented in place which may allow the compilers to
250 * generate better code.
251 * The increment or decrement is known and therefore one boundary check can
252 * be omitted.
253 *
254 * NOTE: These functions are very performance sensitive. Change only
255 * with care.
256 *
257 * Some processors have inc/dec instructions that are atomic vs an interrupt.
258 * However, the code must first determine the differential location in a zone
259 * based on the processor number and then inc/dec the counter. There is no
260 * guarantee without disabling preemption that the processor will not change
261 * in between and therefore the atomicity vs. interrupt cannot be exploited
262 * in a useful way here.
263 */
264void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
265{
266 struct per_cpu_pageset __percpu *pcp = zone->pageset;
267 s8 __percpu *p = pcp->vm_stat_diff + item;
268 s8 v, t;
269
270 v = __this_cpu_inc_return(*p);
271 t = __this_cpu_read(pcp->stat_threshold);
272 if (unlikely(v > t)) {
273 s8 overstep = t >> 1;
274
275 zone_page_state_add(v + overstep, zone, item);
276 __this_cpu_write(*p, -overstep);
277 }
278}
279
280void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
281{
282 __inc_zone_state(page_zone(page), item);
283}
284EXPORT_SYMBOL(__inc_zone_page_state);
285
286void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
287{
288 struct per_cpu_pageset __percpu *pcp = zone->pageset;
289 s8 __percpu *p = pcp->vm_stat_diff + item;
290 s8 v, t;
291
292 v = __this_cpu_dec_return(*p);
293 t = __this_cpu_read(pcp->stat_threshold);
294 if (unlikely(v < - t)) {
295 s8 overstep = t >> 1;
296
297 zone_page_state_add(v - overstep, zone, item);
298 __this_cpu_write(*p, overstep);
299 }
300}
301
302void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
303{
304 __dec_zone_state(page_zone(page), item);
305}
306EXPORT_SYMBOL(__dec_zone_page_state);
307
308#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
309/*
310 * If we have cmpxchg_local support then we do not need to incur the overhead
311 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
312 *
313 * mod_state() modifies the zone counter state through atomic per cpu
314 * operations.
315 *
316 * Overstep mode specifies how overstep should handled:
317 * 0 No overstepping
318 * 1 Overstepping half of threshold
319 * -1 Overstepping minus half of threshold
320*/
321static inline void mod_state(struct zone *zone, enum zone_stat_item item,
322 long delta, int overstep_mode)
323{
324 struct per_cpu_pageset __percpu *pcp = zone->pageset;
325 s8 __percpu *p = pcp->vm_stat_diff + item;
326 long o, n, t, z;
327
328 do {
329 z = 0; /* overflow to zone counters */
330
331 /*
332 * The fetching of the stat_threshold is racy. We may apply
333 * a counter threshold to the wrong the cpu if we get
334 * rescheduled while executing here. However, the next
335 * counter update will apply the threshold again and
336 * therefore bring the counter under the threshold again.
337 *
338 * Most of the time the thresholds are the same anyways
339 * for all cpus in a zone.
340 */
341 t = this_cpu_read(pcp->stat_threshold);
342
343 o = this_cpu_read(*p);
344 n = delta + o;
345
346 if (n > t || n < -t) {
347 int os = overstep_mode * (t >> 1) ;
348
349 /* Overflow must be added to zone counters */
350 z = n + os;
351 n = -os;
352 }
353 } while (this_cpu_cmpxchg(*p, o, n) != o);
354
355 if (z)
356 zone_page_state_add(z, zone, item);
357}
358
359void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
360 long delta)
361{
362 mod_state(zone, item, delta, 0);
363}
364EXPORT_SYMBOL(mod_zone_page_state);
365
366void inc_zone_state(struct zone *zone, enum zone_stat_item item)
367{
368 mod_state(zone, item, 1, 1);
369}
370
371void inc_zone_page_state(struct page *page, enum zone_stat_item item)
372{
373 mod_state(page_zone(page), item, 1, 1);
374}
375EXPORT_SYMBOL(inc_zone_page_state);
376
377void dec_zone_page_state(struct page *page, enum zone_stat_item item)
378{
379 mod_state(page_zone(page), item, -1, -1);
380}
381EXPORT_SYMBOL(dec_zone_page_state);
382#else
383/*
384 * Use interrupt disable to serialize counter updates
385 */
386void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
387 long delta)
388{
389 unsigned long flags;
390
391 local_irq_save(flags);
392 __mod_zone_page_state(zone, item, delta);
393 local_irq_restore(flags);
394}
395EXPORT_SYMBOL(mod_zone_page_state);
396
397void inc_zone_state(struct zone *zone, enum zone_stat_item item)
398{
399 unsigned long flags;
400
401 local_irq_save(flags);
402 __inc_zone_state(zone, item);
403 local_irq_restore(flags);
404}
405
406void inc_zone_page_state(struct page *page, enum zone_stat_item item)
407{
408 unsigned long flags;
409 struct zone *zone;
410
411 zone = page_zone(page);
412 local_irq_save(flags);
413 __inc_zone_state(zone, item);
414 local_irq_restore(flags);
415}
416EXPORT_SYMBOL(inc_zone_page_state);
417
418void dec_zone_page_state(struct page *page, enum zone_stat_item item)
419{
420 unsigned long flags;
421
422 local_irq_save(flags);
423 __dec_zone_page_state(page, item);
424 local_irq_restore(flags);
425}
426EXPORT_SYMBOL(dec_zone_page_state);
427#endif
428
429
430/*
431 * Fold a differential into the global counters.
432 * Returns the number of counters updated.
433 */
434static int fold_diff(int *diff)
435{
436 int i;
437 int changes = 0;
438
439 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
440 if (diff[i]) {
441 atomic_long_add(diff[i], &vm_stat[i]);
442 changes++;
443 }
444 return changes;
445}
446
447/*
448 * Update the zone counters for the current cpu.
449 *
450 * Note that refresh_cpu_vm_stats strives to only access
451 * node local memory. The per cpu pagesets on remote zones are placed
452 * in the memory local to the processor using that pageset. So the
453 * loop over all zones will access a series of cachelines local to
454 * the processor.
455 *
456 * The call to zone_page_state_add updates the cachelines with the
457 * statistics in the remote zone struct as well as the global cachelines
458 * with the global counters. These could cause remote node cache line
459 * bouncing and will have to be only done when necessary.
460 *
461 * The function returns the number of global counters updated.
462 */
463static int refresh_cpu_vm_stats(bool do_pagesets)
464{
465 struct zone *zone;
466 int i;
467 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
468 int changes = 0;
469
470 for_each_populated_zone(zone) {
471 struct per_cpu_pageset __percpu *p = zone->pageset;
472
473 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
474 int v;
475
476 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
477 if (v) {
478
479 atomic_long_add(v, &zone->vm_stat[i]);
480 global_diff[i] += v;
481#ifdef CONFIG_NUMA
482 /* 3 seconds idle till flush */
483 __this_cpu_write(p->expire, 3);
484#endif
485 }
486 }
487#ifdef CONFIG_NUMA
488 if (do_pagesets) {
489 cond_resched();
490 /*
491 * Deal with draining the remote pageset of this
492 * processor
493 *
494 * Check if there are pages remaining in this pageset
495 * if not then there is nothing to expire.
496 */
497 if (!__this_cpu_read(p->expire) ||
498 !__this_cpu_read(p->pcp.count))
499 continue;
500
501 /*
502 * We never drain zones local to this processor.
503 */
504 if (zone_to_nid(zone) == numa_node_id()) {
505 __this_cpu_write(p->expire, 0);
506 continue;
507 }
508
509 if (__this_cpu_dec_return(p->expire))
510 continue;
511
512 if (__this_cpu_read(p->pcp.count)) {
513 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
514 changes++;
515 }
516 }
517#endif
518 }
519 changes += fold_diff(global_diff);
520 return changes;
521}
522
523/*
524 * Fold the data for an offline cpu into the global array.
525 * There cannot be any access by the offline cpu and therefore
526 * synchronization is simplified.
527 */
528void cpu_vm_stats_fold(int cpu)
529{
530 struct zone *zone;
531 int i;
532 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
533
534 for_each_populated_zone(zone) {
535 struct per_cpu_pageset *p;
536
537 p = per_cpu_ptr(zone->pageset, cpu);
538
539 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
540 if (p->vm_stat_diff[i]) {
541 int v;
542
543 v = p->vm_stat_diff[i];
544 p->vm_stat_diff[i] = 0;
545 atomic_long_add(v, &zone->vm_stat[i]);
546 global_diff[i] += v;
547 }
548 }
549
550 fold_diff(global_diff);
551}
552
553/*
554 * this is only called if !populated_zone(zone), which implies no other users of
555 * pset->vm_stat_diff[] exsist.
556 */
557void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
558{
559 int i;
560
561 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
562 if (pset->vm_stat_diff[i]) {
563 int v = pset->vm_stat_diff[i];
564 pset->vm_stat_diff[i] = 0;
565 atomic_long_add(v, &zone->vm_stat[i]);
566 atomic_long_add(v, &vm_stat[i]);
567 }
568}
569#endif
570
571#ifdef CONFIG_NUMA
572/*
573 * zonelist = the list of zones passed to the allocator
574 * z = the zone from which the allocation occurred.
575 *
576 * Must be called with interrupts disabled.
577 *
578 * When __GFP_OTHER_NODE is set assume the node of the preferred
579 * zone is the local node. This is useful for daemons who allocate
580 * memory on behalf of other processes.
581 */
582void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
583{
584 if (z->zone_pgdat == preferred_zone->zone_pgdat) {
585 __inc_zone_state(z, NUMA_HIT);
586 } else {
587 __inc_zone_state(z, NUMA_MISS);
588 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
589 }
590 if (z->node == ((flags & __GFP_OTHER_NODE) ?
591 preferred_zone->node : numa_node_id()))
592 __inc_zone_state(z, NUMA_LOCAL);
593 else
594 __inc_zone_state(z, NUMA_OTHER);
595}
596
597/*
598 * Determine the per node value of a stat item.
599 */
600unsigned long node_page_state(int node, enum zone_stat_item item)
601{
602 struct zone *zones = NODE_DATA(node)->node_zones;
603
604 return
605#ifdef CONFIG_ZONE_DMA
606 zone_page_state(&zones[ZONE_DMA], item) +
607#endif
608#ifdef CONFIG_ZONE_DMA32
609 zone_page_state(&zones[ZONE_DMA32], item) +
610#endif
611#ifdef CONFIG_HIGHMEM
612 zone_page_state(&zones[ZONE_HIGHMEM], item) +
613#endif
614 zone_page_state(&zones[ZONE_NORMAL], item) +
615 zone_page_state(&zones[ZONE_MOVABLE], item);
616}
617
618#endif
619
620#ifdef CONFIG_COMPACTION
621
622struct contig_page_info {
623 unsigned long free_pages;
624 unsigned long free_blocks_total;
625 unsigned long free_blocks_suitable;
626};
627
628/*
629 * Calculate the number of free pages in a zone, how many contiguous
630 * pages are free and how many are large enough to satisfy an allocation of
631 * the target size. Note that this function makes no attempt to estimate
632 * how many suitable free blocks there *might* be if MOVABLE pages were
633 * migrated. Calculating that is possible, but expensive and can be
634 * figured out from userspace
635 */
636static void fill_contig_page_info(struct zone *zone,
637 unsigned int suitable_order,
638 struct contig_page_info *info)
639{
640 unsigned int order;
641
642 info->free_pages = 0;
643 info->free_blocks_total = 0;
644 info->free_blocks_suitable = 0;
645
646 for (order = 0; order < MAX_ORDER; order++) {
647 unsigned long blocks;
648
649 /* Count number of free blocks */
650 blocks = zone->free_area[order].nr_free;
651 info->free_blocks_total += blocks;
652
653 /* Count free base pages */
654 info->free_pages += blocks << order;
655
656 /* Count the suitable free blocks */
657 if (order >= suitable_order)
658 info->free_blocks_suitable += blocks <<
659 (order - suitable_order);
660 }
661}
662
663/*
664 * A fragmentation index only makes sense if an allocation of a requested
665 * size would fail. If that is true, the fragmentation index indicates
666 * whether external fragmentation or a lack of memory was the problem.
667 * The value can be used to determine if page reclaim or compaction
668 * should be used
669 */
670static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
671{
672 unsigned long requested = 1UL << order;
673
674 if (!info->free_blocks_total)
675 return 0;
676
677 /* Fragmentation index only makes sense when a request would fail */
678 if (info->free_blocks_suitable)
679 return -1000;
680
681 /*
682 * Index is between 0 and 1 so return within 3 decimal places
683 *
684 * 0 => allocation would fail due to lack of memory
685 * 1 => allocation would fail due to fragmentation
686 */
687 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
688}
689
690/* Same as __fragmentation index but allocs contig_page_info on stack */
691int fragmentation_index(struct zone *zone, unsigned int order)
692{
693 struct contig_page_info info;
694
695 fill_contig_page_info(zone, order, &info);
696 return __fragmentation_index(order, &info);
697}
698#endif
699
700#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
701#ifdef CONFIG_ZONE_DMA
702#define TEXT_FOR_DMA(xx) xx "_dma",
703#else
704#define TEXT_FOR_DMA(xx)
705#endif
706
707#ifdef CONFIG_ZONE_DMA32
708#define TEXT_FOR_DMA32(xx) xx "_dma32",
709#else
710#define TEXT_FOR_DMA32(xx)
711#endif
712
713#ifdef CONFIG_HIGHMEM
714#define TEXT_FOR_HIGHMEM(xx) xx "_high",
715#else
716#define TEXT_FOR_HIGHMEM(xx)
717#endif
718
719#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
720 TEXT_FOR_HIGHMEM(xx) xx "_movable",
721
722const char * const vmstat_text[] = {
723 /* enum zone_stat_item countes */
724 "nr_free_pages",
725 "nr_alloc_batch",
726 "nr_inactive_anon",
727 "nr_active_anon",
728 "nr_inactive_file",
729 "nr_active_file",
730 "nr_unevictable",
731 "nr_mlock",
732 "nr_anon_pages",
733 "nr_mapped",
734 "nr_file_pages",
735 "nr_dirty",
736 "nr_writeback",
737 "nr_slab_reclaimable",
738 "nr_slab_unreclaimable",
739 "nr_page_table_pages",
740 "nr_kernel_stack",
741 "nr_unstable",
742 "nr_bounce",
743 "nr_vmscan_write",
744 "nr_vmscan_immediate_reclaim",
745 "nr_writeback_temp",
746 "nr_isolated_anon",
747 "nr_isolated_file",
748 "nr_shmem",
749 "nr_dirtied",
750 "nr_written",
751 "nr_pages_scanned",
752
753#ifdef CONFIG_NUMA
754 "numa_hit",
755 "numa_miss",
756 "numa_foreign",
757 "numa_interleave",
758 "numa_local",
759 "numa_other",
760#endif
761 "workingset_refault",
762 "workingset_activate",
763 "workingset_nodereclaim",
764 "nr_anon_transparent_hugepages",
765 "nr_free_cma",
766
767 /* enum writeback_stat_item counters */
768 "nr_dirty_threshold",
769 "nr_dirty_background_threshold",
770
771#ifdef CONFIG_VM_EVENT_COUNTERS
772 /* enum vm_event_item counters */
773 "pgpgin",
774 "pgpgout",
775 "pswpin",
776 "pswpout",
777
778 TEXTS_FOR_ZONES("pgalloc")
779
780 "pgfree",
781 "pgactivate",
782 "pgdeactivate",
783
784 "pgfault",
785 "pgmajfault",
786 "pglazyfreed",
787
788 TEXTS_FOR_ZONES("pgrefill")
789 TEXTS_FOR_ZONES("pgsteal_kswapd")
790 TEXTS_FOR_ZONES("pgsteal_direct")
791 TEXTS_FOR_ZONES("pgscan_kswapd")
792 TEXTS_FOR_ZONES("pgscan_direct")
793 "pgscan_direct_throttle",
794
795#ifdef CONFIG_NUMA
796 "zone_reclaim_failed",
797#endif
798 "pginodesteal",
799 "slabs_scanned",
800 "kswapd_inodesteal",
801 "kswapd_low_wmark_hit_quickly",
802 "kswapd_high_wmark_hit_quickly",
803 "pageoutrun",
804 "allocstall",
805
806 "pgrotated",
807
808 "drop_pagecache",
809 "drop_slab",
810
811#ifdef CONFIG_NUMA_BALANCING
812 "numa_pte_updates",
813 "numa_huge_pte_updates",
814 "numa_hint_faults",
815 "numa_hint_faults_local",
816 "numa_pages_migrated",
817#endif
818#ifdef CONFIG_MIGRATION
819 "pgmigrate_success",
820 "pgmigrate_fail",
821#endif
822#ifdef CONFIG_COMPACTION
823 "compact_migrate_scanned",
824 "compact_free_scanned",
825 "compact_isolated",
826 "compact_stall",
827 "compact_fail",
828 "compact_success",
829 "compact_daemon_wake",
830#endif
831
832#ifdef CONFIG_HUGETLB_PAGE
833 "htlb_buddy_alloc_success",
834 "htlb_buddy_alloc_fail",
835#endif
836 "unevictable_pgs_culled",
837 "unevictable_pgs_scanned",
838 "unevictable_pgs_rescued",
839 "unevictable_pgs_mlocked",
840 "unevictable_pgs_munlocked",
841 "unevictable_pgs_cleared",
842 "unevictable_pgs_stranded",
843
844#ifdef CONFIG_TRANSPARENT_HUGEPAGE
845 "thp_fault_alloc",
846 "thp_fault_fallback",
847 "thp_collapse_alloc",
848 "thp_collapse_alloc_failed",
849 "thp_split_page",
850 "thp_split_page_failed",
851 "thp_deferred_split_page",
852 "thp_split_pmd",
853 "thp_zero_page_alloc",
854 "thp_zero_page_alloc_failed",
855#endif
856#ifdef CONFIG_MEMORY_BALLOON
857 "balloon_inflate",
858 "balloon_deflate",
859#ifdef CONFIG_BALLOON_COMPACTION
860 "balloon_migrate",
861#endif
862#endif /* CONFIG_MEMORY_BALLOON */
863#ifdef CONFIG_DEBUG_TLBFLUSH
864#ifdef CONFIG_SMP
865 "nr_tlb_remote_flush",
866 "nr_tlb_remote_flush_received",
867#endif /* CONFIG_SMP */
868 "nr_tlb_local_flush_all",
869 "nr_tlb_local_flush_one",
870#endif /* CONFIG_DEBUG_TLBFLUSH */
871
872#ifdef CONFIG_DEBUG_VM_VMACACHE
873 "vmacache_find_calls",
874 "vmacache_find_hits",
875 "vmacache_full_flushes",
876#endif
877#endif /* CONFIG_VM_EVENTS_COUNTERS */
878};
879#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
880
881
882#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
883 defined(CONFIG_PROC_FS)
884static void *frag_start(struct seq_file *m, loff_t *pos)
885{
886 pg_data_t *pgdat;
887 loff_t node = *pos;
888
889 for (pgdat = first_online_pgdat();
890 pgdat && node;
891 pgdat = next_online_pgdat(pgdat))
892 --node;
893
894 return pgdat;
895}
896
897static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
898{
899 pg_data_t *pgdat = (pg_data_t *)arg;
900
901 (*pos)++;
902 return next_online_pgdat(pgdat);
903}
904
905static void frag_stop(struct seq_file *m, void *arg)
906{
907}
908
909/* Walk all the zones in a node and print using a callback */
910static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
911 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
912{
913 struct zone *zone;
914 struct zone *node_zones = pgdat->node_zones;
915 unsigned long flags;
916
917 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
918 if (!populated_zone(zone))
919 continue;
920
921 spin_lock_irqsave(&zone->lock, flags);
922 print(m, pgdat, zone);
923 spin_unlock_irqrestore(&zone->lock, flags);
924 }
925}
926#endif
927
928#ifdef CONFIG_PROC_FS
929static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
930 struct zone *zone)
931{
932 int order;
933
934 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
935 for (order = 0; order < MAX_ORDER; ++order)
936 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
937 seq_putc(m, '\n');
938}
939
940/*
941 * This walks the free areas for each zone.
942 */
943static int frag_show(struct seq_file *m, void *arg)
944{
945 pg_data_t *pgdat = (pg_data_t *)arg;
946 walk_zones_in_node(m, pgdat, frag_show_print);
947 return 0;
948}
949
950static void pagetypeinfo_showfree_print(struct seq_file *m,
951 pg_data_t *pgdat, struct zone *zone)
952{
953 int order, mtype;
954
955 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
956 seq_printf(m, "Node %4d, zone %8s, type %12s ",
957 pgdat->node_id,
958 zone->name,
959 migratetype_names[mtype]);
960 for (order = 0; order < MAX_ORDER; ++order) {
961 unsigned long freecount = 0;
962 struct free_area *area;
963 struct list_head *curr;
964
965 area = &(zone->free_area[order]);
966
967 list_for_each(curr, &area->free_list[mtype])
968 freecount++;
969 seq_printf(m, "%6lu ", freecount);
970 }
971 seq_putc(m, '\n');
972 }
973}
974
975/* Print out the free pages at each order for each migatetype */
976static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
977{
978 int order;
979 pg_data_t *pgdat = (pg_data_t *)arg;
980
981 /* Print header */
982 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
983 for (order = 0; order < MAX_ORDER; ++order)
984 seq_printf(m, "%6d ", order);
985 seq_putc(m, '\n');
986
987 walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
988
989 return 0;
990}
991
992static void pagetypeinfo_showblockcount_print(struct seq_file *m,
993 pg_data_t *pgdat, struct zone *zone)
994{
995 int mtype;
996 unsigned long pfn;
997 unsigned long start_pfn = zone->zone_start_pfn;
998 unsigned long end_pfn = zone_end_pfn(zone);
999 unsigned long count[MIGRATE_TYPES] = { 0, };
1000
1001 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1002 struct page *page;
1003
1004 if (!pfn_valid(pfn))
1005 continue;
1006
1007 page = pfn_to_page(pfn);
1008
1009 /* Watch for unexpected holes punched in the memmap */
1010 if (!memmap_valid_within(pfn, page, zone))
1011 continue;
1012
1013 mtype = get_pageblock_migratetype(page);
1014
1015 if (mtype < MIGRATE_TYPES)
1016 count[mtype]++;
1017 }
1018
1019 /* Print counts */
1020 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1021 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1022 seq_printf(m, "%12lu ", count[mtype]);
1023 seq_putc(m, '\n');
1024}
1025
1026/* Print out the free pages at each order for each migratetype */
1027static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1028{
1029 int mtype;
1030 pg_data_t *pgdat = (pg_data_t *)arg;
1031
1032 seq_printf(m, "\n%-23s", "Number of blocks type ");
1033 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1034 seq_printf(m, "%12s ", migratetype_names[mtype]);
1035 seq_putc(m, '\n');
1036 walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1037
1038 return 0;
1039}
1040
1041#ifdef CONFIG_PAGE_OWNER
1042static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1043 pg_data_t *pgdat,
1044 struct zone *zone)
1045{
1046 struct page *page;
1047 struct page_ext *page_ext;
1048 unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1049 unsigned long end_pfn = pfn + zone->spanned_pages;
1050 unsigned long count[MIGRATE_TYPES] = { 0, };
1051 int pageblock_mt, page_mt;
1052 int i;
1053
1054 /* Scan block by block. First and last block may be incomplete */
1055 pfn = zone->zone_start_pfn;
1056
1057 /*
1058 * Walk the zone in pageblock_nr_pages steps. If a page block spans
1059 * a zone boundary, it will be double counted between zones. This does
1060 * not matter as the mixed block count will still be correct
1061 */
1062 for (; pfn < end_pfn; ) {
1063 if (!pfn_valid(pfn)) {
1064 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1065 continue;
1066 }
1067
1068 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1069 block_end_pfn = min(block_end_pfn, end_pfn);
1070
1071 page = pfn_to_page(pfn);
1072 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1073
1074 for (; pfn < block_end_pfn; pfn++) {
1075 if (!pfn_valid_within(pfn))
1076 continue;
1077
1078 page = pfn_to_page(pfn);
1079 if (PageBuddy(page)) {
1080 pfn += (1UL << page_order(page)) - 1;
1081 continue;
1082 }
1083
1084 if (PageReserved(page))
1085 continue;
1086
1087 page_ext = lookup_page_ext(page);
1088
1089 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1090 continue;
1091
1092 page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1093 if (pageblock_mt != page_mt) {
1094 if (is_migrate_cma(pageblock_mt))
1095 count[MIGRATE_MOVABLE]++;
1096 else
1097 count[pageblock_mt]++;
1098
1099 pfn = block_end_pfn;
1100 break;
1101 }
1102 pfn += (1UL << page_ext->order) - 1;
1103 }
1104 }
1105
1106 /* Print counts */
1107 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1108 for (i = 0; i < MIGRATE_TYPES; i++)
1109 seq_printf(m, "%12lu ", count[i]);
1110 seq_putc(m, '\n');
1111}
1112#endif /* CONFIG_PAGE_OWNER */
1113
1114/*
1115 * Print out the number of pageblocks for each migratetype that contain pages
1116 * of other types. This gives an indication of how well fallbacks are being
1117 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1118 * to determine what is going on
1119 */
1120static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1121{
1122#ifdef CONFIG_PAGE_OWNER
1123 int mtype;
1124
1125 if (!static_branch_unlikely(&page_owner_inited))
1126 return;
1127
1128 drain_all_pages(NULL);
1129
1130 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1131 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1132 seq_printf(m, "%12s ", migratetype_names[mtype]);
1133 seq_putc(m, '\n');
1134
1135 walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1136#endif /* CONFIG_PAGE_OWNER */
1137}
1138
1139/*
1140 * This prints out statistics in relation to grouping pages by mobility.
1141 * It is expensive to collect so do not constantly read the file.
1142 */
1143static int pagetypeinfo_show(struct seq_file *m, void *arg)
1144{
1145 pg_data_t *pgdat = (pg_data_t *)arg;
1146
1147 /* check memoryless node */
1148 if (!node_state(pgdat->node_id, N_MEMORY))
1149 return 0;
1150
1151 seq_printf(m, "Page block order: %d\n", pageblock_order);
1152 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1153 seq_putc(m, '\n');
1154 pagetypeinfo_showfree(m, pgdat);
1155 pagetypeinfo_showblockcount(m, pgdat);
1156 pagetypeinfo_showmixedcount(m, pgdat);
1157
1158 return 0;
1159}
1160
1161static const struct seq_operations fragmentation_op = {
1162 .start = frag_start,
1163 .next = frag_next,
1164 .stop = frag_stop,
1165 .show = frag_show,
1166};
1167
1168static int fragmentation_open(struct inode *inode, struct file *file)
1169{
1170 return seq_open(file, &fragmentation_op);
1171}
1172
1173static const struct file_operations fragmentation_file_operations = {
1174 .open = fragmentation_open,
1175 .read = seq_read,
1176 .llseek = seq_lseek,
1177 .release = seq_release,
1178};
1179
1180static const struct seq_operations pagetypeinfo_op = {
1181 .start = frag_start,
1182 .next = frag_next,
1183 .stop = frag_stop,
1184 .show = pagetypeinfo_show,
1185};
1186
1187static int pagetypeinfo_open(struct inode *inode, struct file *file)
1188{
1189 return seq_open(file, &pagetypeinfo_op);
1190}
1191
1192static const struct file_operations pagetypeinfo_file_ops = {
1193 .open = pagetypeinfo_open,
1194 .read = seq_read,
1195 .llseek = seq_lseek,
1196 .release = seq_release,
1197};
1198
1199static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1200 struct zone *zone)
1201{
1202 int i;
1203 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1204 seq_printf(m,
1205 "\n pages free %lu"
1206 "\n min %lu"
1207 "\n low %lu"
1208 "\n high %lu"
1209 "\n scanned %lu"
1210 "\n spanned %lu"
1211 "\n present %lu"
1212 "\n managed %lu",
1213 zone_page_state(zone, NR_FREE_PAGES),
1214 min_wmark_pages(zone),
1215 low_wmark_pages(zone),
1216 high_wmark_pages(zone),
1217 zone_page_state(zone, NR_PAGES_SCANNED),
1218 zone->spanned_pages,
1219 zone->present_pages,
1220 zone->managed_pages);
1221
1222 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1223 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1224 zone_page_state(zone, i));
1225
1226 seq_printf(m,
1227 "\n protection: (%ld",
1228 zone->lowmem_reserve[0]);
1229 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1230 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1231 seq_printf(m,
1232 ")"
1233 "\n pagesets");
1234 for_each_online_cpu(i) {
1235 struct per_cpu_pageset *pageset;
1236
1237 pageset = per_cpu_ptr(zone->pageset, i);
1238 seq_printf(m,
1239 "\n cpu: %i"
1240 "\n count: %i"
1241 "\n high: %i"
1242 "\n batch: %i",
1243 i,
1244 pageset->pcp.count,
1245 pageset->pcp.high,
1246 pageset->pcp.batch);
1247#ifdef CONFIG_SMP
1248 seq_printf(m, "\n vm stats threshold: %d",
1249 pageset->stat_threshold);
1250#endif
1251 }
1252 seq_printf(m,
1253 "\n all_unreclaimable: %u"
1254 "\n start_pfn: %lu"
1255 "\n inactive_ratio: %u",
1256 !zone_reclaimable(zone),
1257 zone->zone_start_pfn,
1258 zone->inactive_ratio);
1259 seq_putc(m, '\n');
1260}
1261
1262/*
1263 * Output information about zones in @pgdat.
1264 */
1265static int zoneinfo_show(struct seq_file *m, void *arg)
1266{
1267 pg_data_t *pgdat = (pg_data_t *)arg;
1268 walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1269 return 0;
1270}
1271
1272static const struct seq_operations zoneinfo_op = {
1273 .start = frag_start, /* iterate over all zones. The same as in
1274 * fragmentation. */
1275 .next = frag_next,
1276 .stop = frag_stop,
1277 .show = zoneinfo_show,
1278};
1279
1280static int zoneinfo_open(struct inode *inode, struct file *file)
1281{
1282 return seq_open(file, &zoneinfo_op);
1283}
1284
1285static const struct file_operations proc_zoneinfo_file_operations = {
1286 .open = zoneinfo_open,
1287 .read = seq_read,
1288 .llseek = seq_lseek,
1289 .release = seq_release,
1290};
1291
1292enum writeback_stat_item {
1293 NR_DIRTY_THRESHOLD,
1294 NR_DIRTY_BG_THRESHOLD,
1295 NR_VM_WRITEBACK_STAT_ITEMS,
1296};
1297
1298static void *vmstat_start(struct seq_file *m, loff_t *pos)
1299{
1300 unsigned long *v;
1301 int i, stat_items_size;
1302
1303 if (*pos >= ARRAY_SIZE(vmstat_text))
1304 return NULL;
1305 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1306 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1307
1308#ifdef CONFIG_VM_EVENT_COUNTERS
1309 stat_items_size += sizeof(struct vm_event_state);
1310#endif
1311
1312 v = kmalloc(stat_items_size, GFP_KERNEL);
1313 m->private = v;
1314 if (!v)
1315 return ERR_PTR(-ENOMEM);
1316 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1317 v[i] = global_page_state(i);
1318 v += NR_VM_ZONE_STAT_ITEMS;
1319
1320 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1321 v + NR_DIRTY_THRESHOLD);
1322 v += NR_VM_WRITEBACK_STAT_ITEMS;
1323
1324#ifdef CONFIG_VM_EVENT_COUNTERS
1325 all_vm_events(v);
1326 v[PGPGIN] /= 2; /* sectors -> kbytes */
1327 v[PGPGOUT] /= 2;
1328#endif
1329 return (unsigned long *)m->private + *pos;
1330}
1331
1332static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1333{
1334 (*pos)++;
1335 if (*pos >= ARRAY_SIZE(vmstat_text))
1336 return NULL;
1337 return (unsigned long *)m->private + *pos;
1338}
1339
1340static int vmstat_show(struct seq_file *m, void *arg)
1341{
1342 unsigned long *l = arg;
1343 unsigned long off = l - (unsigned long *)m->private;
1344
1345 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1346 return 0;
1347}
1348
1349static void vmstat_stop(struct seq_file *m, void *arg)
1350{
1351 kfree(m->private);
1352 m->private = NULL;
1353}
1354
1355static const struct seq_operations vmstat_op = {
1356 .start = vmstat_start,
1357 .next = vmstat_next,
1358 .stop = vmstat_stop,
1359 .show = vmstat_show,
1360};
1361
1362static int vmstat_open(struct inode *inode, struct file *file)
1363{
1364 return seq_open(file, &vmstat_op);
1365}
1366
1367static const struct file_operations proc_vmstat_file_operations = {
1368 .open = vmstat_open,
1369 .read = seq_read,
1370 .llseek = seq_lseek,
1371 .release = seq_release,
1372};
1373#endif /* CONFIG_PROC_FS */
1374
1375#ifdef CONFIG_SMP
1376static struct workqueue_struct *vmstat_wq;
1377static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1378int sysctl_stat_interval __read_mostly = HZ;
1379static cpumask_var_t cpu_stat_off;
1380
1381static void vmstat_update(struct work_struct *w)
1382{
1383 if (refresh_cpu_vm_stats(true)) {
1384 /*
1385 * Counters were updated so we expect more updates
1386 * to occur in the future. Keep on running the
1387 * update worker thread.
1388 * If we were marked on cpu_stat_off clear the flag
1389 * so that vmstat_shepherd doesn't schedule us again.
1390 */
1391 if (!cpumask_test_and_clear_cpu(smp_processor_id(),
1392 cpu_stat_off)) {
1393 queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1394 this_cpu_ptr(&vmstat_work),
1395 round_jiffies_relative(sysctl_stat_interval));
1396 }
1397 } else {
1398 /*
1399 * We did not update any counters so the app may be in
1400 * a mode where it does not cause counter updates.
1401 * We may be uselessly running vmstat_update.
1402 * Defer the checking for differentials to the
1403 * shepherd thread on a different processor.
1404 */
1405 cpumask_set_cpu(smp_processor_id(), cpu_stat_off);
1406 }
1407}
1408
1409/*
1410 * Switch off vmstat processing and then fold all the remaining differentials
1411 * until the diffs stay at zero. The function is used by NOHZ and can only be
1412 * invoked when tick processing is not active.
1413 */
1414/*
1415 * Check if the diffs for a certain cpu indicate that
1416 * an update is needed.
1417 */
1418static bool need_update(int cpu)
1419{
1420 struct zone *zone;
1421
1422 for_each_populated_zone(zone) {
1423 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1424
1425 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1426 /*
1427 * The fast way of checking if there are any vmstat diffs.
1428 * This works because the diffs are byte sized items.
1429 */
1430 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1431 return true;
1432
1433 }
1434 return false;
1435}
1436
1437void quiet_vmstat(void)
1438{
1439 if (system_state != SYSTEM_RUNNING)
1440 return;
1441
1442 /*
1443 * If we are already in hands of the shepherd then there
1444 * is nothing for us to do here.
1445 */
1446 if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
1447 return;
1448
1449 if (!need_update(smp_processor_id()))
1450 return;
1451
1452 /*
1453 * Just refresh counters and do not care about the pending delayed
1454 * vmstat_update. It doesn't fire that often to matter and canceling
1455 * it would be too expensive from this path.
1456 * vmstat_shepherd will take care about that for us.
1457 */
1458 refresh_cpu_vm_stats(false);
1459}
1460
1461
1462/*
1463 * Shepherd worker thread that checks the
1464 * differentials of processors that have their worker
1465 * threads for vm statistics updates disabled because of
1466 * inactivity.
1467 */
1468static void vmstat_shepherd(struct work_struct *w);
1469
1470static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1471
1472static void vmstat_shepherd(struct work_struct *w)
1473{
1474 int cpu;
1475
1476 get_online_cpus();
1477 /* Check processors whose vmstat worker threads have been disabled */
1478 for_each_cpu(cpu, cpu_stat_off) {
1479 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1480
1481 if (need_update(cpu)) {
1482 if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1483 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1484 } else {
1485 /*
1486 * Cancel the work if quiet_vmstat has put this
1487 * cpu on cpu_stat_off because the work item might
1488 * be still scheduled
1489 */
1490 cancel_delayed_work(dw);
1491 }
1492 }
1493 put_online_cpus();
1494
1495 schedule_delayed_work(&shepherd,
1496 round_jiffies_relative(sysctl_stat_interval));
1497}
1498
1499static void __init start_shepherd_timer(void)
1500{
1501 int cpu;
1502
1503 for_each_possible_cpu(cpu)
1504 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1505 vmstat_update);
1506
1507 if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1508 BUG();
1509 cpumask_copy(cpu_stat_off, cpu_online_mask);
1510
1511 vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1512 schedule_delayed_work(&shepherd,
1513 round_jiffies_relative(sysctl_stat_interval));
1514}
1515
1516static void vmstat_cpu_dead(int node)
1517{
1518 int cpu;
1519
1520 get_online_cpus();
1521 for_each_online_cpu(cpu)
1522 if (cpu_to_node(cpu) == node)
1523 goto end;
1524
1525 node_clear_state(node, N_CPU);
1526end:
1527 put_online_cpus();
1528}
1529
1530/*
1531 * Use the cpu notifier to insure that the thresholds are recalculated
1532 * when necessary.
1533 */
1534static int vmstat_cpuup_callback(struct notifier_block *nfb,
1535 unsigned long action,
1536 void *hcpu)
1537{
1538 long cpu = (long)hcpu;
1539
1540 switch (action) {
1541 case CPU_ONLINE:
1542 case CPU_ONLINE_FROZEN:
1543 refresh_zone_stat_thresholds();
1544 node_set_state(cpu_to_node(cpu), N_CPU);
1545 cpumask_set_cpu(cpu, cpu_stat_off);
1546 break;
1547 case CPU_DOWN_PREPARE:
1548 case CPU_DOWN_PREPARE_FROZEN:
1549 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1550 cpumask_clear_cpu(cpu, cpu_stat_off);
1551 break;
1552 case CPU_DOWN_FAILED:
1553 case CPU_DOWN_FAILED_FROZEN:
1554 cpumask_set_cpu(cpu, cpu_stat_off);
1555 break;
1556 case CPU_DEAD:
1557 case CPU_DEAD_FROZEN:
1558 refresh_zone_stat_thresholds();
1559 vmstat_cpu_dead(cpu_to_node(cpu));
1560 break;
1561 default:
1562 break;
1563 }
1564 return NOTIFY_OK;
1565}
1566
1567static struct notifier_block vmstat_notifier =
1568 { &vmstat_cpuup_callback, NULL, 0 };
1569#endif
1570
1571static int __init setup_vmstat(void)
1572{
1573#ifdef CONFIG_SMP
1574 cpu_notifier_register_begin();
1575 __register_cpu_notifier(&vmstat_notifier);
1576
1577 start_shepherd_timer();
1578 cpu_notifier_register_done();
1579#endif
1580#ifdef CONFIG_PROC_FS
1581 proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1582 proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1583 proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1584 proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1585#endif
1586 return 0;
1587}
1588module_init(setup_vmstat)
1589
1590#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1591
1592/*
1593 * Return an index indicating how much of the available free memory is
1594 * unusable for an allocation of the requested size.
1595 */
1596static int unusable_free_index(unsigned int order,
1597 struct contig_page_info *info)
1598{
1599 /* No free memory is interpreted as all free memory is unusable */
1600 if (info->free_pages == 0)
1601 return 1000;
1602
1603 /*
1604 * Index should be a value between 0 and 1. Return a value to 3
1605 * decimal places.
1606 *
1607 * 0 => no fragmentation
1608 * 1 => high fragmentation
1609 */
1610 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1611
1612}
1613
1614static void unusable_show_print(struct seq_file *m,
1615 pg_data_t *pgdat, struct zone *zone)
1616{
1617 unsigned int order;
1618 int index;
1619 struct contig_page_info info;
1620
1621 seq_printf(m, "Node %d, zone %8s ",
1622 pgdat->node_id,
1623 zone->name);
1624 for (order = 0; order < MAX_ORDER; ++order) {
1625 fill_contig_page_info(zone, order, &info);
1626 index = unusable_free_index(order, &info);
1627 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1628 }
1629
1630 seq_putc(m, '\n');
1631}
1632
1633/*
1634 * Display unusable free space index
1635 *
1636 * The unusable free space index measures how much of the available free
1637 * memory cannot be used to satisfy an allocation of a given size and is a
1638 * value between 0 and 1. The higher the value, the more of free memory is
1639 * unusable and by implication, the worse the external fragmentation is. This
1640 * can be expressed as a percentage by multiplying by 100.
1641 */
1642static int unusable_show(struct seq_file *m, void *arg)
1643{
1644 pg_data_t *pgdat = (pg_data_t *)arg;
1645
1646 /* check memoryless node */
1647 if (!node_state(pgdat->node_id, N_MEMORY))
1648 return 0;
1649
1650 walk_zones_in_node(m, pgdat, unusable_show_print);
1651
1652 return 0;
1653}
1654
1655static const struct seq_operations unusable_op = {
1656 .start = frag_start,
1657 .next = frag_next,
1658 .stop = frag_stop,
1659 .show = unusable_show,
1660};
1661
1662static int unusable_open(struct inode *inode, struct file *file)
1663{
1664 return seq_open(file, &unusable_op);
1665}
1666
1667static const struct file_operations unusable_file_ops = {
1668 .open = unusable_open,
1669 .read = seq_read,
1670 .llseek = seq_lseek,
1671 .release = seq_release,
1672};
1673
1674static void extfrag_show_print(struct seq_file *m,
1675 pg_data_t *pgdat, struct zone *zone)
1676{
1677 unsigned int order;
1678 int index;
1679
1680 /* Alloc on stack as interrupts are disabled for zone walk */
1681 struct contig_page_info info;
1682
1683 seq_printf(m, "Node %d, zone %8s ",
1684 pgdat->node_id,
1685 zone->name);
1686 for (order = 0; order < MAX_ORDER; ++order) {
1687 fill_contig_page_info(zone, order, &info);
1688 index = __fragmentation_index(order, &info);
1689 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1690 }
1691
1692 seq_putc(m, '\n');
1693}
1694
1695/*
1696 * Display fragmentation index for orders that allocations would fail for
1697 */
1698static int extfrag_show(struct seq_file *m, void *arg)
1699{
1700 pg_data_t *pgdat = (pg_data_t *)arg;
1701
1702 walk_zones_in_node(m, pgdat, extfrag_show_print);
1703
1704 return 0;
1705}
1706
1707static const struct seq_operations extfrag_op = {
1708 .start = frag_start,
1709 .next = frag_next,
1710 .stop = frag_stop,
1711 .show = extfrag_show,
1712};
1713
1714static int extfrag_open(struct inode *inode, struct file *file)
1715{
1716 return seq_open(file, &extfrag_op);
1717}
1718
1719static const struct file_operations extfrag_file_ops = {
1720 .open = extfrag_open,
1721 .read = seq_read,
1722 .llseek = seq_lseek,
1723 .release = seq_release,
1724};
1725
1726static int __init extfrag_debug_init(void)
1727{
1728 struct dentry *extfrag_debug_root;
1729
1730 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1731 if (!extfrag_debug_root)
1732 return -ENOMEM;
1733
1734 if (!debugfs_create_file("unusable_index", 0444,
1735 extfrag_debug_root, NULL, &unusable_file_ops))
1736 goto fail;
1737
1738 if (!debugfs_create_file("extfrag_index", 0444,
1739 extfrag_debug_root, NULL, &extfrag_file_ops))
1740 goto fail;
1741
1742 return 0;
1743fail:
1744 debugfs_remove_recursive(extfrag_debug_root);
1745 return -ENOMEM;
1746}
1747
1748module_init(extfrag_debug_init);
1749#endif