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 *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 if (vmstat_item_in_bytes(item
)) {
345 VM_WARN_ON_ONCE(delta
& (PAGE_SIZE
- 1));
346 delta
>>= PAGE_SHIFT
;
349 x
= delta
+ __this_cpu_read(*p
);
351 t
= __this_cpu_read(pcp
->stat_threshold
);
353 if (unlikely(x
> t
|| x
< -t
)) {
354 node_page_state_add(x
, pgdat
, item
);
357 __this_cpu_write(*p
, x
);
359 EXPORT_SYMBOL(__mod_node_page_state
);
362 * Optimized increment and decrement functions.
364 * These are only for a single page and therefore can take a struct page *
365 * argument instead of struct zone *. This allows the inclusion of the code
366 * generated for page_zone(page) into the optimized functions.
368 * No overflow check is necessary and therefore the differential can be
369 * incremented or decremented in place which may allow the compilers to
370 * generate better code.
371 * The increment or decrement is known and therefore one boundary check can
374 * NOTE: These functions are very performance sensitive. Change only
377 * Some processors have inc/dec instructions that are atomic vs an interrupt.
378 * However, the code must first determine the differential location in a zone
379 * based on the processor number and then inc/dec the counter. There is no
380 * guarantee without disabling preemption that the processor will not change
381 * in between and therefore the atomicity vs. interrupt cannot be exploited
382 * in a useful way here.
384 void __inc_zone_state(struct zone
*zone
, enum zone_stat_item item
)
386 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
387 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
390 v
= __this_cpu_inc_return(*p
);
391 t
= __this_cpu_read(pcp
->stat_threshold
);
392 if (unlikely(v
> t
)) {
393 s8 overstep
= t
>> 1;
395 zone_page_state_add(v
+ overstep
, zone
, item
);
396 __this_cpu_write(*p
, -overstep
);
400 void __inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
402 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
403 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
406 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
408 v
= __this_cpu_inc_return(*p
);
409 t
= __this_cpu_read(pcp
->stat_threshold
);
410 if (unlikely(v
> t
)) {
411 s8 overstep
= t
>> 1;
413 node_page_state_add(v
+ overstep
, pgdat
, item
);
414 __this_cpu_write(*p
, -overstep
);
418 void __inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
420 __inc_zone_state(page_zone(page
), item
);
422 EXPORT_SYMBOL(__inc_zone_page_state
);
424 void __inc_node_page_state(struct page
*page
, enum node_stat_item item
)
426 __inc_node_state(page_pgdat(page
), item
);
428 EXPORT_SYMBOL(__inc_node_page_state
);
430 void __dec_zone_state(struct zone
*zone
, enum zone_stat_item item
)
432 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
433 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
436 v
= __this_cpu_dec_return(*p
);
437 t
= __this_cpu_read(pcp
->stat_threshold
);
438 if (unlikely(v
< - t
)) {
439 s8 overstep
= t
>> 1;
441 zone_page_state_add(v
- overstep
, zone
, item
);
442 __this_cpu_write(*p
, overstep
);
446 void __dec_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
448 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
449 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
452 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
454 v
= __this_cpu_dec_return(*p
);
455 t
= __this_cpu_read(pcp
->stat_threshold
);
456 if (unlikely(v
< - t
)) {
457 s8 overstep
= t
>> 1;
459 node_page_state_add(v
- overstep
, pgdat
, item
);
460 __this_cpu_write(*p
, overstep
);
464 void __dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
466 __dec_zone_state(page_zone(page
), item
);
468 EXPORT_SYMBOL(__dec_zone_page_state
);
470 void __dec_node_page_state(struct page
*page
, enum node_stat_item item
)
472 __dec_node_state(page_pgdat(page
), item
);
474 EXPORT_SYMBOL(__dec_node_page_state
);
476 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
478 * If we have cmpxchg_local support then we do not need to incur the overhead
479 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
481 * mod_state() modifies the zone counter state through atomic per cpu
484 * Overstep mode specifies how overstep should handled:
486 * 1 Overstepping half of threshold
487 * -1 Overstepping minus half of threshold
489 static inline void mod_zone_state(struct zone
*zone
,
490 enum zone_stat_item item
, long delta
, int overstep_mode
)
492 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
493 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
497 z
= 0; /* overflow to zone counters */
500 * The fetching of the stat_threshold is racy. We may apply
501 * a counter threshold to the wrong the cpu if we get
502 * rescheduled while executing here. However, the next
503 * counter update will apply the threshold again and
504 * therefore bring the counter under the threshold again.
506 * Most of the time the thresholds are the same anyways
507 * for all cpus in a zone.
509 t
= this_cpu_read(pcp
->stat_threshold
);
511 o
= this_cpu_read(*p
);
514 if (n
> t
|| n
< -t
) {
515 int os
= overstep_mode
* (t
>> 1) ;
517 /* Overflow must be added to zone counters */
521 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
524 zone_page_state_add(z
, zone
, item
);
527 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
530 mod_zone_state(zone
, item
, delta
, 0);
532 EXPORT_SYMBOL(mod_zone_page_state
);
534 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
536 mod_zone_state(page_zone(page
), item
, 1, 1);
538 EXPORT_SYMBOL(inc_zone_page_state
);
540 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
542 mod_zone_state(page_zone(page
), item
, -1, -1);
544 EXPORT_SYMBOL(dec_zone_page_state
);
546 static inline void mod_node_state(struct pglist_data
*pgdat
,
547 enum node_stat_item item
, int delta
, int overstep_mode
)
549 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
550 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
553 if (vmstat_item_in_bytes(item
)) {
554 VM_WARN_ON_ONCE(delta
& (PAGE_SIZE
- 1));
555 delta
>>= PAGE_SHIFT
;
559 z
= 0; /* overflow to node counters */
562 * The fetching of the stat_threshold is racy. We may apply
563 * a counter threshold to the wrong the cpu if we get
564 * rescheduled while executing here. However, the next
565 * counter update will apply the threshold again and
566 * therefore bring the counter under the threshold again.
568 * Most of the time the thresholds are the same anyways
569 * for all cpus in a node.
571 t
= this_cpu_read(pcp
->stat_threshold
);
573 o
= this_cpu_read(*p
);
576 if (n
> t
|| n
< -t
) {
577 int os
= overstep_mode
* (t
>> 1) ;
579 /* Overflow must be added to node counters */
583 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
586 node_page_state_add(z
, pgdat
, item
);
589 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
592 mod_node_state(pgdat
, item
, delta
, 0);
594 EXPORT_SYMBOL(mod_node_page_state
);
596 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
598 mod_node_state(pgdat
, item
, 1, 1);
601 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
603 mod_node_state(page_pgdat(page
), item
, 1, 1);
605 EXPORT_SYMBOL(inc_node_page_state
);
607 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
609 mod_node_state(page_pgdat(page
), item
, -1, -1);
611 EXPORT_SYMBOL(dec_node_page_state
);
614 * Use interrupt disable to serialize counter updates
616 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
621 local_irq_save(flags
);
622 __mod_zone_page_state(zone
, item
, delta
);
623 local_irq_restore(flags
);
625 EXPORT_SYMBOL(mod_zone_page_state
);
627 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
632 zone
= page_zone(page
);
633 local_irq_save(flags
);
634 __inc_zone_state(zone
, item
);
635 local_irq_restore(flags
);
637 EXPORT_SYMBOL(inc_zone_page_state
);
639 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
643 local_irq_save(flags
);
644 __dec_zone_page_state(page
, item
);
645 local_irq_restore(flags
);
647 EXPORT_SYMBOL(dec_zone_page_state
);
649 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
653 local_irq_save(flags
);
654 __inc_node_state(pgdat
, item
);
655 local_irq_restore(flags
);
657 EXPORT_SYMBOL(inc_node_state
);
659 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
664 local_irq_save(flags
);
665 __mod_node_page_state(pgdat
, item
, delta
);
666 local_irq_restore(flags
);
668 EXPORT_SYMBOL(mod_node_page_state
);
670 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
673 struct pglist_data
*pgdat
;
675 pgdat
= page_pgdat(page
);
676 local_irq_save(flags
);
677 __inc_node_state(pgdat
, item
);
678 local_irq_restore(flags
);
680 EXPORT_SYMBOL(inc_node_page_state
);
682 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
686 local_irq_save(flags
);
687 __dec_node_page_state(page
, item
);
688 local_irq_restore(flags
);
690 EXPORT_SYMBOL(dec_node_page_state
);
694 * Fold a differential into the global counters.
695 * Returns the number of counters updated.
698 static int fold_diff(int *zone_diff
, int *numa_diff
, int *node_diff
)
703 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
705 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
709 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
711 atomic_long_add(numa_diff
[i
], &vm_numa_stat
[i
]);
715 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
717 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
723 static int fold_diff(int *zone_diff
, int *node_diff
)
728 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
730 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
734 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
736 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
741 #endif /* CONFIG_NUMA */
744 * Update the zone counters for the current cpu.
746 * Note that refresh_cpu_vm_stats strives to only access
747 * node local memory. The per cpu pagesets on remote zones are placed
748 * in the memory local to the processor using that pageset. So the
749 * loop over all zones will access a series of cachelines local to
752 * The call to zone_page_state_add updates the cachelines with the
753 * statistics in the remote zone struct as well as the global cachelines
754 * with the global counters. These could cause remote node cache line
755 * bouncing and will have to be only done when necessary.
757 * The function returns the number of global counters updated.
759 static int refresh_cpu_vm_stats(bool do_pagesets
)
761 struct pglist_data
*pgdat
;
764 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
766 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
768 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
771 for_each_populated_zone(zone
) {
772 struct per_cpu_pageset __percpu
*p
= zone
->pageset
;
774 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
777 v
= this_cpu_xchg(p
->vm_stat_diff
[i
], 0);
780 atomic_long_add(v
, &zone
->vm_stat
[i
]);
781 global_zone_diff
[i
] += v
;
783 /* 3 seconds idle till flush */
784 __this_cpu_write(p
->expire
, 3);
789 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
792 v
= this_cpu_xchg(p
->vm_numa_stat_diff
[i
], 0);
795 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
796 global_numa_diff
[i
] += v
;
797 __this_cpu_write(p
->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(p
->expire
) ||
811 !__this_cpu_read(p
->pcp
.count
))
815 * We never drain zones local to this processor.
817 if (zone_to_nid(zone
) == numa_node_id()) {
818 __this_cpu_write(p
->expire
, 0);
822 if (__this_cpu_dec_return(p
->expire
))
825 if (__this_cpu_read(p
->pcp
.count
)) {
826 drain_zone_pages(zone
, this_cpu_ptr(&p
->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
;
848 changes
+= fold_diff(global_zone_diff
, global_numa_diff
,
851 changes
+= fold_diff(global_zone_diff
, global_node_diff
);
857 * Fold the data for an offline cpu into the global array.
858 * There cannot be any access by the offline cpu and therefore
859 * synchronization is simplified.
861 void cpu_vm_stats_fold(int cpu
)
863 struct pglist_data
*pgdat
;
866 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
868 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
870 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
872 for_each_populated_zone(zone
) {
873 struct per_cpu_pageset
*p
;
875 p
= per_cpu_ptr(zone
->pageset
, cpu
);
877 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
878 if (p
->vm_stat_diff
[i
]) {
881 v
= p
->vm_stat_diff
[i
];
882 p
->vm_stat_diff
[i
] = 0;
883 atomic_long_add(v
, &zone
->vm_stat
[i
]);
884 global_zone_diff
[i
] += v
;
888 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
889 if (p
->vm_numa_stat_diff
[i
]) {
892 v
= p
->vm_numa_stat_diff
[i
];
893 p
->vm_numa_stat_diff
[i
] = 0;
894 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
895 global_numa_diff
[i
] += v
;
900 for_each_online_pgdat(pgdat
) {
901 struct per_cpu_nodestat
*p
;
903 p
= per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
);
905 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
906 if (p
->vm_node_stat_diff
[i
]) {
909 v
= p
->vm_node_stat_diff
[i
];
910 p
->vm_node_stat_diff
[i
] = 0;
911 atomic_long_add(v
, &pgdat
->vm_stat
[i
]);
912 global_node_diff
[i
] += v
;
917 fold_diff(global_zone_diff
, global_numa_diff
, global_node_diff
);
919 fold_diff(global_zone_diff
, global_node_diff
);
924 * this is only called if !populated_zone(zone), which implies no other users of
925 * pset->vm_stat_diff[] exsist.
927 void drain_zonestat(struct zone
*zone
, struct per_cpu_pageset
*pset
)
931 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
932 if (pset
->vm_stat_diff
[i
]) {
933 int v
= pset
->vm_stat_diff
[i
];
934 pset
->vm_stat_diff
[i
] = 0;
935 atomic_long_add(v
, &zone
->vm_stat
[i
]);
936 atomic_long_add(v
, &vm_zone_stat
[i
]);
940 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
941 if (pset
->vm_numa_stat_diff
[i
]) {
942 int v
= pset
->vm_numa_stat_diff
[i
];
944 pset
->vm_numa_stat_diff
[i
] = 0;
945 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
946 atomic_long_add(v
, &vm_numa_stat
[i
]);
953 void __inc_numa_state(struct zone
*zone
,
954 enum numa_stat_item item
)
956 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
957 u16 __percpu
*p
= pcp
->vm_numa_stat_diff
+ item
;
960 v
= __this_cpu_inc_return(*p
);
962 if (unlikely(v
> NUMA_STATS_THRESHOLD
)) {
963 zone_numa_state_add(v
, zone
, item
);
964 __this_cpu_write(*p
, 0);
969 * Determine the per node value of a stat item. This function
970 * is called frequently in a NUMA machine, so try to be as
971 * frugal as possible.
973 unsigned long sum_zone_node_page_state(int node
,
974 enum zone_stat_item item
)
976 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
978 unsigned long count
= 0;
980 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
981 count
+= zone_page_state(zones
+ i
, item
);
987 * Determine the per node value of a numa stat item. To avoid deviation,
988 * the per cpu stat number in vm_numa_stat_diff[] is also included.
990 unsigned long sum_zone_numa_state(int node
,
991 enum numa_stat_item item
)
993 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
995 unsigned long count
= 0;
997 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
998 count
+= zone_numa_state_snapshot(zones
+ i
, item
);
1004 * Determine the per node value of a stat item.
1006 unsigned long node_page_state_pages(struct pglist_data
*pgdat
,
1007 enum node_stat_item item
)
1009 long x
= atomic_long_read(&pgdat
->vm_stat
[item
]);
1017 unsigned long node_page_state(struct pglist_data
*pgdat
,
1018 enum node_stat_item item
)
1020 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
1022 return node_page_state_pages(pgdat
, item
);
1026 #ifdef CONFIG_COMPACTION
1028 struct contig_page_info
{
1029 unsigned long free_pages
;
1030 unsigned long free_blocks_total
;
1031 unsigned long free_blocks_suitable
;
1035 * Calculate the number of free pages in a zone, how many contiguous
1036 * pages are free and how many are large enough to satisfy an allocation of
1037 * the target size. Note that this function makes no attempt to estimate
1038 * how many suitable free blocks there *might* be if MOVABLE pages were
1039 * migrated. Calculating that is possible, but expensive and can be
1040 * figured out from userspace
1042 static void fill_contig_page_info(struct zone
*zone
,
1043 unsigned int suitable_order
,
1044 struct contig_page_info
*info
)
1048 info
->free_pages
= 0;
1049 info
->free_blocks_total
= 0;
1050 info
->free_blocks_suitable
= 0;
1052 for (order
= 0; order
< MAX_ORDER
; order
++) {
1053 unsigned long blocks
;
1055 /* Count number of free blocks */
1056 blocks
= zone
->free_area
[order
].nr_free
;
1057 info
->free_blocks_total
+= blocks
;
1059 /* Count free base pages */
1060 info
->free_pages
+= blocks
<< order
;
1062 /* Count the suitable free blocks */
1063 if (order
>= suitable_order
)
1064 info
->free_blocks_suitable
+= blocks
<<
1065 (order
- suitable_order
);
1070 * A fragmentation index only makes sense if an allocation of a requested
1071 * size would fail. If that is true, the fragmentation index indicates
1072 * whether external fragmentation or a lack of memory was the problem.
1073 * The value can be used to determine if page reclaim or compaction
1076 static int __fragmentation_index(unsigned int order
, struct contig_page_info
*info
)
1078 unsigned long requested
= 1UL << order
;
1080 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
1083 if (!info
->free_blocks_total
)
1086 /* Fragmentation index only makes sense when a request would fail */
1087 if (info
->free_blocks_suitable
)
1091 * Index is between 0 and 1 so return within 3 decimal places
1093 * 0 => allocation would fail due to lack of memory
1094 * 1 => allocation would fail due to fragmentation
1096 return 1000 - div_u64( (1000+(div_u64(info
->free_pages
* 1000ULL, requested
))), info
->free_blocks_total
);
1099 /* Same as __fragmentation index but allocs contig_page_info on stack */
1100 int fragmentation_index(struct zone
*zone
, unsigned int order
)
1102 struct contig_page_info info
;
1104 fill_contig_page_info(zone
, order
, &info
);
1105 return __fragmentation_index(order
, &info
);
1109 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1110 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1111 #ifdef CONFIG_ZONE_DMA
1112 #define TEXT_FOR_DMA(xx) xx "_dma",
1114 #define TEXT_FOR_DMA(xx)
1117 #ifdef CONFIG_ZONE_DMA32
1118 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1120 #define TEXT_FOR_DMA32(xx)
1123 #ifdef CONFIG_HIGHMEM
1124 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1126 #define TEXT_FOR_HIGHMEM(xx)
1129 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1130 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1132 const char * const vmstat_text
[] = {
1133 /* enum zone_stat_item counters */
1135 "nr_zone_inactive_anon",
1136 "nr_zone_active_anon",
1137 "nr_zone_inactive_file",
1138 "nr_zone_active_file",
1139 "nr_zone_unevictable",
1140 "nr_zone_write_pending",
1142 "nr_page_table_pages",
1144 #if IS_ENABLED(CONFIG_ZSMALLOC)
1149 /* enum numa_stat_item counters */
1159 /* enum node_stat_item counters */
1165 "nr_slab_reclaimable",
1166 "nr_slab_unreclaimable",
1170 "workingset_refault",
1171 "workingset_activate",
1172 "workingset_restore",
1173 "workingset_nodereclaim",
1179 "nr_writeback_temp",
1181 "nr_shmem_hugepages",
1182 "nr_shmem_pmdmapped",
1183 "nr_file_hugepages",
1184 "nr_file_pmdmapped",
1185 "nr_anon_transparent_hugepages",
1187 "nr_vmscan_immediate_reclaim",
1190 "nr_kernel_misc_reclaimable",
1191 "nr_foll_pin_acquired",
1192 "nr_foll_pin_released",
1194 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1195 "nr_shadow_call_stack",
1198 /* enum writeback_stat_item counters */
1199 "nr_dirty_threshold",
1200 "nr_dirty_background_threshold",
1202 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1203 /* enum vm_event_item counters */
1209 TEXTS_FOR_ZONES("pgalloc")
1210 TEXTS_FOR_ZONES("allocstall")
1211 TEXTS_FOR_ZONES("pgskip")
1227 "pgscan_direct_throttle",
1234 "zone_reclaim_failed",
1238 "kswapd_inodesteal",
1239 "kswapd_low_wmark_hit_quickly",
1240 "kswapd_high_wmark_hit_quickly",
1249 #ifdef CONFIG_NUMA_BALANCING
1251 "numa_huge_pte_updates",
1253 "numa_hint_faults_local",
1254 "numa_pages_migrated",
1256 #ifdef CONFIG_MIGRATION
1257 "pgmigrate_success",
1260 #ifdef CONFIG_COMPACTION
1261 "compact_migrate_scanned",
1262 "compact_free_scanned",
1267 "compact_daemon_wake",
1268 "compact_daemon_migrate_scanned",
1269 "compact_daemon_free_scanned",
1272 #ifdef CONFIG_HUGETLB_PAGE
1273 "htlb_buddy_alloc_success",
1274 "htlb_buddy_alloc_fail",
1276 "unevictable_pgs_culled",
1277 "unevictable_pgs_scanned",
1278 "unevictable_pgs_rescued",
1279 "unevictable_pgs_mlocked",
1280 "unevictable_pgs_munlocked",
1281 "unevictable_pgs_cleared",
1282 "unevictable_pgs_stranded",
1284 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1286 "thp_fault_fallback",
1287 "thp_fault_fallback_charge",
1288 "thp_collapse_alloc",
1289 "thp_collapse_alloc_failed",
1291 "thp_file_fallback",
1292 "thp_file_fallback_charge",
1295 "thp_split_page_failed",
1296 "thp_deferred_split_page",
1298 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1301 "thp_zero_page_alloc",
1302 "thp_zero_page_alloc_failed",
1304 "thp_swpout_fallback",
1306 #ifdef CONFIG_MEMORY_BALLOON
1309 #ifdef CONFIG_BALLOON_COMPACTION
1312 #endif /* CONFIG_MEMORY_BALLOON */
1313 #ifdef CONFIG_DEBUG_TLBFLUSH
1314 "nr_tlb_remote_flush",
1315 "nr_tlb_remote_flush_received",
1316 "nr_tlb_local_flush_all",
1317 "nr_tlb_local_flush_one",
1318 #endif /* CONFIG_DEBUG_TLBFLUSH */
1320 #ifdef CONFIG_DEBUG_VM_VMACACHE
1321 "vmacache_find_calls",
1322 "vmacache_find_hits",
1328 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1330 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1332 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1333 defined(CONFIG_PROC_FS)
1334 static void *frag_start(struct seq_file
*m
, loff_t
*pos
)
1339 for (pgdat
= first_online_pgdat();
1341 pgdat
= next_online_pgdat(pgdat
))
1347 static void *frag_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1349 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1352 return next_online_pgdat(pgdat
);
1355 static void frag_stop(struct seq_file
*m
, void *arg
)
1360 * Walk zones in a node and print using a callback.
1361 * If @assert_populated is true, only use callback for zones that are populated.
1363 static void walk_zones_in_node(struct seq_file
*m
, pg_data_t
*pgdat
,
1364 bool assert_populated
, bool nolock
,
1365 void (*print
)(struct seq_file
*m
, pg_data_t
*, struct zone
*))
1368 struct zone
*node_zones
= pgdat
->node_zones
;
1369 unsigned long flags
;
1371 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
1372 if (assert_populated
&& !populated_zone(zone
))
1376 spin_lock_irqsave(&zone
->lock
, flags
);
1377 print(m
, pgdat
, zone
);
1379 spin_unlock_irqrestore(&zone
->lock
, flags
);
1384 #ifdef CONFIG_PROC_FS
1385 static void frag_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1390 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1391 for (order
= 0; order
< MAX_ORDER
; ++order
)
1392 seq_printf(m
, "%6lu ", zone
->free_area
[order
].nr_free
);
1397 * This walks the free areas for each zone.
1399 static int frag_show(struct seq_file
*m
, void *arg
)
1401 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1402 walk_zones_in_node(m
, pgdat
, true, false, frag_show_print
);
1406 static void pagetypeinfo_showfree_print(struct seq_file
*m
,
1407 pg_data_t
*pgdat
, struct zone
*zone
)
1411 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++) {
1412 seq_printf(m
, "Node %4d, zone %8s, type %12s ",
1415 migratetype_names
[mtype
]);
1416 for (order
= 0; order
< MAX_ORDER
; ++order
) {
1417 unsigned long freecount
= 0;
1418 struct free_area
*area
;
1419 struct list_head
*curr
;
1420 bool overflow
= false;
1422 area
= &(zone
->free_area
[order
]);
1424 list_for_each(curr
, &area
->free_list
[mtype
]) {
1426 * Cap the free_list iteration because it might
1427 * be really large and we are under a spinlock
1428 * so a long time spent here could trigger a
1429 * hard lockup detector. Anyway this is a
1430 * debugging tool so knowing there is a handful
1431 * of pages of this order should be more than
1434 if (++freecount
>= 100000) {
1439 seq_printf(m
, "%s%6lu ", overflow
? ">" : "", freecount
);
1440 spin_unlock_irq(&zone
->lock
);
1442 spin_lock_irq(&zone
->lock
);
1448 /* Print out the free pages at each order for each migatetype */
1449 static int pagetypeinfo_showfree(struct seq_file
*m
, void *arg
)
1452 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1455 seq_printf(m
, "%-43s ", "Free pages count per migrate type at order");
1456 for (order
= 0; order
< MAX_ORDER
; ++order
)
1457 seq_printf(m
, "%6d ", order
);
1460 walk_zones_in_node(m
, pgdat
, true, false, pagetypeinfo_showfree_print
);
1465 static void pagetypeinfo_showblockcount_print(struct seq_file
*m
,
1466 pg_data_t
*pgdat
, struct zone
*zone
)
1470 unsigned long start_pfn
= zone
->zone_start_pfn
;
1471 unsigned long end_pfn
= zone_end_pfn(zone
);
1472 unsigned long count
[MIGRATE_TYPES
] = { 0, };
1474 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
1477 page
= pfn_to_online_page(pfn
);
1481 /* Watch for unexpected holes punched in the memmap */
1482 if (!memmap_valid_within(pfn
, page
, zone
))
1485 if (page_zone(page
) != zone
)
1488 mtype
= get_pageblock_migratetype(page
);
1490 if (mtype
< MIGRATE_TYPES
)
1495 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1496 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1497 seq_printf(m
, "%12lu ", count
[mtype
]);
1501 /* Print out the number of pageblocks for each migratetype */
1502 static int pagetypeinfo_showblockcount(struct seq_file
*m
, void *arg
)
1505 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1507 seq_printf(m
, "\n%-23s", "Number of blocks type ");
1508 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1509 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1511 walk_zones_in_node(m
, pgdat
, true, false,
1512 pagetypeinfo_showblockcount_print
);
1518 * Print out the number of pageblocks for each migratetype that contain pages
1519 * of other types. This gives an indication of how well fallbacks are being
1520 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1521 * to determine what is going on
1523 static void pagetypeinfo_showmixedcount(struct seq_file
*m
, pg_data_t
*pgdat
)
1525 #ifdef CONFIG_PAGE_OWNER
1528 if (!static_branch_unlikely(&page_owner_inited
))
1531 drain_all_pages(NULL
);
1533 seq_printf(m
, "\n%-23s", "Number of mixed blocks ");
1534 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1535 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1538 walk_zones_in_node(m
, pgdat
, true, true,
1539 pagetypeinfo_showmixedcount_print
);
1540 #endif /* CONFIG_PAGE_OWNER */
1544 * This prints out statistics in relation to grouping pages by mobility.
1545 * It is expensive to collect so do not constantly read the file.
1547 static int pagetypeinfo_show(struct seq_file
*m
, void *arg
)
1549 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1551 /* check memoryless node */
1552 if (!node_state(pgdat
->node_id
, N_MEMORY
))
1555 seq_printf(m
, "Page block order: %d\n", pageblock_order
);
1556 seq_printf(m
, "Pages per block: %lu\n", pageblock_nr_pages
);
1558 pagetypeinfo_showfree(m
, pgdat
);
1559 pagetypeinfo_showblockcount(m
, pgdat
);
1560 pagetypeinfo_showmixedcount(m
, pgdat
);
1565 static const struct seq_operations fragmentation_op
= {
1566 .start
= frag_start
,
1572 static const struct seq_operations pagetypeinfo_op
= {
1573 .start
= frag_start
,
1576 .show
= pagetypeinfo_show
,
1579 static bool is_zone_first_populated(pg_data_t
*pgdat
, struct zone
*zone
)
1583 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1584 struct zone
*compare
= &pgdat
->node_zones
[zid
];
1586 if (populated_zone(compare
))
1587 return zone
== compare
;
1593 static void zoneinfo_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1597 seq_printf(m
, "Node %d, zone %8s", pgdat
->node_id
, zone
->name
);
1598 if (is_zone_first_populated(pgdat
, zone
)) {
1599 seq_printf(m
, "\n per-node stats");
1600 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
1601 seq_printf(m
, "\n %-12s %lu", node_stat_name(i
),
1602 node_page_state_pages(pgdat
, i
));
1613 zone_page_state(zone
, NR_FREE_PAGES
),
1614 min_wmark_pages(zone
),
1615 low_wmark_pages(zone
),
1616 high_wmark_pages(zone
),
1617 zone
->spanned_pages
,
1618 zone
->present_pages
,
1619 zone_managed_pages(zone
));
1621 /* If unpopulated, no other information is useful */
1622 if (!populated_zone(zone
)) {
1628 "\n protection: (%ld",
1629 zone
->lowmem_reserve
[0]);
1630 for (i
= 1; i
< ARRAY_SIZE(zone
->lowmem_reserve
); i
++)
1631 seq_printf(m
, ", %ld", zone
->lowmem_reserve
[i
]);
1634 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1635 seq_printf(m
, "\n %-12s %lu", zone_stat_name(i
),
1636 zone_page_state(zone
, i
));
1639 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1640 seq_printf(m
, "\n %-12s %lu", numa_stat_name(i
),
1641 zone_numa_state_snapshot(zone
, i
));
1644 seq_printf(m
, "\n pagesets");
1645 for_each_online_cpu(i
) {
1646 struct per_cpu_pageset
*pageset
;
1648 pageset
= per_cpu_ptr(zone
->pageset
, i
);
1657 pageset
->pcp
.batch
);
1659 seq_printf(m
, "\n vm stats threshold: %d",
1660 pageset
->stat_threshold
);
1664 "\n node_unreclaimable: %u"
1665 "\n start_pfn: %lu",
1666 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
,
1667 zone
->zone_start_pfn
);
1672 * Output information about zones in @pgdat. All zones are printed regardless
1673 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1674 * set of all zones and userspace would not be aware of such zones if they are
1675 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1677 static int zoneinfo_show(struct seq_file
*m
, void *arg
)
1679 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1680 walk_zones_in_node(m
, pgdat
, false, false, zoneinfo_show_print
);
1684 static const struct seq_operations zoneinfo_op
= {
1685 .start
= frag_start
, /* iterate over all zones. The same as in
1689 .show
= zoneinfo_show
,
1692 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1693 NR_VM_NUMA_STAT_ITEMS + \
1694 NR_VM_NODE_STAT_ITEMS + \
1695 NR_VM_WRITEBACK_STAT_ITEMS + \
1696 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1697 NR_VM_EVENT_ITEMS : 0))
1699 static void *vmstat_start(struct seq_file
*m
, loff_t
*pos
)
1704 if (*pos
>= NR_VMSTAT_ITEMS
)
1707 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text
) < NR_VMSTAT_ITEMS
);
1708 v
= kmalloc_array(NR_VMSTAT_ITEMS
, sizeof(unsigned long), GFP_KERNEL
);
1711 return ERR_PTR(-ENOMEM
);
1712 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1713 v
[i
] = global_zone_page_state(i
);
1714 v
+= NR_VM_ZONE_STAT_ITEMS
;
1717 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1718 v
[i
] = global_numa_state(i
);
1719 v
+= NR_VM_NUMA_STAT_ITEMS
;
1722 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
1723 v
[i
] = global_node_page_state_pages(i
);
1724 v
+= NR_VM_NODE_STAT_ITEMS
;
1726 global_dirty_limits(v
+ NR_DIRTY_BG_THRESHOLD
,
1727 v
+ NR_DIRTY_THRESHOLD
);
1728 v
+= NR_VM_WRITEBACK_STAT_ITEMS
;
1730 #ifdef CONFIG_VM_EVENT_COUNTERS
1732 v
[PGPGIN
] /= 2; /* sectors -> kbytes */
1735 return (unsigned long *)m
->private + *pos
;
1738 static void *vmstat_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1741 if (*pos
>= NR_VMSTAT_ITEMS
)
1743 return (unsigned long *)m
->private + *pos
;
1746 static int vmstat_show(struct seq_file
*m
, void *arg
)
1748 unsigned long *l
= arg
;
1749 unsigned long off
= l
- (unsigned long *)m
->private;
1751 seq_puts(m
, vmstat_text
[off
]);
1752 seq_put_decimal_ull(m
, " ", *l
);
1755 if (off
== NR_VMSTAT_ITEMS
- 1) {
1757 * We've come to the end - add any deprecated counters to avoid
1758 * breaking userspace which might depend on them being present.
1760 seq_puts(m
, "nr_unstable 0\n");
1765 static void vmstat_stop(struct seq_file
*m
, void *arg
)
1771 static const struct seq_operations vmstat_op
= {
1772 .start
= vmstat_start
,
1773 .next
= vmstat_next
,
1774 .stop
= vmstat_stop
,
1775 .show
= vmstat_show
,
1777 #endif /* CONFIG_PROC_FS */
1780 static DEFINE_PER_CPU(struct delayed_work
, vmstat_work
);
1781 int sysctl_stat_interval __read_mostly
= HZ
;
1783 #ifdef CONFIG_PROC_FS
1784 static void refresh_vm_stats(struct work_struct
*work
)
1786 refresh_cpu_vm_stats(true);
1789 int vmstat_refresh(struct ctl_table
*table
, int write
,
1790 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1797 * The regular update, every sysctl_stat_interval, may come later
1798 * than expected: leaving a significant amount in per_cpu buckets.
1799 * This is particularly misleading when checking a quantity of HUGE
1800 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1801 * which can equally be echo'ed to or cat'ted from (by root),
1802 * can be used to update the stats just before reading them.
1804 * Oh, and since global_zone_page_state() etc. are so careful to hide
1805 * transiently negative values, report an error here if any of
1806 * the stats is negative, so we know to go looking for imbalance.
1808 err
= schedule_on_each_cpu(refresh_vm_stats
);
1811 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
1812 val
= atomic_long_read(&vm_zone_stat
[i
]);
1814 pr_warn("%s: %s %ld\n",
1815 __func__
, zone_stat_name(i
), val
);
1820 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
1821 val
= atomic_long_read(&vm_numa_stat
[i
]);
1823 pr_warn("%s: %s %ld\n",
1824 __func__
, numa_stat_name(i
), val
);
1837 #endif /* CONFIG_PROC_FS */
1839 static void vmstat_update(struct work_struct
*w
)
1841 if (refresh_cpu_vm_stats(true)) {
1843 * Counters were updated so we expect more updates
1844 * to occur in the future. Keep on running the
1845 * update worker thread.
1847 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq
,
1848 this_cpu_ptr(&vmstat_work
),
1849 round_jiffies_relative(sysctl_stat_interval
));
1854 * Switch off vmstat processing and then fold all the remaining differentials
1855 * until the diffs stay at zero. The function is used by NOHZ and can only be
1856 * invoked when tick processing is not active.
1859 * Check if the diffs for a certain cpu indicate that
1860 * an update is needed.
1862 static bool need_update(int cpu
)
1866 for_each_populated_zone(zone
) {
1867 struct per_cpu_pageset
*p
= per_cpu_ptr(zone
->pageset
, cpu
);
1869 BUILD_BUG_ON(sizeof(p
->vm_stat_diff
[0]) != 1);
1871 BUILD_BUG_ON(sizeof(p
->vm_numa_stat_diff
[0]) != 2);
1875 * The fast way of checking if there are any vmstat diffs.
1877 if (memchr_inv(p
->vm_stat_diff
, 0, NR_VM_ZONE_STAT_ITEMS
*
1878 sizeof(p
->vm_stat_diff
[0])))
1881 if (memchr_inv(p
->vm_numa_stat_diff
, 0, NR_VM_NUMA_STAT_ITEMS
*
1882 sizeof(p
->vm_numa_stat_diff
[0])))
1890 * Switch off vmstat processing and then fold all the remaining differentials
1891 * until the diffs stay at zero. The function is used by NOHZ and can only be
1892 * invoked when tick processing is not active.
1894 void quiet_vmstat(void)
1896 if (system_state
!= SYSTEM_RUNNING
)
1899 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work
)))
1902 if (!need_update(smp_processor_id()))
1906 * Just refresh counters and do not care about the pending delayed
1907 * vmstat_update. It doesn't fire that often to matter and canceling
1908 * it would be too expensive from this path.
1909 * vmstat_shepherd will take care about that for us.
1911 refresh_cpu_vm_stats(false);
1915 * Shepherd worker thread that checks the
1916 * differentials of processors that have their worker
1917 * threads for vm statistics updates disabled because of
1920 static void vmstat_shepherd(struct work_struct
*w
);
1922 static DECLARE_DEFERRABLE_WORK(shepherd
, vmstat_shepherd
);
1924 static void vmstat_shepherd(struct work_struct
*w
)
1929 /* Check processors whose vmstat worker threads have been disabled */
1930 for_each_online_cpu(cpu
) {
1931 struct delayed_work
*dw
= &per_cpu(vmstat_work
, cpu
);
1933 if (!delayed_work_pending(dw
) && need_update(cpu
))
1934 queue_delayed_work_on(cpu
, mm_percpu_wq
, dw
, 0);
1938 schedule_delayed_work(&shepherd
,
1939 round_jiffies_relative(sysctl_stat_interval
));
1942 static void __init
start_shepherd_timer(void)
1946 for_each_possible_cpu(cpu
)
1947 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work
, cpu
),
1950 schedule_delayed_work(&shepherd
,
1951 round_jiffies_relative(sysctl_stat_interval
));
1954 static void __init
init_cpu_node_state(void)
1958 for_each_online_node(node
) {
1959 if (cpumask_weight(cpumask_of_node(node
)) > 0)
1960 node_set_state(node
, N_CPU
);
1964 static int vmstat_cpu_online(unsigned int cpu
)
1966 refresh_zone_stat_thresholds();
1967 node_set_state(cpu_to_node(cpu
), N_CPU
);
1971 static int vmstat_cpu_down_prep(unsigned int cpu
)
1973 cancel_delayed_work_sync(&per_cpu(vmstat_work
, cpu
));
1977 static int vmstat_cpu_dead(unsigned int cpu
)
1979 const struct cpumask
*node_cpus
;
1982 node
= cpu_to_node(cpu
);
1984 refresh_zone_stat_thresholds();
1985 node_cpus
= cpumask_of_node(node
);
1986 if (cpumask_weight(node_cpus
) > 0)
1989 node_clear_state(node
, N_CPU
);
1995 struct workqueue_struct
*mm_percpu_wq
;
1997 void __init
init_mm_internals(void)
1999 int ret __maybe_unused
;
2001 mm_percpu_wq
= alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM
, 0);
2004 ret
= cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD
, "mm/vmstat:dead",
2005 NULL
, vmstat_cpu_dead
);
2007 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2009 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
, "mm/vmstat:online",
2011 vmstat_cpu_down_prep
);
2013 pr_err("vmstat: failed to register 'online' hotplug state\n");
2016 init_cpu_node_state();
2019 start_shepherd_timer();
2021 #ifdef CONFIG_PROC_FS
2022 proc_create_seq("buddyinfo", 0444, NULL
, &fragmentation_op
);
2023 proc_create_seq("pagetypeinfo", 0400, NULL
, &pagetypeinfo_op
);
2024 proc_create_seq("vmstat", 0444, NULL
, &vmstat_op
);
2025 proc_create_seq("zoneinfo", 0444, NULL
, &zoneinfo_op
);
2029 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2032 * Return an index indicating how much of the available free memory is
2033 * unusable for an allocation of the requested size.
2035 static int unusable_free_index(unsigned int order
,
2036 struct contig_page_info
*info
)
2038 /* No free memory is interpreted as all free memory is unusable */
2039 if (info
->free_pages
== 0)
2043 * Index should be a value between 0 and 1. Return a value to 3
2046 * 0 => no fragmentation
2047 * 1 => high fragmentation
2049 return div_u64((info
->free_pages
- (info
->free_blocks_suitable
<< order
)) * 1000ULL, info
->free_pages
);
2053 static void unusable_show_print(struct seq_file
*m
,
2054 pg_data_t
*pgdat
, struct zone
*zone
)
2058 struct contig_page_info info
;
2060 seq_printf(m
, "Node %d, zone %8s ",
2063 for (order
= 0; order
< MAX_ORDER
; ++order
) {
2064 fill_contig_page_info(zone
, order
, &info
);
2065 index
= unusable_free_index(order
, &info
);
2066 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2073 * Display unusable free space index
2075 * The unusable free space index measures how much of the available free
2076 * memory cannot be used to satisfy an allocation of a given size and is a
2077 * value between 0 and 1. The higher the value, the more of free memory is
2078 * unusable and by implication, the worse the external fragmentation is. This
2079 * can be expressed as a percentage by multiplying by 100.
2081 static int unusable_show(struct seq_file
*m
, void *arg
)
2083 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2085 /* check memoryless node */
2086 if (!node_state(pgdat
->node_id
, N_MEMORY
))
2089 walk_zones_in_node(m
, pgdat
, true, false, unusable_show_print
);
2094 static const struct seq_operations unusable_sops
= {
2095 .start
= frag_start
,
2098 .show
= unusable_show
,
2101 DEFINE_SEQ_ATTRIBUTE(unusable
);
2103 static void extfrag_show_print(struct seq_file
*m
,
2104 pg_data_t
*pgdat
, struct zone
*zone
)
2109 /* Alloc on stack as interrupts are disabled for zone walk */
2110 struct contig_page_info info
;
2112 seq_printf(m
, "Node %d, zone %8s ",
2115 for (order
= 0; order
< MAX_ORDER
; ++order
) {
2116 fill_contig_page_info(zone
, order
, &info
);
2117 index
= __fragmentation_index(order
, &info
);
2118 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2125 * Display fragmentation index for orders that allocations would fail for
2127 static int extfrag_show(struct seq_file
*m
, void *arg
)
2129 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2131 walk_zones_in_node(m
, pgdat
, true, false, extfrag_show_print
);
2136 static const struct seq_operations extfrag_sops
= {
2137 .start
= frag_start
,
2140 .show
= extfrag_show
,
2143 DEFINE_SEQ_ATTRIBUTE(extfrag
);
2145 static int __init
extfrag_debug_init(void)
2147 struct dentry
*extfrag_debug_root
;
2149 extfrag_debug_root
= debugfs_create_dir("extfrag", NULL
);
2151 debugfs_create_file("unusable_index", 0444, extfrag_debug_root
, NULL
,
2154 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root
, NULL
,
2160 module_init(extfrag_debug_init
);