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>
35 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
37 /* zero numa counters within a zone */
38 static void zero_zone_numa_counters(struct zone *zone)
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]
51 /* zero numa counters of all the populated zones */
52 static void zero_zones_numa_counters(void)
56 for_each_populated_zone(zone)
57 zero_zone_numa_counters(zone);
60 /* zero global numa counters */
61 static void zero_global_numa_counters(void)
65 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
66 atomic_long_set(&vm_numa_event[item], 0);
69 static void invalid_numa_statistics(void)
71 zero_zones_numa_counters();
72 zero_global_numa_counters();
75 static DEFINE_MUTEX(vm_numa_stat_lock);
77 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
78 void *buffer, size_t *length, loff_t *ppos)
82 mutex_lock(&vm_numa_stat_lock);
84 oldval = sysctl_vm_numa_stat;
85 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
89 if (oldval == sysctl_vm_numa_stat)
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");
95 static_branch_disable(&vm_numa_stat_key);
96 invalid_numa_statistics();
97 pr_info("disable numa statistics, and clear numa counters\n");
101 mutex_unlock(&vm_numa_stat_lock);
106 #ifdef CONFIG_VM_EVENT_COUNTERS
107 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
108 EXPORT_PER_CPU_SYMBOL(vm_event_states);
110 static void sum_vm_events(unsigned long *ret)
115 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
117 for_each_online_cpu(cpu) {
118 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
120 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
121 ret[i] += this->event[i];
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.
130 void all_vm_events(unsigned long *ret)
136 EXPORT_SYMBOL_GPL(all_vm_events);
139 * Fold the foreign cpu events into our own.
141 * This is adding to the events on one processor
142 * but keeps the global counts constant.
144 void vm_events_fold_cpu(int cpu)
146 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
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;
155 #endif /* CONFIG_VM_EVENT_COUNTERS */
158 * Manage combined zone based / global counters
160 * vm_stat contains the global counters
162 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
163 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
164 atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
165 EXPORT_SYMBOL(vm_zone_stat);
166 EXPORT_SYMBOL(vm_node_stat);
170 int calculate_pressure_threshold(struct zone *zone)
173 int watermark_distance;
176 * As vmstats are not up to date, there is drift between the estimated
177 * and real values. For high thresholds and a high number of CPUs, it
178 * is possible for the min watermark to be breached while the estimated
179 * value looks fine. The pressure threshold is a reduced value such
180 * that even the maximum amount of drift will not accidentally breach
183 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
184 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
187 * Maximum threshold is 125
189 threshold = min(125, threshold);
194 int calculate_normal_threshold(struct zone *zone)
197 int mem; /* memory in 128 MB units */
200 * The threshold scales with the number of processors and the amount
201 * of memory per zone. More memory means that we can defer updates for
202 * longer, more processors could lead to more contention.
203 * fls() is used to have a cheap way of logarithmic scaling.
205 * Some sample thresholds:
207 * Threshold Processors (fls) Zonesize fls(mem+1)
208 * ------------------------------------------------------------------
225 * 125 1024 10 8-16 GB 8
226 * 125 1024 10 16-32 GB 9
229 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
231 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
234 * Maximum threshold is 125
236 threshold = min(125, threshold);
242 * Refresh the thresholds for each zone.
244 void refresh_zone_stat_thresholds(void)
246 struct pglist_data *pgdat;
251 /* Zero current pgdat thresholds */
252 for_each_online_pgdat(pgdat) {
253 for_each_online_cpu(cpu) {
254 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
258 for_each_populated_zone(zone) {
259 struct pglist_data *pgdat = zone->zone_pgdat;
260 unsigned long max_drift, tolerate_drift;
262 threshold = calculate_normal_threshold(zone);
264 for_each_online_cpu(cpu) {
267 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
270 /* Base nodestat threshold on the largest populated zone. */
271 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
272 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
273 = max(threshold, pgdat_threshold);
277 * Only set percpu_drift_mark if there is a danger that
278 * NR_FREE_PAGES reports the low watermark is ok when in fact
279 * the min watermark could be breached by an allocation
281 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
282 max_drift = num_online_cpus() * threshold;
283 if (max_drift > tolerate_drift)
284 zone->percpu_drift_mark = high_wmark_pages(zone) +
289 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
290 int (*calculate_pressure)(struct zone *))
297 for (i = 0; i < pgdat->nr_zones; i++) {
298 zone = &pgdat->node_zones[i];
299 if (!zone->percpu_drift_mark)
302 threshold = (*calculate_pressure)(zone);
303 for_each_online_cpu(cpu)
304 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
310 * For use when we know that interrupts are disabled,
311 * or when we know that preemption is disabled and that
312 * particular counter cannot be updated from interrupt context.
314 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
317 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
318 s8 __percpu *p = pcp->vm_stat_diff + item;
322 x = delta + __this_cpu_read(*p);
324 t = __this_cpu_read(pcp->stat_threshold);
326 if (unlikely(abs(x) > t)) {
327 zone_page_state_add(x, zone, item);
330 __this_cpu_write(*p, x);
332 EXPORT_SYMBOL(__mod_zone_page_state);
334 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
337 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
338 s8 __percpu *p = pcp->vm_node_stat_diff + item;
342 if (vmstat_item_in_bytes(item)) {
344 * Only cgroups use subpage accounting right now; at
345 * the global level, these items still change in
346 * multiples of whole pages. Store them as pages
347 * internally to keep the per-cpu counters compact.
349 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
350 delta >>= PAGE_SHIFT;
353 x = delta + __this_cpu_read(*p);
355 t = __this_cpu_read(pcp->stat_threshold);
357 if (unlikely(abs(x) > t)) {
358 node_page_state_add(x, pgdat, item);
361 __this_cpu_write(*p, x);
363 EXPORT_SYMBOL(__mod_node_page_state);
366 * Optimized increment and decrement functions.
368 * These are only for a single page and therefore can take a struct page *
369 * argument instead of struct zone *. This allows the inclusion of the code
370 * generated for page_zone(page) into the optimized functions.
372 * No overflow check is necessary and therefore the differential can be
373 * incremented or decremented in place which may allow the compilers to
374 * generate better code.
375 * The increment or decrement is known and therefore one boundary check can
378 * NOTE: These functions are very performance sensitive. Change only
381 * Some processors have inc/dec instructions that are atomic vs an interrupt.
382 * However, the code must first determine the differential location in a zone
383 * based on the processor number and then inc/dec the counter. There is no
384 * guarantee without disabling preemption that the processor will not change
385 * in between and therefore the atomicity vs. interrupt cannot be exploited
386 * in a useful way here.
388 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
390 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
391 s8 __percpu *p = pcp->vm_stat_diff + item;
394 v = __this_cpu_inc_return(*p);
395 t = __this_cpu_read(pcp->stat_threshold);
396 if (unlikely(v > t)) {
397 s8 overstep = t >> 1;
399 zone_page_state_add(v + overstep, zone, item);
400 __this_cpu_write(*p, -overstep);
404 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
406 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
407 s8 __percpu *p = pcp->vm_node_stat_diff + item;
410 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
412 v = __this_cpu_inc_return(*p);
413 t = __this_cpu_read(pcp->stat_threshold);
414 if (unlikely(v > t)) {
415 s8 overstep = t >> 1;
417 node_page_state_add(v + overstep, pgdat, item);
418 __this_cpu_write(*p, -overstep);
422 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
424 __inc_zone_state(page_zone(page), item);
426 EXPORT_SYMBOL(__inc_zone_page_state);
428 void __inc_node_page_state(struct page *page, enum node_stat_item item)
430 __inc_node_state(page_pgdat(page), item);
432 EXPORT_SYMBOL(__inc_node_page_state);
434 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
436 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
437 s8 __percpu *p = pcp->vm_stat_diff + item;
440 v = __this_cpu_dec_return(*p);
441 t = __this_cpu_read(pcp->stat_threshold);
442 if (unlikely(v < - t)) {
443 s8 overstep = t >> 1;
445 zone_page_state_add(v - overstep, zone, item);
446 __this_cpu_write(*p, overstep);
450 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
452 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
453 s8 __percpu *p = pcp->vm_node_stat_diff + item;
456 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
458 v = __this_cpu_dec_return(*p);
459 t = __this_cpu_read(pcp->stat_threshold);
460 if (unlikely(v < - t)) {
461 s8 overstep = t >> 1;
463 node_page_state_add(v - overstep, pgdat, item);
464 __this_cpu_write(*p, overstep);
468 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
470 __dec_zone_state(page_zone(page), item);
472 EXPORT_SYMBOL(__dec_zone_page_state);
474 void __dec_node_page_state(struct page *page, enum node_stat_item item)
476 __dec_node_state(page_pgdat(page), item);
478 EXPORT_SYMBOL(__dec_node_page_state);
480 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
482 * If we have cmpxchg_local support then we do not need to incur the overhead
483 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
485 * mod_state() modifies the zone counter state through atomic per cpu
488 * Overstep mode specifies how overstep should handled:
490 * 1 Overstepping half of threshold
491 * -1 Overstepping minus half of threshold
493 static inline void mod_zone_state(struct zone *zone,
494 enum zone_stat_item item, long delta, int overstep_mode)
496 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
497 s8 __percpu *p = pcp->vm_stat_diff + item;
501 z = 0; /* overflow to zone counters */
504 * The fetching of the stat_threshold is racy. We may apply
505 * a counter threshold to the wrong the cpu if we get
506 * rescheduled while executing here. However, the next
507 * counter update will apply the threshold again and
508 * therefore bring the counter under the threshold again.
510 * Most of the time the thresholds are the same anyways
511 * for all cpus in a zone.
513 t = this_cpu_read(pcp->stat_threshold);
515 o = this_cpu_read(*p);
519 int os = overstep_mode * (t >> 1) ;
521 /* Overflow must be added to zone counters */
525 } while (this_cpu_cmpxchg(*p, o, n) != o);
528 zone_page_state_add(z, zone, item);
531 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
534 mod_zone_state(zone, item, delta, 0);
536 EXPORT_SYMBOL(mod_zone_page_state);
538 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
540 mod_zone_state(page_zone(page), item, 1, 1);
542 EXPORT_SYMBOL(inc_zone_page_state);
544 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
546 mod_zone_state(page_zone(page), item, -1, -1);
548 EXPORT_SYMBOL(dec_zone_page_state);
550 static inline void mod_node_state(struct pglist_data *pgdat,
551 enum node_stat_item item, int delta, int overstep_mode)
553 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
554 s8 __percpu *p = pcp->vm_node_stat_diff + item;
557 if (vmstat_item_in_bytes(item)) {
559 * Only cgroups use subpage accounting right now; at
560 * the global level, these items still change in
561 * multiples of whole pages. Store them as pages
562 * internally to keep the per-cpu counters compact.
564 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
565 delta >>= PAGE_SHIFT;
569 z = 0; /* overflow to node counters */
572 * The fetching of the stat_threshold is racy. We may apply
573 * a counter threshold to the wrong the cpu if we get
574 * rescheduled while executing here. However, the next
575 * counter update will apply the threshold again and
576 * therefore bring the counter under the threshold again.
578 * Most of the time the thresholds are the same anyways
579 * for all cpus in a node.
581 t = this_cpu_read(pcp->stat_threshold);
583 o = this_cpu_read(*p);
587 int os = overstep_mode * (t >> 1) ;
589 /* Overflow must be added to node counters */
593 } while (this_cpu_cmpxchg(*p, o, n) != o);
596 node_page_state_add(z, pgdat, item);
599 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
602 mod_node_state(pgdat, item, delta, 0);
604 EXPORT_SYMBOL(mod_node_page_state);
606 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
608 mod_node_state(pgdat, item, 1, 1);
611 void inc_node_page_state(struct page *page, enum node_stat_item item)
613 mod_node_state(page_pgdat(page), item, 1, 1);
615 EXPORT_SYMBOL(inc_node_page_state);
617 void dec_node_page_state(struct page *page, enum node_stat_item item)
619 mod_node_state(page_pgdat(page), item, -1, -1);
621 EXPORT_SYMBOL(dec_node_page_state);
624 * Use interrupt disable to serialize counter updates
626 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
631 local_irq_save(flags);
632 __mod_zone_page_state(zone, item, delta);
633 local_irq_restore(flags);
635 EXPORT_SYMBOL(mod_zone_page_state);
637 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
642 zone = page_zone(page);
643 local_irq_save(flags);
644 __inc_zone_state(zone, item);
645 local_irq_restore(flags);
647 EXPORT_SYMBOL(inc_zone_page_state);
649 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
653 local_irq_save(flags);
654 __dec_zone_page_state(page, item);
655 local_irq_restore(flags);
657 EXPORT_SYMBOL(dec_zone_page_state);
659 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
663 local_irq_save(flags);
664 __inc_node_state(pgdat, item);
665 local_irq_restore(flags);
667 EXPORT_SYMBOL(inc_node_state);
669 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
674 local_irq_save(flags);
675 __mod_node_page_state(pgdat, item, delta);
676 local_irq_restore(flags);
678 EXPORT_SYMBOL(mod_node_page_state);
680 void inc_node_page_state(struct page *page, enum node_stat_item item)
683 struct pglist_data *pgdat;
685 pgdat = page_pgdat(page);
686 local_irq_save(flags);
687 __inc_node_state(pgdat, item);
688 local_irq_restore(flags);
690 EXPORT_SYMBOL(inc_node_page_state);
692 void dec_node_page_state(struct page *page, enum node_stat_item item)
696 local_irq_save(flags);
697 __dec_node_page_state(page, item);
698 local_irq_restore(flags);
700 EXPORT_SYMBOL(dec_node_page_state);
704 * Fold a differential into the global counters.
705 * Returns the number of counters updated.
707 static int fold_diff(int *zone_diff, int *node_diff)
712 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
714 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
718 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
720 atomic_long_add(node_diff[i], &vm_node_stat[i]);
727 static void fold_vm_zone_numa_events(struct zone *zone)
729 unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
731 enum numa_stat_item item;
733 for_each_online_cpu(cpu) {
734 struct per_cpu_zonestat *pzstats;
736 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
737 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
738 zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
741 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
742 zone_numa_event_add(zone_numa_events[item], zone, item);
745 void fold_vm_numa_events(void)
749 for_each_populated_zone(zone)
750 fold_vm_zone_numa_events(zone);
755 * Update the zone counters for the current cpu.
757 * Note that refresh_cpu_vm_stats strives to only access
758 * node local memory. The per cpu pagesets on remote zones are placed
759 * in the memory local to the processor using that pageset. So the
760 * loop over all zones will access a series of cachelines local to
763 * The call to zone_page_state_add updates the cachelines with the
764 * statistics in the remote zone struct as well as the global cachelines
765 * with the global counters. These could cause remote node cache line
766 * bouncing and will have to be only done when necessary.
768 * The function returns the number of global counters updated.
770 static int refresh_cpu_vm_stats(bool do_pagesets)
772 struct pglist_data *pgdat;
775 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
776 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
779 for_each_populated_zone(zone) {
780 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
782 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
785 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
788 v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
791 atomic_long_add(v, &zone->vm_stat[i]);
792 global_zone_diff[i] += v;
794 /* 3 seconds idle till flush */
795 __this_cpu_write(pcp->expire, 3);
804 * Deal with draining the remote pageset of this
807 * Check if there are pages remaining in this pageset
808 * if not then there is nothing to expire.
810 if (!__this_cpu_read(pcp->expire) ||
811 !__this_cpu_read(pcp->count))
815 * We never drain zones local to this processor.
817 if (zone_to_nid(zone) == numa_node_id()) {
818 __this_cpu_write(pcp->expire, 0);
822 if (__this_cpu_dec_return(pcp->expire))
825 if (__this_cpu_read(pcp->count)) {
826 drain_zone_pages(zone, this_cpu_ptr(pcp));
833 for_each_online_pgdat(pgdat) {
834 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
836 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
839 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
841 atomic_long_add(v, &pgdat->vm_stat[i]);
842 global_node_diff[i] += v;
847 changes += fold_diff(global_zone_diff, global_node_diff);
852 * Fold the data for an offline cpu into the global array.
853 * There cannot be any access by the offline cpu and therefore
854 * synchronization is simplified.
856 void cpu_vm_stats_fold(int cpu)
858 struct pglist_data *pgdat;
861 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
862 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
864 for_each_populated_zone(zone) {
865 struct per_cpu_zonestat *pzstats;
867 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
869 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
870 if (pzstats->vm_stat_diff[i]) {
873 v = pzstats->vm_stat_diff[i];
874 pzstats->vm_stat_diff[i] = 0;
875 atomic_long_add(v, &zone->vm_stat[i]);
876 global_zone_diff[i] += v;
880 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
881 if (pzstats->vm_numa_event[i]) {
884 v = pzstats->vm_numa_event[i];
885 pzstats->vm_numa_event[i] = 0;
886 zone_numa_event_add(v, zone, i);
892 for_each_online_pgdat(pgdat) {
893 struct per_cpu_nodestat *p;
895 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
897 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
898 if (p->vm_node_stat_diff[i]) {
901 v = p->vm_node_stat_diff[i];
902 p->vm_node_stat_diff[i] = 0;
903 atomic_long_add(v, &pgdat->vm_stat[i]);
904 global_node_diff[i] += v;
908 fold_diff(global_zone_diff, global_node_diff);
912 * this is only called if !populated_zone(zone), which implies no other users of
913 * pset->vm_stat_diff[] exist.
915 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
920 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
921 if (pzstats->vm_stat_diff[i]) {
922 v = pzstats->vm_stat_diff[i];
923 pzstats->vm_stat_diff[i] = 0;
924 zone_page_state_add(v, zone, i);
929 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
930 if (pzstats->vm_numa_event[i]) {
931 v = pzstats->vm_numa_event[i];
932 pzstats->vm_numa_event[i] = 0;
933 zone_numa_event_add(v, zone, i);
941 /* See __count_vm_event comment on why raw_cpu_inc is used. */
942 void __count_numa_event(struct zone *zone,
943 enum numa_stat_item item)
945 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
947 raw_cpu_inc(pzstats->vm_numa_event[item]);
951 * Determine the per node value of a stat item. This function
952 * is called frequently in a NUMA machine, so try to be as
953 * frugal as possible.
955 unsigned long sum_zone_node_page_state(int node,
956 enum zone_stat_item item)
958 struct zone *zones = NODE_DATA(node)->node_zones;
960 unsigned long count = 0;
962 for (i = 0; i < MAX_NR_ZONES; i++)
963 count += zone_page_state(zones + i, item);
968 /* Determine the per node value of a numa stat item. */
969 unsigned long sum_zone_numa_event_state(int node,
970 enum numa_stat_item item)
972 struct zone *zones = NODE_DATA(node)->node_zones;
973 unsigned long count = 0;
976 for (i = 0; i < MAX_NR_ZONES; i++)
977 count += zone_numa_event_state(zones + i, item);
983 * Determine the per node value of a stat item.
985 unsigned long node_page_state_pages(struct pglist_data *pgdat,
986 enum node_stat_item item)
988 long x = atomic_long_read(&pgdat->vm_stat[item]);
996 unsigned long node_page_state(struct pglist_data *pgdat,
997 enum node_stat_item item)
999 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1001 return node_page_state_pages(pgdat, item);
1005 #ifdef CONFIG_COMPACTION
1007 struct contig_page_info {
1008 unsigned long free_pages;
1009 unsigned long free_blocks_total;
1010 unsigned long free_blocks_suitable;
1014 * Calculate the number of free pages in a zone, how many contiguous
1015 * pages are free and how many are large enough to satisfy an allocation of
1016 * the target size. Note that this function makes no attempt to estimate
1017 * how many suitable free blocks there *might* be if MOVABLE pages were
1018 * migrated. Calculating that is possible, but expensive and can be
1019 * figured out from userspace
1021 static void fill_contig_page_info(struct zone *zone,
1022 unsigned int suitable_order,
1023 struct contig_page_info *info)
1027 info->free_pages = 0;
1028 info->free_blocks_total = 0;
1029 info->free_blocks_suitable = 0;
1031 for (order = 0; order < MAX_ORDER; order++) {
1032 unsigned long blocks;
1034 /* Count number of free blocks */
1035 blocks = zone->free_area[order].nr_free;
1036 info->free_blocks_total += blocks;
1038 /* Count free base pages */
1039 info->free_pages += blocks << order;
1041 /* Count the suitable free blocks */
1042 if (order >= suitable_order)
1043 info->free_blocks_suitable += blocks <<
1044 (order - suitable_order);
1049 * A fragmentation index only makes sense if an allocation of a requested
1050 * size would fail. If that is true, the fragmentation index indicates
1051 * whether external fragmentation or a lack of memory was the problem.
1052 * The value can be used to determine if page reclaim or compaction
1055 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1057 unsigned long requested = 1UL << order;
1059 if (WARN_ON_ONCE(order >= MAX_ORDER))
1062 if (!info->free_blocks_total)
1065 /* Fragmentation index only makes sense when a request would fail */
1066 if (info->free_blocks_suitable)
1070 * Index is between 0 and 1 so return within 3 decimal places
1072 * 0 => allocation would fail due to lack of memory
1073 * 1 => allocation would fail due to fragmentation
1075 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1079 * Calculates external fragmentation within a zone wrt the given order.
1080 * It is defined as the percentage of pages found in blocks of size
1081 * less than 1 << order. It returns values in range [0, 100].
1083 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1085 struct contig_page_info info;
1087 fill_contig_page_info(zone, order, &info);
1088 if (info.free_pages == 0)
1091 return div_u64((info.free_pages -
1092 (info.free_blocks_suitable << order)) * 100,
1096 /* Same as __fragmentation index but allocs contig_page_info on stack */
1097 int fragmentation_index(struct zone *zone, unsigned int order)
1099 struct contig_page_info info;
1101 fill_contig_page_info(zone, order, &info);
1102 return __fragmentation_index(order, &info);
1106 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1107 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1108 #ifdef CONFIG_ZONE_DMA
1109 #define TEXT_FOR_DMA(xx) xx "_dma",
1111 #define TEXT_FOR_DMA(xx)
1114 #ifdef CONFIG_ZONE_DMA32
1115 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1117 #define TEXT_FOR_DMA32(xx)
1120 #ifdef CONFIG_HIGHMEM
1121 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1123 #define TEXT_FOR_HIGHMEM(xx)
1126 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1127 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1129 const char * const vmstat_text[] = {
1130 /* enum zone_stat_item counters */
1132 "nr_zone_inactive_anon",
1133 "nr_zone_active_anon",
1134 "nr_zone_inactive_file",
1135 "nr_zone_active_file",
1136 "nr_zone_unevictable",
1137 "nr_zone_write_pending",
1140 #if IS_ENABLED(CONFIG_ZSMALLOC)
1145 /* enum numa_stat_item counters */
1155 /* enum node_stat_item counters */
1161 "nr_slab_reclaimable",
1162 "nr_slab_unreclaimable",
1166 "workingset_refault_anon",
1167 "workingset_refault_file",
1168 "workingset_activate_anon",
1169 "workingset_activate_file",
1170 "workingset_restore_anon",
1171 "workingset_restore_file",
1172 "workingset_nodereclaim",
1178 "nr_writeback_temp",
1180 "nr_shmem_hugepages",
1181 "nr_shmem_pmdmapped",
1182 "nr_file_hugepages",
1183 "nr_file_pmdmapped",
1184 "nr_anon_transparent_hugepages",
1186 "nr_vmscan_immediate_reclaim",
1189 "nr_kernel_misc_reclaimable",
1190 "nr_foll_pin_acquired",
1191 "nr_foll_pin_released",
1193 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1194 "nr_shadow_call_stack",
1196 "nr_page_table_pages",
1201 /* enum writeback_stat_item counters */
1202 "nr_dirty_threshold",
1203 "nr_dirty_background_threshold",
1205 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1206 /* enum vm_event_item counters */
1212 TEXTS_FOR_ZONES("pgalloc")
1213 TEXTS_FOR_ZONES("allocstall")
1214 TEXTS_FOR_ZONES("pgskip")
1231 "pgscan_direct_throttle",
1238 "zone_reclaim_failed",
1242 "kswapd_inodesteal",
1243 "kswapd_low_wmark_hit_quickly",
1244 "kswapd_high_wmark_hit_quickly",
1253 #ifdef CONFIG_NUMA_BALANCING
1255 "numa_huge_pte_updates",
1257 "numa_hint_faults_local",
1258 "numa_pages_migrated",
1260 #ifdef CONFIG_MIGRATION
1261 "pgmigrate_success",
1263 "thp_migration_success",
1264 "thp_migration_fail",
1265 "thp_migration_split",
1267 #ifdef CONFIG_COMPACTION
1268 "compact_migrate_scanned",
1269 "compact_free_scanned",
1274 "compact_daemon_wake",
1275 "compact_daemon_migrate_scanned",
1276 "compact_daemon_free_scanned",
1279 #ifdef CONFIG_HUGETLB_PAGE
1280 "htlb_buddy_alloc_success",
1281 "htlb_buddy_alloc_fail",
1284 "cma_alloc_success",
1287 "unevictable_pgs_culled",
1288 "unevictable_pgs_scanned",
1289 "unevictable_pgs_rescued",
1290 "unevictable_pgs_mlocked",
1291 "unevictable_pgs_munlocked",
1292 "unevictable_pgs_cleared",
1293 "unevictable_pgs_stranded",
1295 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1297 "thp_fault_fallback",
1298 "thp_fault_fallback_charge",
1299 "thp_collapse_alloc",
1300 "thp_collapse_alloc_failed",
1302 "thp_file_fallback",
1303 "thp_file_fallback_charge",
1306 "thp_split_page_failed",
1307 "thp_deferred_split_page",
1309 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1312 "thp_zero_page_alloc",
1313 "thp_zero_page_alloc_failed",
1315 "thp_swpout_fallback",
1317 #ifdef CONFIG_MEMORY_BALLOON
1320 #ifdef CONFIG_BALLOON_COMPACTION
1323 #endif /* CONFIG_MEMORY_BALLOON */
1324 #ifdef CONFIG_DEBUG_TLBFLUSH
1325 "nr_tlb_remote_flush",
1326 "nr_tlb_remote_flush_received",
1327 "nr_tlb_local_flush_all",
1328 "nr_tlb_local_flush_one",
1329 #endif /* CONFIG_DEBUG_TLBFLUSH */
1331 #ifdef CONFIG_DEBUG_VM_VMACACHE
1332 "vmacache_find_calls",
1333 "vmacache_find_hits",
1340 "direct_map_level2_splits",
1341 "direct_map_level3_splits",
1343 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1345 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1347 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1348 defined(CONFIG_PROC_FS)
1349 static void *frag_start(struct seq_file *m, loff_t *pos)
1354 for (pgdat = first_online_pgdat();
1356 pgdat = next_online_pgdat(pgdat))
1362 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1364 pg_data_t *pgdat = (pg_data_t *)arg;
1367 return next_online_pgdat(pgdat);
1370 static void frag_stop(struct seq_file *m, void *arg)
1375 * Walk zones in a node and print using a callback.
1376 * If @assert_populated is true, only use callback for zones that are populated.
1378 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1379 bool assert_populated, bool nolock,
1380 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1383 struct zone *node_zones = pgdat->node_zones;
1384 unsigned long flags;
1386 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1387 if (assert_populated && !populated_zone(zone))
1391 spin_lock_irqsave(&zone->lock, flags);
1392 print(m, pgdat, zone);
1394 spin_unlock_irqrestore(&zone->lock, flags);
1399 #ifdef CONFIG_PROC_FS
1400 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1405 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1406 for (order = 0; order < MAX_ORDER; ++order)
1407 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1412 * This walks the free areas for each zone.
1414 static int frag_show(struct seq_file *m, void *arg)
1416 pg_data_t *pgdat = (pg_data_t *)arg;
1417 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1421 static void pagetypeinfo_showfree_print(struct seq_file *m,
1422 pg_data_t *pgdat, struct zone *zone)
1426 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1427 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1430 migratetype_names[mtype]);
1431 for (order = 0; order < MAX_ORDER; ++order) {
1432 unsigned long freecount = 0;
1433 struct free_area *area;
1434 struct list_head *curr;
1435 bool overflow = false;
1437 area = &(zone->free_area[order]);
1439 list_for_each(curr, &area->free_list[mtype]) {
1441 * Cap the free_list iteration because it might
1442 * be really large and we are under a spinlock
1443 * so a long time spent here could trigger a
1444 * hard lockup detector. Anyway this is a
1445 * debugging tool so knowing there is a handful
1446 * of pages of this order should be more than
1449 if (++freecount >= 100000) {
1454 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1455 spin_unlock_irq(&zone->lock);
1457 spin_lock_irq(&zone->lock);
1463 /* Print out the free pages at each order for each migatetype */
1464 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1467 pg_data_t *pgdat = (pg_data_t *)arg;
1470 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1471 for (order = 0; order < MAX_ORDER; ++order)
1472 seq_printf(m, "%6d ", order);
1475 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1480 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1481 pg_data_t *pgdat, struct zone *zone)
1485 unsigned long start_pfn = zone->zone_start_pfn;
1486 unsigned long end_pfn = zone_end_pfn(zone);
1487 unsigned long count[MIGRATE_TYPES] = { 0, };
1489 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1492 page = pfn_to_online_page(pfn);
1496 if (page_zone(page) != zone)
1499 mtype = get_pageblock_migratetype(page);
1501 if (mtype < MIGRATE_TYPES)
1506 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1507 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1508 seq_printf(m, "%12lu ", count[mtype]);
1512 /* Print out the number of pageblocks for each migratetype */
1513 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1516 pg_data_t *pgdat = (pg_data_t *)arg;
1518 seq_printf(m, "\n%-23s", "Number of blocks type ");
1519 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1520 seq_printf(m, "%12s ", migratetype_names[mtype]);
1522 walk_zones_in_node(m, pgdat, true, false,
1523 pagetypeinfo_showblockcount_print);
1529 * Print out the number of pageblocks for each migratetype that contain pages
1530 * of other types. This gives an indication of how well fallbacks are being
1531 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1532 * to determine what is going on
1534 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1536 #ifdef CONFIG_PAGE_OWNER
1539 if (!static_branch_unlikely(&page_owner_inited))
1542 drain_all_pages(NULL);
1544 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1545 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1546 seq_printf(m, "%12s ", migratetype_names[mtype]);
1549 walk_zones_in_node(m, pgdat, true, true,
1550 pagetypeinfo_showmixedcount_print);
1551 #endif /* CONFIG_PAGE_OWNER */
1555 * This prints out statistics in relation to grouping pages by mobility.
1556 * It is expensive to collect so do not constantly read the file.
1558 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1560 pg_data_t *pgdat = (pg_data_t *)arg;
1562 /* check memoryless node */
1563 if (!node_state(pgdat->node_id, N_MEMORY))
1566 seq_printf(m, "Page block order: %d\n", pageblock_order);
1567 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1569 pagetypeinfo_showfree(m, pgdat);
1570 pagetypeinfo_showblockcount(m, pgdat);
1571 pagetypeinfo_showmixedcount(m, pgdat);
1576 static const struct seq_operations fragmentation_op = {
1577 .start = frag_start,
1583 static const struct seq_operations pagetypeinfo_op = {
1584 .start = frag_start,
1587 .show = pagetypeinfo_show,
1590 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1594 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1595 struct zone *compare = &pgdat->node_zones[zid];
1597 if (populated_zone(compare))
1598 return zone == compare;
1604 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1608 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1609 if (is_zone_first_populated(pgdat, zone)) {
1610 seq_printf(m, "\n per-node stats");
1611 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1612 unsigned long pages = node_page_state_pages(pgdat, i);
1614 if (vmstat_item_print_in_thp(i))
1615 pages /= HPAGE_PMD_NR;
1616 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1629 zone_page_state(zone, NR_FREE_PAGES),
1630 min_wmark_pages(zone),
1631 low_wmark_pages(zone),
1632 high_wmark_pages(zone),
1633 zone->spanned_pages,
1634 zone->present_pages,
1635 zone_managed_pages(zone),
1636 zone_cma_pages(zone));
1639 "\n protection: (%ld",
1640 zone->lowmem_reserve[0]);
1641 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1642 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1645 /* If unpopulated, no other information is useful */
1646 if (!populated_zone(zone)) {
1651 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1652 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1653 zone_page_state(zone, i));
1656 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1657 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1658 zone_numa_event_state(zone, i));
1661 seq_printf(m, "\n pagesets");
1662 for_each_online_cpu(i) {
1663 struct per_cpu_pages *pcp;
1664 struct per_cpu_zonestat __maybe_unused *pzstats;
1666 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1677 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1678 seq_printf(m, "\n vm stats threshold: %d",
1679 pzstats->stat_threshold);
1683 "\n node_unreclaimable: %u"
1684 "\n start_pfn: %lu",
1685 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1686 zone->zone_start_pfn);
1691 * Output information about zones in @pgdat. All zones are printed regardless
1692 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1693 * set of all zones and userspace would not be aware of such zones if they are
1694 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1696 static int zoneinfo_show(struct seq_file *m, void *arg)
1698 pg_data_t *pgdat = (pg_data_t *)arg;
1699 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1703 static const struct seq_operations zoneinfo_op = {
1704 .start = frag_start, /* iterate over all zones. The same as in
1708 .show = zoneinfo_show,
1711 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1712 NR_VM_NUMA_EVENT_ITEMS + \
1713 NR_VM_NODE_STAT_ITEMS + \
1714 NR_VM_WRITEBACK_STAT_ITEMS + \
1715 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1716 NR_VM_EVENT_ITEMS : 0))
1718 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1723 if (*pos >= NR_VMSTAT_ITEMS)
1726 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1727 fold_vm_numa_events();
1728 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1731 return ERR_PTR(-ENOMEM);
1732 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1733 v[i] = global_zone_page_state(i);
1734 v += NR_VM_ZONE_STAT_ITEMS;
1737 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1738 v[i] = global_numa_event_state(i);
1739 v += NR_VM_NUMA_EVENT_ITEMS;
1742 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1743 v[i] = global_node_page_state_pages(i);
1744 if (vmstat_item_print_in_thp(i))
1745 v[i] /= HPAGE_PMD_NR;
1747 v += NR_VM_NODE_STAT_ITEMS;
1749 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1750 v + NR_DIRTY_THRESHOLD);
1751 v += NR_VM_WRITEBACK_STAT_ITEMS;
1753 #ifdef CONFIG_VM_EVENT_COUNTERS
1755 v[PGPGIN] /= 2; /* sectors -> kbytes */
1758 return (unsigned long *)m->private + *pos;
1761 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1764 if (*pos >= NR_VMSTAT_ITEMS)
1766 return (unsigned long *)m->private + *pos;
1769 static int vmstat_show(struct seq_file *m, void *arg)
1771 unsigned long *l = arg;
1772 unsigned long off = l - (unsigned long *)m->private;
1774 seq_puts(m, vmstat_text[off]);
1775 seq_put_decimal_ull(m, " ", *l);
1778 if (off == NR_VMSTAT_ITEMS - 1) {
1780 * We've come to the end - add any deprecated counters to avoid
1781 * breaking userspace which might depend on them being present.
1783 seq_puts(m, "nr_unstable 0\n");
1788 static void vmstat_stop(struct seq_file *m, void *arg)
1794 static const struct seq_operations vmstat_op = {
1795 .start = vmstat_start,
1796 .next = vmstat_next,
1797 .stop = vmstat_stop,
1798 .show = vmstat_show,
1800 #endif /* CONFIG_PROC_FS */
1803 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1804 int sysctl_stat_interval __read_mostly = HZ;
1806 #ifdef CONFIG_PROC_FS
1807 static void refresh_vm_stats(struct work_struct *work)
1809 refresh_cpu_vm_stats(true);
1812 int vmstat_refresh(struct ctl_table *table, int write,
1813 void *buffer, size_t *lenp, loff_t *ppos)
1820 * The regular update, every sysctl_stat_interval, may come later
1821 * than expected: leaving a significant amount in per_cpu buckets.
1822 * This is particularly misleading when checking a quantity of HUGE
1823 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1824 * which can equally be echo'ed to or cat'ted from (by root),
1825 * can be used to update the stats just before reading them.
1827 * Oh, and since global_zone_page_state() etc. are so careful to hide
1828 * transiently negative values, report an error here if any of
1829 * the stats is negative, so we know to go looking for imbalance.
1831 err = schedule_on_each_cpu(refresh_vm_stats);
1834 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1836 * Skip checking stats known to go negative occasionally.
1839 case NR_ZONE_WRITE_PENDING:
1840 case NR_FREE_CMA_PAGES:
1843 val = atomic_long_read(&vm_zone_stat[i]);
1845 pr_warn("%s: %s %ld\n",
1846 __func__, zone_stat_name(i), val);
1849 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1851 * Skip checking stats known to go negative occasionally.
1857 val = atomic_long_read(&vm_node_stat[i]);
1859 pr_warn("%s: %s %ld\n",
1860 __func__, node_stat_name(i), val);
1869 #endif /* CONFIG_PROC_FS */
1871 static void vmstat_update(struct work_struct *w)
1873 if (refresh_cpu_vm_stats(true)) {
1875 * Counters were updated so we expect more updates
1876 * to occur in the future. Keep on running the
1877 * update worker thread.
1879 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1880 this_cpu_ptr(&vmstat_work),
1881 round_jiffies_relative(sysctl_stat_interval));
1886 * Switch off vmstat processing and then fold all the remaining differentials
1887 * until the diffs stay at zero. The function is used by NOHZ and can only be
1888 * invoked when tick processing is not active.
1891 * Check if the diffs for a certain cpu indicate that
1892 * an update is needed.
1894 static bool need_update(int cpu)
1896 pg_data_t *last_pgdat = NULL;
1899 for_each_populated_zone(zone) {
1900 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1901 struct per_cpu_nodestat *n;
1904 * The fast way of checking if there are any vmstat diffs.
1906 if (memchr_inv(pzstats->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1907 sizeof(pzstats->vm_stat_diff[0])))
1910 if (last_pgdat == zone->zone_pgdat)
1912 last_pgdat = zone->zone_pgdat;
1913 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1914 if (memchr_inv(n->vm_node_stat_diff, 0, NR_VM_NODE_STAT_ITEMS *
1915 sizeof(n->vm_node_stat_diff[0])))
1922 * Switch off vmstat processing and then fold all the remaining differentials
1923 * until the diffs stay at zero. The function is used by NOHZ and can only be
1924 * invoked when tick processing is not active.
1926 void quiet_vmstat(void)
1928 if (system_state != SYSTEM_RUNNING)
1931 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1934 if (!need_update(smp_processor_id()))
1938 * Just refresh counters and do not care about the pending delayed
1939 * vmstat_update. It doesn't fire that often to matter and canceling
1940 * it would be too expensive from this path.
1941 * vmstat_shepherd will take care about that for us.
1943 refresh_cpu_vm_stats(false);
1947 * Shepherd worker thread that checks the
1948 * differentials of processors that have their worker
1949 * threads for vm statistics updates disabled because of
1952 static void vmstat_shepherd(struct work_struct *w);
1954 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1956 static void vmstat_shepherd(struct work_struct *w)
1961 /* Check processors whose vmstat worker threads have been disabled */
1962 for_each_online_cpu(cpu) {
1963 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1965 if (!delayed_work_pending(dw) && need_update(cpu))
1966 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1972 schedule_delayed_work(&shepherd,
1973 round_jiffies_relative(sysctl_stat_interval));
1976 static void __init start_shepherd_timer(void)
1980 for_each_possible_cpu(cpu)
1981 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1984 schedule_delayed_work(&shepherd,
1985 round_jiffies_relative(sysctl_stat_interval));
1988 static void __init init_cpu_node_state(void)
1992 for_each_online_node(node) {
1993 if (cpumask_weight(cpumask_of_node(node)) > 0)
1994 node_set_state(node, N_CPU);
1998 static int vmstat_cpu_online(unsigned int cpu)
2000 refresh_zone_stat_thresholds();
2001 node_set_state(cpu_to_node(cpu), N_CPU);
2005 static int vmstat_cpu_down_prep(unsigned int cpu)
2007 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2011 static int vmstat_cpu_dead(unsigned int cpu)
2013 const struct cpumask *node_cpus;
2016 node = cpu_to_node(cpu);
2018 refresh_zone_stat_thresholds();
2019 node_cpus = cpumask_of_node(node);
2020 if (cpumask_weight(node_cpus) > 0)
2023 node_clear_state(node, N_CPU);
2029 struct workqueue_struct *mm_percpu_wq;
2031 void __init init_mm_internals(void)
2033 int ret __maybe_unused;
2035 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2038 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2039 NULL, vmstat_cpu_dead);
2041 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2043 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2045 vmstat_cpu_down_prep);
2047 pr_err("vmstat: failed to register 'online' hotplug state\n");
2050 init_cpu_node_state();
2053 start_shepherd_timer();
2055 #ifdef CONFIG_PROC_FS
2056 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2057 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2058 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2059 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2063 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2066 * Return an index indicating how much of the available free memory is
2067 * unusable for an allocation of the requested size.
2069 static int unusable_free_index(unsigned int order,
2070 struct contig_page_info *info)
2072 /* No free memory is interpreted as all free memory is unusable */
2073 if (info->free_pages == 0)
2077 * Index should be a value between 0 and 1. Return a value to 3
2080 * 0 => no fragmentation
2081 * 1 => high fragmentation
2083 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2087 static void unusable_show_print(struct seq_file *m,
2088 pg_data_t *pgdat, struct zone *zone)
2092 struct contig_page_info info;
2094 seq_printf(m, "Node %d, zone %8s ",
2097 for (order = 0; order < MAX_ORDER; ++order) {
2098 fill_contig_page_info(zone, order, &info);
2099 index = unusable_free_index(order, &info);
2100 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2107 * Display unusable free space index
2109 * The unusable free space index measures how much of the available free
2110 * memory cannot be used to satisfy an allocation of a given size and is a
2111 * value between 0 and 1. The higher the value, the more of free memory is
2112 * unusable and by implication, the worse the external fragmentation is. This
2113 * can be expressed as a percentage by multiplying by 100.
2115 static int unusable_show(struct seq_file *m, void *arg)
2117 pg_data_t *pgdat = (pg_data_t *)arg;
2119 /* check memoryless node */
2120 if (!node_state(pgdat->node_id, N_MEMORY))
2123 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2128 static const struct seq_operations unusable_sops = {
2129 .start = frag_start,
2132 .show = unusable_show,
2135 DEFINE_SEQ_ATTRIBUTE(unusable);
2137 static void extfrag_show_print(struct seq_file *m,
2138 pg_data_t *pgdat, struct zone *zone)
2143 /* Alloc on stack as interrupts are disabled for zone walk */
2144 struct contig_page_info info;
2146 seq_printf(m, "Node %d, zone %8s ",
2149 for (order = 0; order < MAX_ORDER; ++order) {
2150 fill_contig_page_info(zone, order, &info);
2151 index = __fragmentation_index(order, &info);
2152 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2159 * Display fragmentation index for orders that allocations would fail for
2161 static int extfrag_show(struct seq_file *m, void *arg)
2163 pg_data_t *pgdat = (pg_data_t *)arg;
2165 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2170 static const struct seq_operations extfrag_sops = {
2171 .start = frag_start,
2174 .show = extfrag_show,
2177 DEFINE_SEQ_ATTRIBUTE(extfrag);
2179 static int __init extfrag_debug_init(void)
2181 struct dentry *extfrag_debug_root;
2183 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2185 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2188 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2194 module_init(extfrag_debug_init);