4 * Manages VM statistics
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Copyright (C) 2006 Silicon Graphics, Inc.,
9 * Christoph Lameter <christoph@lameter.com>
10 * Copyright (C) 2008-2014 Christoph Lameter
14 #include <linux/err.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/cpu.h>
18 #include <linux/cpumask.h>
19 #include <linux/vmstat.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/debugfs.h>
23 #include <linux/sched.h>
24 #include <linux/math64.h>
25 #include <linux/writeback.h>
26 #include <linux/compaction.h>
27 #include <linux/mm_inline.h>
28 #include <linux/page_ext.h>
29 #include <linux/page_owner.h>
33 #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
35 #ifdef CONFIG_VM_EVENT_COUNTERS
36 DEFINE_PER_CPU(struct vm_event_state
, vm_event_states
) = {{0}};
37 EXPORT_PER_CPU_SYMBOL(vm_event_states
);
39 static void sum_vm_events(unsigned long *ret
)
44 memset(ret
, 0, NR_VM_EVENT_ITEMS
* sizeof(unsigned long));
46 for_each_online_cpu(cpu
) {
47 struct vm_event_state
*this = &per_cpu(vm_event_states
, cpu
);
49 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
50 ret
[i
] += this->event
[i
];
55 * Accumulate the vm event counters across all CPUs.
56 * The result is unavoidably approximate - it can change
57 * during and after execution of this function.
59 void all_vm_events(unsigned long *ret
)
65 EXPORT_SYMBOL_GPL(all_vm_events
);
68 * Fold the foreign cpu events into our own.
70 * This is adding to the events on one processor
71 * but keeps the global counts constant.
73 void vm_events_fold_cpu(int cpu
)
75 struct vm_event_state
*fold_state
= &per_cpu(vm_event_states
, cpu
);
78 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
79 count_vm_events(i
, fold_state
->event
[i
]);
80 fold_state
->event
[i
] = 0;
84 #endif /* CONFIG_VM_EVENT_COUNTERS */
87 * Manage combined zone based / global counters
89 * vm_stat contains the global counters
91 atomic_long_t vm_zone_stat
[NR_VM_ZONE_STAT_ITEMS
] __cacheline_aligned_in_smp
;
92 atomic_long_t vm_numa_stat
[NR_VM_NUMA_STAT_ITEMS
] __cacheline_aligned_in_smp
;
93 atomic_long_t vm_node_stat
[NR_VM_NODE_STAT_ITEMS
] __cacheline_aligned_in_smp
;
94 EXPORT_SYMBOL(vm_zone_stat
);
95 EXPORT_SYMBOL(vm_numa_stat
);
96 EXPORT_SYMBOL(vm_node_stat
);
100 int calculate_pressure_threshold(struct zone
*zone
)
103 int watermark_distance
;
106 * As vmstats are not up to date, there is drift between the estimated
107 * and real values. For high thresholds and a high number of CPUs, it
108 * is possible for the min watermark to be breached while the estimated
109 * value looks fine. The pressure threshold is a reduced value such
110 * that even the maximum amount of drift will not accidentally breach
113 watermark_distance
= low_wmark_pages(zone
) - min_wmark_pages(zone
);
114 threshold
= max(1, (int)(watermark_distance
/ num_online_cpus()));
117 * Maximum threshold is 125
119 threshold
= min(125, threshold
);
124 int calculate_normal_threshold(struct zone
*zone
)
127 int mem
; /* memory in 128 MB units */
130 * The threshold scales with the number of processors and the amount
131 * of memory per zone. More memory means that we can defer updates for
132 * longer, more processors could lead to more contention.
133 * fls() is used to have a cheap way of logarithmic scaling.
135 * Some sample thresholds:
137 * Threshold Processors (fls) Zonesize fls(mem+1)
138 * ------------------------------------------------------------------
155 * 125 1024 10 8-16 GB 8
156 * 125 1024 10 16-32 GB 9
159 mem
= zone
->managed_pages
>> (27 - PAGE_SHIFT
);
161 threshold
= 2 * fls(num_online_cpus()) * (1 + fls(mem
));
164 * Maximum threshold is 125
166 threshold
= min(125, threshold
);
172 * Refresh the thresholds for each zone.
174 void refresh_zone_stat_thresholds(void)
176 struct pglist_data
*pgdat
;
181 /* Zero current pgdat thresholds */
182 for_each_online_pgdat(pgdat
) {
183 for_each_online_cpu(cpu
) {
184 per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
)->stat_threshold
= 0;
188 for_each_populated_zone(zone
) {
189 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
190 unsigned long max_drift
, tolerate_drift
;
192 threshold
= calculate_normal_threshold(zone
);
194 for_each_online_cpu(cpu
) {
197 per_cpu_ptr(zone
->pageset
, cpu
)->stat_threshold
200 /* Base nodestat threshold on the largest populated zone. */
201 pgdat_threshold
= per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
)->stat_threshold
;
202 per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
)->stat_threshold
203 = max(threshold
, pgdat_threshold
);
207 * Only set percpu_drift_mark if there is a danger that
208 * NR_FREE_PAGES reports the low watermark is ok when in fact
209 * the min watermark could be breached by an allocation
211 tolerate_drift
= low_wmark_pages(zone
) - min_wmark_pages(zone
);
212 max_drift
= num_online_cpus() * threshold
;
213 if (max_drift
> tolerate_drift
)
214 zone
->percpu_drift_mark
= high_wmark_pages(zone
) +
219 void set_pgdat_percpu_threshold(pg_data_t
*pgdat
,
220 int (*calculate_pressure
)(struct zone
*))
227 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
228 zone
= &pgdat
->node_zones
[i
];
229 if (!zone
->percpu_drift_mark
)
232 threshold
= (*calculate_pressure
)(zone
);
233 for_each_online_cpu(cpu
)
234 per_cpu_ptr(zone
->pageset
, cpu
)->stat_threshold
240 * For use when we know that interrupts are disabled,
241 * or when we know that preemption is disabled and that
242 * particular counter cannot be updated from interrupt context.
244 void __mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
247 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
248 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
252 x
= delta
+ __this_cpu_read(*p
);
254 t
= __this_cpu_read(pcp
->stat_threshold
);
256 if (unlikely(x
> t
|| x
< -t
)) {
257 zone_page_state_add(x
, zone
, item
);
260 __this_cpu_write(*p
, x
);
262 EXPORT_SYMBOL(__mod_zone_page_state
);
264 void __mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
267 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
268 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
272 x
= delta
+ __this_cpu_read(*p
);
274 t
= __this_cpu_read(pcp
->stat_threshold
);
276 if (unlikely(x
> t
|| x
< -t
)) {
277 node_page_state_add(x
, pgdat
, item
);
280 __this_cpu_write(*p
, x
);
282 EXPORT_SYMBOL(__mod_node_page_state
);
285 * Optimized increment and decrement functions.
287 * These are only for a single page and therefore can take a struct page *
288 * argument instead of struct zone *. This allows the inclusion of the code
289 * generated for page_zone(page) into the optimized functions.
291 * No overflow check is necessary and therefore the differential can be
292 * incremented or decremented in place which may allow the compilers to
293 * generate better code.
294 * The increment or decrement is known and therefore one boundary check can
297 * NOTE: These functions are very performance sensitive. Change only
300 * Some processors have inc/dec instructions that are atomic vs an interrupt.
301 * However, the code must first determine the differential location in a zone
302 * based on the processor number and then inc/dec the counter. There is no
303 * guarantee without disabling preemption that the processor will not change
304 * in between and therefore the atomicity vs. interrupt cannot be exploited
305 * in a useful way here.
307 void __inc_zone_state(struct zone
*zone
, enum zone_stat_item item
)
309 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
310 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
313 v
= __this_cpu_inc_return(*p
);
314 t
= __this_cpu_read(pcp
->stat_threshold
);
315 if (unlikely(v
> t
)) {
316 s8 overstep
= t
>> 1;
318 zone_page_state_add(v
+ overstep
, zone
, item
);
319 __this_cpu_write(*p
, -overstep
);
323 void __inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
325 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
326 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
329 v
= __this_cpu_inc_return(*p
);
330 t
= __this_cpu_read(pcp
->stat_threshold
);
331 if (unlikely(v
> t
)) {
332 s8 overstep
= t
>> 1;
334 node_page_state_add(v
+ overstep
, pgdat
, item
);
335 __this_cpu_write(*p
, -overstep
);
339 void __inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
341 __inc_zone_state(page_zone(page
), item
);
343 EXPORT_SYMBOL(__inc_zone_page_state
);
345 void __inc_node_page_state(struct page
*page
, enum node_stat_item item
)
347 __inc_node_state(page_pgdat(page
), item
);
349 EXPORT_SYMBOL(__inc_node_page_state
);
351 void __dec_zone_state(struct zone
*zone
, enum zone_stat_item item
)
353 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
354 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
357 v
= __this_cpu_dec_return(*p
);
358 t
= __this_cpu_read(pcp
->stat_threshold
);
359 if (unlikely(v
< - t
)) {
360 s8 overstep
= t
>> 1;
362 zone_page_state_add(v
- overstep
, zone
, item
);
363 __this_cpu_write(*p
, overstep
);
367 void __dec_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
369 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
370 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
373 v
= __this_cpu_dec_return(*p
);
374 t
= __this_cpu_read(pcp
->stat_threshold
);
375 if (unlikely(v
< - t
)) {
376 s8 overstep
= t
>> 1;
378 node_page_state_add(v
- overstep
, pgdat
, item
);
379 __this_cpu_write(*p
, overstep
);
383 void __dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
385 __dec_zone_state(page_zone(page
), item
);
387 EXPORT_SYMBOL(__dec_zone_page_state
);
389 void __dec_node_page_state(struct page
*page
, enum node_stat_item item
)
391 __dec_node_state(page_pgdat(page
), item
);
393 EXPORT_SYMBOL(__dec_node_page_state
);
395 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
397 * If we have cmpxchg_local support then we do not need to incur the overhead
398 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
400 * mod_state() modifies the zone counter state through atomic per cpu
403 * Overstep mode specifies how overstep should handled:
405 * 1 Overstepping half of threshold
406 * -1 Overstepping minus half of threshold
408 static inline void mod_zone_state(struct zone
*zone
,
409 enum zone_stat_item item
, long delta
, int overstep_mode
)
411 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
412 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
416 z
= 0; /* overflow to zone counters */
419 * The fetching of the stat_threshold is racy. We may apply
420 * a counter threshold to the wrong the cpu if we get
421 * rescheduled while executing here. However, the next
422 * counter update will apply the threshold again and
423 * therefore bring the counter under the threshold again.
425 * Most of the time the thresholds are the same anyways
426 * for all cpus in a zone.
428 t
= this_cpu_read(pcp
->stat_threshold
);
430 o
= this_cpu_read(*p
);
433 if (n
> t
|| n
< -t
) {
434 int os
= overstep_mode
* (t
>> 1) ;
436 /* Overflow must be added to zone counters */
440 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
443 zone_page_state_add(z
, zone
, item
);
446 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
449 mod_zone_state(zone
, item
, delta
, 0);
451 EXPORT_SYMBOL(mod_zone_page_state
);
453 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
455 mod_zone_state(page_zone(page
), item
, 1, 1);
457 EXPORT_SYMBOL(inc_zone_page_state
);
459 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
461 mod_zone_state(page_zone(page
), item
, -1, -1);
463 EXPORT_SYMBOL(dec_zone_page_state
);
465 static inline void mod_node_state(struct pglist_data
*pgdat
,
466 enum node_stat_item item
, int delta
, int overstep_mode
)
468 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
469 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
473 z
= 0; /* overflow to node counters */
476 * The fetching of the stat_threshold is racy. We may apply
477 * a counter threshold to the wrong the cpu if we get
478 * rescheduled while executing here. However, the next
479 * counter update will apply the threshold again and
480 * therefore bring the counter under the threshold again.
482 * Most of the time the thresholds are the same anyways
483 * for all cpus in a node.
485 t
= this_cpu_read(pcp
->stat_threshold
);
487 o
= this_cpu_read(*p
);
490 if (n
> t
|| n
< -t
) {
491 int os
= overstep_mode
* (t
>> 1) ;
493 /* Overflow must be added to node counters */
497 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
500 node_page_state_add(z
, pgdat
, item
);
503 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
506 mod_node_state(pgdat
, item
, delta
, 0);
508 EXPORT_SYMBOL(mod_node_page_state
);
510 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
512 mod_node_state(pgdat
, item
, 1, 1);
515 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
517 mod_node_state(page_pgdat(page
), item
, 1, 1);
519 EXPORT_SYMBOL(inc_node_page_state
);
521 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
523 mod_node_state(page_pgdat(page
), item
, -1, -1);
525 EXPORT_SYMBOL(dec_node_page_state
);
528 * Use interrupt disable to serialize counter updates
530 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
535 local_irq_save(flags
);
536 __mod_zone_page_state(zone
, item
, delta
);
537 local_irq_restore(flags
);
539 EXPORT_SYMBOL(mod_zone_page_state
);
541 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
546 zone
= page_zone(page
);
547 local_irq_save(flags
);
548 __inc_zone_state(zone
, item
);
549 local_irq_restore(flags
);
551 EXPORT_SYMBOL(inc_zone_page_state
);
553 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
557 local_irq_save(flags
);
558 __dec_zone_page_state(page
, item
);
559 local_irq_restore(flags
);
561 EXPORT_SYMBOL(dec_zone_page_state
);
563 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
567 local_irq_save(flags
);
568 __inc_node_state(pgdat
, item
);
569 local_irq_restore(flags
);
571 EXPORT_SYMBOL(inc_node_state
);
573 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
578 local_irq_save(flags
);
579 __mod_node_page_state(pgdat
, item
, delta
);
580 local_irq_restore(flags
);
582 EXPORT_SYMBOL(mod_node_page_state
);
584 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
587 struct pglist_data
*pgdat
;
589 pgdat
= page_pgdat(page
);
590 local_irq_save(flags
);
591 __inc_node_state(pgdat
, item
);
592 local_irq_restore(flags
);
594 EXPORT_SYMBOL(inc_node_page_state
);
596 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
600 local_irq_save(flags
);
601 __dec_node_page_state(page
, item
);
602 local_irq_restore(flags
);
604 EXPORT_SYMBOL(dec_node_page_state
);
608 * Fold a differential into the global counters.
609 * Returns the number of counters updated.
612 static int fold_diff(int *zone_diff
, int *numa_diff
, int *node_diff
)
617 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
619 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
623 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
625 atomic_long_add(numa_diff
[i
], &vm_numa_stat
[i
]);
629 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
631 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
637 static int fold_diff(int *zone_diff
, int *node_diff
)
642 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
644 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
648 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
650 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
655 #endif /* CONFIG_NUMA */
658 * Update the zone counters for the current cpu.
660 * Note that refresh_cpu_vm_stats strives to only access
661 * node local memory. The per cpu pagesets on remote zones are placed
662 * in the memory local to the processor using that pageset. So the
663 * loop over all zones will access a series of cachelines local to
666 * The call to zone_page_state_add updates the cachelines with the
667 * statistics in the remote zone struct as well as the global cachelines
668 * with the global counters. These could cause remote node cache line
669 * bouncing and will have to be only done when necessary.
671 * The function returns the number of global counters updated.
673 static int refresh_cpu_vm_stats(bool do_pagesets
)
675 struct pglist_data
*pgdat
;
678 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
680 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
682 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
685 for_each_populated_zone(zone
) {
686 struct per_cpu_pageset __percpu
*p
= zone
->pageset
;
688 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
691 v
= this_cpu_xchg(p
->vm_stat_diff
[i
], 0);
694 atomic_long_add(v
, &zone
->vm_stat
[i
]);
695 global_zone_diff
[i
] += v
;
697 /* 3 seconds idle till flush */
698 __this_cpu_write(p
->expire
, 3);
703 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
706 v
= this_cpu_xchg(p
->vm_numa_stat_diff
[i
], 0);
709 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
710 global_numa_diff
[i
] += v
;
711 __this_cpu_write(p
->expire
, 3);
718 * Deal with draining the remote pageset of this
721 * Check if there are pages remaining in this pageset
722 * if not then there is nothing to expire.
724 if (!__this_cpu_read(p
->expire
) ||
725 !__this_cpu_read(p
->pcp
.count
))
729 * We never drain zones local to this processor.
731 if (zone_to_nid(zone
) == numa_node_id()) {
732 __this_cpu_write(p
->expire
, 0);
736 if (__this_cpu_dec_return(p
->expire
))
739 if (__this_cpu_read(p
->pcp
.count
)) {
740 drain_zone_pages(zone
, this_cpu_ptr(&p
->pcp
));
747 for_each_online_pgdat(pgdat
) {
748 struct per_cpu_nodestat __percpu
*p
= pgdat
->per_cpu_nodestats
;
750 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
753 v
= this_cpu_xchg(p
->vm_node_stat_diff
[i
], 0);
755 atomic_long_add(v
, &pgdat
->vm_stat
[i
]);
756 global_node_diff
[i
] += v
;
762 changes
+= fold_diff(global_zone_diff
, global_numa_diff
,
765 changes
+= fold_diff(global_zone_diff
, global_node_diff
);
771 * Fold the data for an offline cpu into the global array.
772 * There cannot be any access by the offline cpu and therefore
773 * synchronization is simplified.
775 void cpu_vm_stats_fold(int cpu
)
777 struct pglist_data
*pgdat
;
780 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
782 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
784 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
786 for_each_populated_zone(zone
) {
787 struct per_cpu_pageset
*p
;
789 p
= per_cpu_ptr(zone
->pageset
, cpu
);
791 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
792 if (p
->vm_stat_diff
[i
]) {
795 v
= p
->vm_stat_diff
[i
];
796 p
->vm_stat_diff
[i
] = 0;
797 atomic_long_add(v
, &zone
->vm_stat
[i
]);
798 global_zone_diff
[i
] += v
;
802 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
803 if (p
->vm_numa_stat_diff
[i
]) {
806 v
= p
->vm_numa_stat_diff
[i
];
807 p
->vm_numa_stat_diff
[i
] = 0;
808 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
809 global_numa_diff
[i
] += v
;
814 for_each_online_pgdat(pgdat
) {
815 struct per_cpu_nodestat
*p
;
817 p
= per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
);
819 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
820 if (p
->vm_node_stat_diff
[i
]) {
823 v
= p
->vm_node_stat_diff
[i
];
824 p
->vm_node_stat_diff
[i
] = 0;
825 atomic_long_add(v
, &pgdat
->vm_stat
[i
]);
826 global_node_diff
[i
] += v
;
831 fold_diff(global_zone_diff
, global_numa_diff
, global_node_diff
);
833 fold_diff(global_zone_diff
, global_node_diff
);
838 * this is only called if !populated_zone(zone), which implies no other users of
839 * pset->vm_stat_diff[] exsist.
841 void drain_zonestat(struct zone
*zone
, struct per_cpu_pageset
*pset
)
845 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
846 if (pset
->vm_stat_diff
[i
]) {
847 int v
= pset
->vm_stat_diff
[i
];
848 pset
->vm_stat_diff
[i
] = 0;
849 atomic_long_add(v
, &zone
->vm_stat
[i
]);
850 atomic_long_add(v
, &vm_zone_stat
[i
]);
854 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
855 if (pset
->vm_numa_stat_diff
[i
]) {
856 int v
= pset
->vm_numa_stat_diff
[i
];
858 pset
->vm_numa_stat_diff
[i
] = 0;
859 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
860 atomic_long_add(v
, &vm_numa_stat
[i
]);
867 void __inc_numa_state(struct zone
*zone
,
868 enum numa_stat_item item
)
870 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
871 u16 __percpu
*p
= pcp
->vm_numa_stat_diff
+ item
;
874 v
= __this_cpu_inc_return(*p
);
876 if (unlikely(v
> NUMA_STATS_THRESHOLD
)) {
877 zone_numa_state_add(v
, zone
, item
);
878 __this_cpu_write(*p
, 0);
883 * Determine the per node value of a stat item. This function
884 * is called frequently in a NUMA machine, so try to be as
885 * frugal as possible.
887 unsigned long sum_zone_node_page_state(int node
,
888 enum zone_stat_item item
)
890 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
892 unsigned long count
= 0;
894 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
895 count
+= zone_page_state(zones
+ i
, item
);
901 * Determine the per node value of a numa stat item. To avoid deviation,
902 * the per cpu stat number in vm_numa_stat_diff[] is also included.
904 unsigned long sum_zone_numa_state(int node
,
905 enum numa_stat_item item
)
907 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
909 unsigned long count
= 0;
911 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
912 count
+= zone_numa_state_snapshot(zones
+ i
, item
);
918 * Determine the per node value of a stat item.
920 unsigned long node_page_state(struct pglist_data
*pgdat
,
921 enum node_stat_item item
)
923 long x
= atomic_long_read(&pgdat
->vm_stat
[item
]);
932 #ifdef CONFIG_COMPACTION
934 struct contig_page_info
{
935 unsigned long free_pages
;
936 unsigned long free_blocks_total
;
937 unsigned long free_blocks_suitable
;
941 * Calculate the number of free pages in a zone, how many contiguous
942 * pages are free and how many are large enough to satisfy an allocation of
943 * the target size. Note that this function makes no attempt to estimate
944 * how many suitable free blocks there *might* be if MOVABLE pages were
945 * migrated. Calculating that is possible, but expensive and can be
946 * figured out from userspace
948 static void fill_contig_page_info(struct zone
*zone
,
949 unsigned int suitable_order
,
950 struct contig_page_info
*info
)
954 info
->free_pages
= 0;
955 info
->free_blocks_total
= 0;
956 info
->free_blocks_suitable
= 0;
958 for (order
= 0; order
< MAX_ORDER
; order
++) {
959 unsigned long blocks
;
961 /* Count number of free blocks */
962 blocks
= zone
->free_area
[order
].nr_free
;
963 info
->free_blocks_total
+= blocks
;
965 /* Count free base pages */
966 info
->free_pages
+= blocks
<< order
;
968 /* Count the suitable free blocks */
969 if (order
>= suitable_order
)
970 info
->free_blocks_suitable
+= blocks
<<
971 (order
- suitable_order
);
976 * A fragmentation index only makes sense if an allocation of a requested
977 * size would fail. If that is true, the fragmentation index indicates
978 * whether external fragmentation or a lack of memory was the problem.
979 * The value can be used to determine if page reclaim or compaction
982 static int __fragmentation_index(unsigned int order
, struct contig_page_info
*info
)
984 unsigned long requested
= 1UL << order
;
986 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
989 if (!info
->free_blocks_total
)
992 /* Fragmentation index only makes sense when a request would fail */
993 if (info
->free_blocks_suitable
)
997 * Index is between 0 and 1 so return within 3 decimal places
999 * 0 => allocation would fail due to lack of memory
1000 * 1 => allocation would fail due to fragmentation
1002 return 1000 - div_u64( (1000+(div_u64(info
->free_pages
* 1000ULL, requested
))), info
->free_blocks_total
);
1005 /* Same as __fragmentation index but allocs contig_page_info on stack */
1006 int fragmentation_index(struct zone
*zone
, unsigned int order
)
1008 struct contig_page_info info
;
1010 fill_contig_page_info(zone
, order
, &info
);
1011 return __fragmentation_index(order
, &info
);
1015 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
1016 #ifdef CONFIG_ZONE_DMA
1017 #define TEXT_FOR_DMA(xx) xx "_dma",
1019 #define TEXT_FOR_DMA(xx)
1022 #ifdef CONFIG_ZONE_DMA32
1023 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1025 #define TEXT_FOR_DMA32(xx)
1028 #ifdef CONFIG_HIGHMEM
1029 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1031 #define TEXT_FOR_HIGHMEM(xx)
1034 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1035 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1037 const char * const vmstat_text
[] = {
1038 /* enum zone_stat_item countes */
1040 "nr_zone_inactive_anon",
1041 "nr_zone_active_anon",
1042 "nr_zone_inactive_file",
1043 "nr_zone_active_file",
1044 "nr_zone_unevictable",
1045 "nr_zone_write_pending",
1047 "nr_page_table_pages",
1050 #if IS_ENABLED(CONFIG_ZSMALLOC)
1055 /* enum numa_stat_item counters */
1065 /* Node-based counters */
1071 "nr_slab_reclaimable",
1072 "nr_slab_unreclaimable",
1075 "workingset_refault",
1076 "workingset_activate",
1077 "workingset_nodereclaim",
1083 "nr_writeback_temp",
1085 "nr_shmem_hugepages",
1086 "nr_shmem_pmdmapped",
1087 "nr_anon_transparent_hugepages",
1090 "nr_vmscan_immediate_reclaim",
1093 "", /* nr_indirectly_reclaimable */
1095 /* enum writeback_stat_item counters */
1096 "nr_dirty_threshold",
1097 "nr_dirty_background_threshold",
1099 #ifdef CONFIG_VM_EVENT_COUNTERS
1100 /* enum vm_event_item counters */
1106 TEXTS_FOR_ZONES("pgalloc")
1107 TEXTS_FOR_ZONES("allocstall")
1108 TEXTS_FOR_ZONES("pgskip")
1124 "pgscan_direct_throttle",
1127 "zone_reclaim_failed",
1131 "kswapd_inodesteal",
1132 "kswapd_low_wmark_hit_quickly",
1133 "kswapd_high_wmark_hit_quickly",
1142 #ifdef CONFIG_NUMA_BALANCING
1144 "numa_huge_pte_updates",
1146 "numa_hint_faults_local",
1147 "numa_pages_migrated",
1149 #ifdef CONFIG_MIGRATION
1150 "pgmigrate_success",
1153 #ifdef CONFIG_COMPACTION
1154 "compact_migrate_scanned",
1155 "compact_free_scanned",
1160 "compact_daemon_wake",
1161 "compact_daemon_migrate_scanned",
1162 "compact_daemon_free_scanned",
1165 #ifdef CONFIG_HUGETLB_PAGE
1166 "htlb_buddy_alloc_success",
1167 "htlb_buddy_alloc_fail",
1169 "unevictable_pgs_culled",
1170 "unevictable_pgs_scanned",
1171 "unevictable_pgs_rescued",
1172 "unevictable_pgs_mlocked",
1173 "unevictable_pgs_munlocked",
1174 "unevictable_pgs_cleared",
1175 "unevictable_pgs_stranded",
1177 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1179 "thp_fault_fallback",
1180 "thp_collapse_alloc",
1181 "thp_collapse_alloc_failed",
1185 "thp_split_page_failed",
1186 "thp_deferred_split_page",
1188 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1191 "thp_zero_page_alloc",
1192 "thp_zero_page_alloc_failed",
1194 "thp_swpout_fallback",
1196 #ifdef CONFIG_MEMORY_BALLOON
1199 #ifdef CONFIG_BALLOON_COMPACTION
1202 #endif /* CONFIG_MEMORY_BALLOON */
1203 #ifdef CONFIG_DEBUG_TLBFLUSH
1204 "nr_tlb_remote_flush",
1205 "nr_tlb_remote_flush_received",
1206 "nr_tlb_local_flush_all",
1207 "nr_tlb_local_flush_one",
1208 #endif /* CONFIG_DEBUG_TLBFLUSH */
1210 #ifdef CONFIG_DEBUG_VM_VMACACHE
1211 "vmacache_find_calls",
1212 "vmacache_find_hits",
1218 #endif /* CONFIG_VM_EVENTS_COUNTERS */
1220 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
1222 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1223 defined(CONFIG_PROC_FS)
1224 static void *frag_start(struct seq_file
*m
, loff_t
*pos
)
1229 for (pgdat
= first_online_pgdat();
1231 pgdat
= next_online_pgdat(pgdat
))
1237 static void *frag_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1239 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1242 return next_online_pgdat(pgdat
);
1245 static void frag_stop(struct seq_file
*m
, void *arg
)
1250 * Walk zones in a node and print using a callback.
1251 * If @assert_populated is true, only use callback for zones that are populated.
1253 static void walk_zones_in_node(struct seq_file
*m
, pg_data_t
*pgdat
,
1254 bool assert_populated
, bool nolock
,
1255 void (*print
)(struct seq_file
*m
, pg_data_t
*, struct zone
*))
1258 struct zone
*node_zones
= pgdat
->node_zones
;
1259 unsigned long flags
;
1261 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
1262 if (assert_populated
&& !populated_zone(zone
))
1266 spin_lock_irqsave(&zone
->lock
, flags
);
1267 print(m
, pgdat
, zone
);
1269 spin_unlock_irqrestore(&zone
->lock
, flags
);
1274 #ifdef CONFIG_PROC_FS
1275 static void frag_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1280 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1281 for (order
= 0; order
< MAX_ORDER
; ++order
)
1282 seq_printf(m
, "%6lu ", zone
->free_area
[order
].nr_free
);
1287 * This walks the free areas for each zone.
1289 static int frag_show(struct seq_file
*m
, void *arg
)
1291 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1292 walk_zones_in_node(m
, pgdat
, true, false, frag_show_print
);
1296 static void pagetypeinfo_showfree_print(struct seq_file
*m
,
1297 pg_data_t
*pgdat
, struct zone
*zone
)
1301 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++) {
1302 seq_printf(m
, "Node %4d, zone %8s, type %12s ",
1305 migratetype_names
[mtype
]);
1306 for (order
= 0; order
< MAX_ORDER
; ++order
) {
1307 unsigned long freecount
= 0;
1308 struct free_area
*area
;
1309 struct list_head
*curr
;
1311 area
= &(zone
->free_area
[order
]);
1313 list_for_each(curr
, &area
->free_list
[mtype
])
1315 seq_printf(m
, "%6lu ", freecount
);
1321 /* Print out the free pages at each order for each migatetype */
1322 static int pagetypeinfo_showfree(struct seq_file
*m
, void *arg
)
1325 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1328 seq_printf(m
, "%-43s ", "Free pages count per migrate type at order");
1329 for (order
= 0; order
< MAX_ORDER
; ++order
)
1330 seq_printf(m
, "%6d ", order
);
1333 walk_zones_in_node(m
, pgdat
, true, false, pagetypeinfo_showfree_print
);
1338 static void pagetypeinfo_showblockcount_print(struct seq_file
*m
,
1339 pg_data_t
*pgdat
, struct zone
*zone
)
1343 unsigned long start_pfn
= zone
->zone_start_pfn
;
1344 unsigned long end_pfn
= zone_end_pfn(zone
);
1345 unsigned long count
[MIGRATE_TYPES
] = { 0, };
1347 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
1350 page
= pfn_to_online_page(pfn
);
1354 /* Watch for unexpected holes punched in the memmap */
1355 if (!memmap_valid_within(pfn
, page
, zone
))
1358 if (page_zone(page
) != zone
)
1361 mtype
= get_pageblock_migratetype(page
);
1363 if (mtype
< MIGRATE_TYPES
)
1368 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1369 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1370 seq_printf(m
, "%12lu ", count
[mtype
]);
1374 /* Print out the number of pageblocks for each migratetype */
1375 static int pagetypeinfo_showblockcount(struct seq_file
*m
, void *arg
)
1378 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1380 seq_printf(m
, "\n%-23s", "Number of blocks type ");
1381 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1382 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1384 walk_zones_in_node(m
, pgdat
, true, false,
1385 pagetypeinfo_showblockcount_print
);
1391 * Print out the number of pageblocks for each migratetype that contain pages
1392 * of other types. This gives an indication of how well fallbacks are being
1393 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1394 * to determine what is going on
1396 static void pagetypeinfo_showmixedcount(struct seq_file
*m
, pg_data_t
*pgdat
)
1398 #ifdef CONFIG_PAGE_OWNER
1401 if (!static_branch_unlikely(&page_owner_inited
))
1404 drain_all_pages(NULL
);
1406 seq_printf(m
, "\n%-23s", "Number of mixed blocks ");
1407 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1408 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1411 walk_zones_in_node(m
, pgdat
, true, true,
1412 pagetypeinfo_showmixedcount_print
);
1413 #endif /* CONFIG_PAGE_OWNER */
1417 * This prints out statistics in relation to grouping pages by mobility.
1418 * It is expensive to collect so do not constantly read the file.
1420 static int pagetypeinfo_show(struct seq_file
*m
, void *arg
)
1422 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1424 /* check memoryless node */
1425 if (!node_state(pgdat
->node_id
, N_MEMORY
))
1428 seq_printf(m
, "Page block order: %d\n", pageblock_order
);
1429 seq_printf(m
, "Pages per block: %lu\n", pageblock_nr_pages
);
1431 pagetypeinfo_showfree(m
, pgdat
);
1432 pagetypeinfo_showblockcount(m
, pgdat
);
1433 pagetypeinfo_showmixedcount(m
, pgdat
);
1438 static const struct seq_operations fragmentation_op
= {
1439 .start
= frag_start
,
1445 static int fragmentation_open(struct inode
*inode
, struct file
*file
)
1447 return seq_open(file
, &fragmentation_op
);
1450 static const struct file_operations buddyinfo_file_operations
= {
1451 .open
= fragmentation_open
,
1453 .llseek
= seq_lseek
,
1454 .release
= seq_release
,
1457 static const struct seq_operations pagetypeinfo_op
= {
1458 .start
= frag_start
,
1461 .show
= pagetypeinfo_show
,
1464 static int pagetypeinfo_open(struct inode
*inode
, struct file
*file
)
1466 return seq_open(file
, &pagetypeinfo_op
);
1469 static const struct file_operations pagetypeinfo_file_operations
= {
1470 .open
= pagetypeinfo_open
,
1472 .llseek
= seq_lseek
,
1473 .release
= seq_release
,
1476 static bool is_zone_first_populated(pg_data_t
*pgdat
, struct zone
*zone
)
1480 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1481 struct zone
*compare
= &pgdat
->node_zones
[zid
];
1483 if (populated_zone(compare
))
1484 return zone
== compare
;
1490 static void zoneinfo_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1494 seq_printf(m
, "Node %d, zone %8s", pgdat
->node_id
, zone
->name
);
1495 if (is_zone_first_populated(pgdat
, zone
)) {
1496 seq_printf(m
, "\n per-node stats");
1497 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
1498 /* Skip hidden vmstat items. */
1499 if (*vmstat_text
[i
+ NR_VM_ZONE_STAT_ITEMS
+
1500 NR_VM_NUMA_STAT_ITEMS
] == '\0')
1502 seq_printf(m
, "\n %-12s %lu",
1503 vmstat_text
[i
+ NR_VM_ZONE_STAT_ITEMS
+
1504 NR_VM_NUMA_STAT_ITEMS
],
1505 node_page_state(pgdat
, i
));
1516 zone_page_state(zone
, NR_FREE_PAGES
),
1517 min_wmark_pages(zone
),
1518 low_wmark_pages(zone
),
1519 high_wmark_pages(zone
),
1520 zone
->spanned_pages
,
1521 zone
->present_pages
,
1522 zone
->managed_pages
);
1525 "\n protection: (%ld",
1526 zone
->lowmem_reserve
[0]);
1527 for (i
= 1; i
< ARRAY_SIZE(zone
->lowmem_reserve
); i
++)
1528 seq_printf(m
, ", %ld", zone
->lowmem_reserve
[i
]);
1531 /* If unpopulated, no other information is useful */
1532 if (!populated_zone(zone
)) {
1537 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1538 seq_printf(m
, "\n %-12s %lu", vmstat_text
[i
],
1539 zone_page_state(zone
, i
));
1542 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1543 seq_printf(m
, "\n %-12s %lu",
1544 vmstat_text
[i
+ NR_VM_ZONE_STAT_ITEMS
],
1545 zone_numa_state_snapshot(zone
, i
));
1548 seq_printf(m
, "\n pagesets");
1549 for_each_online_cpu(i
) {
1550 struct per_cpu_pageset
*pageset
;
1552 pageset
= per_cpu_ptr(zone
->pageset
, i
);
1561 pageset
->pcp
.batch
);
1563 seq_printf(m
, "\n vm stats threshold: %d",
1564 pageset
->stat_threshold
);
1568 "\n node_unreclaimable: %u"
1570 "\n node_inactive_ratio: %u",
1571 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
,
1572 zone
->zone_start_pfn
,
1573 zone
->zone_pgdat
->inactive_ratio
);
1578 * Output information about zones in @pgdat. All zones are printed regardless
1579 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1580 * set of all zones and userspace would not be aware of such zones if they are
1581 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1583 static int zoneinfo_show(struct seq_file
*m
, void *arg
)
1585 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1586 walk_zones_in_node(m
, pgdat
, false, false, zoneinfo_show_print
);
1590 static const struct seq_operations zoneinfo_op
= {
1591 .start
= frag_start
, /* iterate over all zones. The same as in
1595 .show
= zoneinfo_show
,
1598 static int zoneinfo_open(struct inode
*inode
, struct file
*file
)
1600 return seq_open(file
, &zoneinfo_op
);
1603 static const struct file_operations zoneinfo_file_operations
= {
1604 .open
= zoneinfo_open
,
1606 .llseek
= seq_lseek
,
1607 .release
= seq_release
,
1610 enum writeback_stat_item
{
1612 NR_DIRTY_BG_THRESHOLD
,
1613 NR_VM_WRITEBACK_STAT_ITEMS
,
1616 static void *vmstat_start(struct seq_file
*m
, loff_t
*pos
)
1619 int i
, stat_items_size
;
1621 if (*pos
>= ARRAY_SIZE(vmstat_text
))
1623 stat_items_size
= NR_VM_ZONE_STAT_ITEMS
* sizeof(unsigned long) +
1624 NR_VM_NUMA_STAT_ITEMS
* sizeof(unsigned long) +
1625 NR_VM_NODE_STAT_ITEMS
* sizeof(unsigned long) +
1626 NR_VM_WRITEBACK_STAT_ITEMS
* sizeof(unsigned long);
1628 #ifdef CONFIG_VM_EVENT_COUNTERS
1629 stat_items_size
+= sizeof(struct vm_event_state
);
1632 v
= kmalloc(stat_items_size
, GFP_KERNEL
);
1635 return ERR_PTR(-ENOMEM
);
1636 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1637 v
[i
] = global_zone_page_state(i
);
1638 v
+= NR_VM_ZONE_STAT_ITEMS
;
1641 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1642 v
[i
] = global_numa_state(i
);
1643 v
+= NR_VM_NUMA_STAT_ITEMS
;
1646 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
1647 v
[i
] = global_node_page_state(i
);
1648 v
+= NR_VM_NODE_STAT_ITEMS
;
1650 global_dirty_limits(v
+ NR_DIRTY_BG_THRESHOLD
,
1651 v
+ NR_DIRTY_THRESHOLD
);
1652 v
+= NR_VM_WRITEBACK_STAT_ITEMS
;
1654 #ifdef CONFIG_VM_EVENT_COUNTERS
1656 v
[PGPGIN
] /= 2; /* sectors -> kbytes */
1659 return (unsigned long *)m
->private + *pos
;
1662 static void *vmstat_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1665 if (*pos
>= ARRAY_SIZE(vmstat_text
))
1667 return (unsigned long *)m
->private + *pos
;
1670 static int vmstat_show(struct seq_file
*m
, void *arg
)
1672 unsigned long *l
= arg
;
1673 unsigned long off
= l
- (unsigned long *)m
->private;
1675 /* Skip hidden vmstat items. */
1676 if (*vmstat_text
[off
] == '\0')
1679 seq_puts(m
, vmstat_text
[off
]);
1680 seq_put_decimal_ull(m
, " ", *l
);
1685 static void vmstat_stop(struct seq_file
*m
, void *arg
)
1691 static const struct seq_operations vmstat_op
= {
1692 .start
= vmstat_start
,
1693 .next
= vmstat_next
,
1694 .stop
= vmstat_stop
,
1695 .show
= vmstat_show
,
1698 static int vmstat_open(struct inode
*inode
, struct file
*file
)
1700 return seq_open(file
, &vmstat_op
);
1703 static const struct file_operations vmstat_file_operations
= {
1704 .open
= vmstat_open
,
1706 .llseek
= seq_lseek
,
1707 .release
= seq_release
,
1709 #endif /* CONFIG_PROC_FS */
1712 static DEFINE_PER_CPU(struct delayed_work
, vmstat_work
);
1713 int sysctl_stat_interval __read_mostly
= HZ
;
1715 #ifdef CONFIG_PROC_FS
1716 static void refresh_vm_stats(struct work_struct
*work
)
1718 refresh_cpu_vm_stats(true);
1721 int vmstat_refresh(struct ctl_table
*table
, int write
,
1722 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1729 * The regular update, every sysctl_stat_interval, may come later
1730 * than expected: leaving a significant amount in per_cpu buckets.
1731 * This is particularly misleading when checking a quantity of HUGE
1732 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1733 * which can equally be echo'ed to or cat'ted from (by root),
1734 * can be used to update the stats just before reading them.
1736 * Oh, and since global_zone_page_state() etc. are so careful to hide
1737 * transiently negative values, report an error here if any of
1738 * the stats is negative, so we know to go looking for imbalance.
1740 err
= schedule_on_each_cpu(refresh_vm_stats
);
1743 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
1744 val
= atomic_long_read(&vm_zone_stat
[i
]);
1746 pr_warn("%s: %s %ld\n",
1747 __func__
, vmstat_text
[i
], val
);
1752 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
1753 val
= atomic_long_read(&vm_numa_stat
[i
]);
1755 pr_warn("%s: %s %ld\n",
1756 __func__
, vmstat_text
[i
+ NR_VM_ZONE_STAT_ITEMS
], val
);
1769 #endif /* CONFIG_PROC_FS */
1771 static void vmstat_update(struct work_struct
*w
)
1773 if (refresh_cpu_vm_stats(true)) {
1775 * Counters were updated so we expect more updates
1776 * to occur in the future. Keep on running the
1777 * update worker thread.
1779 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq
,
1780 this_cpu_ptr(&vmstat_work
),
1781 round_jiffies_relative(sysctl_stat_interval
));
1786 * Switch off vmstat processing and then fold all the remaining differentials
1787 * until the diffs stay at zero. The function is used by NOHZ and can only be
1788 * invoked when tick processing is not active.
1791 * Check if the diffs for a certain cpu indicate that
1792 * an update is needed.
1794 static bool need_update(int cpu
)
1798 for_each_populated_zone(zone
) {
1799 struct per_cpu_pageset
*p
= per_cpu_ptr(zone
->pageset
, cpu
);
1801 BUILD_BUG_ON(sizeof(p
->vm_stat_diff
[0]) != 1);
1803 BUILD_BUG_ON(sizeof(p
->vm_numa_stat_diff
[0]) != 2);
1807 * The fast way of checking if there are any vmstat diffs.
1808 * This works because the diffs are byte sized items.
1810 if (memchr_inv(p
->vm_stat_diff
, 0, NR_VM_ZONE_STAT_ITEMS
))
1813 if (memchr_inv(p
->vm_numa_stat_diff
, 0, NR_VM_NUMA_STAT_ITEMS
))
1821 * Switch off vmstat processing and then fold all the remaining differentials
1822 * until the diffs stay at zero. The function is used by NOHZ and can only be
1823 * invoked when tick processing is not active.
1825 void quiet_vmstat(void)
1827 if (system_state
!= SYSTEM_RUNNING
)
1830 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work
)))
1833 if (!need_update(smp_processor_id()))
1837 * Just refresh counters and do not care about the pending delayed
1838 * vmstat_update. It doesn't fire that often to matter and canceling
1839 * it would be too expensive from this path.
1840 * vmstat_shepherd will take care about that for us.
1842 refresh_cpu_vm_stats(false);
1846 * Shepherd worker thread that checks the
1847 * differentials of processors that have their worker
1848 * threads for vm statistics updates disabled because of
1851 static void vmstat_shepherd(struct work_struct
*w
);
1853 static DECLARE_DEFERRABLE_WORK(shepherd
, vmstat_shepherd
);
1855 static void vmstat_shepherd(struct work_struct
*w
)
1860 /* Check processors whose vmstat worker threads have been disabled */
1861 for_each_online_cpu(cpu
) {
1862 struct delayed_work
*dw
= &per_cpu(vmstat_work
, cpu
);
1864 if (!delayed_work_pending(dw
) && need_update(cpu
))
1865 queue_delayed_work_on(cpu
, mm_percpu_wq
, dw
, 0);
1869 schedule_delayed_work(&shepherd
,
1870 round_jiffies_relative(sysctl_stat_interval
));
1873 static void __init
start_shepherd_timer(void)
1877 for_each_possible_cpu(cpu
)
1878 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work
, cpu
),
1881 schedule_delayed_work(&shepherd
,
1882 round_jiffies_relative(sysctl_stat_interval
));
1885 static void __init
init_cpu_node_state(void)
1889 for_each_online_node(node
) {
1890 if (cpumask_weight(cpumask_of_node(node
)) > 0)
1891 node_set_state(node
, N_CPU
);
1895 static int vmstat_cpu_online(unsigned int cpu
)
1897 refresh_zone_stat_thresholds();
1898 node_set_state(cpu_to_node(cpu
), N_CPU
);
1902 static int vmstat_cpu_down_prep(unsigned int cpu
)
1904 cancel_delayed_work_sync(&per_cpu(vmstat_work
, cpu
));
1908 static int vmstat_cpu_dead(unsigned int cpu
)
1910 const struct cpumask
*node_cpus
;
1913 node
= cpu_to_node(cpu
);
1915 refresh_zone_stat_thresholds();
1916 node_cpus
= cpumask_of_node(node
);
1917 if (cpumask_weight(node_cpus
) > 0)
1920 node_clear_state(node
, N_CPU
);
1926 struct workqueue_struct
*mm_percpu_wq
;
1928 void __init
init_mm_internals(void)
1930 int ret __maybe_unused
;
1932 mm_percpu_wq
= alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM
, 0);
1935 ret
= cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD
, "mm/vmstat:dead",
1936 NULL
, vmstat_cpu_dead
);
1938 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1940 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
, "mm/vmstat:online",
1942 vmstat_cpu_down_prep
);
1944 pr_err("vmstat: failed to register 'online' hotplug state\n");
1947 init_cpu_node_state();
1950 start_shepherd_timer();
1952 #ifdef CONFIG_PROC_FS
1953 proc_create("buddyinfo", 0444, NULL
, &buddyinfo_file_operations
);
1954 proc_create("pagetypeinfo", 0444, NULL
, &pagetypeinfo_file_operations
);
1955 proc_create("vmstat", 0444, NULL
, &vmstat_file_operations
);
1956 proc_create("zoneinfo", 0444, NULL
, &zoneinfo_file_operations
);
1960 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1963 * Return an index indicating how much of the available free memory is
1964 * unusable for an allocation of the requested size.
1966 static int unusable_free_index(unsigned int order
,
1967 struct contig_page_info
*info
)
1969 /* No free memory is interpreted as all free memory is unusable */
1970 if (info
->free_pages
== 0)
1974 * Index should be a value between 0 and 1. Return a value to 3
1977 * 0 => no fragmentation
1978 * 1 => high fragmentation
1980 return div_u64((info
->free_pages
- (info
->free_blocks_suitable
<< order
)) * 1000ULL, info
->free_pages
);
1984 static void unusable_show_print(struct seq_file
*m
,
1985 pg_data_t
*pgdat
, struct zone
*zone
)
1989 struct contig_page_info info
;
1991 seq_printf(m
, "Node %d, zone %8s ",
1994 for (order
= 0; order
< MAX_ORDER
; ++order
) {
1995 fill_contig_page_info(zone
, order
, &info
);
1996 index
= unusable_free_index(order
, &info
);
1997 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2004 * Display unusable free space index
2006 * The unusable free space index measures how much of the available free
2007 * memory cannot be used to satisfy an allocation of a given size and is a
2008 * value between 0 and 1. The higher the value, the more of free memory is
2009 * unusable and by implication, the worse the external fragmentation is. This
2010 * can be expressed as a percentage by multiplying by 100.
2012 static int unusable_show(struct seq_file
*m
, void *arg
)
2014 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2016 /* check memoryless node */
2017 if (!node_state(pgdat
->node_id
, N_MEMORY
))
2020 walk_zones_in_node(m
, pgdat
, true, false, unusable_show_print
);
2025 static const struct seq_operations unusable_op
= {
2026 .start
= frag_start
,
2029 .show
= unusable_show
,
2032 static int unusable_open(struct inode
*inode
, struct file
*file
)
2034 return seq_open(file
, &unusable_op
);
2037 static const struct file_operations unusable_file_ops
= {
2038 .open
= unusable_open
,
2040 .llseek
= seq_lseek
,
2041 .release
= seq_release
,
2044 static void extfrag_show_print(struct seq_file
*m
,
2045 pg_data_t
*pgdat
, struct zone
*zone
)
2050 /* Alloc on stack as interrupts are disabled for zone walk */
2051 struct contig_page_info info
;
2053 seq_printf(m
, "Node %d, zone %8s ",
2056 for (order
= 0; order
< MAX_ORDER
; ++order
) {
2057 fill_contig_page_info(zone
, order
, &info
);
2058 index
= __fragmentation_index(order
, &info
);
2059 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2066 * Display fragmentation index for orders that allocations would fail for
2068 static int extfrag_show(struct seq_file
*m
, void *arg
)
2070 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2072 walk_zones_in_node(m
, pgdat
, true, false, extfrag_show_print
);
2077 static const struct seq_operations extfrag_op
= {
2078 .start
= frag_start
,
2081 .show
= extfrag_show
,
2084 static int extfrag_open(struct inode
*inode
, struct file
*file
)
2086 return seq_open(file
, &extfrag_op
);
2089 static const struct file_operations extfrag_file_ops
= {
2090 .open
= extfrag_open
,
2092 .llseek
= seq_lseek
,
2093 .release
= seq_release
,
2096 static int __init
extfrag_debug_init(void)
2098 struct dentry
*extfrag_debug_root
;
2100 extfrag_debug_root
= debugfs_create_dir("extfrag", NULL
);
2101 if (!extfrag_debug_root
)
2104 if (!debugfs_create_file("unusable_index", 0444,
2105 extfrag_debug_root
, NULL
, &unusable_file_ops
))
2108 if (!debugfs_create_file("extfrag_index", 0444,
2109 extfrag_debug_root
, NULL
, &extfrag_file_ops
))
2114 debugfs_remove_recursive(extfrag_debug_root
);
2118 module_init(extfrag_debug_init
);