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(zone) >> (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_nodes",
1147 "workingset_refault",
1148 "workingset_activate",
1149 "workingset_restore",
1150 "workingset_nodereclaim",
1151 "nr_anon_pages",
1152 "nr_mapped",
1153 "nr_file_pages",
1154 "nr_dirty",
1155 "nr_writeback",
1156 "nr_writeback_temp",
1157 "nr_shmem",
1158 "nr_shmem_hugepages",
1159 "nr_shmem_pmdmapped",
1160 "nr_anon_transparent_hugepages",
1161 "nr_unstable",
1162 "nr_vmscan_write",
1163 "nr_vmscan_immediate_reclaim",
1164 "nr_dirtied",
1165 "nr_written",
1166 "nr_kernel_misc_reclaimable",
1167
1168 /* enum writeback_stat_item counters */
1169 "nr_dirty_threshold",
1170 "nr_dirty_background_threshold",
1171
1172#ifdef CONFIG_VM_EVENT_COUNTERS
1173 /* enum vm_event_item counters */
1174 "pgpgin",
1175 "pgpgout",
1176 "pswpin",
1177 "pswpout",
1178
1179 TEXTS_FOR_ZONES("pgalloc")
1180 TEXTS_FOR_ZONES("allocstall")
1181 TEXTS_FOR_ZONES("pgskip")
1182
1183 "pgfree",
1184 "pgactivate",
1185 "pgdeactivate",
1186 "pglazyfree",
1187
1188 "pgfault",
1189 "pgmajfault",
1190 "pglazyfreed",
1191
1192 "pgrefill",
1193 "pgsteal_kswapd",
1194 "pgsteal_direct",
1195 "pgscan_kswapd",
1196 "pgscan_direct",
1197 "pgscan_direct_throttle",
1198
1199#ifdef CONFIG_NUMA
1200 "zone_reclaim_failed",
1201#endif
1202 "pginodesteal",
1203 "slabs_scanned",
1204 "kswapd_inodesteal",
1205 "kswapd_low_wmark_hit_quickly",
1206 "kswapd_high_wmark_hit_quickly",
1207 "pageoutrun",
1208
1209 "pgrotated",
1210
1211 "drop_pagecache",
1212 "drop_slab",
1213 "oom_kill",
1214
1215#ifdef CONFIG_NUMA_BALANCING
1216 "numa_pte_updates",
1217 "numa_huge_pte_updates",
1218 "numa_hint_faults",
1219 "numa_hint_faults_local",
1220 "numa_pages_migrated",
1221#endif
1222#ifdef CONFIG_MIGRATION
1223 "pgmigrate_success",
1224 "pgmigrate_fail",
1225#endif
1226#ifdef CONFIG_COMPACTION
1227 "compact_migrate_scanned",
1228 "compact_free_scanned",
1229 "compact_isolated",
1230 "compact_stall",
1231 "compact_fail",
1232 "compact_success",
1233 "compact_daemon_wake",
1234 "compact_daemon_migrate_scanned",
1235 "compact_daemon_free_scanned",
1236#endif
1237
1238#ifdef CONFIG_HUGETLB_PAGE
1239 "htlb_buddy_alloc_success",
1240 "htlb_buddy_alloc_fail",
1241#endif
1242 "unevictable_pgs_culled",
1243 "unevictable_pgs_scanned",
1244 "unevictable_pgs_rescued",
1245 "unevictable_pgs_mlocked",
1246 "unevictable_pgs_munlocked",
1247 "unevictable_pgs_cleared",
1248 "unevictable_pgs_stranded",
1249
1250#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1251 "thp_fault_alloc",
1252 "thp_fault_fallback",
1253 "thp_collapse_alloc",
1254 "thp_collapse_alloc_failed",
1255 "thp_file_alloc",
1256 "thp_file_mapped",
1257 "thp_split_page",
1258 "thp_split_page_failed",
1259 "thp_deferred_split_page",
1260 "thp_split_pmd",
1261#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1262 "thp_split_pud",
1263#endif
1264 "thp_zero_page_alloc",
1265 "thp_zero_page_alloc_failed",
1266 "thp_swpout",
1267 "thp_swpout_fallback",
1268#endif
1269#ifdef CONFIG_MEMORY_BALLOON
1270 "balloon_inflate",
1271 "balloon_deflate",
1272#ifdef CONFIG_BALLOON_COMPACTION
1273 "balloon_migrate",
1274#endif
1275#endif /* CONFIG_MEMORY_BALLOON */
1276#ifdef CONFIG_DEBUG_TLBFLUSH
1277#ifdef CONFIG_SMP
1278 "nr_tlb_remote_flush",
1279 "nr_tlb_remote_flush_received",
1280#else
1281 "", /* nr_tlb_remote_flush */
1282 "", /* nr_tlb_remote_flush_received */
1283#endif /* CONFIG_SMP */
1284 "nr_tlb_local_flush_all",
1285 "nr_tlb_local_flush_one",
1286#endif /* CONFIG_DEBUG_TLBFLUSH */
1287
1288#ifdef CONFIG_DEBUG_VM_VMACACHE
1289 "vmacache_find_calls",
1290 "vmacache_find_hits",
1291#endif
1292#ifdef CONFIG_SWAP
1293 "swap_ra",
1294 "swap_ra_hit",
1295#endif
1296#endif /* CONFIG_VM_EVENTS_COUNTERS */
1297};
1298#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
1299
1300#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1301 defined(CONFIG_PROC_FS)
1302static void *frag_start(struct seq_file *m, loff_t *pos)
1303{
1304 pg_data_t *pgdat;
1305 loff_t node = *pos;
1306
1307 for (pgdat = first_online_pgdat();
1308 pgdat && node;
1309 pgdat = next_online_pgdat(pgdat))
1310 --node;
1311
1312 return pgdat;
1313}
1314
1315static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1316{
1317 pg_data_t *pgdat = (pg_data_t *)arg;
1318
1319 (*pos)++;
1320 return next_online_pgdat(pgdat);
1321}
1322
1323static void frag_stop(struct seq_file *m, void *arg)
1324{
1325}
1326
1327/*
1328 * Walk zones in a node and print using a callback.
1329 * If @assert_populated is true, only use callback for zones that are populated.
1330 */
1331static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1332 bool assert_populated, bool nolock,
1333 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1334{
1335 struct zone *zone;
1336 struct zone *node_zones = pgdat->node_zones;
1337 unsigned long flags;
1338
1339 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1340 if (assert_populated && !populated_zone(zone))
1341 continue;
1342
1343 if (!nolock)
1344 spin_lock_irqsave(&zone->lock, flags);
1345 print(m, pgdat, zone);
1346 if (!nolock)
1347 spin_unlock_irqrestore(&zone->lock, flags);
1348 }
1349}
1350#endif
1351
1352#ifdef CONFIG_PROC_FS
1353static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1354 struct zone *zone)
1355{
1356 int order;
1357
1358 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1359 for (order = 0; order < MAX_ORDER; ++order)
1360 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1361 seq_putc(m, '\n');
1362}
1363
1364/*
1365 * This walks the free areas for each zone.
1366 */
1367static int frag_show(struct seq_file *m, void *arg)
1368{
1369 pg_data_t *pgdat = (pg_data_t *)arg;
1370 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1371 return 0;
1372}
1373
1374static void pagetypeinfo_showfree_print(struct seq_file *m,
1375 pg_data_t *pgdat, struct zone *zone)
1376{
1377 int order, mtype;
1378
1379 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1380 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1381 pgdat->node_id,
1382 zone->name,
1383 migratetype_names[mtype]);
1384 for (order = 0; order < MAX_ORDER; ++order) {
1385 unsigned long freecount = 0;
1386 struct free_area *area;
1387 struct list_head *curr;
1388
1389 area = &(zone->free_area[order]);
1390
1391 list_for_each(curr, &area->free_list[mtype])
1392 freecount++;
1393 seq_printf(m, "%6lu ", freecount);
1394 }
1395 seq_putc(m, '\n');
1396 }
1397}
1398
1399/* Print out the free pages at each order for each migatetype */
1400static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1401{
1402 int order;
1403 pg_data_t *pgdat = (pg_data_t *)arg;
1404
1405 /* Print header */
1406 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1407 for (order = 0; order < MAX_ORDER; ++order)
1408 seq_printf(m, "%6d ", order);
1409 seq_putc(m, '\n');
1410
1411 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1412
1413 return 0;
1414}
1415
1416static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1417 pg_data_t *pgdat, struct zone *zone)
1418{
1419 int mtype;
1420 unsigned long pfn;
1421 unsigned long start_pfn = zone->zone_start_pfn;
1422 unsigned long end_pfn = zone_end_pfn(zone);
1423 unsigned long count[MIGRATE_TYPES] = { 0, };
1424
1425 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1426 struct page *page;
1427
1428 page = pfn_to_online_page(pfn);
1429 if (!page)
1430 continue;
1431
1432 /* Watch for unexpected holes punched in the memmap */
1433 if (!memmap_valid_within(pfn, page, zone))
1434 continue;
1435
1436 if (page_zone(page) != zone)
1437 continue;
1438
1439 mtype = get_pageblock_migratetype(page);
1440
1441 if (mtype < MIGRATE_TYPES)
1442 count[mtype]++;
1443 }
1444
1445 /* Print counts */
1446 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1447 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1448 seq_printf(m, "%12lu ", count[mtype]);
1449 seq_putc(m, '\n');
1450}
1451
1452/* Print out the number of pageblocks for each migratetype */
1453static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1454{
1455 int mtype;
1456 pg_data_t *pgdat = (pg_data_t *)arg;
1457
1458 seq_printf(m, "\n%-23s", "Number of blocks type ");
1459 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1460 seq_printf(m, "%12s ", migratetype_names[mtype]);
1461 seq_putc(m, '\n');
1462 walk_zones_in_node(m, pgdat, true, false,
1463 pagetypeinfo_showblockcount_print);
1464
1465 return 0;
1466}
1467
1468/*
1469 * Print out the number of pageblocks for each migratetype that contain pages
1470 * of other types. This gives an indication of how well fallbacks are being
1471 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1472 * to determine what is going on
1473 */
1474static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1475{
1476#ifdef CONFIG_PAGE_OWNER
1477 int mtype;
1478
1479 if (!static_branch_unlikely(&page_owner_inited))
1480 return;
1481
1482 drain_all_pages(NULL);
1483
1484 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1485 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1486 seq_printf(m, "%12s ", migratetype_names[mtype]);
1487 seq_putc(m, '\n');
1488
1489 walk_zones_in_node(m, pgdat, true, true,
1490 pagetypeinfo_showmixedcount_print);
1491#endif /* CONFIG_PAGE_OWNER */
1492}
1493
1494/*
1495 * This prints out statistics in relation to grouping pages by mobility.
1496 * It is expensive to collect so do not constantly read the file.
1497 */
1498static int pagetypeinfo_show(struct seq_file *m, void *arg)
1499{
1500 pg_data_t *pgdat = (pg_data_t *)arg;
1501
1502 /* check memoryless node */
1503 if (!node_state(pgdat->node_id, N_MEMORY))
1504 return 0;
1505
1506 seq_printf(m, "Page block order: %d\n", pageblock_order);
1507 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1508 seq_putc(m, '\n');
1509 pagetypeinfo_showfree(m, pgdat);
1510 pagetypeinfo_showblockcount(m, pgdat);
1511 pagetypeinfo_showmixedcount(m, pgdat);
1512
1513 return 0;
1514}
1515
1516static const struct seq_operations fragmentation_op = {
1517 .start = frag_start,
1518 .next = frag_next,
1519 .stop = frag_stop,
1520 .show = frag_show,
1521};
1522
1523static const struct seq_operations pagetypeinfo_op = {
1524 .start = frag_start,
1525 .next = frag_next,
1526 .stop = frag_stop,
1527 .show = pagetypeinfo_show,
1528};
1529
1530static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1531{
1532 int zid;
1533
1534 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1535 struct zone *compare = &pgdat->node_zones[zid];
1536
1537 if (populated_zone(compare))
1538 return zone == compare;
1539 }
1540
1541 return false;
1542}
1543
1544static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1545 struct zone *zone)
1546{
1547 int i;
1548 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1549 if (is_zone_first_populated(pgdat, zone)) {
1550 seq_printf(m, "\n per-node stats");
1551 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1552 seq_printf(m, "\n %-12s %lu",
1553 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS +
1554 NR_VM_NUMA_STAT_ITEMS],
1555 node_page_state(pgdat, i));
1556 }
1557 }
1558 seq_printf(m,
1559 "\n pages free %lu"
1560 "\n min %lu"
1561 "\n low %lu"
1562 "\n high %lu"
1563 "\n spanned %lu"
1564 "\n present %lu"
1565 "\n managed %lu",
1566 zone_page_state(zone, NR_FREE_PAGES),
1567 min_wmark_pages(zone),
1568 low_wmark_pages(zone),
1569 high_wmark_pages(zone),
1570 zone->spanned_pages,
1571 zone->present_pages,
1572 zone_managed_pages(zone));
1573
1574 seq_printf(m,
1575 "\n protection: (%ld",
1576 zone->lowmem_reserve[0]);
1577 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1578 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1579 seq_putc(m, ')');
1580
1581 /* If unpopulated, no other information is useful */
1582 if (!populated_zone(zone)) {
1583 seq_putc(m, '\n');
1584 return;
1585 }
1586
1587 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1588 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1589 zone_page_state(zone, i));
1590
1591#ifdef CONFIG_NUMA
1592 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1593 seq_printf(m, "\n %-12s %lu",
1594 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
1595 zone_numa_state_snapshot(zone, i));
1596#endif
1597
1598 seq_printf(m, "\n pagesets");
1599 for_each_online_cpu(i) {
1600 struct per_cpu_pageset *pageset;
1601
1602 pageset = per_cpu_ptr(zone->pageset, i);
1603 seq_printf(m,
1604 "\n cpu: %i"
1605 "\n count: %i"
1606 "\n high: %i"
1607 "\n batch: %i",
1608 i,
1609 pageset->pcp.count,
1610 pageset->pcp.high,
1611 pageset->pcp.batch);
1612#ifdef CONFIG_SMP
1613 seq_printf(m, "\n vm stats threshold: %d",
1614 pageset->stat_threshold);
1615#endif
1616 }
1617 seq_printf(m,
1618 "\n node_unreclaimable: %u"
1619 "\n start_pfn: %lu",
1620 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1621 zone->zone_start_pfn);
1622 seq_putc(m, '\n');
1623}
1624
1625/*
1626 * Output information about zones in @pgdat. All zones are printed regardless
1627 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1628 * set of all zones and userspace would not be aware of such zones if they are
1629 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1630 */
1631static int zoneinfo_show(struct seq_file *m, void *arg)
1632{
1633 pg_data_t *pgdat = (pg_data_t *)arg;
1634 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1635 return 0;
1636}
1637
1638static const struct seq_operations zoneinfo_op = {
1639 .start = frag_start, /* iterate over all zones. The same as in
1640 * fragmentation. */
1641 .next = frag_next,
1642 .stop = frag_stop,
1643 .show = zoneinfo_show,
1644};
1645
1646enum writeback_stat_item {
1647 NR_DIRTY_THRESHOLD,
1648 NR_DIRTY_BG_THRESHOLD,
1649 NR_VM_WRITEBACK_STAT_ITEMS,
1650};
1651
1652static void *vmstat_start(struct seq_file *m, loff_t *pos)
1653{
1654 unsigned long *v;
1655 int i, stat_items_size;
1656
1657 if (*pos >= ARRAY_SIZE(vmstat_text))
1658 return NULL;
1659 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1660 NR_VM_NUMA_STAT_ITEMS * sizeof(unsigned long) +
1661 NR_VM_NODE_STAT_ITEMS * sizeof(unsigned long) +
1662 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1663
1664#ifdef CONFIG_VM_EVENT_COUNTERS
1665 stat_items_size += sizeof(struct vm_event_state);
1666#endif
1667
1668 BUILD_BUG_ON(stat_items_size !=
1669 ARRAY_SIZE(vmstat_text) * sizeof(unsigned long));
1670 v = kmalloc(stat_items_size, GFP_KERNEL);
1671 m->private = v;
1672 if (!v)
1673 return ERR_PTR(-ENOMEM);
1674 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1675 v[i] = global_zone_page_state(i);
1676 v += NR_VM_ZONE_STAT_ITEMS;
1677
1678#ifdef CONFIG_NUMA
1679 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1680 v[i] = global_numa_state(i);
1681 v += NR_VM_NUMA_STAT_ITEMS;
1682#endif
1683
1684 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1685 v[i] = global_node_page_state(i);
1686 v += NR_VM_NODE_STAT_ITEMS;
1687
1688 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1689 v + NR_DIRTY_THRESHOLD);
1690 v += NR_VM_WRITEBACK_STAT_ITEMS;
1691
1692#ifdef CONFIG_VM_EVENT_COUNTERS
1693 all_vm_events(v);
1694 v[PGPGIN] /= 2; /* sectors -> kbytes */
1695 v[PGPGOUT] /= 2;
1696#endif
1697 return (unsigned long *)m->private + *pos;
1698}
1699
1700static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1701{
1702 (*pos)++;
1703 if (*pos >= ARRAY_SIZE(vmstat_text))
1704 return NULL;
1705 return (unsigned long *)m->private + *pos;
1706}
1707
1708static int vmstat_show(struct seq_file *m, void *arg)
1709{
1710 unsigned long *l = arg;
1711 unsigned long off = l - (unsigned long *)m->private;
1712
1713 seq_puts(m, vmstat_text[off]);
1714 seq_put_decimal_ull(m, " ", *l);
1715 seq_putc(m, '\n');
1716 return 0;
1717}
1718
1719static void vmstat_stop(struct seq_file *m, void *arg)
1720{
1721 kfree(m->private);
1722 m->private = NULL;
1723}
1724
1725static const struct seq_operations vmstat_op = {
1726 .start = vmstat_start,
1727 .next = vmstat_next,
1728 .stop = vmstat_stop,
1729 .show = vmstat_show,
1730};
1731#endif /* CONFIG_PROC_FS */
1732
1733#ifdef CONFIG_SMP
1734static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1735int sysctl_stat_interval __read_mostly = HZ;
1736
1737#ifdef CONFIG_PROC_FS
1738static void refresh_vm_stats(struct work_struct *work)
1739{
1740 refresh_cpu_vm_stats(true);
1741}
1742
1743int vmstat_refresh(struct ctl_table *table, int write,
1744 void __user *buffer, size_t *lenp, loff_t *ppos)
1745{
1746 long val;
1747 int err;
1748 int i;
1749
1750 /*
1751 * The regular update, every sysctl_stat_interval, may come later
1752 * than expected: leaving a significant amount in per_cpu buckets.
1753 * This is particularly misleading when checking a quantity of HUGE
1754 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1755 * which can equally be echo'ed to or cat'ted from (by root),
1756 * can be used to update the stats just before reading them.
1757 *
1758 * Oh, and since global_zone_page_state() etc. are so careful to hide
1759 * transiently negative values, report an error here if any of
1760 * the stats is negative, so we know to go looking for imbalance.
1761 */
1762 err = schedule_on_each_cpu(refresh_vm_stats);
1763 if (err)
1764 return err;
1765 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1766 val = atomic_long_read(&vm_zone_stat[i]);
1767 if (val < 0) {
1768 pr_warn("%s: %s %ld\n",
1769 __func__, vmstat_text[i], val);
1770 err = -EINVAL;
1771 }
1772 }
1773#ifdef CONFIG_NUMA
1774 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1775 val = atomic_long_read(&vm_numa_stat[i]);
1776 if (val < 0) {
1777 pr_warn("%s: %s %ld\n",
1778 __func__, vmstat_text[i + NR_VM_ZONE_STAT_ITEMS], val);
1779 err = -EINVAL;
1780 }
1781 }
1782#endif
1783 if (err)
1784 return err;
1785 if (write)
1786 *ppos += *lenp;
1787 else
1788 *lenp = 0;
1789 return 0;
1790}
1791#endif /* CONFIG_PROC_FS */
1792
1793static void vmstat_update(struct work_struct *w)
1794{
1795 if (refresh_cpu_vm_stats(true)) {
1796 /*
1797 * Counters were updated so we expect more updates
1798 * to occur in the future. Keep on running the
1799 * update worker thread.
1800 */
1801 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1802 this_cpu_ptr(&vmstat_work),
1803 round_jiffies_relative(sysctl_stat_interval));
1804 }
1805}
1806
1807/*
1808 * Switch off vmstat processing and then fold all the remaining differentials
1809 * until the diffs stay at zero. The function is used by NOHZ and can only be
1810 * invoked when tick processing is not active.
1811 */
1812/*
1813 * Check if the diffs for a certain cpu indicate that
1814 * an update is needed.
1815 */
1816static bool need_update(int cpu)
1817{
1818 struct zone *zone;
1819
1820 for_each_populated_zone(zone) {
1821 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1822
1823 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1824#ifdef CONFIG_NUMA
1825 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1826#endif
1827
1828 /*
1829 * The fast way of checking if there are any vmstat diffs.
1830 */
1831 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1832 sizeof(p->vm_stat_diff[0])))
1833 return true;
1834#ifdef CONFIG_NUMA
1835 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1836 sizeof(p->vm_numa_stat_diff[0])))
1837 return true;
1838#endif
1839 }
1840 return false;
1841}
1842
1843/*
1844 * Switch off vmstat processing and then fold all the remaining differentials
1845 * until the diffs stay at zero. The function is used by NOHZ and can only be
1846 * invoked when tick processing is not active.
1847 */
1848void quiet_vmstat(void)
1849{
1850 if (system_state != SYSTEM_RUNNING)
1851 return;
1852
1853 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1854 return;
1855
1856 if (!need_update(smp_processor_id()))
1857 return;
1858
1859 /*
1860 * Just refresh counters and do not care about the pending delayed
1861 * vmstat_update. It doesn't fire that often to matter and canceling
1862 * it would be too expensive from this path.
1863 * vmstat_shepherd will take care about that for us.
1864 */
1865 refresh_cpu_vm_stats(false);
1866}
1867
1868/*
1869 * Shepherd worker thread that checks the
1870 * differentials of processors that have their worker
1871 * threads for vm statistics updates disabled because of
1872 * inactivity.
1873 */
1874static void vmstat_shepherd(struct work_struct *w);
1875
1876static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1877
1878static void vmstat_shepherd(struct work_struct *w)
1879{
1880 int cpu;
1881
1882 get_online_cpus();
1883 /* Check processors whose vmstat worker threads have been disabled */
1884 for_each_online_cpu(cpu) {
1885 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1886
1887 if (!delayed_work_pending(dw) && need_update(cpu))
1888 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1889 }
1890 put_online_cpus();
1891
1892 schedule_delayed_work(&shepherd,
1893 round_jiffies_relative(sysctl_stat_interval));
1894}
1895
1896static void __init start_shepherd_timer(void)
1897{
1898 int cpu;
1899
1900 for_each_possible_cpu(cpu)
1901 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1902 vmstat_update);
1903
1904 schedule_delayed_work(&shepherd,
1905 round_jiffies_relative(sysctl_stat_interval));
1906}
1907
1908static void __init init_cpu_node_state(void)
1909{
1910 int node;
1911
1912 for_each_online_node(node) {
1913 if (cpumask_weight(cpumask_of_node(node)) > 0)
1914 node_set_state(node, N_CPU);
1915 }
1916}
1917
1918static int vmstat_cpu_online(unsigned int cpu)
1919{
1920 refresh_zone_stat_thresholds();
1921 node_set_state(cpu_to_node(cpu), N_CPU);
1922 return 0;
1923}
1924
1925static int vmstat_cpu_down_prep(unsigned int cpu)
1926{
1927 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1928 return 0;
1929}
1930
1931static int vmstat_cpu_dead(unsigned int cpu)
1932{
1933 const struct cpumask *node_cpus;
1934 int node;
1935
1936 node = cpu_to_node(cpu);
1937
1938 refresh_zone_stat_thresholds();
1939 node_cpus = cpumask_of_node(node);
1940 if (cpumask_weight(node_cpus) > 0)
1941 return 0;
1942
1943 node_clear_state(node, N_CPU);
1944 return 0;
1945}
1946
1947#endif
1948
1949struct workqueue_struct *mm_percpu_wq;
1950
1951void __init init_mm_internals(void)
1952{
1953 int ret __maybe_unused;
1954
1955 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
1956
1957#ifdef CONFIG_SMP
1958 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
1959 NULL, vmstat_cpu_dead);
1960 if (ret < 0)
1961 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1962
1963 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
1964 vmstat_cpu_online,
1965 vmstat_cpu_down_prep);
1966 if (ret < 0)
1967 pr_err("vmstat: failed to register 'online' hotplug state\n");
1968
1969 get_online_cpus();
1970 init_cpu_node_state();
1971 put_online_cpus();
1972
1973 start_shepherd_timer();
1974#endif
1975#ifdef CONFIG_PROC_FS
1976 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
1977 proc_create_seq("pagetypeinfo", 0444, NULL, &pagetypeinfo_op);
1978 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
1979 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
1980#endif
1981}
1982
1983#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1984
1985/*
1986 * Return an index indicating how much of the available free memory is
1987 * unusable for an allocation of the requested size.
1988 */
1989static int unusable_free_index(unsigned int order,
1990 struct contig_page_info *info)
1991{
1992 /* No free memory is interpreted as all free memory is unusable */
1993 if (info->free_pages == 0)
1994 return 1000;
1995
1996 /*
1997 * Index should be a value between 0 and 1. Return a value to 3
1998 * decimal places.
1999 *
2000 * 0 => no fragmentation
2001 * 1 => high fragmentation
2002 */
2003 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2004
2005}
2006
2007static void unusable_show_print(struct seq_file *m,
2008 pg_data_t *pgdat, struct zone *zone)
2009{
2010 unsigned int order;
2011 int index;
2012 struct contig_page_info info;
2013
2014 seq_printf(m, "Node %d, zone %8s ",
2015 pgdat->node_id,
2016 zone->name);
2017 for (order = 0; order < MAX_ORDER; ++order) {
2018 fill_contig_page_info(zone, order, &info);
2019 index = unusable_free_index(order, &info);
2020 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2021 }
2022
2023 seq_putc(m, '\n');
2024}
2025
2026/*
2027 * Display unusable free space index
2028 *
2029 * The unusable free space index measures how much of the available free
2030 * memory cannot be used to satisfy an allocation of a given size and is a
2031 * value between 0 and 1. The higher the value, the more of free memory is
2032 * unusable and by implication, the worse the external fragmentation is. This
2033 * can be expressed as a percentage by multiplying by 100.
2034 */
2035static int unusable_show(struct seq_file *m, void *arg)
2036{
2037 pg_data_t *pgdat = (pg_data_t *)arg;
2038
2039 /* check memoryless node */
2040 if (!node_state(pgdat->node_id, N_MEMORY))
2041 return 0;
2042
2043 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2044
2045 return 0;
2046}
2047
2048static const struct seq_operations unusable_op = {
2049 .start = frag_start,
2050 .next = frag_next,
2051 .stop = frag_stop,
2052 .show = unusable_show,
2053};
2054
2055static int unusable_open(struct inode *inode, struct file *file)
2056{
2057 return seq_open(file, &unusable_op);
2058}
2059
2060static const struct file_operations unusable_file_ops = {
2061 .open = unusable_open,
2062 .read = seq_read,
2063 .llseek = seq_lseek,
2064 .release = seq_release,
2065};
2066
2067static void extfrag_show_print(struct seq_file *m,
2068 pg_data_t *pgdat, struct zone *zone)
2069{
2070 unsigned int order;
2071 int index;
2072
2073 /* Alloc on stack as interrupts are disabled for zone walk */
2074 struct contig_page_info info;
2075
2076 seq_printf(m, "Node %d, zone %8s ",
2077 pgdat->node_id,
2078 zone->name);
2079 for (order = 0; order < MAX_ORDER; ++order) {
2080 fill_contig_page_info(zone, order, &info);
2081 index = __fragmentation_index(order, &info);
2082 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2083 }
2084
2085 seq_putc(m, '\n');
2086}
2087
2088/*
2089 * Display fragmentation index for orders that allocations would fail for
2090 */
2091static int extfrag_show(struct seq_file *m, void *arg)
2092{
2093 pg_data_t *pgdat = (pg_data_t *)arg;
2094
2095 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2096
2097 return 0;
2098}
2099
2100static const struct seq_operations extfrag_op = {
2101 .start = frag_start,
2102 .next = frag_next,
2103 .stop = frag_stop,
2104 .show = extfrag_show,
2105};
2106
2107static int extfrag_open(struct inode *inode, struct file *file)
2108{
2109 return seq_open(file, &extfrag_op);
2110}
2111
2112static const struct file_operations extfrag_file_ops = {
2113 .open = extfrag_open,
2114 .read = seq_read,
2115 .llseek = seq_lseek,
2116 .release = seq_release,
2117};
2118
2119static int __init extfrag_debug_init(void)
2120{
2121 struct dentry *extfrag_debug_root;
2122
2123 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2124
2125 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2126 &unusable_file_ops);
2127
2128 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2129 &extfrag_file_ops);
2130
2131 return 0;
2132}
2133
2134module_init(extfrag_debug_init);
2135#endif
2136