Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmstat.c
4 *
5 * Manages VM statistics
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 *
8 * zoned VM statistics
9 * Copyright (C) 2006 Silicon Graphics, Inc.,
10 * Christoph Lameter <christoph@lameter.com>
11 * Copyright (C) 2008-2014 Christoph Lameter
12 */
13#include <linux/fs.h>
14#include <linux/mm.h>
15#include <linux/err.h>
16#include <linux/module.h>
17#include <linux/slab.h>
18#include <linux/cpu.h>
19#include <linux/cpumask.h>
20#include <linux/vmstat.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/debugfs.h>
24#include <linux/sched.h>
25#include <linux/math64.h>
26#include <linux/writeback.h>
27#include <linux/compaction.h>
28#include <linux/mm_inline.h>
29#include <linux/page_ext.h>
30#include <linux/page_owner.h>
31
32#include "internal.h"
33
34#define NUMA_STATS_THRESHOLD (U16_MAX - 2)
35
36#ifdef CONFIG_NUMA
37int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
38
39/* zero numa counters within a zone */
40static void zero_zone_numa_counters(struct zone *zone)
41{
42 int item, cpu;
43
44 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
45 atomic_long_set(&zone->vm_numa_stat[item], 0);
46 for_each_online_cpu(cpu)
47 per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
48 = 0;
49 }
50}
51
52/* zero numa counters of all the populated zones */
53static void zero_zones_numa_counters(void)
54{
55 struct zone *zone;
56
57 for_each_populated_zone(zone)
58 zero_zone_numa_counters(zone);
59}
60
61/* zero global numa counters */
62static void zero_global_numa_counters(void)
63{
64 int item;
65
66 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
67 atomic_long_set(&vm_numa_stat[item], 0);
68}
69
70static void invalid_numa_statistics(void)
71{
72 zero_zones_numa_counters();
73 zero_global_numa_counters();
74}
75
76static DEFINE_MUTEX(vm_numa_stat_lock);
77
78int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
79 void *buffer, size_t *length, loff_t *ppos)
80{
81 int ret, oldval;
82
83 mutex_lock(&vm_numa_stat_lock);
84 if (write)
85 oldval = sysctl_vm_numa_stat;
86 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
87 if (ret || !write)
88 goto out;
89
90 if (oldval == sysctl_vm_numa_stat)
91 goto out;
92 else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
93 static_branch_enable(&vm_numa_stat_key);
94 pr_info("enable numa statistics\n");
95 } else {
96 static_branch_disable(&vm_numa_stat_key);
97 invalid_numa_statistics();
98 pr_info("disable numa statistics, and clear numa counters\n");
99 }
100
101out:
102 mutex_unlock(&vm_numa_stat_lock);
103 return ret;
104}
105#endif
106
107#ifdef CONFIG_VM_EVENT_COUNTERS
108DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
109EXPORT_PER_CPU_SYMBOL(vm_event_states);
110
111static void sum_vm_events(unsigned long *ret)
112{
113 int cpu;
114 int i;
115
116 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
117
118 for_each_online_cpu(cpu) {
119 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
120
121 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
122 ret[i] += this->event[i];
123 }
124}
125
126/*
127 * Accumulate the vm event counters across all CPUs.
128 * The result is unavoidably approximate - it can change
129 * during and after execution of this function.
130*/
131void all_vm_events(unsigned long *ret)
132{
133 get_online_cpus();
134 sum_vm_events(ret);
135 put_online_cpus();
136}
137EXPORT_SYMBOL_GPL(all_vm_events);
138
139/*
140 * Fold the foreign cpu events into our own.
141 *
142 * This is adding to the events on one processor
143 * but keeps the global counts constant.
144 */
145void vm_events_fold_cpu(int cpu)
146{
147 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
148 int i;
149
150 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
151 count_vm_events(i, fold_state->event[i]);
152 fold_state->event[i] = 0;
153 }
154}
155
156#endif /* CONFIG_VM_EVENT_COUNTERS */
157
158/*
159 * Manage combined zone based / global counters
160 *
161 * vm_stat contains the global counters
162 */
163atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
165atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
166EXPORT_SYMBOL(vm_zone_stat);
167EXPORT_SYMBOL(vm_numa_stat);
168EXPORT_SYMBOL(vm_node_stat);
169
170#ifdef CONFIG_SMP
171
172int calculate_pressure_threshold(struct zone *zone)
173{
174 int threshold;
175 int watermark_distance;
176
177 /*
178 * As vmstats are not up to date, there is drift between the estimated
179 * and real values. For high thresholds and a high number of CPUs, it
180 * is possible for the min watermark to be breached while the estimated
181 * value looks fine. The pressure threshold is a reduced value such
182 * that even the maximum amount of drift will not accidentally breach
183 * the min watermark
184 */
185 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
186 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
187
188 /*
189 * Maximum threshold is 125
190 */
191 threshold = min(125, threshold);
192
193 return threshold;
194}
195
196int calculate_normal_threshold(struct zone *zone)
197{
198 int threshold;
199 int mem; /* memory in 128 MB units */
200
201 /*
202 * The threshold scales with the number of processors and the amount
203 * of memory per zone. More memory means that we can defer updates for
204 * longer, more processors could lead to more contention.
205 * fls() is used to have a cheap way of logarithmic scaling.
206 *
207 * Some sample thresholds:
208 *
209 * Threshold Processors (fls) Zonesize fls(mem+1)
210 * ------------------------------------------------------------------
211 * 8 1 1 0.9-1 GB 4
212 * 16 2 2 0.9-1 GB 4
213 * 20 2 2 1-2 GB 5
214 * 24 2 2 2-4 GB 6
215 * 28 2 2 4-8 GB 7
216 * 32 2 2 8-16 GB 8
217 * 4 2 2 <128M 1
218 * 30 4 3 2-4 GB 5
219 * 48 4 3 8-16 GB 8
220 * 32 8 4 1-2 GB 4
221 * 32 8 4 0.9-1GB 4
222 * 10 16 5 <128M 1
223 * 40 16 5 900M 4
224 * 70 64 7 2-4 GB 5
225 * 84 64 7 4-8 GB 6
226 * 108 512 9 4-8 GB 6
227 * 125 1024 10 8-16 GB 8
228 * 125 1024 10 16-32 GB 9
229 */
230
231 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
232
233 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
234
235 /*
236 * Maximum threshold is 125
237 */
238 threshold = min(125, threshold);
239
240 return threshold;
241}
242
243/*
244 * Refresh the thresholds for each zone.
245 */
246void refresh_zone_stat_thresholds(void)
247{
248 struct pglist_data *pgdat;
249 struct zone *zone;
250 int cpu;
251 int threshold;
252
253 /* Zero current pgdat thresholds */
254 for_each_online_pgdat(pgdat) {
255 for_each_online_cpu(cpu) {
256 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
257 }
258 }
259
260 for_each_populated_zone(zone) {
261 struct pglist_data *pgdat = zone->zone_pgdat;
262 unsigned long max_drift, tolerate_drift;
263
264 threshold = calculate_normal_threshold(zone);
265
266 for_each_online_cpu(cpu) {
267 int pgdat_threshold;
268
269 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
270 = threshold;
271
272 /* Base nodestat threshold on the largest populated zone. */
273 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
274 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
275 = max(threshold, pgdat_threshold);
276 }
277
278 /*
279 * Only set percpu_drift_mark if there is a danger that
280 * NR_FREE_PAGES reports the low watermark is ok when in fact
281 * the min watermark could be breached by an allocation
282 */
283 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
284 max_drift = num_online_cpus() * threshold;
285 if (max_drift > tolerate_drift)
286 zone->percpu_drift_mark = high_wmark_pages(zone) +
287 max_drift;
288 }
289}
290
291void set_pgdat_percpu_threshold(pg_data_t *pgdat,
292 int (*calculate_pressure)(struct zone *))
293{
294 struct zone *zone;
295 int cpu;
296 int threshold;
297 int i;
298
299 for (i = 0; i < pgdat->nr_zones; i++) {
300 zone = &pgdat->node_zones[i];
301 if (!zone->percpu_drift_mark)
302 continue;
303
304 threshold = (*calculate_pressure)(zone);
305 for_each_online_cpu(cpu)
306 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
307 = threshold;
308 }
309}
310
311/*
312 * For use when we know that interrupts are disabled,
313 * or when we know that preemption is disabled and that
314 * particular counter cannot be updated from interrupt context.
315 */
316void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
317 long delta)
318{
319 struct per_cpu_pageset __percpu *pcp = zone->pageset;
320 s8 __percpu *p = pcp->vm_stat_diff + item;
321 long x;
322 long t;
323
324 x = delta + __this_cpu_read(*p);
325
326 t = __this_cpu_read(pcp->stat_threshold);
327
328 if (unlikely(x > t || x < -t)) {
329 zone_page_state_add(x, zone, item);
330 x = 0;
331 }
332 __this_cpu_write(*p, x);
333}
334EXPORT_SYMBOL(__mod_zone_page_state);
335
336void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
337 long delta)
338{
339 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
340 s8 __percpu *p = pcp->vm_node_stat_diff + item;
341 long x;
342 long t;
343
344 if (vmstat_item_in_bytes(item)) {
345 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
346 delta >>= PAGE_SHIFT;
347 }
348
349 x = delta + __this_cpu_read(*p);
350
351 t = __this_cpu_read(pcp->stat_threshold);
352
353 if (unlikely(x > t || x < -t)) {
354 node_page_state_add(x, pgdat, item);
355 x = 0;
356 }
357 __this_cpu_write(*p, x);
358}
359EXPORT_SYMBOL(__mod_node_page_state);
360
361/*
362 * Optimized increment and decrement functions.
363 *
364 * These are only for a single page and therefore can take a struct page *
365 * argument instead of struct zone *. This allows the inclusion of the code
366 * generated for page_zone(page) into the optimized functions.
367 *
368 * No overflow check is necessary and therefore the differential can be
369 * incremented or decremented in place which may allow the compilers to
370 * generate better code.
371 * The increment or decrement is known and therefore one boundary check can
372 * be omitted.
373 *
374 * NOTE: These functions are very performance sensitive. Change only
375 * with care.
376 *
377 * Some processors have inc/dec instructions that are atomic vs an interrupt.
378 * However, the code must first determine the differential location in a zone
379 * based on the processor number and then inc/dec the counter. There is no
380 * guarantee without disabling preemption that the processor will not change
381 * in between and therefore the atomicity vs. interrupt cannot be exploited
382 * in a useful way here.
383 */
384void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
385{
386 struct per_cpu_pageset __percpu *pcp = zone->pageset;
387 s8 __percpu *p = pcp->vm_stat_diff + item;
388 s8 v, t;
389
390 v = __this_cpu_inc_return(*p);
391 t = __this_cpu_read(pcp->stat_threshold);
392 if (unlikely(v > t)) {
393 s8 overstep = t >> 1;
394
395 zone_page_state_add(v + overstep, zone, item);
396 __this_cpu_write(*p, -overstep);
397 }
398}
399
400void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
401{
402 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
403 s8 __percpu *p = pcp->vm_node_stat_diff + item;
404 s8 v, t;
405
406 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
407
408 v = __this_cpu_inc_return(*p);
409 t = __this_cpu_read(pcp->stat_threshold);
410 if (unlikely(v > t)) {
411 s8 overstep = t >> 1;
412
413 node_page_state_add(v + overstep, pgdat, item);
414 __this_cpu_write(*p, -overstep);
415 }
416}
417
418void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
419{
420 __inc_zone_state(page_zone(page), item);
421}
422EXPORT_SYMBOL(__inc_zone_page_state);
423
424void __inc_node_page_state(struct page *page, enum node_stat_item item)
425{
426 __inc_node_state(page_pgdat(page), item);
427}
428EXPORT_SYMBOL(__inc_node_page_state);
429
430void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
431{
432 struct per_cpu_pageset __percpu *pcp = zone->pageset;
433 s8 __percpu *p = pcp->vm_stat_diff + item;
434 s8 v, t;
435
436 v = __this_cpu_dec_return(*p);
437 t = __this_cpu_read(pcp->stat_threshold);
438 if (unlikely(v < - t)) {
439 s8 overstep = t >> 1;
440
441 zone_page_state_add(v - overstep, zone, item);
442 __this_cpu_write(*p, overstep);
443 }
444}
445
446void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
447{
448 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
449 s8 __percpu *p = pcp->vm_node_stat_diff + item;
450 s8 v, t;
451
452 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
453
454 v = __this_cpu_dec_return(*p);
455 t = __this_cpu_read(pcp->stat_threshold);
456 if (unlikely(v < - t)) {
457 s8 overstep = t >> 1;
458
459 node_page_state_add(v - overstep, pgdat, item);
460 __this_cpu_write(*p, overstep);
461 }
462}
463
464void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
465{
466 __dec_zone_state(page_zone(page), item);
467}
468EXPORT_SYMBOL(__dec_zone_page_state);
469
470void __dec_node_page_state(struct page *page, enum node_stat_item item)
471{
472 __dec_node_state(page_pgdat(page), item);
473}
474EXPORT_SYMBOL(__dec_node_page_state);
475
476#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
477/*
478 * If we have cmpxchg_local support then we do not need to incur the overhead
479 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
480 *
481 * mod_state() modifies the zone counter state through atomic per cpu
482 * operations.
483 *
484 * Overstep mode specifies how overstep should handled:
485 * 0 No overstepping
486 * 1 Overstepping half of threshold
487 * -1 Overstepping minus half of threshold
488*/
489static inline void mod_zone_state(struct zone *zone,
490 enum zone_stat_item item, long delta, int overstep_mode)
491{
492 struct per_cpu_pageset __percpu *pcp = zone->pageset;
493 s8 __percpu *p = pcp->vm_stat_diff + item;
494 long o, n, t, z;
495
496 do {
497 z = 0; /* overflow to zone counters */
498
499 /*
500 * The fetching of the stat_threshold is racy. We may apply
501 * a counter threshold to the wrong the cpu if we get
502 * rescheduled while executing here. However, the next
503 * counter update will apply the threshold again and
504 * therefore bring the counter under the threshold again.
505 *
506 * Most of the time the thresholds are the same anyways
507 * for all cpus in a zone.
508 */
509 t = this_cpu_read(pcp->stat_threshold);
510
511 o = this_cpu_read(*p);
512 n = delta + o;
513
514 if (n > t || n < -t) {
515 int os = overstep_mode * (t >> 1) ;
516
517 /* Overflow must be added to zone counters */
518 z = n + os;
519 n = -os;
520 }
521 } while (this_cpu_cmpxchg(*p, o, n) != o);
522
523 if (z)
524 zone_page_state_add(z, zone, item);
525}
526
527void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
528 long delta)
529{
530 mod_zone_state(zone, item, delta, 0);
531}
532EXPORT_SYMBOL(mod_zone_page_state);
533
534void inc_zone_page_state(struct page *page, enum zone_stat_item item)
535{
536 mod_zone_state(page_zone(page), item, 1, 1);
537}
538EXPORT_SYMBOL(inc_zone_page_state);
539
540void dec_zone_page_state(struct page *page, enum zone_stat_item item)
541{
542 mod_zone_state(page_zone(page), item, -1, -1);
543}
544EXPORT_SYMBOL(dec_zone_page_state);
545
546static inline void mod_node_state(struct pglist_data *pgdat,
547 enum node_stat_item item, int delta, int overstep_mode)
548{
549 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
550 s8 __percpu *p = pcp->vm_node_stat_diff + item;
551 long o, n, t, z;
552
553 if (vmstat_item_in_bytes(item)) {
554 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
555 delta >>= PAGE_SHIFT;
556 }
557
558 do {
559 z = 0; /* overflow to node counters */
560
561 /*
562 * The fetching of the stat_threshold is racy. We may apply
563 * a counter threshold to the wrong the cpu if we get
564 * rescheduled while executing here. However, the next
565 * counter update will apply the threshold again and
566 * therefore bring the counter under the threshold again.
567 *
568 * Most of the time the thresholds are the same anyways
569 * for all cpus in a node.
570 */
571 t = this_cpu_read(pcp->stat_threshold);
572
573 o = this_cpu_read(*p);
574 n = delta + o;
575
576 if (n > t || n < -t) {
577 int os = overstep_mode * (t >> 1) ;
578
579 /* Overflow must be added to node counters */
580 z = n + os;
581 n = -os;
582 }
583 } while (this_cpu_cmpxchg(*p, o, n) != o);
584
585 if (z)
586 node_page_state_add(z, pgdat, item);
587}
588
589void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
590 long delta)
591{
592 mod_node_state(pgdat, item, delta, 0);
593}
594EXPORT_SYMBOL(mod_node_page_state);
595
596void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
597{
598 mod_node_state(pgdat, item, 1, 1);
599}
600
601void inc_node_page_state(struct page *page, enum node_stat_item item)
602{
603 mod_node_state(page_pgdat(page), item, 1, 1);
604}
605EXPORT_SYMBOL(inc_node_page_state);
606
607void dec_node_page_state(struct page *page, enum node_stat_item item)
608{
609 mod_node_state(page_pgdat(page), item, -1, -1);
610}
611EXPORT_SYMBOL(dec_node_page_state);
612#else
613/*
614 * Use interrupt disable to serialize counter updates
615 */
616void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
617 long delta)
618{
619 unsigned long flags;
620
621 local_irq_save(flags);
622 __mod_zone_page_state(zone, item, delta);
623 local_irq_restore(flags);
624}
625EXPORT_SYMBOL(mod_zone_page_state);
626
627void inc_zone_page_state(struct page *page, enum zone_stat_item item)
628{
629 unsigned long flags;
630 struct zone *zone;
631
632 zone = page_zone(page);
633 local_irq_save(flags);
634 __inc_zone_state(zone, item);
635 local_irq_restore(flags);
636}
637EXPORT_SYMBOL(inc_zone_page_state);
638
639void dec_zone_page_state(struct page *page, enum zone_stat_item item)
640{
641 unsigned long flags;
642
643 local_irq_save(flags);
644 __dec_zone_page_state(page, item);
645 local_irq_restore(flags);
646}
647EXPORT_SYMBOL(dec_zone_page_state);
648
649void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
650{
651 unsigned long flags;
652
653 local_irq_save(flags);
654 __inc_node_state(pgdat, item);
655 local_irq_restore(flags);
656}
657EXPORT_SYMBOL(inc_node_state);
658
659void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
660 long delta)
661{
662 unsigned long flags;
663
664 local_irq_save(flags);
665 __mod_node_page_state(pgdat, item, delta);
666 local_irq_restore(flags);
667}
668EXPORT_SYMBOL(mod_node_page_state);
669
670void inc_node_page_state(struct page *page, enum node_stat_item item)
671{
672 unsigned long flags;
673 struct pglist_data *pgdat;
674
675 pgdat = page_pgdat(page);
676 local_irq_save(flags);
677 __inc_node_state(pgdat, item);
678 local_irq_restore(flags);
679}
680EXPORT_SYMBOL(inc_node_page_state);
681
682void dec_node_page_state(struct page *page, enum node_stat_item item)
683{
684 unsigned long flags;
685
686 local_irq_save(flags);
687 __dec_node_page_state(page, item);
688 local_irq_restore(flags);
689}
690EXPORT_SYMBOL(dec_node_page_state);
691#endif
692
693/*
694 * Fold a differential into the global counters.
695 * Returns the number of counters updated.
696 */
697#ifdef CONFIG_NUMA
698static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
699{
700 int i;
701 int changes = 0;
702
703 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
704 if (zone_diff[i]) {
705 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
706 changes++;
707 }
708
709 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
710 if (numa_diff[i]) {
711 atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
712 changes++;
713 }
714
715 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
716 if (node_diff[i]) {
717 atomic_long_add(node_diff[i], &vm_node_stat[i]);
718 changes++;
719 }
720 return changes;
721}
722#else
723static int fold_diff(int *zone_diff, int *node_diff)
724{
725 int i;
726 int changes = 0;
727
728 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
729 if (zone_diff[i]) {
730 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
731 changes++;
732 }
733
734 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
735 if (node_diff[i]) {
736 atomic_long_add(node_diff[i], &vm_node_stat[i]);
737 changes++;
738 }
739 return changes;
740}
741#endif /* CONFIG_NUMA */
742
743/*
744 * Update the zone counters for the current cpu.
745 *
746 * Note that refresh_cpu_vm_stats strives to only access
747 * node local memory. The per cpu pagesets on remote zones are placed
748 * in the memory local to the processor using that pageset. So the
749 * loop over all zones will access a series of cachelines local to
750 * the processor.
751 *
752 * The call to zone_page_state_add updates the cachelines with the
753 * statistics in the remote zone struct as well as the global cachelines
754 * with the global counters. These could cause remote node cache line
755 * bouncing and will have to be only done when necessary.
756 *
757 * The function returns the number of global counters updated.
758 */
759static int refresh_cpu_vm_stats(bool do_pagesets)
760{
761 struct pglist_data *pgdat;
762 struct zone *zone;
763 int i;
764 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
765#ifdef CONFIG_NUMA
766 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
767#endif
768 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
769 int changes = 0;
770
771 for_each_populated_zone(zone) {
772 struct per_cpu_pageset __percpu *p = zone->pageset;
773
774 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
775 int v;
776
777 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
778 if (v) {
779
780 atomic_long_add(v, &zone->vm_stat[i]);
781 global_zone_diff[i] += v;
782#ifdef CONFIG_NUMA
783 /* 3 seconds idle till flush */
784 __this_cpu_write(p->expire, 3);
785#endif
786 }
787 }
788#ifdef CONFIG_NUMA
789 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
790 int v;
791
792 v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
793 if (v) {
794
795 atomic_long_add(v, &zone->vm_numa_stat[i]);
796 global_numa_diff[i] += v;
797 __this_cpu_write(p->expire, 3);
798 }
799 }
800
801 if (do_pagesets) {
802 cond_resched();
803 /*
804 * Deal with draining the remote pageset of this
805 * processor
806 *
807 * Check if there are pages remaining in this pageset
808 * if not then there is nothing to expire.
809 */
810 if (!__this_cpu_read(p->expire) ||
811 !__this_cpu_read(p->pcp.count))
812 continue;
813
814 /*
815 * We never drain zones local to this processor.
816 */
817 if (zone_to_nid(zone) == numa_node_id()) {
818 __this_cpu_write(p->expire, 0);
819 continue;
820 }
821
822 if (__this_cpu_dec_return(p->expire))
823 continue;
824
825 if (__this_cpu_read(p->pcp.count)) {
826 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
827 changes++;
828 }
829 }
830#endif
831 }
832
833 for_each_online_pgdat(pgdat) {
834 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
835
836 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
837 int v;
838
839 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
840 if (v) {
841 atomic_long_add(v, &pgdat->vm_stat[i]);
842 global_node_diff[i] += v;
843 }
844 }
845 }
846
847#ifdef CONFIG_NUMA
848 changes += fold_diff(global_zone_diff, global_numa_diff,
849 global_node_diff);
850#else
851 changes += fold_diff(global_zone_diff, global_node_diff);
852#endif
853 return changes;
854}
855
856/*
857 * Fold the data for an offline cpu into the global array.
858 * There cannot be any access by the offline cpu and therefore
859 * synchronization is simplified.
860 */
861void cpu_vm_stats_fold(int cpu)
862{
863 struct pglist_data *pgdat;
864 struct zone *zone;
865 int i;
866 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
867#ifdef CONFIG_NUMA
868 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
869#endif
870 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
871
872 for_each_populated_zone(zone) {
873 struct per_cpu_pageset *p;
874
875 p = per_cpu_ptr(zone->pageset, cpu);
876
877 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
878 if (p->vm_stat_diff[i]) {
879 int v;
880
881 v = p->vm_stat_diff[i];
882 p->vm_stat_diff[i] = 0;
883 atomic_long_add(v, &zone->vm_stat[i]);
884 global_zone_diff[i] += v;
885 }
886
887#ifdef CONFIG_NUMA
888 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
889 if (p->vm_numa_stat_diff[i]) {
890 int v;
891
892 v = p->vm_numa_stat_diff[i];
893 p->vm_numa_stat_diff[i] = 0;
894 atomic_long_add(v, &zone->vm_numa_stat[i]);
895 global_numa_diff[i] += v;
896 }
897#endif
898 }
899
900 for_each_online_pgdat(pgdat) {
901 struct per_cpu_nodestat *p;
902
903 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
904
905 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
906 if (p->vm_node_stat_diff[i]) {
907 int v;
908
909 v = p->vm_node_stat_diff[i];
910 p->vm_node_stat_diff[i] = 0;
911 atomic_long_add(v, &pgdat->vm_stat[i]);
912 global_node_diff[i] += v;
913 }
914 }
915
916#ifdef CONFIG_NUMA
917 fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
918#else
919 fold_diff(global_zone_diff, global_node_diff);
920#endif
921}
922
923/*
924 * this is only called if !populated_zone(zone), which implies no other users of
925 * pset->vm_stat_diff[] exsist.
926 */
927void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
928{
929 int i;
930
931 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
932 if (pset->vm_stat_diff[i]) {
933 int v = pset->vm_stat_diff[i];
934 pset->vm_stat_diff[i] = 0;
935 atomic_long_add(v, &zone->vm_stat[i]);
936 atomic_long_add(v, &vm_zone_stat[i]);
937 }
938
939#ifdef CONFIG_NUMA
940 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
941 if (pset->vm_numa_stat_diff[i]) {
942 int v = pset->vm_numa_stat_diff[i];
943
944 pset->vm_numa_stat_diff[i] = 0;
945 atomic_long_add(v, &zone->vm_numa_stat[i]);
946 atomic_long_add(v, &vm_numa_stat[i]);
947 }
948#endif
949}
950#endif
951
952#ifdef CONFIG_NUMA
953void __inc_numa_state(struct zone *zone,
954 enum numa_stat_item item)
955{
956 struct per_cpu_pageset __percpu *pcp = zone->pageset;
957 u16 __percpu *p = pcp->vm_numa_stat_diff + item;
958 u16 v;
959
960 v = __this_cpu_inc_return(*p);
961
962 if (unlikely(v > NUMA_STATS_THRESHOLD)) {
963 zone_numa_state_add(v, zone, item);
964 __this_cpu_write(*p, 0);
965 }
966}
967
968/*
969 * Determine the per node value of a stat item. This function
970 * is called frequently in a NUMA machine, so try to be as
971 * frugal as possible.
972 */
973unsigned long sum_zone_node_page_state(int node,
974 enum zone_stat_item item)
975{
976 struct zone *zones = NODE_DATA(node)->node_zones;
977 int i;
978 unsigned long count = 0;
979
980 for (i = 0; i < MAX_NR_ZONES; i++)
981 count += zone_page_state(zones + i, item);
982
983 return count;
984}
985
986/*
987 * Determine the per node value of a numa stat item. To avoid deviation,
988 * the per cpu stat number in vm_numa_stat_diff[] is also included.
989 */
990unsigned long sum_zone_numa_state(int node,
991 enum numa_stat_item item)
992{
993 struct zone *zones = NODE_DATA(node)->node_zones;
994 int i;
995 unsigned long count = 0;
996
997 for (i = 0; i < MAX_NR_ZONES; i++)
998 count += zone_numa_state_snapshot(zones + i, item);
999
1000 return count;
1001}
1002
1003/*
1004 * Determine the per node value of a stat item.
1005 */
1006unsigned long node_page_state_pages(struct pglist_data *pgdat,
1007 enum node_stat_item item)
1008{
1009 long x = atomic_long_read(&pgdat->vm_stat[item]);
1010#ifdef CONFIG_SMP
1011 if (x < 0)
1012 x = 0;
1013#endif
1014 return x;
1015}
1016
1017unsigned long node_page_state(struct pglist_data *pgdat,
1018 enum node_stat_item item)
1019{
1020 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1021
1022 return node_page_state_pages(pgdat, item);
1023}
1024#endif
1025
1026#ifdef CONFIG_COMPACTION
1027
1028struct contig_page_info {
1029 unsigned long free_pages;
1030 unsigned long free_blocks_total;
1031 unsigned long free_blocks_suitable;
1032};
1033
1034/*
1035 * Calculate the number of free pages in a zone, how many contiguous
1036 * pages are free and how many are large enough to satisfy an allocation of
1037 * the target size. Note that this function makes no attempt to estimate
1038 * how many suitable free blocks there *might* be if MOVABLE pages were
1039 * migrated. Calculating that is possible, but expensive and can be
1040 * figured out from userspace
1041 */
1042static void fill_contig_page_info(struct zone *zone,
1043 unsigned int suitable_order,
1044 struct contig_page_info *info)
1045{
1046 unsigned int order;
1047
1048 info->free_pages = 0;
1049 info->free_blocks_total = 0;
1050 info->free_blocks_suitable = 0;
1051
1052 for (order = 0; order < MAX_ORDER; order++) {
1053 unsigned long blocks;
1054
1055 /* Count number of free blocks */
1056 blocks = zone->free_area[order].nr_free;
1057 info->free_blocks_total += blocks;
1058
1059 /* Count free base pages */
1060 info->free_pages += blocks << order;
1061
1062 /* Count the suitable free blocks */
1063 if (order >= suitable_order)
1064 info->free_blocks_suitable += blocks <<
1065 (order - suitable_order);
1066 }
1067}
1068
1069/*
1070 * A fragmentation index only makes sense if an allocation of a requested
1071 * size would fail. If that is true, the fragmentation index indicates
1072 * whether external fragmentation or a lack of memory was the problem.
1073 * The value can be used to determine if page reclaim or compaction
1074 * should be used
1075 */
1076static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1077{
1078 unsigned long requested = 1UL << order;
1079
1080 if (WARN_ON_ONCE(order >= MAX_ORDER))
1081 return 0;
1082
1083 if (!info->free_blocks_total)
1084 return 0;
1085
1086 /* Fragmentation index only makes sense when a request would fail */
1087 if (info->free_blocks_suitable)
1088 return -1000;
1089
1090 /*
1091 * Index is between 0 and 1 so return within 3 decimal places
1092 *
1093 * 0 => allocation would fail due to lack of memory
1094 * 1 => allocation would fail due to fragmentation
1095 */
1096 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1097}
1098
1099/*
1100 * Calculates external fragmentation within a zone wrt the given order.
1101 * It is defined as the percentage of pages found in blocks of size
1102 * less than 1 << order. It returns values in range [0, 100].
1103 */
1104unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1105{
1106 struct contig_page_info info;
1107
1108 fill_contig_page_info(zone, order, &info);
1109 if (info.free_pages == 0)
1110 return 0;
1111
1112 return div_u64((info.free_pages -
1113 (info.free_blocks_suitable << order)) * 100,
1114 info.free_pages);
1115}
1116
1117/* Same as __fragmentation index but allocs contig_page_info on stack */
1118int fragmentation_index(struct zone *zone, unsigned int order)
1119{
1120 struct contig_page_info info;
1121
1122 fill_contig_page_info(zone, order, &info);
1123 return __fragmentation_index(order, &info);
1124}
1125#endif
1126
1127#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1128 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1129#ifdef CONFIG_ZONE_DMA
1130#define TEXT_FOR_DMA(xx) xx "_dma",
1131#else
1132#define TEXT_FOR_DMA(xx)
1133#endif
1134
1135#ifdef CONFIG_ZONE_DMA32
1136#define TEXT_FOR_DMA32(xx) xx "_dma32",
1137#else
1138#define TEXT_FOR_DMA32(xx)
1139#endif
1140
1141#ifdef CONFIG_HIGHMEM
1142#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1143#else
1144#define TEXT_FOR_HIGHMEM(xx)
1145#endif
1146
1147#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1148 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1149
1150const char * const vmstat_text[] = {
1151 /* enum zone_stat_item counters */
1152 "nr_free_pages",
1153 "nr_zone_inactive_anon",
1154 "nr_zone_active_anon",
1155 "nr_zone_inactive_file",
1156 "nr_zone_active_file",
1157 "nr_zone_unevictable",
1158 "nr_zone_write_pending",
1159 "nr_mlock",
1160 "nr_page_table_pages",
1161 "nr_bounce",
1162#if IS_ENABLED(CONFIG_ZSMALLOC)
1163 "nr_zspages",
1164#endif
1165 "nr_free_cma",
1166
1167 /* enum numa_stat_item counters */
1168#ifdef CONFIG_NUMA
1169 "numa_hit",
1170 "numa_miss",
1171 "numa_foreign",
1172 "numa_interleave",
1173 "numa_local",
1174 "numa_other",
1175#endif
1176
1177 /* enum node_stat_item counters */
1178 "nr_inactive_anon",
1179 "nr_active_anon",
1180 "nr_inactive_file",
1181 "nr_active_file",
1182 "nr_unevictable",
1183 "nr_slab_reclaimable",
1184 "nr_slab_unreclaimable",
1185 "nr_isolated_anon",
1186 "nr_isolated_file",
1187 "workingset_nodes",
1188 "workingset_refault_anon",
1189 "workingset_refault_file",
1190 "workingset_activate_anon",
1191 "workingset_activate_file",
1192 "workingset_restore_anon",
1193 "workingset_restore_file",
1194 "workingset_nodereclaim",
1195 "nr_anon_pages",
1196 "nr_mapped",
1197 "nr_file_pages",
1198 "nr_dirty",
1199 "nr_writeback",
1200 "nr_writeback_temp",
1201 "nr_shmem",
1202 "nr_shmem_hugepages",
1203 "nr_shmem_pmdmapped",
1204 "nr_file_hugepages",
1205 "nr_file_pmdmapped",
1206 "nr_anon_transparent_hugepages",
1207 "nr_vmscan_write",
1208 "nr_vmscan_immediate_reclaim",
1209 "nr_dirtied",
1210 "nr_written",
1211 "nr_kernel_misc_reclaimable",
1212 "nr_foll_pin_acquired",
1213 "nr_foll_pin_released",
1214 "nr_kernel_stack",
1215#if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1216 "nr_shadow_call_stack",
1217#endif
1218
1219 /* enum writeback_stat_item counters */
1220 "nr_dirty_threshold",
1221 "nr_dirty_background_threshold",
1222
1223#if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1224 /* enum vm_event_item counters */
1225 "pgpgin",
1226 "pgpgout",
1227 "pswpin",
1228 "pswpout",
1229
1230 TEXTS_FOR_ZONES("pgalloc")
1231 TEXTS_FOR_ZONES("allocstall")
1232 TEXTS_FOR_ZONES("pgskip")
1233
1234 "pgfree",
1235 "pgactivate",
1236 "pgdeactivate",
1237 "pglazyfree",
1238
1239 "pgfault",
1240 "pgmajfault",
1241 "pglazyfreed",
1242
1243 "pgrefill",
1244 "pgreuse",
1245 "pgsteal_kswapd",
1246 "pgsteal_direct",
1247 "pgscan_kswapd",
1248 "pgscan_direct",
1249 "pgscan_direct_throttle",
1250 "pgscan_anon",
1251 "pgscan_file",
1252 "pgsteal_anon",
1253 "pgsteal_file",
1254
1255#ifdef CONFIG_NUMA
1256 "zone_reclaim_failed",
1257#endif
1258 "pginodesteal",
1259 "slabs_scanned",
1260 "kswapd_inodesteal",
1261 "kswapd_low_wmark_hit_quickly",
1262 "kswapd_high_wmark_hit_quickly",
1263 "pageoutrun",
1264
1265 "pgrotated",
1266
1267 "drop_pagecache",
1268 "drop_slab",
1269 "oom_kill",
1270
1271#ifdef CONFIG_NUMA_BALANCING
1272 "numa_pte_updates",
1273 "numa_huge_pte_updates",
1274 "numa_hint_faults",
1275 "numa_hint_faults_local",
1276 "numa_pages_migrated",
1277#endif
1278#ifdef CONFIG_MIGRATION
1279 "pgmigrate_success",
1280 "pgmigrate_fail",
1281 "thp_migration_success",
1282 "thp_migration_fail",
1283 "thp_migration_split",
1284#endif
1285#ifdef CONFIG_COMPACTION
1286 "compact_migrate_scanned",
1287 "compact_free_scanned",
1288 "compact_isolated",
1289 "compact_stall",
1290 "compact_fail",
1291 "compact_success",
1292 "compact_daemon_wake",
1293 "compact_daemon_migrate_scanned",
1294 "compact_daemon_free_scanned",
1295#endif
1296
1297#ifdef CONFIG_HUGETLB_PAGE
1298 "htlb_buddy_alloc_success",
1299 "htlb_buddy_alloc_fail",
1300#endif
1301 "unevictable_pgs_culled",
1302 "unevictable_pgs_scanned",
1303 "unevictable_pgs_rescued",
1304 "unevictable_pgs_mlocked",
1305 "unevictable_pgs_munlocked",
1306 "unevictable_pgs_cleared",
1307 "unevictable_pgs_stranded",
1308
1309#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1310 "thp_fault_alloc",
1311 "thp_fault_fallback",
1312 "thp_fault_fallback_charge",
1313 "thp_collapse_alloc",
1314 "thp_collapse_alloc_failed",
1315 "thp_file_alloc",
1316 "thp_file_fallback",
1317 "thp_file_fallback_charge",
1318 "thp_file_mapped",
1319 "thp_split_page",
1320 "thp_split_page_failed",
1321 "thp_deferred_split_page",
1322 "thp_split_pmd",
1323#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1324 "thp_split_pud",
1325#endif
1326 "thp_zero_page_alloc",
1327 "thp_zero_page_alloc_failed",
1328 "thp_swpout",
1329 "thp_swpout_fallback",
1330#endif
1331#ifdef CONFIG_MEMORY_BALLOON
1332 "balloon_inflate",
1333 "balloon_deflate",
1334#ifdef CONFIG_BALLOON_COMPACTION
1335 "balloon_migrate",
1336#endif
1337#endif /* CONFIG_MEMORY_BALLOON */
1338#ifdef CONFIG_DEBUG_TLBFLUSH
1339 "nr_tlb_remote_flush",
1340 "nr_tlb_remote_flush_received",
1341 "nr_tlb_local_flush_all",
1342 "nr_tlb_local_flush_one",
1343#endif /* CONFIG_DEBUG_TLBFLUSH */
1344
1345#ifdef CONFIG_DEBUG_VM_VMACACHE
1346 "vmacache_find_calls",
1347 "vmacache_find_hits",
1348#endif
1349#ifdef CONFIG_SWAP
1350 "swap_ra",
1351 "swap_ra_hit",
1352#endif
1353#endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1354};
1355#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1356
1357#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1358 defined(CONFIG_PROC_FS)
1359static void *frag_start(struct seq_file *m, loff_t *pos)
1360{
1361 pg_data_t *pgdat;
1362 loff_t node = *pos;
1363
1364 for (pgdat = first_online_pgdat();
1365 pgdat && node;
1366 pgdat = next_online_pgdat(pgdat))
1367 --node;
1368
1369 return pgdat;
1370}
1371
1372static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1373{
1374 pg_data_t *pgdat = (pg_data_t *)arg;
1375
1376 (*pos)++;
1377 return next_online_pgdat(pgdat);
1378}
1379
1380static void frag_stop(struct seq_file *m, void *arg)
1381{
1382}
1383
1384/*
1385 * Walk zones in a node and print using a callback.
1386 * If @assert_populated is true, only use callback for zones that are populated.
1387 */
1388static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1389 bool assert_populated, bool nolock,
1390 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1391{
1392 struct zone *zone;
1393 struct zone *node_zones = pgdat->node_zones;
1394 unsigned long flags;
1395
1396 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1397 if (assert_populated && !populated_zone(zone))
1398 continue;
1399
1400 if (!nolock)
1401 spin_lock_irqsave(&zone->lock, flags);
1402 print(m, pgdat, zone);
1403 if (!nolock)
1404 spin_unlock_irqrestore(&zone->lock, flags);
1405 }
1406}
1407#endif
1408
1409#ifdef CONFIG_PROC_FS
1410static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1411 struct zone *zone)
1412{
1413 int order;
1414
1415 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1416 for (order = 0; order < MAX_ORDER; ++order)
1417 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1418 seq_putc(m, '\n');
1419}
1420
1421/*
1422 * This walks the free areas for each zone.
1423 */
1424static int frag_show(struct seq_file *m, void *arg)
1425{
1426 pg_data_t *pgdat = (pg_data_t *)arg;
1427 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1428 return 0;
1429}
1430
1431static void pagetypeinfo_showfree_print(struct seq_file *m,
1432 pg_data_t *pgdat, struct zone *zone)
1433{
1434 int order, mtype;
1435
1436 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1437 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1438 pgdat->node_id,
1439 zone->name,
1440 migratetype_names[mtype]);
1441 for (order = 0; order < MAX_ORDER; ++order) {
1442 unsigned long freecount = 0;
1443 struct free_area *area;
1444 struct list_head *curr;
1445 bool overflow = false;
1446
1447 area = &(zone->free_area[order]);
1448
1449 list_for_each(curr, &area->free_list[mtype]) {
1450 /*
1451 * Cap the free_list iteration because it might
1452 * be really large and we are under a spinlock
1453 * so a long time spent here could trigger a
1454 * hard lockup detector. Anyway this is a
1455 * debugging tool so knowing there is a handful
1456 * of pages of this order should be more than
1457 * sufficient.
1458 */
1459 if (++freecount >= 100000) {
1460 overflow = true;
1461 break;
1462 }
1463 }
1464 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1465 spin_unlock_irq(&zone->lock);
1466 cond_resched();
1467 spin_lock_irq(&zone->lock);
1468 }
1469 seq_putc(m, '\n');
1470 }
1471}
1472
1473/* Print out the free pages at each order for each migatetype */
1474static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1475{
1476 int order;
1477 pg_data_t *pgdat = (pg_data_t *)arg;
1478
1479 /* Print header */
1480 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1481 for (order = 0; order < MAX_ORDER; ++order)
1482 seq_printf(m, "%6d ", order);
1483 seq_putc(m, '\n');
1484
1485 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1486
1487 return 0;
1488}
1489
1490static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1491 pg_data_t *pgdat, struct zone *zone)
1492{
1493 int mtype;
1494 unsigned long pfn;
1495 unsigned long start_pfn = zone->zone_start_pfn;
1496 unsigned long end_pfn = zone_end_pfn(zone);
1497 unsigned long count[MIGRATE_TYPES] = { 0, };
1498
1499 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1500 struct page *page;
1501
1502 page = pfn_to_online_page(pfn);
1503 if (!page)
1504 continue;
1505
1506 /* Watch for unexpected holes punched in the memmap */
1507 if (!memmap_valid_within(pfn, page, zone))
1508 continue;
1509
1510 if (page_zone(page) != zone)
1511 continue;
1512
1513 mtype = get_pageblock_migratetype(page);
1514
1515 if (mtype < MIGRATE_TYPES)
1516 count[mtype]++;
1517 }
1518
1519 /* Print counts */
1520 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1521 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1522 seq_printf(m, "%12lu ", count[mtype]);
1523 seq_putc(m, '\n');
1524}
1525
1526/* Print out the number of pageblocks for each migratetype */
1527static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1528{
1529 int mtype;
1530 pg_data_t *pgdat = (pg_data_t *)arg;
1531
1532 seq_printf(m, "\n%-23s", "Number of blocks type ");
1533 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1534 seq_printf(m, "%12s ", migratetype_names[mtype]);
1535 seq_putc(m, '\n');
1536 walk_zones_in_node(m, pgdat, true, false,
1537 pagetypeinfo_showblockcount_print);
1538
1539 return 0;
1540}
1541
1542/*
1543 * Print out the number of pageblocks for each migratetype that contain pages
1544 * of other types. This gives an indication of how well fallbacks are being
1545 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1546 * to determine what is going on
1547 */
1548static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1549{
1550#ifdef CONFIG_PAGE_OWNER
1551 int mtype;
1552
1553 if (!static_branch_unlikely(&page_owner_inited))
1554 return;
1555
1556 drain_all_pages(NULL);
1557
1558 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1559 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1560 seq_printf(m, "%12s ", migratetype_names[mtype]);
1561 seq_putc(m, '\n');
1562
1563 walk_zones_in_node(m, pgdat, true, true,
1564 pagetypeinfo_showmixedcount_print);
1565#endif /* CONFIG_PAGE_OWNER */
1566}
1567
1568/*
1569 * This prints out statistics in relation to grouping pages by mobility.
1570 * It is expensive to collect so do not constantly read the file.
1571 */
1572static int pagetypeinfo_show(struct seq_file *m, void *arg)
1573{
1574 pg_data_t *pgdat = (pg_data_t *)arg;
1575
1576 /* check memoryless node */
1577 if (!node_state(pgdat->node_id, N_MEMORY))
1578 return 0;
1579
1580 seq_printf(m, "Page block order: %d\n", pageblock_order);
1581 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1582 seq_putc(m, '\n');
1583 pagetypeinfo_showfree(m, pgdat);
1584 pagetypeinfo_showblockcount(m, pgdat);
1585 pagetypeinfo_showmixedcount(m, pgdat);
1586
1587 return 0;
1588}
1589
1590static const struct seq_operations fragmentation_op = {
1591 .start = frag_start,
1592 .next = frag_next,
1593 .stop = frag_stop,
1594 .show = frag_show,
1595};
1596
1597static const struct seq_operations pagetypeinfo_op = {
1598 .start = frag_start,
1599 .next = frag_next,
1600 .stop = frag_stop,
1601 .show = pagetypeinfo_show,
1602};
1603
1604static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1605{
1606 int zid;
1607
1608 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1609 struct zone *compare = &pgdat->node_zones[zid];
1610
1611 if (populated_zone(compare))
1612 return zone == compare;
1613 }
1614
1615 return false;
1616}
1617
1618static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1619 struct zone *zone)
1620{
1621 int i;
1622 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1623 if (is_zone_first_populated(pgdat, zone)) {
1624 seq_printf(m, "\n per-node stats");
1625 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1626 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1627 node_page_state_pages(pgdat, i));
1628 }
1629 }
1630 seq_printf(m,
1631 "\n pages free %lu"
1632 "\n min %lu"
1633 "\n low %lu"
1634 "\n high %lu"
1635 "\n spanned %lu"
1636 "\n present %lu"
1637 "\n managed %lu",
1638 zone_page_state(zone, NR_FREE_PAGES),
1639 min_wmark_pages(zone),
1640 low_wmark_pages(zone),
1641 high_wmark_pages(zone),
1642 zone->spanned_pages,
1643 zone->present_pages,
1644 zone_managed_pages(zone));
1645
1646 seq_printf(m,
1647 "\n protection: (%ld",
1648 zone->lowmem_reserve[0]);
1649 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1650 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1651 seq_putc(m, ')');
1652
1653 /* If unpopulated, no other information is useful */
1654 if (!populated_zone(zone)) {
1655 seq_putc(m, '\n');
1656 return;
1657 }
1658
1659 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1660 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1661 zone_page_state(zone, i));
1662
1663#ifdef CONFIG_NUMA
1664 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1665 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1666 zone_numa_state_snapshot(zone, i));
1667#endif
1668
1669 seq_printf(m, "\n pagesets");
1670 for_each_online_cpu(i) {
1671 struct per_cpu_pageset *pageset;
1672
1673 pageset = per_cpu_ptr(zone->pageset, i);
1674 seq_printf(m,
1675 "\n cpu: %i"
1676 "\n count: %i"
1677 "\n high: %i"
1678 "\n batch: %i",
1679 i,
1680 pageset->pcp.count,
1681 pageset->pcp.high,
1682 pageset->pcp.batch);
1683#ifdef CONFIG_SMP
1684 seq_printf(m, "\n vm stats threshold: %d",
1685 pageset->stat_threshold);
1686#endif
1687 }
1688 seq_printf(m,
1689 "\n node_unreclaimable: %u"
1690 "\n start_pfn: %lu",
1691 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1692 zone->zone_start_pfn);
1693 seq_putc(m, '\n');
1694}
1695
1696/*
1697 * Output information about zones in @pgdat. All zones are printed regardless
1698 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1699 * set of all zones and userspace would not be aware of such zones if they are
1700 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1701 */
1702static int zoneinfo_show(struct seq_file *m, void *arg)
1703{
1704 pg_data_t *pgdat = (pg_data_t *)arg;
1705 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1706 return 0;
1707}
1708
1709static const struct seq_operations zoneinfo_op = {
1710 .start = frag_start, /* iterate over all zones. The same as in
1711 * fragmentation. */
1712 .next = frag_next,
1713 .stop = frag_stop,
1714 .show = zoneinfo_show,
1715};
1716
1717#define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1718 NR_VM_NUMA_STAT_ITEMS + \
1719 NR_VM_NODE_STAT_ITEMS + \
1720 NR_VM_WRITEBACK_STAT_ITEMS + \
1721 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1722 NR_VM_EVENT_ITEMS : 0))
1723
1724static void *vmstat_start(struct seq_file *m, loff_t *pos)
1725{
1726 unsigned long *v;
1727 int i;
1728
1729 if (*pos >= NR_VMSTAT_ITEMS)
1730 return NULL;
1731
1732 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1733 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1734 m->private = v;
1735 if (!v)
1736 return ERR_PTR(-ENOMEM);
1737 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1738 v[i] = global_zone_page_state(i);
1739 v += NR_VM_ZONE_STAT_ITEMS;
1740
1741#ifdef CONFIG_NUMA
1742 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1743 v[i] = global_numa_state(i);
1744 v += NR_VM_NUMA_STAT_ITEMS;
1745#endif
1746
1747 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1748 v[i] = global_node_page_state_pages(i);
1749 v += NR_VM_NODE_STAT_ITEMS;
1750
1751 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1752 v + NR_DIRTY_THRESHOLD);
1753 v += NR_VM_WRITEBACK_STAT_ITEMS;
1754
1755#ifdef CONFIG_VM_EVENT_COUNTERS
1756 all_vm_events(v);
1757 v[PGPGIN] /= 2; /* sectors -> kbytes */
1758 v[PGPGOUT] /= 2;
1759#endif
1760 return (unsigned long *)m->private + *pos;
1761}
1762
1763static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1764{
1765 (*pos)++;
1766 if (*pos >= NR_VMSTAT_ITEMS)
1767 return NULL;
1768 return (unsigned long *)m->private + *pos;
1769}
1770
1771static int vmstat_show(struct seq_file *m, void *arg)
1772{
1773 unsigned long *l = arg;
1774 unsigned long off = l - (unsigned long *)m->private;
1775
1776 seq_puts(m, vmstat_text[off]);
1777 seq_put_decimal_ull(m, " ", *l);
1778 seq_putc(m, '\n');
1779
1780 if (off == NR_VMSTAT_ITEMS - 1) {
1781 /*
1782 * We've come to the end - add any deprecated counters to avoid
1783 * breaking userspace which might depend on them being present.
1784 */
1785 seq_puts(m, "nr_unstable 0\n");
1786 }
1787 return 0;
1788}
1789
1790static void vmstat_stop(struct seq_file *m, void *arg)
1791{
1792 kfree(m->private);
1793 m->private = NULL;
1794}
1795
1796static const struct seq_operations vmstat_op = {
1797 .start = vmstat_start,
1798 .next = vmstat_next,
1799 .stop = vmstat_stop,
1800 .show = vmstat_show,
1801};
1802#endif /* CONFIG_PROC_FS */
1803
1804#ifdef CONFIG_SMP
1805static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1806int sysctl_stat_interval __read_mostly = HZ;
1807
1808#ifdef CONFIG_PROC_FS
1809static void refresh_vm_stats(struct work_struct *work)
1810{
1811 refresh_cpu_vm_stats(true);
1812}
1813
1814int vmstat_refresh(struct ctl_table *table, int write,
1815 void *buffer, size_t *lenp, loff_t *ppos)
1816{
1817 long val;
1818 int err;
1819 int i;
1820
1821 /*
1822 * The regular update, every sysctl_stat_interval, may come later
1823 * than expected: leaving a significant amount in per_cpu buckets.
1824 * This is particularly misleading when checking a quantity of HUGE
1825 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1826 * which can equally be echo'ed to or cat'ted from (by root),
1827 * can be used to update the stats just before reading them.
1828 *
1829 * Oh, and since global_zone_page_state() etc. are so careful to hide
1830 * transiently negative values, report an error here if any of
1831 * the stats is negative, so we know to go looking for imbalance.
1832 */
1833 err = schedule_on_each_cpu(refresh_vm_stats);
1834 if (err)
1835 return err;
1836 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1837 val = atomic_long_read(&vm_zone_stat[i]);
1838 if (val < 0) {
1839 pr_warn("%s: %s %ld\n",
1840 __func__, zone_stat_name(i), val);
1841 err = -EINVAL;
1842 }
1843 }
1844#ifdef CONFIG_NUMA
1845 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1846 val = atomic_long_read(&vm_numa_stat[i]);
1847 if (val < 0) {
1848 pr_warn("%s: %s %ld\n",
1849 __func__, numa_stat_name(i), val);
1850 err = -EINVAL;
1851 }
1852 }
1853#endif
1854 if (err)
1855 return err;
1856 if (write)
1857 *ppos += *lenp;
1858 else
1859 *lenp = 0;
1860 return 0;
1861}
1862#endif /* CONFIG_PROC_FS */
1863
1864static void vmstat_update(struct work_struct *w)
1865{
1866 if (refresh_cpu_vm_stats(true)) {
1867 /*
1868 * Counters were updated so we expect more updates
1869 * to occur in the future. Keep on running the
1870 * update worker thread.
1871 */
1872 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1873 this_cpu_ptr(&vmstat_work),
1874 round_jiffies_relative(sysctl_stat_interval));
1875 }
1876}
1877
1878/*
1879 * Switch off vmstat processing and then fold all the remaining differentials
1880 * until the diffs stay at zero. The function is used by NOHZ and can only be
1881 * invoked when tick processing is not active.
1882 */
1883/*
1884 * Check if the diffs for a certain cpu indicate that
1885 * an update is needed.
1886 */
1887static bool need_update(int cpu)
1888{
1889 struct zone *zone;
1890
1891 for_each_populated_zone(zone) {
1892 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1893
1894 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1895#ifdef CONFIG_NUMA
1896 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1897#endif
1898
1899 /*
1900 * The fast way of checking if there are any vmstat diffs.
1901 */
1902 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1903 sizeof(p->vm_stat_diff[0])))
1904 return true;
1905#ifdef CONFIG_NUMA
1906 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1907 sizeof(p->vm_numa_stat_diff[0])))
1908 return true;
1909#endif
1910 }
1911 return false;
1912}
1913
1914/*
1915 * Switch off vmstat processing and then fold all the remaining differentials
1916 * until the diffs stay at zero. The function is used by NOHZ and can only be
1917 * invoked when tick processing is not active.
1918 */
1919void quiet_vmstat(void)
1920{
1921 if (system_state != SYSTEM_RUNNING)
1922 return;
1923
1924 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1925 return;
1926
1927 if (!need_update(smp_processor_id()))
1928 return;
1929
1930 /*
1931 * Just refresh counters and do not care about the pending delayed
1932 * vmstat_update. It doesn't fire that often to matter and canceling
1933 * it would be too expensive from this path.
1934 * vmstat_shepherd will take care about that for us.
1935 */
1936 refresh_cpu_vm_stats(false);
1937}
1938
1939/*
1940 * Shepherd worker thread that checks the
1941 * differentials of processors that have their worker
1942 * threads for vm statistics updates disabled because of
1943 * inactivity.
1944 */
1945static void vmstat_shepherd(struct work_struct *w);
1946
1947static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1948
1949static void vmstat_shepherd(struct work_struct *w)
1950{
1951 int cpu;
1952
1953 get_online_cpus();
1954 /* Check processors whose vmstat worker threads have been disabled */
1955 for_each_online_cpu(cpu) {
1956 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1957
1958 if (!delayed_work_pending(dw) && need_update(cpu))
1959 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1960 }
1961 put_online_cpus();
1962
1963 schedule_delayed_work(&shepherd,
1964 round_jiffies_relative(sysctl_stat_interval));
1965}
1966
1967static void __init start_shepherd_timer(void)
1968{
1969 int cpu;
1970
1971 for_each_possible_cpu(cpu)
1972 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1973 vmstat_update);
1974
1975 schedule_delayed_work(&shepherd,
1976 round_jiffies_relative(sysctl_stat_interval));
1977}
1978
1979static void __init init_cpu_node_state(void)
1980{
1981 int node;
1982
1983 for_each_online_node(node) {
1984 if (cpumask_weight(cpumask_of_node(node)) > 0)
1985 node_set_state(node, N_CPU);
1986 }
1987}
1988
1989static int vmstat_cpu_online(unsigned int cpu)
1990{
1991 refresh_zone_stat_thresholds();
1992 node_set_state(cpu_to_node(cpu), N_CPU);
1993 return 0;
1994}
1995
1996static int vmstat_cpu_down_prep(unsigned int cpu)
1997{
1998 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1999 return 0;
2000}
2001
2002static int vmstat_cpu_dead(unsigned int cpu)
2003{
2004 const struct cpumask *node_cpus;
2005 int node;
2006
2007 node = cpu_to_node(cpu);
2008
2009 refresh_zone_stat_thresholds();
2010 node_cpus = cpumask_of_node(node);
2011 if (cpumask_weight(node_cpus) > 0)
2012 return 0;
2013
2014 node_clear_state(node, N_CPU);
2015 return 0;
2016}
2017
2018#endif
2019
2020struct workqueue_struct *mm_percpu_wq;
2021
2022void __init init_mm_internals(void)
2023{
2024 int ret __maybe_unused;
2025
2026 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2027
2028#ifdef CONFIG_SMP
2029 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2030 NULL, vmstat_cpu_dead);
2031 if (ret < 0)
2032 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2033
2034 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2035 vmstat_cpu_online,
2036 vmstat_cpu_down_prep);
2037 if (ret < 0)
2038 pr_err("vmstat: failed to register 'online' hotplug state\n");
2039
2040 get_online_cpus();
2041 init_cpu_node_state();
2042 put_online_cpus();
2043
2044 start_shepherd_timer();
2045#endif
2046#ifdef CONFIG_PROC_FS
2047 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2048 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2049 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2050 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2051#endif
2052}
2053
2054#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2055
2056/*
2057 * Return an index indicating how much of the available free memory is
2058 * unusable for an allocation of the requested size.
2059 */
2060static int unusable_free_index(unsigned int order,
2061 struct contig_page_info *info)
2062{
2063 /* No free memory is interpreted as all free memory is unusable */
2064 if (info->free_pages == 0)
2065 return 1000;
2066
2067 /*
2068 * Index should be a value between 0 and 1. Return a value to 3
2069 * decimal places.
2070 *
2071 * 0 => no fragmentation
2072 * 1 => high fragmentation
2073 */
2074 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2075
2076}
2077
2078static void unusable_show_print(struct seq_file *m,
2079 pg_data_t *pgdat, struct zone *zone)
2080{
2081 unsigned int order;
2082 int index;
2083 struct contig_page_info info;
2084
2085 seq_printf(m, "Node %d, zone %8s ",
2086 pgdat->node_id,
2087 zone->name);
2088 for (order = 0; order < MAX_ORDER; ++order) {
2089 fill_contig_page_info(zone, order, &info);
2090 index = unusable_free_index(order, &info);
2091 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2092 }
2093
2094 seq_putc(m, '\n');
2095}
2096
2097/*
2098 * Display unusable free space index
2099 *
2100 * The unusable free space index measures how much of the available free
2101 * memory cannot be used to satisfy an allocation of a given size and is a
2102 * value between 0 and 1. The higher the value, the more of free memory is
2103 * unusable and by implication, the worse the external fragmentation is. This
2104 * can be expressed as a percentage by multiplying by 100.
2105 */
2106static int unusable_show(struct seq_file *m, void *arg)
2107{
2108 pg_data_t *pgdat = (pg_data_t *)arg;
2109
2110 /* check memoryless node */
2111 if (!node_state(pgdat->node_id, N_MEMORY))
2112 return 0;
2113
2114 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2115
2116 return 0;
2117}
2118
2119static const struct seq_operations unusable_sops = {
2120 .start = frag_start,
2121 .next = frag_next,
2122 .stop = frag_stop,
2123 .show = unusable_show,
2124};
2125
2126DEFINE_SEQ_ATTRIBUTE(unusable);
2127
2128static void extfrag_show_print(struct seq_file *m,
2129 pg_data_t *pgdat, struct zone *zone)
2130{
2131 unsigned int order;
2132 int index;
2133
2134 /* Alloc on stack as interrupts are disabled for zone walk */
2135 struct contig_page_info info;
2136
2137 seq_printf(m, "Node %d, zone %8s ",
2138 pgdat->node_id,
2139 zone->name);
2140 for (order = 0; order < MAX_ORDER; ++order) {
2141 fill_contig_page_info(zone, order, &info);
2142 index = __fragmentation_index(order, &info);
2143 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2144 }
2145
2146 seq_putc(m, '\n');
2147}
2148
2149/*
2150 * Display fragmentation index for orders that allocations would fail for
2151 */
2152static int extfrag_show(struct seq_file *m, void *arg)
2153{
2154 pg_data_t *pgdat = (pg_data_t *)arg;
2155
2156 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2157
2158 return 0;
2159}
2160
2161static const struct seq_operations extfrag_sops = {
2162 .start = frag_start,
2163 .next = frag_next,
2164 .stop = frag_stop,
2165 .show = extfrag_show,
2166};
2167
2168DEFINE_SEQ_ATTRIBUTE(extfrag);
2169
2170static int __init extfrag_debug_init(void)
2171{
2172 struct dentry *extfrag_debug_root;
2173
2174 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2175
2176 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2177 &unusable_fops);
2178
2179 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2180 &extfrag_fops);
2181
2182 return 0;
2183}
2184
2185module_init(extfrag_debug_init);
2186#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmstat.c
4 *
5 * Manages VM statistics
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 *
8 * zoned VM statistics
9 * Copyright (C) 2006 Silicon Graphics, Inc.,
10 * Christoph Lameter <christoph@lameter.com>
11 * Copyright (C) 2008-2014 Christoph Lameter
12 */
13#include <linux/fs.h>
14#include <linux/mm.h>
15#include <linux/err.h>
16#include <linux/module.h>
17#include <linux/slab.h>
18#include <linux/cpu.h>
19#include <linux/cpumask.h>
20#include <linux/vmstat.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/debugfs.h>
24#include <linux/sched.h>
25#include <linux/math64.h>
26#include <linux/writeback.h>
27#include <linux/compaction.h>
28#include <linux/mm_inline.h>
29#include <linux/page_owner.h>
30#include <linux/sched/isolation.h>
31
32#include "internal.h"
33
34#ifdef CONFIG_NUMA
35int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
36
37/* zero numa counters within a zone */
38static void zero_zone_numa_counters(struct zone *zone)
39{
40 int item, cpu;
41
42 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
43 atomic_long_set(&zone->vm_numa_event[item], 0);
44 for_each_online_cpu(cpu) {
45 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
46 = 0;
47 }
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_EVENT_ITEMS; item++)
66 atomic_long_set(&vm_numa_event[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(const struct ctl_table *table, int write,
78 void *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 cpus_read_lock();
133 sum_vm_events(ret);
134 cpus_read_unlock();
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_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
165EXPORT_SYMBOL(vm_zone_stat);
166EXPORT_SYMBOL(vm_node_stat);
167
168#ifdef CONFIG_NUMA
169static void fold_vm_zone_numa_events(struct zone *zone)
170{
171 unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
172 int cpu;
173 enum numa_stat_item item;
174
175 for_each_online_cpu(cpu) {
176 struct per_cpu_zonestat *pzstats;
177
178 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
179 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
180 zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
181 }
182
183 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
184 zone_numa_event_add(zone_numa_events[item], zone, item);
185}
186
187void fold_vm_numa_events(void)
188{
189 struct zone *zone;
190
191 for_each_populated_zone(zone)
192 fold_vm_zone_numa_events(zone);
193}
194#endif
195
196#ifdef CONFIG_SMP
197
198int calculate_pressure_threshold(struct zone *zone)
199{
200 int threshold;
201 int watermark_distance;
202
203 /*
204 * As vmstats are not up to date, there is drift between the estimated
205 * and real values. For high thresholds and a high number of CPUs, it
206 * is possible for the min watermark to be breached while the estimated
207 * value looks fine. The pressure threshold is a reduced value such
208 * that even the maximum amount of drift will not accidentally breach
209 * the min watermark
210 */
211 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
212 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
213
214 /*
215 * Maximum threshold is 125
216 */
217 threshold = min(125, threshold);
218
219 return threshold;
220}
221
222int calculate_normal_threshold(struct zone *zone)
223{
224 int threshold;
225 int mem; /* memory in 128 MB units */
226
227 /*
228 * The threshold scales with the number of processors and the amount
229 * of memory per zone. More memory means that we can defer updates for
230 * longer, more processors could lead to more contention.
231 * fls() is used to have a cheap way of logarithmic scaling.
232 *
233 * Some sample thresholds:
234 *
235 * Threshold Processors (fls) Zonesize fls(mem)+1
236 * ------------------------------------------------------------------
237 * 8 1 1 0.9-1 GB 4
238 * 16 2 2 0.9-1 GB 4
239 * 20 2 2 1-2 GB 5
240 * 24 2 2 2-4 GB 6
241 * 28 2 2 4-8 GB 7
242 * 32 2 2 8-16 GB 8
243 * 4 2 2 <128M 1
244 * 30 4 3 2-4 GB 5
245 * 48 4 3 8-16 GB 8
246 * 32 8 4 1-2 GB 4
247 * 32 8 4 0.9-1GB 4
248 * 10 16 5 <128M 1
249 * 40 16 5 900M 4
250 * 70 64 7 2-4 GB 5
251 * 84 64 7 4-8 GB 6
252 * 108 512 9 4-8 GB 6
253 * 125 1024 10 8-16 GB 8
254 * 125 1024 10 16-32 GB 9
255 */
256
257 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
258
259 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
260
261 /*
262 * Maximum threshold is 125
263 */
264 threshold = min(125, threshold);
265
266 return threshold;
267}
268
269/*
270 * Refresh the thresholds for each zone.
271 */
272void refresh_zone_stat_thresholds(void)
273{
274 struct pglist_data *pgdat;
275 struct zone *zone;
276 int cpu;
277 int threshold;
278
279 /* Zero current pgdat thresholds */
280 for_each_online_pgdat(pgdat) {
281 for_each_online_cpu(cpu) {
282 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
283 }
284 }
285
286 for_each_populated_zone(zone) {
287 struct pglist_data *pgdat = zone->zone_pgdat;
288 unsigned long max_drift, tolerate_drift;
289
290 threshold = calculate_normal_threshold(zone);
291
292 for_each_online_cpu(cpu) {
293 int pgdat_threshold;
294
295 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
296 = threshold;
297
298 /* Base nodestat threshold on the largest populated zone. */
299 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
300 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
301 = max(threshold, pgdat_threshold);
302 }
303
304 /*
305 * Only set percpu_drift_mark if there is a danger that
306 * NR_FREE_PAGES reports the low watermark is ok when in fact
307 * the min watermark could be breached by an allocation
308 */
309 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
310 max_drift = num_online_cpus() * threshold;
311 if (max_drift > tolerate_drift)
312 zone->percpu_drift_mark = high_wmark_pages(zone) +
313 max_drift;
314 }
315}
316
317void set_pgdat_percpu_threshold(pg_data_t *pgdat,
318 int (*calculate_pressure)(struct zone *))
319{
320 struct zone *zone;
321 int cpu;
322 int threshold;
323 int i;
324
325 for (i = 0; i < pgdat->nr_zones; i++) {
326 zone = &pgdat->node_zones[i];
327 if (!zone->percpu_drift_mark)
328 continue;
329
330 threshold = (*calculate_pressure)(zone);
331 for_each_online_cpu(cpu)
332 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
333 = threshold;
334 }
335}
336
337/*
338 * For use when we know that interrupts are disabled,
339 * or when we know that preemption is disabled and that
340 * particular counter cannot be updated from interrupt context.
341 */
342void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
343 long delta)
344{
345 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
346 s8 __percpu *p = pcp->vm_stat_diff + item;
347 long x;
348 long t;
349
350 /*
351 * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
352 * atomicity is provided by IRQs being disabled -- either explicitly
353 * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
354 * CPU migrations and preemption potentially corrupts a counter so
355 * disable preemption.
356 */
357 preempt_disable_nested();
358
359 x = delta + __this_cpu_read(*p);
360
361 t = __this_cpu_read(pcp->stat_threshold);
362
363 if (unlikely(abs(x) > t)) {
364 zone_page_state_add(x, zone, item);
365 x = 0;
366 }
367 __this_cpu_write(*p, x);
368
369 preempt_enable_nested();
370}
371EXPORT_SYMBOL(__mod_zone_page_state);
372
373void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
374 long delta)
375{
376 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
377 s8 __percpu *p = pcp->vm_node_stat_diff + item;
378 long x;
379 long t;
380
381 if (vmstat_item_in_bytes(item)) {
382 /*
383 * Only cgroups use subpage accounting right now; at
384 * the global level, these items still change in
385 * multiples of whole pages. Store them as pages
386 * internally to keep the per-cpu counters compact.
387 */
388 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
389 delta >>= PAGE_SHIFT;
390 }
391
392 /* See __mod_node_page_state */
393 preempt_disable_nested();
394
395 x = delta + __this_cpu_read(*p);
396
397 t = __this_cpu_read(pcp->stat_threshold);
398
399 if (unlikely(abs(x) > t)) {
400 node_page_state_add(x, pgdat, item);
401 x = 0;
402 }
403 __this_cpu_write(*p, x);
404
405 preempt_enable_nested();
406}
407EXPORT_SYMBOL(__mod_node_page_state);
408
409/*
410 * Optimized increment and decrement functions.
411 *
412 * These are only for a single page and therefore can take a struct page *
413 * argument instead of struct zone *. This allows the inclusion of the code
414 * generated for page_zone(page) into the optimized functions.
415 *
416 * No overflow check is necessary and therefore the differential can be
417 * incremented or decremented in place which may allow the compilers to
418 * generate better code.
419 * The increment or decrement is known and therefore one boundary check can
420 * be omitted.
421 *
422 * NOTE: These functions are very performance sensitive. Change only
423 * with care.
424 *
425 * Some processors have inc/dec instructions that are atomic vs an interrupt.
426 * However, the code must first determine the differential location in a zone
427 * based on the processor number and then inc/dec the counter. There is no
428 * guarantee without disabling preemption that the processor will not change
429 * in between and therefore the atomicity vs. interrupt cannot be exploited
430 * in a useful way here.
431 */
432void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
433{
434 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
435 s8 __percpu *p = pcp->vm_stat_diff + item;
436 s8 v, t;
437
438 /* See __mod_node_page_state */
439 preempt_disable_nested();
440
441 v = __this_cpu_inc_return(*p);
442 t = __this_cpu_read(pcp->stat_threshold);
443 if (unlikely(v > t)) {
444 s8 overstep = t >> 1;
445
446 zone_page_state_add(v + overstep, zone, item);
447 __this_cpu_write(*p, -overstep);
448 }
449
450 preempt_enable_nested();
451}
452
453void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
454{
455 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
456 s8 __percpu *p = pcp->vm_node_stat_diff + item;
457 s8 v, t;
458
459 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
460
461 /* See __mod_node_page_state */
462 preempt_disable_nested();
463
464 v = __this_cpu_inc_return(*p);
465 t = __this_cpu_read(pcp->stat_threshold);
466 if (unlikely(v > t)) {
467 s8 overstep = t >> 1;
468
469 node_page_state_add(v + overstep, pgdat, item);
470 __this_cpu_write(*p, -overstep);
471 }
472
473 preempt_enable_nested();
474}
475
476void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
477{
478 __inc_zone_state(page_zone(page), item);
479}
480EXPORT_SYMBOL(__inc_zone_page_state);
481
482void __inc_node_page_state(struct page *page, enum node_stat_item item)
483{
484 __inc_node_state(page_pgdat(page), item);
485}
486EXPORT_SYMBOL(__inc_node_page_state);
487
488void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
489{
490 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
491 s8 __percpu *p = pcp->vm_stat_diff + item;
492 s8 v, t;
493
494 /* See __mod_node_page_state */
495 preempt_disable_nested();
496
497 v = __this_cpu_dec_return(*p);
498 t = __this_cpu_read(pcp->stat_threshold);
499 if (unlikely(v < - t)) {
500 s8 overstep = t >> 1;
501
502 zone_page_state_add(v - overstep, zone, item);
503 __this_cpu_write(*p, overstep);
504 }
505
506 preempt_enable_nested();
507}
508
509void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
510{
511 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
512 s8 __percpu *p = pcp->vm_node_stat_diff + item;
513 s8 v, t;
514
515 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
516
517 /* See __mod_node_page_state */
518 preempt_disable_nested();
519
520 v = __this_cpu_dec_return(*p);
521 t = __this_cpu_read(pcp->stat_threshold);
522 if (unlikely(v < - t)) {
523 s8 overstep = t >> 1;
524
525 node_page_state_add(v - overstep, pgdat, item);
526 __this_cpu_write(*p, overstep);
527 }
528
529 preempt_enable_nested();
530}
531
532void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
533{
534 __dec_zone_state(page_zone(page), item);
535}
536EXPORT_SYMBOL(__dec_zone_page_state);
537
538void __dec_node_page_state(struct page *page, enum node_stat_item item)
539{
540 __dec_node_state(page_pgdat(page), item);
541}
542EXPORT_SYMBOL(__dec_node_page_state);
543
544#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
545/*
546 * If we have cmpxchg_local support then we do not need to incur the overhead
547 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
548 *
549 * mod_state() modifies the zone counter state through atomic per cpu
550 * operations.
551 *
552 * Overstep mode specifies how overstep should handled:
553 * 0 No overstepping
554 * 1 Overstepping half of threshold
555 * -1 Overstepping minus half of threshold
556*/
557static inline void mod_zone_state(struct zone *zone,
558 enum zone_stat_item item, long delta, int overstep_mode)
559{
560 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
561 s8 __percpu *p = pcp->vm_stat_diff + item;
562 long n, t, z;
563 s8 o;
564
565 o = this_cpu_read(*p);
566 do {
567 z = 0; /* overflow to zone counters */
568
569 /*
570 * The fetching of the stat_threshold is racy. We may apply
571 * a counter threshold to the wrong the cpu if we get
572 * rescheduled while executing here. However, the next
573 * counter update will apply the threshold again and
574 * therefore bring the counter under the threshold again.
575 *
576 * Most of the time the thresholds are the same anyways
577 * for all cpus in a zone.
578 */
579 t = this_cpu_read(pcp->stat_threshold);
580
581 n = delta + (long)o;
582
583 if (abs(n) > t) {
584 int os = overstep_mode * (t >> 1) ;
585
586 /* Overflow must be added to zone counters */
587 z = n + os;
588 n = -os;
589 }
590 } while (!this_cpu_try_cmpxchg(*p, &o, n));
591
592 if (z)
593 zone_page_state_add(z, zone, item);
594}
595
596void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
597 long delta)
598{
599 mod_zone_state(zone, item, delta, 0);
600}
601EXPORT_SYMBOL(mod_zone_page_state);
602
603void inc_zone_page_state(struct page *page, enum zone_stat_item item)
604{
605 mod_zone_state(page_zone(page), item, 1, 1);
606}
607EXPORT_SYMBOL(inc_zone_page_state);
608
609void dec_zone_page_state(struct page *page, enum zone_stat_item item)
610{
611 mod_zone_state(page_zone(page), item, -1, -1);
612}
613EXPORT_SYMBOL(dec_zone_page_state);
614
615static inline void mod_node_state(struct pglist_data *pgdat,
616 enum node_stat_item item, int delta, int overstep_mode)
617{
618 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
619 s8 __percpu *p = pcp->vm_node_stat_diff + item;
620 long n, t, z;
621 s8 o;
622
623 if (vmstat_item_in_bytes(item)) {
624 /*
625 * Only cgroups use subpage accounting right now; at
626 * the global level, these items still change in
627 * multiples of whole pages. Store them as pages
628 * internally to keep the per-cpu counters compact.
629 */
630 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
631 delta >>= PAGE_SHIFT;
632 }
633
634 o = this_cpu_read(*p);
635 do {
636 z = 0; /* overflow to node counters */
637
638 /*
639 * The fetching of the stat_threshold is racy. We may apply
640 * a counter threshold to the wrong the cpu if we get
641 * rescheduled while executing here. However, the next
642 * counter update will apply the threshold again and
643 * therefore bring the counter under the threshold again.
644 *
645 * Most of the time the thresholds are the same anyways
646 * for all cpus in a node.
647 */
648 t = this_cpu_read(pcp->stat_threshold);
649
650 n = delta + (long)o;
651
652 if (abs(n) > t) {
653 int os = overstep_mode * (t >> 1) ;
654
655 /* Overflow must be added to node counters */
656 z = n + os;
657 n = -os;
658 }
659 } while (!this_cpu_try_cmpxchg(*p, &o, n));
660
661 if (z)
662 node_page_state_add(z, pgdat, item);
663}
664
665void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
666 long delta)
667{
668 mod_node_state(pgdat, item, delta, 0);
669}
670EXPORT_SYMBOL(mod_node_page_state);
671
672void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
673{
674 mod_node_state(pgdat, item, 1, 1);
675}
676
677void inc_node_page_state(struct page *page, enum node_stat_item item)
678{
679 mod_node_state(page_pgdat(page), item, 1, 1);
680}
681EXPORT_SYMBOL(inc_node_page_state);
682
683void dec_node_page_state(struct page *page, enum node_stat_item item)
684{
685 mod_node_state(page_pgdat(page), item, -1, -1);
686}
687EXPORT_SYMBOL(dec_node_page_state);
688#else
689/*
690 * Use interrupt disable to serialize counter updates
691 */
692void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
693 long delta)
694{
695 unsigned long flags;
696
697 local_irq_save(flags);
698 __mod_zone_page_state(zone, item, delta);
699 local_irq_restore(flags);
700}
701EXPORT_SYMBOL(mod_zone_page_state);
702
703void inc_zone_page_state(struct page *page, enum zone_stat_item item)
704{
705 unsigned long flags;
706 struct zone *zone;
707
708 zone = page_zone(page);
709 local_irq_save(flags);
710 __inc_zone_state(zone, item);
711 local_irq_restore(flags);
712}
713EXPORT_SYMBOL(inc_zone_page_state);
714
715void dec_zone_page_state(struct page *page, enum zone_stat_item item)
716{
717 unsigned long flags;
718
719 local_irq_save(flags);
720 __dec_zone_page_state(page, item);
721 local_irq_restore(flags);
722}
723EXPORT_SYMBOL(dec_zone_page_state);
724
725void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
726{
727 unsigned long flags;
728
729 local_irq_save(flags);
730 __inc_node_state(pgdat, item);
731 local_irq_restore(flags);
732}
733EXPORT_SYMBOL(inc_node_state);
734
735void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
736 long delta)
737{
738 unsigned long flags;
739
740 local_irq_save(flags);
741 __mod_node_page_state(pgdat, item, delta);
742 local_irq_restore(flags);
743}
744EXPORT_SYMBOL(mod_node_page_state);
745
746void inc_node_page_state(struct page *page, enum node_stat_item item)
747{
748 unsigned long flags;
749 struct pglist_data *pgdat;
750
751 pgdat = page_pgdat(page);
752 local_irq_save(flags);
753 __inc_node_state(pgdat, item);
754 local_irq_restore(flags);
755}
756EXPORT_SYMBOL(inc_node_page_state);
757
758void dec_node_page_state(struct page *page, enum node_stat_item item)
759{
760 unsigned long flags;
761
762 local_irq_save(flags);
763 __dec_node_page_state(page, item);
764 local_irq_restore(flags);
765}
766EXPORT_SYMBOL(dec_node_page_state);
767#endif
768
769/*
770 * Fold a differential into the global counters.
771 * Returns the number of counters updated.
772 */
773static int fold_diff(int *zone_diff, int *node_diff)
774{
775 int i;
776 int changes = 0;
777
778 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
779 if (zone_diff[i]) {
780 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
781 changes++;
782 }
783
784 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
785 if (node_diff[i]) {
786 atomic_long_add(node_diff[i], &vm_node_stat[i]);
787 changes++;
788 }
789 return changes;
790}
791
792/*
793 * Update the zone counters for the current cpu.
794 *
795 * Note that refresh_cpu_vm_stats strives to only access
796 * node local memory. The per cpu pagesets on remote zones are placed
797 * in the memory local to the processor using that pageset. So the
798 * loop over all zones will access a series of cachelines local to
799 * the processor.
800 *
801 * The call to zone_page_state_add updates the cachelines with the
802 * statistics in the remote zone struct as well as the global cachelines
803 * with the global counters. These could cause remote node cache line
804 * bouncing and will have to be only done when necessary.
805 *
806 * The function returns the number of global counters updated.
807 */
808static int refresh_cpu_vm_stats(bool do_pagesets)
809{
810 struct pglist_data *pgdat;
811 struct zone *zone;
812 int i;
813 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
814 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
815 int changes = 0;
816
817 for_each_populated_zone(zone) {
818 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
819 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
820
821 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
822 int v;
823
824 v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
825 if (v) {
826
827 atomic_long_add(v, &zone->vm_stat[i]);
828 global_zone_diff[i] += v;
829#ifdef CONFIG_NUMA
830 /* 3 seconds idle till flush */
831 __this_cpu_write(pcp->expire, 3);
832#endif
833 }
834 }
835
836 if (do_pagesets) {
837 cond_resched();
838
839 changes += decay_pcp_high(zone, this_cpu_ptr(pcp));
840#ifdef CONFIG_NUMA
841 /*
842 * Deal with draining the remote pageset of this
843 * processor
844 *
845 * Check if there are pages remaining in this pageset
846 * if not then there is nothing to expire.
847 */
848 if (!__this_cpu_read(pcp->expire) ||
849 !__this_cpu_read(pcp->count))
850 continue;
851
852 /*
853 * We never drain zones local to this processor.
854 */
855 if (zone_to_nid(zone) == numa_node_id()) {
856 __this_cpu_write(pcp->expire, 0);
857 continue;
858 }
859
860 if (__this_cpu_dec_return(pcp->expire)) {
861 changes++;
862 continue;
863 }
864
865 if (__this_cpu_read(pcp->count)) {
866 drain_zone_pages(zone, this_cpu_ptr(pcp));
867 changes++;
868 }
869#endif
870 }
871 }
872
873 for_each_online_pgdat(pgdat) {
874 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
875
876 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
877 int v;
878
879 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
880 if (v) {
881 atomic_long_add(v, &pgdat->vm_stat[i]);
882 global_node_diff[i] += v;
883 }
884 }
885 }
886
887 changes += fold_diff(global_zone_diff, global_node_diff);
888 return changes;
889}
890
891/*
892 * Fold the data for an offline cpu into the global array.
893 * There cannot be any access by the offline cpu and therefore
894 * synchronization is simplified.
895 */
896void cpu_vm_stats_fold(int cpu)
897{
898 struct pglist_data *pgdat;
899 struct zone *zone;
900 int i;
901 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
902 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
903
904 for_each_populated_zone(zone) {
905 struct per_cpu_zonestat *pzstats;
906
907 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
908
909 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
910 if (pzstats->vm_stat_diff[i]) {
911 int v;
912
913 v = pzstats->vm_stat_diff[i];
914 pzstats->vm_stat_diff[i] = 0;
915 atomic_long_add(v, &zone->vm_stat[i]);
916 global_zone_diff[i] += v;
917 }
918 }
919#ifdef CONFIG_NUMA
920 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
921 if (pzstats->vm_numa_event[i]) {
922 unsigned long v;
923
924 v = pzstats->vm_numa_event[i];
925 pzstats->vm_numa_event[i] = 0;
926 zone_numa_event_add(v, zone, i);
927 }
928 }
929#endif
930 }
931
932 for_each_online_pgdat(pgdat) {
933 struct per_cpu_nodestat *p;
934
935 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
936
937 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
938 if (p->vm_node_stat_diff[i]) {
939 int v;
940
941 v = p->vm_node_stat_diff[i];
942 p->vm_node_stat_diff[i] = 0;
943 atomic_long_add(v, &pgdat->vm_stat[i]);
944 global_node_diff[i] += v;
945 }
946 }
947
948 fold_diff(global_zone_diff, global_node_diff);
949}
950
951/*
952 * this is only called if !populated_zone(zone), which implies no other users of
953 * pset->vm_stat_diff[] exist.
954 */
955void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
956{
957 unsigned long v;
958 int i;
959
960 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
961 if (pzstats->vm_stat_diff[i]) {
962 v = pzstats->vm_stat_diff[i];
963 pzstats->vm_stat_diff[i] = 0;
964 zone_page_state_add(v, zone, i);
965 }
966 }
967
968#ifdef CONFIG_NUMA
969 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
970 if (pzstats->vm_numa_event[i]) {
971 v = pzstats->vm_numa_event[i];
972 pzstats->vm_numa_event[i] = 0;
973 zone_numa_event_add(v, zone, i);
974 }
975 }
976#endif
977}
978#endif
979
980#ifdef CONFIG_NUMA
981/*
982 * Determine the per node value of a stat item. This function
983 * is called frequently in a NUMA machine, so try to be as
984 * frugal as possible.
985 */
986unsigned long sum_zone_node_page_state(int node,
987 enum zone_stat_item item)
988{
989 struct zone *zones = NODE_DATA(node)->node_zones;
990 int i;
991 unsigned long count = 0;
992
993 for (i = 0; i < MAX_NR_ZONES; i++)
994 count += zone_page_state(zones + i, item);
995
996 return count;
997}
998
999/* Determine the per node value of a numa stat item. */
1000unsigned long sum_zone_numa_event_state(int node,
1001 enum numa_stat_item item)
1002{
1003 struct zone *zones = NODE_DATA(node)->node_zones;
1004 unsigned long count = 0;
1005 int i;
1006
1007 for (i = 0; i < MAX_NR_ZONES; i++)
1008 count += zone_numa_event_state(zones + i, item);
1009
1010 return count;
1011}
1012
1013/*
1014 * Determine the per node value of a stat item.
1015 */
1016unsigned long node_page_state_pages(struct pglist_data *pgdat,
1017 enum node_stat_item item)
1018{
1019 long x = atomic_long_read(&pgdat->vm_stat[item]);
1020#ifdef CONFIG_SMP
1021 if (x < 0)
1022 x = 0;
1023#endif
1024 return x;
1025}
1026
1027unsigned long node_page_state(struct pglist_data *pgdat,
1028 enum node_stat_item item)
1029{
1030 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1031
1032 return node_page_state_pages(pgdat, item);
1033}
1034#endif
1035
1036/*
1037 * Count number of pages "struct page" and "struct page_ext" consume.
1038 * nr_memmap_boot_pages: # of pages allocated by boot allocator
1039 * nr_memmap_pages: # of pages that were allocated by buddy allocator
1040 */
1041static atomic_long_t nr_memmap_boot_pages = ATOMIC_LONG_INIT(0);
1042static atomic_long_t nr_memmap_pages = ATOMIC_LONG_INIT(0);
1043
1044void memmap_boot_pages_add(long delta)
1045{
1046 atomic_long_add(delta, &nr_memmap_boot_pages);
1047}
1048
1049void memmap_pages_add(long delta)
1050{
1051 atomic_long_add(delta, &nr_memmap_pages);
1052}
1053
1054#ifdef CONFIG_COMPACTION
1055
1056struct contig_page_info {
1057 unsigned long free_pages;
1058 unsigned long free_blocks_total;
1059 unsigned long free_blocks_suitable;
1060};
1061
1062/*
1063 * Calculate the number of free pages in a zone, how many contiguous
1064 * pages are free and how many are large enough to satisfy an allocation of
1065 * the target size. Note that this function makes no attempt to estimate
1066 * how many suitable free blocks there *might* be if MOVABLE pages were
1067 * migrated. Calculating that is possible, but expensive and can be
1068 * figured out from userspace
1069 */
1070static void fill_contig_page_info(struct zone *zone,
1071 unsigned int suitable_order,
1072 struct contig_page_info *info)
1073{
1074 unsigned int order;
1075
1076 info->free_pages = 0;
1077 info->free_blocks_total = 0;
1078 info->free_blocks_suitable = 0;
1079
1080 for (order = 0; order < NR_PAGE_ORDERS; order++) {
1081 unsigned long blocks;
1082
1083 /*
1084 * Count number of free blocks.
1085 *
1086 * Access to nr_free is lockless as nr_free is used only for
1087 * diagnostic purposes. Use data_race to avoid KCSAN warning.
1088 */
1089 blocks = data_race(zone->free_area[order].nr_free);
1090 info->free_blocks_total += blocks;
1091
1092 /* Count free base pages */
1093 info->free_pages += blocks << order;
1094
1095 /* Count the suitable free blocks */
1096 if (order >= suitable_order)
1097 info->free_blocks_suitable += blocks <<
1098 (order - suitable_order);
1099 }
1100}
1101
1102/*
1103 * A fragmentation index only makes sense if an allocation of a requested
1104 * size would fail. If that is true, the fragmentation index indicates
1105 * whether external fragmentation or a lack of memory was the problem.
1106 * The value can be used to determine if page reclaim or compaction
1107 * should be used
1108 */
1109static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1110{
1111 unsigned long requested = 1UL << order;
1112
1113 if (WARN_ON_ONCE(order > MAX_PAGE_ORDER))
1114 return 0;
1115
1116 if (!info->free_blocks_total)
1117 return 0;
1118
1119 /* Fragmentation index only makes sense when a request would fail */
1120 if (info->free_blocks_suitable)
1121 return -1000;
1122
1123 /*
1124 * Index is between 0 and 1 so return within 3 decimal places
1125 *
1126 * 0 => allocation would fail due to lack of memory
1127 * 1 => allocation would fail due to fragmentation
1128 */
1129 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1130}
1131
1132/*
1133 * Calculates external fragmentation within a zone wrt the given order.
1134 * It is defined as the percentage of pages found in blocks of size
1135 * less than 1 << order. It returns values in range [0, 100].
1136 */
1137unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1138{
1139 struct contig_page_info info;
1140
1141 fill_contig_page_info(zone, order, &info);
1142 if (info.free_pages == 0)
1143 return 0;
1144
1145 return div_u64((info.free_pages -
1146 (info.free_blocks_suitable << order)) * 100,
1147 info.free_pages);
1148}
1149
1150/* Same as __fragmentation index but allocs contig_page_info on stack */
1151int fragmentation_index(struct zone *zone, unsigned int order)
1152{
1153 struct contig_page_info info;
1154
1155 fill_contig_page_info(zone, order, &info);
1156 return __fragmentation_index(order, &info);
1157}
1158#endif
1159
1160#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1161 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1162#ifdef CONFIG_ZONE_DMA
1163#define TEXT_FOR_DMA(xx) xx "_dma",
1164#else
1165#define TEXT_FOR_DMA(xx)
1166#endif
1167
1168#ifdef CONFIG_ZONE_DMA32
1169#define TEXT_FOR_DMA32(xx) xx "_dma32",
1170#else
1171#define TEXT_FOR_DMA32(xx)
1172#endif
1173
1174#ifdef CONFIG_HIGHMEM
1175#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1176#else
1177#define TEXT_FOR_HIGHMEM(xx)
1178#endif
1179
1180#ifdef CONFIG_ZONE_DEVICE
1181#define TEXT_FOR_DEVICE(xx) xx "_device",
1182#else
1183#define TEXT_FOR_DEVICE(xx)
1184#endif
1185
1186#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1187 TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1188 TEXT_FOR_DEVICE(xx)
1189
1190const char * const vmstat_text[] = {
1191 /* enum zone_stat_item counters */
1192 "nr_free_pages",
1193 "nr_zone_inactive_anon",
1194 "nr_zone_active_anon",
1195 "nr_zone_inactive_file",
1196 "nr_zone_active_file",
1197 "nr_zone_unevictable",
1198 "nr_zone_write_pending",
1199 "nr_mlock",
1200 "nr_bounce",
1201#if IS_ENABLED(CONFIG_ZSMALLOC)
1202 "nr_zspages",
1203#endif
1204 "nr_free_cma",
1205#ifdef CONFIG_UNACCEPTED_MEMORY
1206 "nr_unaccepted",
1207#endif
1208
1209 /* enum numa_stat_item counters */
1210#ifdef CONFIG_NUMA
1211 "numa_hit",
1212 "numa_miss",
1213 "numa_foreign",
1214 "numa_interleave",
1215 "numa_local",
1216 "numa_other",
1217#endif
1218
1219 /* enum node_stat_item counters */
1220 "nr_inactive_anon",
1221 "nr_active_anon",
1222 "nr_inactive_file",
1223 "nr_active_file",
1224 "nr_unevictable",
1225 "nr_slab_reclaimable",
1226 "nr_slab_unreclaimable",
1227 "nr_isolated_anon",
1228 "nr_isolated_file",
1229 "workingset_nodes",
1230 "workingset_refault_anon",
1231 "workingset_refault_file",
1232 "workingset_activate_anon",
1233 "workingset_activate_file",
1234 "workingset_restore_anon",
1235 "workingset_restore_file",
1236 "workingset_nodereclaim",
1237 "nr_anon_pages",
1238 "nr_mapped",
1239 "nr_file_pages",
1240 "nr_dirty",
1241 "nr_writeback",
1242 "nr_writeback_temp",
1243 "nr_shmem",
1244 "nr_shmem_hugepages",
1245 "nr_shmem_pmdmapped",
1246 "nr_file_hugepages",
1247 "nr_file_pmdmapped",
1248 "nr_anon_transparent_hugepages",
1249 "nr_vmscan_write",
1250 "nr_vmscan_immediate_reclaim",
1251 "nr_dirtied",
1252 "nr_written",
1253 "nr_throttled_written",
1254 "nr_kernel_misc_reclaimable",
1255 "nr_foll_pin_acquired",
1256 "nr_foll_pin_released",
1257 "nr_kernel_stack",
1258#if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1259 "nr_shadow_call_stack",
1260#endif
1261 "nr_page_table_pages",
1262 "nr_sec_page_table_pages",
1263#ifdef CONFIG_IOMMU_SUPPORT
1264 "nr_iommu_pages",
1265#endif
1266#ifdef CONFIG_SWAP
1267 "nr_swapcached",
1268#endif
1269#ifdef CONFIG_NUMA_BALANCING
1270 "pgpromote_success",
1271 "pgpromote_candidate",
1272#endif
1273 "pgdemote_kswapd",
1274 "pgdemote_direct",
1275 "pgdemote_khugepaged",
1276#ifdef CONFIG_HUGETLB_PAGE
1277 "nr_hugetlb",
1278#endif
1279 /* system-wide enum vm_stat_item counters */
1280 "nr_dirty_threshold",
1281 "nr_dirty_background_threshold",
1282 "nr_memmap_pages",
1283 "nr_memmap_boot_pages",
1284
1285#if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1286 /* enum vm_event_item counters */
1287 "pgpgin",
1288 "pgpgout",
1289 "pswpin",
1290 "pswpout",
1291
1292 TEXTS_FOR_ZONES("pgalloc")
1293 TEXTS_FOR_ZONES("allocstall")
1294 TEXTS_FOR_ZONES("pgskip")
1295
1296 "pgfree",
1297 "pgactivate",
1298 "pgdeactivate",
1299 "pglazyfree",
1300
1301 "pgfault",
1302 "pgmajfault",
1303 "pglazyfreed",
1304
1305 "pgrefill",
1306 "pgreuse",
1307 "pgsteal_kswapd",
1308 "pgsteal_direct",
1309 "pgsteal_khugepaged",
1310 "pgscan_kswapd",
1311 "pgscan_direct",
1312 "pgscan_khugepaged",
1313 "pgscan_direct_throttle",
1314 "pgscan_anon",
1315 "pgscan_file",
1316 "pgsteal_anon",
1317 "pgsteal_file",
1318
1319#ifdef CONFIG_NUMA
1320 "zone_reclaim_success",
1321 "zone_reclaim_failed",
1322#endif
1323 "pginodesteal",
1324 "slabs_scanned",
1325 "kswapd_inodesteal",
1326 "kswapd_low_wmark_hit_quickly",
1327 "kswapd_high_wmark_hit_quickly",
1328 "pageoutrun",
1329
1330 "pgrotated",
1331
1332 "drop_pagecache",
1333 "drop_slab",
1334 "oom_kill",
1335
1336#ifdef CONFIG_NUMA_BALANCING
1337 "numa_pte_updates",
1338 "numa_huge_pte_updates",
1339 "numa_hint_faults",
1340 "numa_hint_faults_local",
1341 "numa_pages_migrated",
1342#endif
1343#ifdef CONFIG_MIGRATION
1344 "pgmigrate_success",
1345 "pgmigrate_fail",
1346 "thp_migration_success",
1347 "thp_migration_fail",
1348 "thp_migration_split",
1349#endif
1350#ifdef CONFIG_COMPACTION
1351 "compact_migrate_scanned",
1352 "compact_free_scanned",
1353 "compact_isolated",
1354 "compact_stall",
1355 "compact_fail",
1356 "compact_success",
1357 "compact_daemon_wake",
1358 "compact_daemon_migrate_scanned",
1359 "compact_daemon_free_scanned",
1360#endif
1361
1362#ifdef CONFIG_HUGETLB_PAGE
1363 "htlb_buddy_alloc_success",
1364 "htlb_buddy_alloc_fail",
1365#endif
1366#ifdef CONFIG_CMA
1367 "cma_alloc_success",
1368 "cma_alloc_fail",
1369#endif
1370 "unevictable_pgs_culled",
1371 "unevictable_pgs_scanned",
1372 "unevictable_pgs_rescued",
1373 "unevictable_pgs_mlocked",
1374 "unevictable_pgs_munlocked",
1375 "unevictable_pgs_cleared",
1376 "unevictable_pgs_stranded",
1377
1378#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1379 "thp_fault_alloc",
1380 "thp_fault_fallback",
1381 "thp_fault_fallback_charge",
1382 "thp_collapse_alloc",
1383 "thp_collapse_alloc_failed",
1384 "thp_file_alloc",
1385 "thp_file_fallback",
1386 "thp_file_fallback_charge",
1387 "thp_file_mapped",
1388 "thp_split_page",
1389 "thp_split_page_failed",
1390 "thp_deferred_split_page",
1391 "thp_underused_split_page",
1392 "thp_split_pmd",
1393 "thp_scan_exceed_none_pte",
1394 "thp_scan_exceed_swap_pte",
1395 "thp_scan_exceed_share_pte",
1396#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1397 "thp_split_pud",
1398#endif
1399 "thp_zero_page_alloc",
1400 "thp_zero_page_alloc_failed",
1401 "thp_swpout",
1402 "thp_swpout_fallback",
1403#endif
1404#ifdef CONFIG_MEMORY_BALLOON
1405 "balloon_inflate",
1406 "balloon_deflate",
1407#ifdef CONFIG_BALLOON_COMPACTION
1408 "balloon_migrate",
1409#endif
1410#endif /* CONFIG_MEMORY_BALLOON */
1411#ifdef CONFIG_DEBUG_TLBFLUSH
1412 "nr_tlb_remote_flush",
1413 "nr_tlb_remote_flush_received",
1414 "nr_tlb_local_flush_all",
1415 "nr_tlb_local_flush_one",
1416#endif /* CONFIG_DEBUG_TLBFLUSH */
1417
1418#ifdef CONFIG_SWAP
1419 "swap_ra",
1420 "swap_ra_hit",
1421 "swpin_zero",
1422 "swpout_zero",
1423#ifdef CONFIG_KSM
1424 "ksm_swpin_copy",
1425#endif
1426#endif
1427#ifdef CONFIG_KSM
1428 "cow_ksm",
1429#endif
1430#ifdef CONFIG_ZSWAP
1431 "zswpin",
1432 "zswpout",
1433 "zswpwb",
1434#endif
1435#ifdef CONFIG_X86
1436 "direct_map_level2_splits",
1437 "direct_map_level3_splits",
1438#endif
1439#ifdef CONFIG_PER_VMA_LOCK_STATS
1440 "vma_lock_success",
1441 "vma_lock_abort",
1442 "vma_lock_retry",
1443 "vma_lock_miss",
1444#endif
1445#ifdef CONFIG_DEBUG_STACK_USAGE
1446 "kstack_1k",
1447#if THREAD_SIZE > 1024
1448 "kstack_2k",
1449#endif
1450#if THREAD_SIZE > 2048
1451 "kstack_4k",
1452#endif
1453#if THREAD_SIZE > 4096
1454 "kstack_8k",
1455#endif
1456#if THREAD_SIZE > 8192
1457 "kstack_16k",
1458#endif
1459#if THREAD_SIZE > 16384
1460 "kstack_32k",
1461#endif
1462#if THREAD_SIZE > 32768
1463 "kstack_64k",
1464#endif
1465#if THREAD_SIZE > 65536
1466 "kstack_rest",
1467#endif
1468#endif
1469#endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1470};
1471#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1472
1473#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1474 defined(CONFIG_PROC_FS)
1475static void *frag_start(struct seq_file *m, loff_t *pos)
1476{
1477 pg_data_t *pgdat;
1478 loff_t node = *pos;
1479
1480 for (pgdat = first_online_pgdat();
1481 pgdat && node;
1482 pgdat = next_online_pgdat(pgdat))
1483 --node;
1484
1485 return pgdat;
1486}
1487
1488static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1489{
1490 pg_data_t *pgdat = (pg_data_t *)arg;
1491
1492 (*pos)++;
1493 return next_online_pgdat(pgdat);
1494}
1495
1496static void frag_stop(struct seq_file *m, void *arg)
1497{
1498}
1499
1500/*
1501 * Walk zones in a node and print using a callback.
1502 * If @assert_populated is true, only use callback for zones that are populated.
1503 */
1504static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1505 bool assert_populated, bool nolock,
1506 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1507{
1508 struct zone *zone;
1509 struct zone *node_zones = pgdat->node_zones;
1510 unsigned long flags;
1511
1512 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1513 if (assert_populated && !populated_zone(zone))
1514 continue;
1515
1516 if (!nolock)
1517 spin_lock_irqsave(&zone->lock, flags);
1518 print(m, pgdat, zone);
1519 if (!nolock)
1520 spin_unlock_irqrestore(&zone->lock, flags);
1521 }
1522}
1523#endif
1524
1525#ifdef CONFIG_PROC_FS
1526static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1527 struct zone *zone)
1528{
1529 int order;
1530
1531 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1532 for (order = 0; order < NR_PAGE_ORDERS; ++order)
1533 /*
1534 * Access to nr_free is lockless as nr_free is used only for
1535 * printing purposes. Use data_race to avoid KCSAN warning.
1536 */
1537 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1538 seq_putc(m, '\n');
1539}
1540
1541/*
1542 * This walks the free areas for each zone.
1543 */
1544static int frag_show(struct seq_file *m, void *arg)
1545{
1546 pg_data_t *pgdat = (pg_data_t *)arg;
1547 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1548 return 0;
1549}
1550
1551static void pagetypeinfo_showfree_print(struct seq_file *m,
1552 pg_data_t *pgdat, struct zone *zone)
1553{
1554 int order, mtype;
1555
1556 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1557 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1558 pgdat->node_id,
1559 zone->name,
1560 migratetype_names[mtype]);
1561 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1562 unsigned long freecount = 0;
1563 struct free_area *area;
1564 struct list_head *curr;
1565 bool overflow = false;
1566
1567 area = &(zone->free_area[order]);
1568
1569 list_for_each(curr, &area->free_list[mtype]) {
1570 /*
1571 * Cap the free_list iteration because it might
1572 * be really large and we are under a spinlock
1573 * so a long time spent here could trigger a
1574 * hard lockup detector. Anyway this is a
1575 * debugging tool so knowing there is a handful
1576 * of pages of this order should be more than
1577 * sufficient.
1578 */
1579 if (++freecount >= 100000) {
1580 overflow = true;
1581 break;
1582 }
1583 }
1584 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1585 spin_unlock_irq(&zone->lock);
1586 cond_resched();
1587 spin_lock_irq(&zone->lock);
1588 }
1589 seq_putc(m, '\n');
1590 }
1591}
1592
1593/* Print out the free pages at each order for each migatetype */
1594static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1595{
1596 int order;
1597 pg_data_t *pgdat = (pg_data_t *)arg;
1598
1599 /* Print header */
1600 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1601 for (order = 0; order < NR_PAGE_ORDERS; ++order)
1602 seq_printf(m, "%6d ", order);
1603 seq_putc(m, '\n');
1604
1605 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1606}
1607
1608static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1609 pg_data_t *pgdat, struct zone *zone)
1610{
1611 int mtype;
1612 unsigned long pfn;
1613 unsigned long start_pfn = zone->zone_start_pfn;
1614 unsigned long end_pfn = zone_end_pfn(zone);
1615 unsigned long count[MIGRATE_TYPES] = { 0, };
1616
1617 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1618 struct page *page;
1619
1620 page = pfn_to_online_page(pfn);
1621 if (!page)
1622 continue;
1623
1624 if (page_zone(page) != zone)
1625 continue;
1626
1627 mtype = get_pageblock_migratetype(page);
1628
1629 if (mtype < MIGRATE_TYPES)
1630 count[mtype]++;
1631 }
1632
1633 /* Print counts */
1634 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1635 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1636 seq_printf(m, "%12lu ", count[mtype]);
1637 seq_putc(m, '\n');
1638}
1639
1640/* Print out the number of pageblocks for each migratetype */
1641static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1642{
1643 int mtype;
1644 pg_data_t *pgdat = (pg_data_t *)arg;
1645
1646 seq_printf(m, "\n%-23s", "Number of blocks type ");
1647 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1648 seq_printf(m, "%12s ", migratetype_names[mtype]);
1649 seq_putc(m, '\n');
1650 walk_zones_in_node(m, pgdat, true, false,
1651 pagetypeinfo_showblockcount_print);
1652}
1653
1654/*
1655 * Print out the number of pageblocks for each migratetype that contain pages
1656 * of other types. This gives an indication of how well fallbacks are being
1657 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1658 * to determine what is going on
1659 */
1660static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1661{
1662#ifdef CONFIG_PAGE_OWNER
1663 int mtype;
1664
1665 if (!static_branch_unlikely(&page_owner_inited))
1666 return;
1667
1668 drain_all_pages(NULL);
1669
1670 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1671 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1672 seq_printf(m, "%12s ", migratetype_names[mtype]);
1673 seq_putc(m, '\n');
1674
1675 walk_zones_in_node(m, pgdat, true, true,
1676 pagetypeinfo_showmixedcount_print);
1677#endif /* CONFIG_PAGE_OWNER */
1678}
1679
1680/*
1681 * This prints out statistics in relation to grouping pages by mobility.
1682 * It is expensive to collect so do not constantly read the file.
1683 */
1684static int pagetypeinfo_show(struct seq_file *m, void *arg)
1685{
1686 pg_data_t *pgdat = (pg_data_t *)arg;
1687
1688 /* check memoryless node */
1689 if (!node_state(pgdat->node_id, N_MEMORY))
1690 return 0;
1691
1692 seq_printf(m, "Page block order: %d\n", pageblock_order);
1693 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1694 seq_putc(m, '\n');
1695 pagetypeinfo_showfree(m, pgdat);
1696 pagetypeinfo_showblockcount(m, pgdat);
1697 pagetypeinfo_showmixedcount(m, pgdat);
1698
1699 return 0;
1700}
1701
1702static const struct seq_operations fragmentation_op = {
1703 .start = frag_start,
1704 .next = frag_next,
1705 .stop = frag_stop,
1706 .show = frag_show,
1707};
1708
1709static const struct seq_operations pagetypeinfo_op = {
1710 .start = frag_start,
1711 .next = frag_next,
1712 .stop = frag_stop,
1713 .show = pagetypeinfo_show,
1714};
1715
1716static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1717{
1718 int zid;
1719
1720 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1721 struct zone *compare = &pgdat->node_zones[zid];
1722
1723 if (populated_zone(compare))
1724 return zone == compare;
1725 }
1726
1727 return false;
1728}
1729
1730static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1731 struct zone *zone)
1732{
1733 int i;
1734 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1735 if (is_zone_first_populated(pgdat, zone)) {
1736 seq_printf(m, "\n per-node stats");
1737 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1738 unsigned long pages = node_page_state_pages(pgdat, i);
1739
1740 if (vmstat_item_print_in_thp(i))
1741 pages /= HPAGE_PMD_NR;
1742 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1743 pages);
1744 }
1745 }
1746 seq_printf(m,
1747 "\n pages free %lu"
1748 "\n boost %lu"
1749 "\n min %lu"
1750 "\n low %lu"
1751 "\n high %lu"
1752 "\n promo %lu"
1753 "\n spanned %lu"
1754 "\n present %lu"
1755 "\n managed %lu"
1756 "\n cma %lu",
1757 zone_page_state(zone, NR_FREE_PAGES),
1758 zone->watermark_boost,
1759 min_wmark_pages(zone),
1760 low_wmark_pages(zone),
1761 high_wmark_pages(zone),
1762 promo_wmark_pages(zone),
1763 zone->spanned_pages,
1764 zone->present_pages,
1765 zone_managed_pages(zone),
1766 zone_cma_pages(zone));
1767
1768 seq_printf(m,
1769 "\n protection: (%ld",
1770 zone->lowmem_reserve[0]);
1771 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1772 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1773 seq_putc(m, ')');
1774
1775 /* If unpopulated, no other information is useful */
1776 if (!populated_zone(zone)) {
1777 seq_putc(m, '\n');
1778 return;
1779 }
1780
1781 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1782 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1783 zone_page_state(zone, i));
1784
1785#ifdef CONFIG_NUMA
1786 fold_vm_zone_numa_events(zone);
1787 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1788 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1789 zone_numa_event_state(zone, i));
1790#endif
1791
1792 seq_printf(m, "\n pagesets");
1793 for_each_online_cpu(i) {
1794 struct per_cpu_pages *pcp;
1795 struct per_cpu_zonestat __maybe_unused *pzstats;
1796
1797 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1798 seq_printf(m,
1799 "\n cpu: %i"
1800 "\n count: %i"
1801 "\n high: %i"
1802 "\n batch: %i"
1803 "\n high_min: %i"
1804 "\n high_max: %i",
1805 i,
1806 pcp->count,
1807 pcp->high,
1808 pcp->batch,
1809 pcp->high_min,
1810 pcp->high_max);
1811#ifdef CONFIG_SMP
1812 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1813 seq_printf(m, "\n vm stats threshold: %d",
1814 pzstats->stat_threshold);
1815#endif
1816 }
1817 seq_printf(m,
1818 "\n node_unreclaimable: %u"
1819 "\n start_pfn: %lu",
1820 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1821 zone->zone_start_pfn);
1822 seq_putc(m, '\n');
1823}
1824
1825/*
1826 * Output information about zones in @pgdat. All zones are printed regardless
1827 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1828 * set of all zones and userspace would not be aware of such zones if they are
1829 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1830 */
1831static int zoneinfo_show(struct seq_file *m, void *arg)
1832{
1833 pg_data_t *pgdat = (pg_data_t *)arg;
1834 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1835 return 0;
1836}
1837
1838static const struct seq_operations zoneinfo_op = {
1839 .start = frag_start, /* iterate over all zones. The same as in
1840 * fragmentation. */
1841 .next = frag_next,
1842 .stop = frag_stop,
1843 .show = zoneinfo_show,
1844};
1845
1846#define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1847 NR_VM_NUMA_EVENT_ITEMS + \
1848 NR_VM_NODE_STAT_ITEMS + \
1849 NR_VM_STAT_ITEMS + \
1850 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1851 NR_VM_EVENT_ITEMS : 0))
1852
1853static void *vmstat_start(struct seq_file *m, loff_t *pos)
1854{
1855 unsigned long *v;
1856 int i;
1857
1858 if (*pos >= NR_VMSTAT_ITEMS)
1859 return NULL;
1860
1861 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1862 fold_vm_numa_events();
1863 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1864 m->private = v;
1865 if (!v)
1866 return ERR_PTR(-ENOMEM);
1867 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1868 v[i] = global_zone_page_state(i);
1869 v += NR_VM_ZONE_STAT_ITEMS;
1870
1871#ifdef CONFIG_NUMA
1872 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1873 v[i] = global_numa_event_state(i);
1874 v += NR_VM_NUMA_EVENT_ITEMS;
1875#endif
1876
1877 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1878 v[i] = global_node_page_state_pages(i);
1879 if (vmstat_item_print_in_thp(i))
1880 v[i] /= HPAGE_PMD_NR;
1881 }
1882 v += NR_VM_NODE_STAT_ITEMS;
1883
1884 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1885 v + NR_DIRTY_THRESHOLD);
1886 v[NR_MEMMAP_PAGES] = atomic_long_read(&nr_memmap_pages);
1887 v[NR_MEMMAP_BOOT_PAGES] = atomic_long_read(&nr_memmap_boot_pages);
1888 v += NR_VM_STAT_ITEMS;
1889
1890#ifdef CONFIG_VM_EVENT_COUNTERS
1891 all_vm_events(v);
1892 v[PGPGIN] /= 2; /* sectors -> kbytes */
1893 v[PGPGOUT] /= 2;
1894#endif
1895 return (unsigned long *)m->private + *pos;
1896}
1897
1898static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1899{
1900 (*pos)++;
1901 if (*pos >= NR_VMSTAT_ITEMS)
1902 return NULL;
1903 return (unsigned long *)m->private + *pos;
1904}
1905
1906static int vmstat_show(struct seq_file *m, void *arg)
1907{
1908 unsigned long *l = arg;
1909 unsigned long off = l - (unsigned long *)m->private;
1910
1911 seq_puts(m, vmstat_text[off]);
1912 seq_put_decimal_ull(m, " ", *l);
1913 seq_putc(m, '\n');
1914
1915 if (off == NR_VMSTAT_ITEMS - 1) {
1916 /*
1917 * We've come to the end - add any deprecated counters to avoid
1918 * breaking userspace which might depend on them being present.
1919 */
1920 seq_puts(m, "nr_unstable 0\n");
1921 }
1922 return 0;
1923}
1924
1925static void vmstat_stop(struct seq_file *m, void *arg)
1926{
1927 kfree(m->private);
1928 m->private = NULL;
1929}
1930
1931static const struct seq_operations vmstat_op = {
1932 .start = vmstat_start,
1933 .next = vmstat_next,
1934 .stop = vmstat_stop,
1935 .show = vmstat_show,
1936};
1937#endif /* CONFIG_PROC_FS */
1938
1939#ifdef CONFIG_SMP
1940static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1941int sysctl_stat_interval __read_mostly = HZ;
1942static int vmstat_late_init_done;
1943
1944#ifdef CONFIG_PROC_FS
1945static void refresh_vm_stats(struct work_struct *work)
1946{
1947 refresh_cpu_vm_stats(true);
1948}
1949
1950int vmstat_refresh(const struct ctl_table *table, int write,
1951 void *buffer, size_t *lenp, loff_t *ppos)
1952{
1953 long val;
1954 int err;
1955 int i;
1956
1957 /*
1958 * The regular update, every sysctl_stat_interval, may come later
1959 * than expected: leaving a significant amount in per_cpu buckets.
1960 * This is particularly misleading when checking a quantity of HUGE
1961 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1962 * which can equally be echo'ed to or cat'ted from (by root),
1963 * can be used to update the stats just before reading them.
1964 *
1965 * Oh, and since global_zone_page_state() etc. are so careful to hide
1966 * transiently negative values, report an error here if any of
1967 * the stats is negative, so we know to go looking for imbalance.
1968 */
1969 err = schedule_on_each_cpu(refresh_vm_stats);
1970 if (err)
1971 return err;
1972 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1973 /*
1974 * Skip checking stats known to go negative occasionally.
1975 */
1976 switch (i) {
1977 case NR_ZONE_WRITE_PENDING:
1978 case NR_FREE_CMA_PAGES:
1979 continue;
1980 }
1981 val = atomic_long_read(&vm_zone_stat[i]);
1982 if (val < 0) {
1983 pr_warn("%s: %s %ld\n",
1984 __func__, zone_stat_name(i), val);
1985 }
1986 }
1987 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1988 /*
1989 * Skip checking stats known to go negative occasionally.
1990 */
1991 switch (i) {
1992 case NR_WRITEBACK:
1993 continue;
1994 }
1995 val = atomic_long_read(&vm_node_stat[i]);
1996 if (val < 0) {
1997 pr_warn("%s: %s %ld\n",
1998 __func__, node_stat_name(i), val);
1999 }
2000 }
2001 if (write)
2002 *ppos += *lenp;
2003 else
2004 *lenp = 0;
2005 return 0;
2006}
2007#endif /* CONFIG_PROC_FS */
2008
2009static void vmstat_update(struct work_struct *w)
2010{
2011 if (refresh_cpu_vm_stats(true)) {
2012 /*
2013 * Counters were updated so we expect more updates
2014 * to occur in the future. Keep on running the
2015 * update worker thread.
2016 */
2017 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
2018 this_cpu_ptr(&vmstat_work),
2019 round_jiffies_relative(sysctl_stat_interval));
2020 }
2021}
2022
2023/*
2024 * Check if the diffs for a certain cpu indicate that
2025 * an update is needed.
2026 */
2027static bool need_update(int cpu)
2028{
2029 pg_data_t *last_pgdat = NULL;
2030 struct zone *zone;
2031
2032 for_each_populated_zone(zone) {
2033 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
2034 struct per_cpu_nodestat *n;
2035
2036 /*
2037 * The fast way of checking if there are any vmstat diffs.
2038 */
2039 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
2040 return true;
2041
2042 if (last_pgdat == zone->zone_pgdat)
2043 continue;
2044 last_pgdat = zone->zone_pgdat;
2045 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
2046 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
2047 return true;
2048 }
2049 return false;
2050}
2051
2052/*
2053 * Switch off vmstat processing and then fold all the remaining differentials
2054 * until the diffs stay at zero. The function is used by NOHZ and can only be
2055 * invoked when tick processing is not active.
2056 */
2057void quiet_vmstat(void)
2058{
2059 if (system_state != SYSTEM_RUNNING)
2060 return;
2061
2062 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
2063 return;
2064
2065 if (!need_update(smp_processor_id()))
2066 return;
2067
2068 /*
2069 * Just refresh counters and do not care about the pending delayed
2070 * vmstat_update. It doesn't fire that often to matter and canceling
2071 * it would be too expensive from this path.
2072 * vmstat_shepherd will take care about that for us.
2073 */
2074 refresh_cpu_vm_stats(false);
2075}
2076
2077/*
2078 * Shepherd worker thread that checks the
2079 * differentials of processors that have their worker
2080 * threads for vm statistics updates disabled because of
2081 * inactivity.
2082 */
2083static void vmstat_shepherd(struct work_struct *w);
2084
2085static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2086
2087static void vmstat_shepherd(struct work_struct *w)
2088{
2089 int cpu;
2090
2091 cpus_read_lock();
2092 /* Check processors whose vmstat worker threads have been disabled */
2093 for_each_online_cpu(cpu) {
2094 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2095
2096 /*
2097 * In kernel users of vmstat counters either require the precise value and
2098 * they are using zone_page_state_snapshot interface or they can live with
2099 * an imprecision as the regular flushing can happen at arbitrary time and
2100 * cumulative error can grow (see calculate_normal_threshold).
2101 *
2102 * From that POV the regular flushing can be postponed for CPUs that have
2103 * been isolated from the kernel interference without critical
2104 * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2105 * for all isolated CPUs to avoid interference with the isolated workload.
2106 */
2107 if (cpu_is_isolated(cpu))
2108 continue;
2109
2110 if (!delayed_work_pending(dw) && need_update(cpu))
2111 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2112
2113 cond_resched();
2114 }
2115 cpus_read_unlock();
2116
2117 schedule_delayed_work(&shepherd,
2118 round_jiffies_relative(sysctl_stat_interval));
2119}
2120
2121static void __init start_shepherd_timer(void)
2122{
2123 int cpu;
2124
2125 for_each_possible_cpu(cpu) {
2126 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2127 vmstat_update);
2128
2129 /*
2130 * For secondary CPUs during CPU hotplug scenarios,
2131 * vmstat_cpu_online() will enable the work.
2132 * mm/vmstat:online enables and disables vmstat_work
2133 * symmetrically during CPU hotplug events.
2134 */
2135 if (!cpu_online(cpu))
2136 disable_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2137 }
2138
2139 schedule_delayed_work(&shepherd,
2140 round_jiffies_relative(sysctl_stat_interval));
2141}
2142
2143static void __init init_cpu_node_state(void)
2144{
2145 int node;
2146
2147 for_each_online_node(node) {
2148 if (!cpumask_empty(cpumask_of_node(node)))
2149 node_set_state(node, N_CPU);
2150 }
2151}
2152
2153static int vmstat_cpu_online(unsigned int cpu)
2154{
2155 if (vmstat_late_init_done)
2156 refresh_zone_stat_thresholds();
2157
2158 if (!node_state(cpu_to_node(cpu), N_CPU)) {
2159 node_set_state(cpu_to_node(cpu), N_CPU);
2160 }
2161 enable_delayed_work(&per_cpu(vmstat_work, cpu));
2162
2163 return 0;
2164}
2165
2166static int vmstat_cpu_down_prep(unsigned int cpu)
2167{
2168 disable_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2169 return 0;
2170}
2171
2172static int vmstat_cpu_dead(unsigned int cpu)
2173{
2174 const struct cpumask *node_cpus;
2175 int node;
2176
2177 node = cpu_to_node(cpu);
2178
2179 refresh_zone_stat_thresholds();
2180 node_cpus = cpumask_of_node(node);
2181 if (!cpumask_empty(node_cpus))
2182 return 0;
2183
2184 node_clear_state(node, N_CPU);
2185
2186 return 0;
2187}
2188
2189static int __init vmstat_late_init(void)
2190{
2191 refresh_zone_stat_thresholds();
2192 vmstat_late_init_done = 1;
2193
2194 return 0;
2195}
2196late_initcall(vmstat_late_init);
2197#endif
2198
2199struct workqueue_struct *mm_percpu_wq;
2200
2201void __init init_mm_internals(void)
2202{
2203 int ret __maybe_unused;
2204
2205 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2206
2207#ifdef CONFIG_SMP
2208 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2209 NULL, vmstat_cpu_dead);
2210 if (ret < 0)
2211 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2212
2213 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2214 vmstat_cpu_online,
2215 vmstat_cpu_down_prep);
2216 if (ret < 0)
2217 pr_err("vmstat: failed to register 'online' hotplug state\n");
2218
2219 cpus_read_lock();
2220 init_cpu_node_state();
2221 cpus_read_unlock();
2222
2223 start_shepherd_timer();
2224#endif
2225#ifdef CONFIG_PROC_FS
2226 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2227 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2228 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2229 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2230#endif
2231}
2232
2233#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2234
2235/*
2236 * Return an index indicating how much of the available free memory is
2237 * unusable for an allocation of the requested size.
2238 */
2239static int unusable_free_index(unsigned int order,
2240 struct contig_page_info *info)
2241{
2242 /* No free memory is interpreted as all free memory is unusable */
2243 if (info->free_pages == 0)
2244 return 1000;
2245
2246 /*
2247 * Index should be a value between 0 and 1. Return a value to 3
2248 * decimal places.
2249 *
2250 * 0 => no fragmentation
2251 * 1 => high fragmentation
2252 */
2253 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2254
2255}
2256
2257static void unusable_show_print(struct seq_file *m,
2258 pg_data_t *pgdat, struct zone *zone)
2259{
2260 unsigned int order;
2261 int index;
2262 struct contig_page_info info;
2263
2264 seq_printf(m, "Node %d, zone %8s ",
2265 pgdat->node_id,
2266 zone->name);
2267 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2268 fill_contig_page_info(zone, order, &info);
2269 index = unusable_free_index(order, &info);
2270 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2271 }
2272
2273 seq_putc(m, '\n');
2274}
2275
2276/*
2277 * Display unusable free space index
2278 *
2279 * The unusable free space index measures how much of the available free
2280 * memory cannot be used to satisfy an allocation of a given size and is a
2281 * value between 0 and 1. The higher the value, the more of free memory is
2282 * unusable and by implication, the worse the external fragmentation is. This
2283 * can be expressed as a percentage by multiplying by 100.
2284 */
2285static int unusable_show(struct seq_file *m, void *arg)
2286{
2287 pg_data_t *pgdat = (pg_data_t *)arg;
2288
2289 /* check memoryless node */
2290 if (!node_state(pgdat->node_id, N_MEMORY))
2291 return 0;
2292
2293 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2294
2295 return 0;
2296}
2297
2298static const struct seq_operations unusable_sops = {
2299 .start = frag_start,
2300 .next = frag_next,
2301 .stop = frag_stop,
2302 .show = unusable_show,
2303};
2304
2305DEFINE_SEQ_ATTRIBUTE(unusable);
2306
2307static void extfrag_show_print(struct seq_file *m,
2308 pg_data_t *pgdat, struct zone *zone)
2309{
2310 unsigned int order;
2311 int index;
2312
2313 /* Alloc on stack as interrupts are disabled for zone walk */
2314 struct contig_page_info info;
2315
2316 seq_printf(m, "Node %d, zone %8s ",
2317 pgdat->node_id,
2318 zone->name);
2319 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2320 fill_contig_page_info(zone, order, &info);
2321 index = __fragmentation_index(order, &info);
2322 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2323 }
2324
2325 seq_putc(m, '\n');
2326}
2327
2328/*
2329 * Display fragmentation index for orders that allocations would fail for
2330 */
2331static int extfrag_show(struct seq_file *m, void *arg)
2332{
2333 pg_data_t *pgdat = (pg_data_t *)arg;
2334
2335 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2336
2337 return 0;
2338}
2339
2340static const struct seq_operations extfrag_sops = {
2341 .start = frag_start,
2342 .next = frag_next,
2343 .stop = frag_stop,
2344 .show = extfrag_show,
2345};
2346
2347DEFINE_SEQ_ATTRIBUTE(extfrag);
2348
2349static int __init extfrag_debug_init(void)
2350{
2351 struct dentry *extfrag_debug_root;
2352
2353 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2354
2355 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2356 &unusable_fops);
2357
2358 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2359 &extfrag_fops);
2360
2361 return 0;
2362}
2363
2364module_init(extfrag_debug_init);
2365
2366#endif