1 // SPDX-License-Identifier: GPL-2.0-only
5 * Manages VM statistics
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Copyright (C) 2006 Silicon Graphics, Inc.,
10 * Christoph Lameter <christoph@lameter.com>
11 * Copyright (C) 2008-2014 Christoph Lameter
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>
34 #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
37 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
39 /* zero numa counters within a zone */
40 static void zero_zone_numa_counters(struct zone *zone)
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]
52 /* zero numa counters of all the populated zones */
53 static void zero_zones_numa_counters(void)
57 for_each_populated_zone(zone)
58 zero_zone_numa_counters(zone);
61 /* zero global numa counters */
62 static void zero_global_numa_counters(void)
66 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
67 atomic_long_set(&vm_numa_stat[item], 0);
70 static void invalid_numa_statistics(void)
72 zero_zones_numa_counters();
73 zero_global_numa_counters();
76 static DEFINE_MUTEX(vm_numa_stat_lock);
78 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
79 void __user *buffer, size_t *length, loff_t *ppos)
83 mutex_lock(&vm_numa_stat_lock);
85 oldval = sysctl_vm_numa_stat;
86 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
90 if (oldval == sysctl_vm_numa_stat)
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");
96 static_branch_disable(&vm_numa_stat_key);
97 invalid_numa_statistics();
98 pr_info("disable numa statistics, and clear numa counters\n");
102 mutex_unlock(&vm_numa_stat_lock);
107 #ifdef CONFIG_VM_EVENT_COUNTERS
108 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
109 EXPORT_PER_CPU_SYMBOL(vm_event_states);
111 static void sum_vm_events(unsigned long *ret)
116 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
118 for_each_online_cpu(cpu) {
119 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
121 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
122 ret[i] += this->event[i];
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.
131 void all_vm_events(unsigned long *ret)
137 EXPORT_SYMBOL_GPL(all_vm_events);
140 * Fold the foreign cpu events into our own.
142 * This is adding to the events on one processor
143 * but keeps the global counts constant.
145 void vm_events_fold_cpu(int cpu)
147 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
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;
156 #endif /* CONFIG_VM_EVENT_COUNTERS */
159 * Manage combined zone based / global counters
161 * vm_stat contains the global counters
163 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
164 atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
165 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
166 EXPORT_SYMBOL(vm_zone_stat);
167 EXPORT_SYMBOL(vm_numa_stat);
168 EXPORT_SYMBOL(vm_node_stat);
172 int calculate_pressure_threshold(struct zone *zone)
175 int watermark_distance;
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
185 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
186 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
189 * Maximum threshold is 125
191 threshold = min(125, threshold);
196 int calculate_normal_threshold(struct zone *zone)
199 int mem; /* memory in 128 MB units */
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.
207 * Some sample thresholds:
209 * Threshold Processors (fls) Zonesize fls(mem+1)
210 * ------------------------------------------------------------------
227 * 125 1024 10 8-16 GB 8
228 * 125 1024 10 16-32 GB 9
231 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
233 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
236 * Maximum threshold is 125
238 threshold = min(125, threshold);
244 * Refresh the thresholds for each zone.
246 void refresh_zone_stat_thresholds(void)
248 struct pglist_data *pgdat;
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;
260 for_each_populated_zone(zone) {
261 struct pglist_data *pgdat = zone->zone_pgdat;
262 unsigned long max_drift, tolerate_drift;
264 threshold = calculate_normal_threshold(zone);
266 for_each_online_cpu(cpu) {
269 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
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);
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
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) +
291 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
292 int (*calculate_pressure)(struct zone *))
299 for (i = 0; i < pgdat->nr_zones; i++) {
300 zone = &pgdat->node_zones[i];
301 if (!zone->percpu_drift_mark)
304 threshold = (*calculate_pressure)(zone);
305 for_each_online_cpu(cpu)
306 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
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.
316 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
319 struct per_cpu_pageset __percpu *pcp = zone->pageset;
320 s8 __percpu *p = pcp->vm_stat_diff + item;
324 x = delta + __this_cpu_read(*p);
326 t = __this_cpu_read(pcp->stat_threshold);
328 if (unlikely(x > t || x < -t)) {
329 zone_page_state_add(x, zone, item);
332 __this_cpu_write(*p, x);
334 EXPORT_SYMBOL(__mod_zone_page_state);
336 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
339 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
340 s8 __percpu *p = pcp->vm_node_stat_diff + item;
344 x = delta + __this_cpu_read(*p);
346 t = __this_cpu_read(pcp->stat_threshold);
348 if (unlikely(x > t || x < -t)) {
349 node_page_state_add(x, pgdat, item);
352 __this_cpu_write(*p, x);
354 EXPORT_SYMBOL(__mod_node_page_state);
357 * Optimized increment and decrement functions.
359 * These are only for a single page and therefore can take a struct page *
360 * argument instead of struct zone *. This allows the inclusion of the code
361 * generated for page_zone(page) into the optimized functions.
363 * No overflow check is necessary and therefore the differential can be
364 * incremented or decremented in place which may allow the compilers to
365 * generate better code.
366 * The increment or decrement is known and therefore one boundary check can
369 * NOTE: These functions are very performance sensitive. Change only
372 * Some processors have inc/dec instructions that are atomic vs an interrupt.
373 * However, the code must first determine the differential location in a zone
374 * based on the processor number and then inc/dec the counter. There is no
375 * guarantee without disabling preemption that the processor will not change
376 * in between and therefore the atomicity vs. interrupt cannot be exploited
377 * in a useful way here.
379 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
381 struct per_cpu_pageset __percpu *pcp = zone->pageset;
382 s8 __percpu *p = pcp->vm_stat_diff + item;
385 v = __this_cpu_inc_return(*p);
386 t = __this_cpu_read(pcp->stat_threshold);
387 if (unlikely(v > t)) {
388 s8 overstep = t >> 1;
390 zone_page_state_add(v + overstep, zone, item);
391 __this_cpu_write(*p, -overstep);
395 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
397 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
398 s8 __percpu *p = pcp->vm_node_stat_diff + item;
401 v = __this_cpu_inc_return(*p);
402 t = __this_cpu_read(pcp->stat_threshold);
403 if (unlikely(v > t)) {
404 s8 overstep = t >> 1;
406 node_page_state_add(v + overstep, pgdat, item);
407 __this_cpu_write(*p, -overstep);
411 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
413 __inc_zone_state(page_zone(page), item);
415 EXPORT_SYMBOL(__inc_zone_page_state);
417 void __inc_node_page_state(struct page *page, enum node_stat_item item)
419 __inc_node_state(page_pgdat(page), item);
421 EXPORT_SYMBOL(__inc_node_page_state);
423 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
425 struct per_cpu_pageset __percpu *pcp = zone->pageset;
426 s8 __percpu *p = pcp->vm_stat_diff + item;
429 v = __this_cpu_dec_return(*p);
430 t = __this_cpu_read(pcp->stat_threshold);
431 if (unlikely(v < - t)) {
432 s8 overstep = t >> 1;
434 zone_page_state_add(v - overstep, zone, item);
435 __this_cpu_write(*p, overstep);
439 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
441 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
442 s8 __percpu *p = pcp->vm_node_stat_diff + item;
445 v = __this_cpu_dec_return(*p);
446 t = __this_cpu_read(pcp->stat_threshold);
447 if (unlikely(v < - t)) {
448 s8 overstep = t >> 1;
450 node_page_state_add(v - overstep, pgdat, item);
451 __this_cpu_write(*p, overstep);
455 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
457 __dec_zone_state(page_zone(page), item);
459 EXPORT_SYMBOL(__dec_zone_page_state);
461 void __dec_node_page_state(struct page *page, enum node_stat_item item)
463 __dec_node_state(page_pgdat(page), item);
465 EXPORT_SYMBOL(__dec_node_page_state);
467 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
469 * If we have cmpxchg_local support then we do not need to incur the overhead
470 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
472 * mod_state() modifies the zone counter state through atomic per cpu
475 * Overstep mode specifies how overstep should handled:
477 * 1 Overstepping half of threshold
478 * -1 Overstepping minus half of threshold
480 static inline void mod_zone_state(struct zone *zone,
481 enum zone_stat_item item, long delta, int overstep_mode)
483 struct per_cpu_pageset __percpu *pcp = zone->pageset;
484 s8 __percpu *p = pcp->vm_stat_diff + item;
488 z = 0; /* overflow to zone counters */
491 * The fetching of the stat_threshold is racy. We may apply
492 * a counter threshold to the wrong the cpu if we get
493 * rescheduled while executing here. However, the next
494 * counter update will apply the threshold again and
495 * therefore bring the counter under the threshold again.
497 * Most of the time the thresholds are the same anyways
498 * for all cpus in a zone.
500 t = this_cpu_read(pcp->stat_threshold);
502 o = this_cpu_read(*p);
505 if (n > t || n < -t) {
506 int os = overstep_mode * (t >> 1) ;
508 /* Overflow must be added to zone counters */
512 } while (this_cpu_cmpxchg(*p, o, n) != o);
515 zone_page_state_add(z, zone, item);
518 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
521 mod_zone_state(zone, item, delta, 0);
523 EXPORT_SYMBOL(mod_zone_page_state);
525 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
527 mod_zone_state(page_zone(page), item, 1, 1);
529 EXPORT_SYMBOL(inc_zone_page_state);
531 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
533 mod_zone_state(page_zone(page), item, -1, -1);
535 EXPORT_SYMBOL(dec_zone_page_state);
537 static inline void mod_node_state(struct pglist_data *pgdat,
538 enum node_stat_item item, int delta, int overstep_mode)
540 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
541 s8 __percpu *p = pcp->vm_node_stat_diff + item;
545 z = 0; /* overflow to node counters */
548 * The fetching of the stat_threshold is racy. We may apply
549 * a counter threshold to the wrong the cpu if we get
550 * rescheduled while executing here. However, the next
551 * counter update will apply the threshold again and
552 * therefore bring the counter under the threshold again.
554 * Most of the time the thresholds are the same anyways
555 * for all cpus in a node.
557 t = this_cpu_read(pcp->stat_threshold);
559 o = this_cpu_read(*p);
562 if (n > t || n < -t) {
563 int os = overstep_mode * (t >> 1) ;
565 /* Overflow must be added to node counters */
569 } while (this_cpu_cmpxchg(*p, o, n) != o);
572 node_page_state_add(z, pgdat, item);
575 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
578 mod_node_state(pgdat, item, delta, 0);
580 EXPORT_SYMBOL(mod_node_page_state);
582 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
584 mod_node_state(pgdat, item, 1, 1);
587 void inc_node_page_state(struct page *page, enum node_stat_item item)
589 mod_node_state(page_pgdat(page), item, 1, 1);
591 EXPORT_SYMBOL(inc_node_page_state);
593 void dec_node_page_state(struct page *page, enum node_stat_item item)
595 mod_node_state(page_pgdat(page), item, -1, -1);
597 EXPORT_SYMBOL(dec_node_page_state);
600 * Use interrupt disable to serialize counter updates
602 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
607 local_irq_save(flags);
608 __mod_zone_page_state(zone, item, delta);
609 local_irq_restore(flags);
611 EXPORT_SYMBOL(mod_zone_page_state);
613 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
618 zone = page_zone(page);
619 local_irq_save(flags);
620 __inc_zone_state(zone, item);
621 local_irq_restore(flags);
623 EXPORT_SYMBOL(inc_zone_page_state);
625 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
629 local_irq_save(flags);
630 __dec_zone_page_state(page, item);
631 local_irq_restore(flags);
633 EXPORT_SYMBOL(dec_zone_page_state);
635 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
639 local_irq_save(flags);
640 __inc_node_state(pgdat, item);
641 local_irq_restore(flags);
643 EXPORT_SYMBOL(inc_node_state);
645 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
650 local_irq_save(flags);
651 __mod_node_page_state(pgdat, item, delta);
652 local_irq_restore(flags);
654 EXPORT_SYMBOL(mod_node_page_state);
656 void inc_node_page_state(struct page *page, enum node_stat_item item)
659 struct pglist_data *pgdat;
661 pgdat = page_pgdat(page);
662 local_irq_save(flags);
663 __inc_node_state(pgdat, item);
664 local_irq_restore(flags);
666 EXPORT_SYMBOL(inc_node_page_state);
668 void dec_node_page_state(struct page *page, enum node_stat_item item)
672 local_irq_save(flags);
673 __dec_node_page_state(page, item);
674 local_irq_restore(flags);
676 EXPORT_SYMBOL(dec_node_page_state);
680 * Fold a differential into the global counters.
681 * Returns the number of counters updated.
684 static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
689 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
691 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
695 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
697 atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
701 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
703 atomic_long_add(node_diff[i], &vm_node_stat[i]);
709 static int fold_diff(int *zone_diff, int *node_diff)
714 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
716 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
720 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
722 atomic_long_add(node_diff[i], &vm_node_stat[i]);
727 #endif /* CONFIG_NUMA */
730 * Update the zone counters for the current cpu.
732 * Note that refresh_cpu_vm_stats strives to only access
733 * node local memory. The per cpu pagesets on remote zones are placed
734 * in the memory local to the processor using that pageset. So the
735 * loop over all zones will access a series of cachelines local to
738 * The call to zone_page_state_add updates the cachelines with the
739 * statistics in the remote zone struct as well as the global cachelines
740 * with the global counters. These could cause remote node cache line
741 * bouncing and will have to be only done when necessary.
743 * The function returns the number of global counters updated.
745 static int refresh_cpu_vm_stats(bool do_pagesets)
747 struct pglist_data *pgdat;
750 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
752 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
754 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
757 for_each_populated_zone(zone) {
758 struct per_cpu_pageset __percpu *p = zone->pageset;
760 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
763 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
766 atomic_long_add(v, &zone->vm_stat[i]);
767 global_zone_diff[i] += v;
769 /* 3 seconds idle till flush */
770 __this_cpu_write(p->expire, 3);
775 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
778 v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
781 atomic_long_add(v, &zone->vm_numa_stat[i]);
782 global_numa_diff[i] += v;
783 __this_cpu_write(p->expire, 3);
790 * Deal with draining the remote pageset of this
793 * Check if there are pages remaining in this pageset
794 * if not then there is nothing to expire.
796 if (!__this_cpu_read(p->expire) ||
797 !__this_cpu_read(p->pcp.count))
801 * We never drain zones local to this processor.
803 if (zone_to_nid(zone) == numa_node_id()) {
804 __this_cpu_write(p->expire, 0);
808 if (__this_cpu_dec_return(p->expire))
811 if (__this_cpu_read(p->pcp.count)) {
812 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
819 for_each_online_pgdat(pgdat) {
820 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
822 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
825 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
827 atomic_long_add(v, &pgdat->vm_stat[i]);
828 global_node_diff[i] += v;
834 changes += fold_diff(global_zone_diff, global_numa_diff,
837 changes += fold_diff(global_zone_diff, global_node_diff);
843 * Fold the data for an offline cpu into the global array.
844 * There cannot be any access by the offline cpu and therefore
845 * synchronization is simplified.
847 void cpu_vm_stats_fold(int cpu)
849 struct pglist_data *pgdat;
852 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
854 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
856 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
858 for_each_populated_zone(zone) {
859 struct per_cpu_pageset *p;
861 p = per_cpu_ptr(zone->pageset, cpu);
863 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
864 if (p->vm_stat_diff[i]) {
867 v = p->vm_stat_diff[i];
868 p->vm_stat_diff[i] = 0;
869 atomic_long_add(v, &zone->vm_stat[i]);
870 global_zone_diff[i] += v;
874 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
875 if (p->vm_numa_stat_diff[i]) {
878 v = p->vm_numa_stat_diff[i];
879 p->vm_numa_stat_diff[i] = 0;
880 atomic_long_add(v, &zone->vm_numa_stat[i]);
881 global_numa_diff[i] += v;
886 for_each_online_pgdat(pgdat) {
887 struct per_cpu_nodestat *p;
889 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
891 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
892 if (p->vm_node_stat_diff[i]) {
895 v = p->vm_node_stat_diff[i];
896 p->vm_node_stat_diff[i] = 0;
897 atomic_long_add(v, &pgdat->vm_stat[i]);
898 global_node_diff[i] += v;
903 fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
905 fold_diff(global_zone_diff, global_node_diff);
910 * this is only called if !populated_zone(zone), which implies no other users of
911 * pset->vm_stat_diff[] exsist.
913 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
917 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
918 if (pset->vm_stat_diff[i]) {
919 int v = pset->vm_stat_diff[i];
920 pset->vm_stat_diff[i] = 0;
921 atomic_long_add(v, &zone->vm_stat[i]);
922 atomic_long_add(v, &vm_zone_stat[i]);
926 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
927 if (pset->vm_numa_stat_diff[i]) {
928 int v = pset->vm_numa_stat_diff[i];
930 pset->vm_numa_stat_diff[i] = 0;
931 atomic_long_add(v, &zone->vm_numa_stat[i]);
932 atomic_long_add(v, &vm_numa_stat[i]);
939 void __inc_numa_state(struct zone *zone,
940 enum numa_stat_item item)
942 struct per_cpu_pageset __percpu *pcp = zone->pageset;
943 u16 __percpu *p = pcp->vm_numa_stat_diff + item;
946 v = __this_cpu_inc_return(*p);
948 if (unlikely(v > NUMA_STATS_THRESHOLD)) {
949 zone_numa_state_add(v, zone, item);
950 __this_cpu_write(*p, 0);
955 * Determine the per node value of a stat item. This function
956 * is called frequently in a NUMA machine, so try to be as
957 * frugal as possible.
959 unsigned long sum_zone_node_page_state(int node,
960 enum zone_stat_item item)
962 struct zone *zones = NODE_DATA(node)->node_zones;
964 unsigned long count = 0;
966 for (i = 0; i < MAX_NR_ZONES; i++)
967 count += zone_page_state(zones + i, item);
973 * Determine the per node value of a numa stat item. To avoid deviation,
974 * the per cpu stat number in vm_numa_stat_diff[] is also included.
976 unsigned long sum_zone_numa_state(int node,
977 enum numa_stat_item item)
979 struct zone *zones = NODE_DATA(node)->node_zones;
981 unsigned long count = 0;
983 for (i = 0; i < MAX_NR_ZONES; i++)
984 count += zone_numa_state_snapshot(zones + i, item);
990 * Determine the per node value of a stat item.
992 unsigned long node_page_state(struct pglist_data *pgdat,
993 enum node_stat_item item)
995 long x = atomic_long_read(&pgdat->vm_stat[item]);
1004 #ifdef CONFIG_COMPACTION
1006 struct contig_page_info {
1007 unsigned long free_pages;
1008 unsigned long free_blocks_total;
1009 unsigned long free_blocks_suitable;
1013 * Calculate the number of free pages in a zone, how many contiguous
1014 * pages are free and how many are large enough to satisfy an allocation of
1015 * the target size. Note that this function makes no attempt to estimate
1016 * how many suitable free blocks there *might* be if MOVABLE pages were
1017 * migrated. Calculating that is possible, but expensive and can be
1018 * figured out from userspace
1020 static void fill_contig_page_info(struct zone *zone,
1021 unsigned int suitable_order,
1022 struct contig_page_info *info)
1026 info->free_pages = 0;
1027 info->free_blocks_total = 0;
1028 info->free_blocks_suitable = 0;
1030 for (order = 0; order < MAX_ORDER; order++) {
1031 unsigned long blocks;
1033 /* Count number of free blocks */
1034 blocks = zone->free_area[order].nr_free;
1035 info->free_blocks_total += blocks;
1037 /* Count free base pages */
1038 info->free_pages += blocks << order;
1040 /* Count the suitable free blocks */
1041 if (order >= suitable_order)
1042 info->free_blocks_suitable += blocks <<
1043 (order - suitable_order);
1048 * A fragmentation index only makes sense if an allocation of a requested
1049 * size would fail. If that is true, the fragmentation index indicates
1050 * whether external fragmentation or a lack of memory was the problem.
1051 * The value can be used to determine if page reclaim or compaction
1054 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1056 unsigned long requested = 1UL << order;
1058 if (WARN_ON_ONCE(order >= MAX_ORDER))
1061 if (!info->free_blocks_total)
1064 /* Fragmentation index only makes sense when a request would fail */
1065 if (info->free_blocks_suitable)
1069 * Index is between 0 and 1 so return within 3 decimal places
1071 * 0 => allocation would fail due to lack of memory
1072 * 1 => allocation would fail due to fragmentation
1074 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1077 /* Same as __fragmentation index but allocs contig_page_info on stack */
1078 int fragmentation_index(struct zone *zone, unsigned int order)
1080 struct contig_page_info info;
1082 fill_contig_page_info(zone, order, &info);
1083 return __fragmentation_index(order, &info);
1087 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1088 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1089 #ifdef CONFIG_ZONE_DMA
1090 #define TEXT_FOR_DMA(xx) xx "_dma",
1092 #define TEXT_FOR_DMA(xx)
1095 #ifdef CONFIG_ZONE_DMA32
1096 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1098 #define TEXT_FOR_DMA32(xx)
1101 #ifdef CONFIG_HIGHMEM
1102 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1104 #define TEXT_FOR_HIGHMEM(xx)
1107 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1108 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1110 const char * const vmstat_text[] = {
1111 /* enum zone_stat_item countes */
1113 "nr_zone_inactive_anon",
1114 "nr_zone_active_anon",
1115 "nr_zone_inactive_file",
1116 "nr_zone_active_file",
1117 "nr_zone_unevictable",
1118 "nr_zone_write_pending",
1120 "nr_page_table_pages",
1122 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1123 "nr_shadow_call_stack",
1126 #if IS_ENABLED(CONFIG_ZSMALLOC)
1131 /* enum numa_stat_item counters */
1141 /* enum node_stat_item counters */
1147 "nr_slab_reclaimable",
1148 "nr_slab_unreclaimable",
1152 "workingset_refault",
1153 "workingset_activate",
1154 "workingset_restore",
1155 "workingset_nodereclaim",
1161 "nr_writeback_temp",
1163 "nr_shmem_hugepages",
1164 "nr_shmem_pmdmapped",
1165 "nr_file_hugepages",
1166 "nr_file_pmdmapped",
1167 "nr_anon_transparent_hugepages",
1170 "nr_vmscan_immediate_reclaim",
1173 "nr_kernel_misc_reclaimable",
1174 "nr_foll_pin_acquired",
1175 "nr_foll_pin_released",
1177 /* enum writeback_stat_item counters */
1178 "nr_dirty_threshold",
1179 "nr_dirty_background_threshold",
1181 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1182 /* enum vm_event_item counters */
1188 TEXTS_FOR_ZONES("pgalloc")
1189 TEXTS_FOR_ZONES("allocstall")
1190 TEXTS_FOR_ZONES("pgskip")
1206 "pgscan_direct_throttle",
1209 "zone_reclaim_failed",
1213 "kswapd_inodesteal",
1214 "kswapd_low_wmark_hit_quickly",
1215 "kswapd_high_wmark_hit_quickly",
1224 #ifdef CONFIG_NUMA_BALANCING
1226 "numa_huge_pte_updates",
1228 "numa_hint_faults_local",
1229 "numa_pages_migrated",
1231 #ifdef CONFIG_MIGRATION
1232 "pgmigrate_success",
1235 #ifdef CONFIG_COMPACTION
1236 "compact_migrate_scanned",
1237 "compact_free_scanned",
1242 "compact_daemon_wake",
1243 "compact_daemon_migrate_scanned",
1244 "compact_daemon_free_scanned",
1247 #ifdef CONFIG_HUGETLB_PAGE
1248 "htlb_buddy_alloc_success",
1249 "htlb_buddy_alloc_fail",
1251 "unevictable_pgs_culled",
1252 "unevictable_pgs_scanned",
1253 "unevictable_pgs_rescued",
1254 "unevictable_pgs_mlocked",
1255 "unevictable_pgs_munlocked",
1256 "unevictable_pgs_cleared",
1257 "unevictable_pgs_stranded",
1259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1261 "thp_fault_fallback",
1262 "thp_fault_fallback_charge",
1263 "thp_collapse_alloc",
1264 "thp_collapse_alloc_failed",
1266 "thp_file_fallback",
1267 "thp_file_fallback_charge",
1270 "thp_split_page_failed",
1271 "thp_deferred_split_page",
1273 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1276 "thp_zero_page_alloc",
1277 "thp_zero_page_alloc_failed",
1279 "thp_swpout_fallback",
1281 #ifdef CONFIG_MEMORY_BALLOON
1284 #ifdef CONFIG_BALLOON_COMPACTION
1287 #endif /* CONFIG_MEMORY_BALLOON */
1288 #ifdef CONFIG_DEBUG_TLBFLUSH
1289 "nr_tlb_remote_flush",
1290 "nr_tlb_remote_flush_received",
1291 "nr_tlb_local_flush_all",
1292 "nr_tlb_local_flush_one",
1293 #endif /* CONFIG_DEBUG_TLBFLUSH */
1295 #ifdef CONFIG_DEBUG_VM_VMACACHE
1296 "vmacache_find_calls",
1297 "vmacache_find_hits",
1303 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1305 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1307 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1308 defined(CONFIG_PROC_FS)
1309 static void *frag_start(struct seq_file *m, loff_t *pos)
1314 for (pgdat = first_online_pgdat();
1316 pgdat = next_online_pgdat(pgdat))
1322 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1324 pg_data_t *pgdat = (pg_data_t *)arg;
1327 return next_online_pgdat(pgdat);
1330 static void frag_stop(struct seq_file *m, void *arg)
1335 * Walk zones in a node and print using a callback.
1336 * If @assert_populated is true, only use callback for zones that are populated.
1338 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1339 bool assert_populated, bool nolock,
1340 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1343 struct zone *node_zones = pgdat->node_zones;
1344 unsigned long flags;
1346 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1347 if (assert_populated && !populated_zone(zone))
1351 spin_lock_irqsave(&zone->lock, flags);
1352 print(m, pgdat, zone);
1354 spin_unlock_irqrestore(&zone->lock, flags);
1359 #ifdef CONFIG_PROC_FS
1360 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1365 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1366 for (order = 0; order < MAX_ORDER; ++order)
1367 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1372 * This walks the free areas for each zone.
1374 static int frag_show(struct seq_file *m, void *arg)
1376 pg_data_t *pgdat = (pg_data_t *)arg;
1377 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1381 static void pagetypeinfo_showfree_print(struct seq_file *m,
1382 pg_data_t *pgdat, struct zone *zone)
1386 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1387 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1390 migratetype_names[mtype]);
1391 for (order = 0; order < MAX_ORDER; ++order) {
1392 unsigned long freecount = 0;
1393 struct free_area *area;
1394 struct list_head *curr;
1395 bool overflow = false;
1397 area = &(zone->free_area[order]);
1399 list_for_each(curr, &area->free_list[mtype]) {
1401 * Cap the free_list iteration because it might
1402 * be really large and we are under a spinlock
1403 * so a long time spent here could trigger a
1404 * hard lockup detector. Anyway this is a
1405 * debugging tool so knowing there is a handful
1406 * of pages of this order should be more than
1409 if (++freecount >= 100000) {
1414 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1415 spin_unlock_irq(&zone->lock);
1417 spin_lock_irq(&zone->lock);
1423 /* Print out the free pages at each order for each migatetype */
1424 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1427 pg_data_t *pgdat = (pg_data_t *)arg;
1430 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1431 for (order = 0; order < MAX_ORDER; ++order)
1432 seq_printf(m, "%6d ", order);
1435 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1440 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1441 pg_data_t *pgdat, struct zone *zone)
1445 unsigned long start_pfn = zone->zone_start_pfn;
1446 unsigned long end_pfn = zone_end_pfn(zone);
1447 unsigned long count[MIGRATE_TYPES] = { 0, };
1449 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1452 page = pfn_to_online_page(pfn);
1456 /* Watch for unexpected holes punched in the memmap */
1457 if (!memmap_valid_within(pfn, page, zone))
1460 if (page_zone(page) != zone)
1463 mtype = get_pageblock_migratetype(page);
1465 if (mtype < MIGRATE_TYPES)
1470 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1471 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1472 seq_printf(m, "%12lu ", count[mtype]);
1476 /* Print out the number of pageblocks for each migratetype */
1477 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1480 pg_data_t *pgdat = (pg_data_t *)arg;
1482 seq_printf(m, "\n%-23s", "Number of blocks type ");
1483 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1484 seq_printf(m, "%12s ", migratetype_names[mtype]);
1486 walk_zones_in_node(m, pgdat, true, false,
1487 pagetypeinfo_showblockcount_print);
1493 * Print out the number of pageblocks for each migratetype that contain pages
1494 * of other types. This gives an indication of how well fallbacks are being
1495 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1496 * to determine what is going on
1498 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1500 #ifdef CONFIG_PAGE_OWNER
1503 if (!static_branch_unlikely(&page_owner_inited))
1506 drain_all_pages(NULL);
1508 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1509 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1510 seq_printf(m, "%12s ", migratetype_names[mtype]);
1513 walk_zones_in_node(m, pgdat, true, true,
1514 pagetypeinfo_showmixedcount_print);
1515 #endif /* CONFIG_PAGE_OWNER */
1519 * This prints out statistics in relation to grouping pages by mobility.
1520 * It is expensive to collect so do not constantly read the file.
1522 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1524 pg_data_t *pgdat = (pg_data_t *)arg;
1526 /* check memoryless node */
1527 if (!node_state(pgdat->node_id, N_MEMORY))
1530 seq_printf(m, "Page block order: %d\n", pageblock_order);
1531 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1533 pagetypeinfo_showfree(m, pgdat);
1534 pagetypeinfo_showblockcount(m, pgdat);
1535 pagetypeinfo_showmixedcount(m, pgdat);
1540 static const struct seq_operations fragmentation_op = {
1541 .start = frag_start,
1547 static const struct seq_operations pagetypeinfo_op = {
1548 .start = frag_start,
1551 .show = pagetypeinfo_show,
1554 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1558 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1559 struct zone *compare = &pgdat->node_zones[zid];
1561 if (populated_zone(compare))
1562 return zone == compare;
1568 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1572 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1573 if (is_zone_first_populated(pgdat, zone)) {
1574 seq_printf(m, "\n per-node stats");
1575 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1576 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1577 node_page_state(pgdat, i));
1588 zone_page_state(zone, NR_FREE_PAGES),
1589 min_wmark_pages(zone),
1590 low_wmark_pages(zone),
1591 high_wmark_pages(zone),
1592 zone->spanned_pages,
1593 zone->present_pages,
1594 zone_managed_pages(zone));
1597 "\n protection: (%ld",
1598 zone->lowmem_reserve[0]);
1599 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1600 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1603 /* If unpopulated, no other information is useful */
1604 if (!populated_zone(zone)) {
1609 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1610 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1611 zone_page_state(zone, i));
1614 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1615 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1616 zone_numa_state_snapshot(zone, i));
1619 seq_printf(m, "\n pagesets");
1620 for_each_online_cpu(i) {
1621 struct per_cpu_pageset *pageset;
1623 pageset = per_cpu_ptr(zone->pageset, i);
1632 pageset->pcp.batch);
1634 seq_printf(m, "\n vm stats threshold: %d",
1635 pageset->stat_threshold);
1639 "\n node_unreclaimable: %u"
1640 "\n start_pfn: %lu",
1641 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1642 zone->zone_start_pfn);
1647 * Output information about zones in @pgdat. All zones are printed regardless
1648 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1649 * set of all zones and userspace would not be aware of such zones if they are
1650 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1652 static int zoneinfo_show(struct seq_file *m, void *arg)
1654 pg_data_t *pgdat = (pg_data_t *)arg;
1655 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1659 static const struct seq_operations zoneinfo_op = {
1660 .start = frag_start, /* iterate over all zones. The same as in
1664 .show = zoneinfo_show,
1667 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1668 NR_VM_NUMA_STAT_ITEMS + \
1669 NR_VM_NODE_STAT_ITEMS + \
1670 NR_VM_WRITEBACK_STAT_ITEMS + \
1671 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1672 NR_VM_EVENT_ITEMS : 0))
1674 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1679 if (*pos >= NR_VMSTAT_ITEMS)
1682 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1683 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1686 return ERR_PTR(-ENOMEM);
1687 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1688 v[i] = global_zone_page_state(i);
1689 v += NR_VM_ZONE_STAT_ITEMS;
1692 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1693 v[i] = global_numa_state(i);
1694 v += NR_VM_NUMA_STAT_ITEMS;
1697 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1698 v[i] = global_node_page_state(i);
1699 v += NR_VM_NODE_STAT_ITEMS;
1701 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1702 v + NR_DIRTY_THRESHOLD);
1703 v += NR_VM_WRITEBACK_STAT_ITEMS;
1705 #ifdef CONFIG_VM_EVENT_COUNTERS
1707 v[PGPGIN] /= 2; /* sectors -> kbytes */
1710 return (unsigned long *)m->private + *pos;
1713 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1716 if (*pos >= NR_VMSTAT_ITEMS)
1718 return (unsigned long *)m->private + *pos;
1721 static int vmstat_show(struct seq_file *m, void *arg)
1723 unsigned long *l = arg;
1724 unsigned long off = l - (unsigned long *)m->private;
1726 seq_puts(m, vmstat_text[off]);
1727 seq_put_decimal_ull(m, " ", *l);
1732 static void vmstat_stop(struct seq_file *m, void *arg)
1738 static const struct seq_operations vmstat_op = {
1739 .start = vmstat_start,
1740 .next = vmstat_next,
1741 .stop = vmstat_stop,
1742 .show = vmstat_show,
1744 #endif /* CONFIG_PROC_FS */
1747 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1748 int sysctl_stat_interval __read_mostly = HZ;
1750 #ifdef CONFIG_PROC_FS
1751 static void refresh_vm_stats(struct work_struct *work)
1753 refresh_cpu_vm_stats(true);
1756 int vmstat_refresh(struct ctl_table *table, int write,
1757 void __user *buffer, size_t *lenp, loff_t *ppos)
1764 * The regular update, every sysctl_stat_interval, may come later
1765 * than expected: leaving a significant amount in per_cpu buckets.
1766 * This is particularly misleading when checking a quantity of HUGE
1767 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1768 * which can equally be echo'ed to or cat'ted from (by root),
1769 * can be used to update the stats just before reading them.
1771 * Oh, and since global_zone_page_state() etc. are so careful to hide
1772 * transiently negative values, report an error here if any of
1773 * the stats is negative, so we know to go looking for imbalance.
1775 err = schedule_on_each_cpu(refresh_vm_stats);
1778 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1779 val = atomic_long_read(&vm_zone_stat[i]);
1781 pr_warn("%s: %s %ld\n",
1782 __func__, zone_stat_name(i), val);
1787 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1788 val = atomic_long_read(&vm_numa_stat[i]);
1790 pr_warn("%s: %s %ld\n",
1791 __func__, numa_stat_name(i), val);
1804 #endif /* CONFIG_PROC_FS */
1806 static void vmstat_update(struct work_struct *w)
1808 if (refresh_cpu_vm_stats(true)) {
1810 * Counters were updated so we expect more updates
1811 * to occur in the future. Keep on running the
1812 * update worker thread.
1814 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1815 this_cpu_ptr(&vmstat_work),
1816 round_jiffies_relative(sysctl_stat_interval));
1821 * Switch off vmstat processing and then fold all the remaining differentials
1822 * until the diffs stay at zero. The function is used by NOHZ and can only be
1823 * invoked when tick processing is not active.
1826 * Check if the diffs for a certain cpu indicate that
1827 * an update is needed.
1829 static bool need_update(int cpu)
1833 for_each_populated_zone(zone) {
1834 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1836 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1838 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1842 * The fast way of checking if there are any vmstat diffs.
1844 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1845 sizeof(p->vm_stat_diff[0])))
1848 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1849 sizeof(p->vm_numa_stat_diff[0])))
1857 * Switch off vmstat processing and then fold all the remaining differentials
1858 * until the diffs stay at zero. The function is used by NOHZ and can only be
1859 * invoked when tick processing is not active.
1861 void quiet_vmstat(void)
1863 if (system_state != SYSTEM_RUNNING)
1866 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1869 if (!need_update(smp_processor_id()))
1873 * Just refresh counters and do not care about the pending delayed
1874 * vmstat_update. It doesn't fire that often to matter and canceling
1875 * it would be too expensive from this path.
1876 * vmstat_shepherd will take care about that for us.
1878 refresh_cpu_vm_stats(false);
1882 * Shepherd worker thread that checks the
1883 * differentials of processors that have their worker
1884 * threads for vm statistics updates disabled because of
1887 static void vmstat_shepherd(struct work_struct *w);
1889 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1891 static void vmstat_shepherd(struct work_struct *w)
1896 /* Check processors whose vmstat worker threads have been disabled */
1897 for_each_online_cpu(cpu) {
1898 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1900 if (!delayed_work_pending(dw) && need_update(cpu))
1901 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1905 schedule_delayed_work(&shepherd,
1906 round_jiffies_relative(sysctl_stat_interval));
1909 static void __init start_shepherd_timer(void)
1913 for_each_possible_cpu(cpu)
1914 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1917 schedule_delayed_work(&shepherd,
1918 round_jiffies_relative(sysctl_stat_interval));
1921 static void __init init_cpu_node_state(void)
1925 for_each_online_node(node) {
1926 if (cpumask_weight(cpumask_of_node(node)) > 0)
1927 node_set_state(node, N_CPU);
1931 static int vmstat_cpu_online(unsigned int cpu)
1933 refresh_zone_stat_thresholds();
1934 node_set_state(cpu_to_node(cpu), N_CPU);
1938 static int vmstat_cpu_down_prep(unsigned int cpu)
1940 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1944 static int vmstat_cpu_dead(unsigned int cpu)
1946 const struct cpumask *node_cpus;
1949 node = cpu_to_node(cpu);
1951 refresh_zone_stat_thresholds();
1952 node_cpus = cpumask_of_node(node);
1953 if (cpumask_weight(node_cpus) > 0)
1956 node_clear_state(node, N_CPU);
1962 struct workqueue_struct *mm_percpu_wq;
1964 void __init init_mm_internals(void)
1966 int ret __maybe_unused;
1968 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
1971 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
1972 NULL, vmstat_cpu_dead);
1974 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1976 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
1978 vmstat_cpu_down_prep);
1980 pr_err("vmstat: failed to register 'online' hotplug state\n");
1983 init_cpu_node_state();
1986 start_shepherd_timer();
1988 #ifdef CONFIG_PROC_FS
1989 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
1990 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
1991 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
1992 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
1996 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1999 * Return an index indicating how much of the available free memory is
2000 * unusable for an allocation of the requested size.
2002 static int unusable_free_index(unsigned int order,
2003 struct contig_page_info *info)
2005 /* No free memory is interpreted as all free memory is unusable */
2006 if (info->free_pages == 0)
2010 * Index should be a value between 0 and 1. Return a value to 3
2013 * 0 => no fragmentation
2014 * 1 => high fragmentation
2016 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2020 static void unusable_show_print(struct seq_file *m,
2021 pg_data_t *pgdat, struct zone *zone)
2025 struct contig_page_info info;
2027 seq_printf(m, "Node %d, zone %8s ",
2030 for (order = 0; order < MAX_ORDER; ++order) {
2031 fill_contig_page_info(zone, order, &info);
2032 index = unusable_free_index(order, &info);
2033 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2040 * Display unusable free space index
2042 * The unusable free space index measures how much of the available free
2043 * memory cannot be used to satisfy an allocation of a given size and is a
2044 * value between 0 and 1. The higher the value, the more of free memory is
2045 * unusable and by implication, the worse the external fragmentation is. This
2046 * can be expressed as a percentage by multiplying by 100.
2048 static int unusable_show(struct seq_file *m, void *arg)
2050 pg_data_t *pgdat = (pg_data_t *)arg;
2052 /* check memoryless node */
2053 if (!node_state(pgdat->node_id, N_MEMORY))
2056 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2061 static const struct seq_operations unusable_op = {
2062 .start = frag_start,
2065 .show = unusable_show,
2068 static int unusable_open(struct inode *inode, struct file *file)
2070 return seq_open(file, &unusable_op);
2073 static const struct file_operations unusable_file_ops = {
2074 .open = unusable_open,
2076 .llseek = seq_lseek,
2077 .release = seq_release,
2080 static void extfrag_show_print(struct seq_file *m,
2081 pg_data_t *pgdat, struct zone *zone)
2086 /* Alloc on stack as interrupts are disabled for zone walk */
2087 struct contig_page_info info;
2089 seq_printf(m, "Node %d, zone %8s ",
2092 for (order = 0; order < MAX_ORDER; ++order) {
2093 fill_contig_page_info(zone, order, &info);
2094 index = __fragmentation_index(order, &info);
2095 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2102 * Display fragmentation index for orders that allocations would fail for
2104 static int extfrag_show(struct seq_file *m, void *arg)
2106 pg_data_t *pgdat = (pg_data_t *)arg;
2108 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2113 static const struct seq_operations extfrag_op = {
2114 .start = frag_start,
2117 .show = extfrag_show,
2120 static int extfrag_open(struct inode *inode, struct file *file)
2122 return seq_open(file, &extfrag_op);
2125 static const struct file_operations extfrag_file_ops = {
2126 .open = extfrag_open,
2128 .llseek = seq_lseek,
2129 .release = seq_release,
2132 static int __init extfrag_debug_init(void)
2134 struct dentry *extfrag_debug_root;
2136 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2138 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2139 &unusable_file_ops);
2141 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2147 module_init(extfrag_debug_init);