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(abs(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
)) {
346 * Only cgroups use subpage accounting right now; at
347 * the global level, these items still change in
348 * multiples of whole pages. Store them as pages
349 * internally to keep the per-cpu counters compact.
351 VM_WARN_ON_ONCE(delta
& (PAGE_SIZE
- 1));
352 delta
>>= PAGE_SHIFT
;
355 x
= delta
+ __this_cpu_read(*p
);
357 t
= __this_cpu_read(pcp
->stat_threshold
);
359 if (unlikely(abs(x
) > t
)) {
360 node_page_state_add(x
, pgdat
, item
);
363 __this_cpu_write(*p
, x
);
365 EXPORT_SYMBOL(__mod_node_page_state
);
368 * Optimized increment and decrement functions.
370 * These are only for a single page and therefore can take a struct page *
371 * argument instead of struct zone *. This allows the inclusion of the code
372 * generated for page_zone(page) into the optimized functions.
374 * No overflow check is necessary and therefore the differential can be
375 * incremented or decremented in place which may allow the compilers to
376 * generate better code.
377 * The increment or decrement is known and therefore one boundary check can
380 * NOTE: These functions are very performance sensitive. Change only
383 * Some processors have inc/dec instructions that are atomic vs an interrupt.
384 * However, the code must first determine the differential location in a zone
385 * based on the processor number and then inc/dec the counter. There is no
386 * guarantee without disabling preemption that the processor will not change
387 * in between and therefore the atomicity vs. interrupt cannot be exploited
388 * in a useful way here.
390 void __inc_zone_state(struct zone
*zone
, enum zone_stat_item item
)
392 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
393 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
396 v
= __this_cpu_inc_return(*p
);
397 t
= __this_cpu_read(pcp
->stat_threshold
);
398 if (unlikely(v
> t
)) {
399 s8 overstep
= t
>> 1;
401 zone_page_state_add(v
+ overstep
, zone
, item
);
402 __this_cpu_write(*p
, -overstep
);
406 void __inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
408 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
409 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
412 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
414 v
= __this_cpu_inc_return(*p
);
415 t
= __this_cpu_read(pcp
->stat_threshold
);
416 if (unlikely(v
> t
)) {
417 s8 overstep
= t
>> 1;
419 node_page_state_add(v
+ overstep
, pgdat
, item
);
420 __this_cpu_write(*p
, -overstep
);
424 void __inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
426 __inc_zone_state(page_zone(page
), item
);
428 EXPORT_SYMBOL(__inc_zone_page_state
);
430 void __inc_node_page_state(struct page
*page
, enum node_stat_item item
)
432 __inc_node_state(page_pgdat(page
), item
);
434 EXPORT_SYMBOL(__inc_node_page_state
);
436 void __dec_zone_state(struct zone
*zone
, enum zone_stat_item item
)
438 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
439 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
442 v
= __this_cpu_dec_return(*p
);
443 t
= __this_cpu_read(pcp
->stat_threshold
);
444 if (unlikely(v
< - t
)) {
445 s8 overstep
= t
>> 1;
447 zone_page_state_add(v
- overstep
, zone
, item
);
448 __this_cpu_write(*p
, overstep
);
452 void __dec_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
454 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
455 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
458 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
460 v
= __this_cpu_dec_return(*p
);
461 t
= __this_cpu_read(pcp
->stat_threshold
);
462 if (unlikely(v
< - t
)) {
463 s8 overstep
= t
>> 1;
465 node_page_state_add(v
- overstep
, pgdat
, item
);
466 __this_cpu_write(*p
, overstep
);
470 void __dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
472 __dec_zone_state(page_zone(page
), item
);
474 EXPORT_SYMBOL(__dec_zone_page_state
);
476 void __dec_node_page_state(struct page
*page
, enum node_stat_item item
)
478 __dec_node_state(page_pgdat(page
), item
);
480 EXPORT_SYMBOL(__dec_node_page_state
);
482 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
484 * If we have cmpxchg_local support then we do not need to incur the overhead
485 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
487 * mod_state() modifies the zone counter state through atomic per cpu
490 * Overstep mode specifies how overstep should handled:
492 * 1 Overstepping half of threshold
493 * -1 Overstepping minus half of threshold
495 static inline void mod_zone_state(struct zone
*zone
,
496 enum zone_stat_item item
, long delta
, int overstep_mode
)
498 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
499 s8 __percpu
*p
= pcp
->vm_stat_diff
+ item
;
503 z
= 0; /* overflow to zone counters */
506 * The fetching of the stat_threshold is racy. We may apply
507 * a counter threshold to the wrong the cpu if we get
508 * rescheduled while executing here. However, the next
509 * counter update will apply the threshold again and
510 * therefore bring the counter under the threshold again.
512 * Most of the time the thresholds are the same anyways
513 * for all cpus in a zone.
515 t
= this_cpu_read(pcp
->stat_threshold
);
517 o
= this_cpu_read(*p
);
521 int os
= overstep_mode
* (t
>> 1) ;
523 /* Overflow must be added to zone counters */
527 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
530 zone_page_state_add(z
, zone
, item
);
533 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
536 mod_zone_state(zone
, item
, delta
, 0);
538 EXPORT_SYMBOL(mod_zone_page_state
);
540 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
542 mod_zone_state(page_zone(page
), item
, 1, 1);
544 EXPORT_SYMBOL(inc_zone_page_state
);
546 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
548 mod_zone_state(page_zone(page
), item
, -1, -1);
550 EXPORT_SYMBOL(dec_zone_page_state
);
552 static inline void mod_node_state(struct pglist_data
*pgdat
,
553 enum node_stat_item item
, int delta
, int overstep_mode
)
555 struct per_cpu_nodestat __percpu
*pcp
= pgdat
->per_cpu_nodestats
;
556 s8 __percpu
*p
= pcp
->vm_node_stat_diff
+ item
;
559 if (vmstat_item_in_bytes(item
)) {
561 * Only cgroups use subpage accounting right now; at
562 * the global level, these items still change in
563 * multiples of whole pages. Store them as pages
564 * internally to keep the per-cpu counters compact.
566 VM_WARN_ON_ONCE(delta
& (PAGE_SIZE
- 1));
567 delta
>>= PAGE_SHIFT
;
571 z
= 0; /* overflow to node counters */
574 * The fetching of the stat_threshold is racy. We may apply
575 * a counter threshold to the wrong the cpu if we get
576 * rescheduled while executing here. However, the next
577 * counter update will apply the threshold again and
578 * therefore bring the counter under the threshold again.
580 * Most of the time the thresholds are the same anyways
581 * for all cpus in a node.
583 t
= this_cpu_read(pcp
->stat_threshold
);
585 o
= this_cpu_read(*p
);
589 int os
= overstep_mode
* (t
>> 1) ;
591 /* Overflow must be added to node counters */
595 } while (this_cpu_cmpxchg(*p
, o
, n
) != o
);
598 node_page_state_add(z
, pgdat
, item
);
601 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
604 mod_node_state(pgdat
, item
, delta
, 0);
606 EXPORT_SYMBOL(mod_node_page_state
);
608 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
610 mod_node_state(pgdat
, item
, 1, 1);
613 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
615 mod_node_state(page_pgdat(page
), item
, 1, 1);
617 EXPORT_SYMBOL(inc_node_page_state
);
619 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
621 mod_node_state(page_pgdat(page
), item
, -1, -1);
623 EXPORT_SYMBOL(dec_node_page_state
);
626 * Use interrupt disable to serialize counter updates
628 void mod_zone_page_state(struct zone
*zone
, enum zone_stat_item item
,
633 local_irq_save(flags
);
634 __mod_zone_page_state(zone
, item
, delta
);
635 local_irq_restore(flags
);
637 EXPORT_SYMBOL(mod_zone_page_state
);
639 void inc_zone_page_state(struct page
*page
, enum zone_stat_item item
)
644 zone
= page_zone(page
);
645 local_irq_save(flags
);
646 __inc_zone_state(zone
, item
);
647 local_irq_restore(flags
);
649 EXPORT_SYMBOL(inc_zone_page_state
);
651 void dec_zone_page_state(struct page
*page
, enum zone_stat_item item
)
655 local_irq_save(flags
);
656 __dec_zone_page_state(page
, item
);
657 local_irq_restore(flags
);
659 EXPORT_SYMBOL(dec_zone_page_state
);
661 void inc_node_state(struct pglist_data
*pgdat
, enum node_stat_item item
)
665 local_irq_save(flags
);
666 __inc_node_state(pgdat
, item
);
667 local_irq_restore(flags
);
669 EXPORT_SYMBOL(inc_node_state
);
671 void mod_node_page_state(struct pglist_data
*pgdat
, enum node_stat_item item
,
676 local_irq_save(flags
);
677 __mod_node_page_state(pgdat
, item
, delta
);
678 local_irq_restore(flags
);
680 EXPORT_SYMBOL(mod_node_page_state
);
682 void inc_node_page_state(struct page
*page
, enum node_stat_item item
)
685 struct pglist_data
*pgdat
;
687 pgdat
= page_pgdat(page
);
688 local_irq_save(flags
);
689 __inc_node_state(pgdat
, item
);
690 local_irq_restore(flags
);
692 EXPORT_SYMBOL(inc_node_page_state
);
694 void dec_node_page_state(struct page
*page
, enum node_stat_item item
)
698 local_irq_save(flags
);
699 __dec_node_page_state(page
, item
);
700 local_irq_restore(flags
);
702 EXPORT_SYMBOL(dec_node_page_state
);
706 * Fold a differential into the global counters.
707 * Returns the number of counters updated.
710 static int fold_diff(int *zone_diff
, int *numa_diff
, int *node_diff
)
715 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
717 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
721 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
723 atomic_long_add(numa_diff
[i
], &vm_numa_stat
[i
]);
727 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
729 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
735 static int fold_diff(int *zone_diff
, int *node_diff
)
740 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
742 atomic_long_add(zone_diff
[i
], &vm_zone_stat
[i
]);
746 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
748 atomic_long_add(node_diff
[i
], &vm_node_stat
[i
]);
753 #endif /* CONFIG_NUMA */
756 * Update the zone counters for the current cpu.
758 * Note that refresh_cpu_vm_stats strives to only access
759 * node local memory. The per cpu pagesets on remote zones are placed
760 * in the memory local to the processor using that pageset. So the
761 * loop over all zones will access a series of cachelines local to
764 * The call to zone_page_state_add updates the cachelines with the
765 * statistics in the remote zone struct as well as the global cachelines
766 * with the global counters. These could cause remote node cache line
767 * bouncing and will have to be only done when necessary.
769 * The function returns the number of global counters updated.
771 static int refresh_cpu_vm_stats(bool do_pagesets
)
773 struct pglist_data
*pgdat
;
776 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
778 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
780 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
783 for_each_populated_zone(zone
) {
784 struct per_cpu_pageset __percpu
*p
= zone
->pageset
;
786 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
789 v
= this_cpu_xchg(p
->vm_stat_diff
[i
], 0);
792 atomic_long_add(v
, &zone
->vm_stat
[i
]);
793 global_zone_diff
[i
] += v
;
795 /* 3 seconds idle till flush */
796 __this_cpu_write(p
->expire
, 3);
801 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
804 v
= this_cpu_xchg(p
->vm_numa_stat_diff
[i
], 0);
807 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
808 global_numa_diff
[i
] += v
;
809 __this_cpu_write(p
->expire
, 3);
816 * Deal with draining the remote pageset of this
819 * Check if there are pages remaining in this pageset
820 * if not then there is nothing to expire.
822 if (!__this_cpu_read(p
->expire
) ||
823 !__this_cpu_read(p
->pcp
.count
))
827 * We never drain zones local to this processor.
829 if (zone_to_nid(zone
) == numa_node_id()) {
830 __this_cpu_write(p
->expire
, 0);
834 if (__this_cpu_dec_return(p
->expire
))
837 if (__this_cpu_read(p
->pcp
.count
)) {
838 drain_zone_pages(zone
, this_cpu_ptr(&p
->pcp
));
845 for_each_online_pgdat(pgdat
) {
846 struct per_cpu_nodestat __percpu
*p
= pgdat
->per_cpu_nodestats
;
848 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
851 v
= this_cpu_xchg(p
->vm_node_stat_diff
[i
], 0);
853 atomic_long_add(v
, &pgdat
->vm_stat
[i
]);
854 global_node_diff
[i
] += v
;
860 changes
+= fold_diff(global_zone_diff
, global_numa_diff
,
863 changes
+= fold_diff(global_zone_diff
, global_node_diff
);
869 * Fold the data for an offline cpu into the global array.
870 * There cannot be any access by the offline cpu and therefore
871 * synchronization is simplified.
873 void cpu_vm_stats_fold(int cpu
)
875 struct pglist_data
*pgdat
;
878 int global_zone_diff
[NR_VM_ZONE_STAT_ITEMS
] = { 0, };
880 int global_numa_diff
[NR_VM_NUMA_STAT_ITEMS
] = { 0, };
882 int global_node_diff
[NR_VM_NODE_STAT_ITEMS
] = { 0, };
884 for_each_populated_zone(zone
) {
885 struct per_cpu_pageset
*p
;
887 p
= per_cpu_ptr(zone
->pageset
, cpu
);
889 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
890 if (p
->vm_stat_diff
[i
]) {
893 v
= p
->vm_stat_diff
[i
];
894 p
->vm_stat_diff
[i
] = 0;
895 atomic_long_add(v
, &zone
->vm_stat
[i
]);
896 global_zone_diff
[i
] += v
;
900 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
901 if (p
->vm_numa_stat_diff
[i
]) {
904 v
= p
->vm_numa_stat_diff
[i
];
905 p
->vm_numa_stat_diff
[i
] = 0;
906 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
907 global_numa_diff
[i
] += v
;
912 for_each_online_pgdat(pgdat
) {
913 struct per_cpu_nodestat
*p
;
915 p
= per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
);
917 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
918 if (p
->vm_node_stat_diff
[i
]) {
921 v
= p
->vm_node_stat_diff
[i
];
922 p
->vm_node_stat_diff
[i
] = 0;
923 atomic_long_add(v
, &pgdat
->vm_stat
[i
]);
924 global_node_diff
[i
] += v
;
929 fold_diff(global_zone_diff
, global_numa_diff
, global_node_diff
);
931 fold_diff(global_zone_diff
, global_node_diff
);
936 * this is only called if !populated_zone(zone), which implies no other users of
937 * pset->vm_stat_diff[] exsist.
939 void drain_zonestat(struct zone
*zone
, struct per_cpu_pageset
*pset
)
943 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
944 if (pset
->vm_stat_diff
[i
]) {
945 int v
= pset
->vm_stat_diff
[i
];
946 pset
->vm_stat_diff
[i
] = 0;
947 atomic_long_add(v
, &zone
->vm_stat
[i
]);
948 atomic_long_add(v
, &vm_zone_stat
[i
]);
952 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
953 if (pset
->vm_numa_stat_diff
[i
]) {
954 int v
= pset
->vm_numa_stat_diff
[i
];
956 pset
->vm_numa_stat_diff
[i
] = 0;
957 atomic_long_add(v
, &zone
->vm_numa_stat
[i
]);
958 atomic_long_add(v
, &vm_numa_stat
[i
]);
965 void __inc_numa_state(struct zone
*zone
,
966 enum numa_stat_item item
)
968 struct per_cpu_pageset __percpu
*pcp
= zone
->pageset
;
969 u16 __percpu
*p
= pcp
->vm_numa_stat_diff
+ item
;
972 v
= __this_cpu_inc_return(*p
);
974 if (unlikely(v
> NUMA_STATS_THRESHOLD
)) {
975 zone_numa_state_add(v
, zone
, item
);
976 __this_cpu_write(*p
, 0);
981 * Determine the per node value of a stat item. This function
982 * is called frequently in a NUMA machine, so try to be as
983 * frugal as possible.
985 unsigned long sum_zone_node_page_state(int node
,
986 enum zone_stat_item item
)
988 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
990 unsigned long count
= 0;
992 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
993 count
+= zone_page_state(zones
+ i
, item
);
999 * Determine the per node value of a numa stat item. To avoid deviation,
1000 * the per cpu stat number in vm_numa_stat_diff[] is also included.
1002 unsigned long sum_zone_numa_state(int node
,
1003 enum numa_stat_item item
)
1005 struct zone
*zones
= NODE_DATA(node
)->node_zones
;
1007 unsigned long count
= 0;
1009 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
1010 count
+= zone_numa_state_snapshot(zones
+ i
, item
);
1016 * Determine the per node value of a stat item.
1018 unsigned long node_page_state_pages(struct pglist_data
*pgdat
,
1019 enum node_stat_item item
)
1021 long x
= atomic_long_read(&pgdat
->vm_stat
[item
]);
1029 unsigned long node_page_state(struct pglist_data
*pgdat
,
1030 enum node_stat_item item
)
1032 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item
));
1034 return node_page_state_pages(pgdat
, item
);
1038 #ifdef CONFIG_COMPACTION
1040 struct contig_page_info
{
1041 unsigned long free_pages
;
1042 unsigned long free_blocks_total
;
1043 unsigned long free_blocks_suitable
;
1047 * Calculate the number of free pages in a zone, how many contiguous
1048 * pages are free and how many are large enough to satisfy an allocation of
1049 * the target size. Note that this function makes no attempt to estimate
1050 * how many suitable free blocks there *might* be if MOVABLE pages were
1051 * migrated. Calculating that is possible, but expensive and can be
1052 * figured out from userspace
1054 static void fill_contig_page_info(struct zone
*zone
,
1055 unsigned int suitable_order
,
1056 struct contig_page_info
*info
)
1060 info
->free_pages
= 0;
1061 info
->free_blocks_total
= 0;
1062 info
->free_blocks_suitable
= 0;
1064 for (order
= 0; order
< MAX_ORDER
; order
++) {
1065 unsigned long blocks
;
1067 /* Count number of free blocks */
1068 blocks
= zone
->free_area
[order
].nr_free
;
1069 info
->free_blocks_total
+= blocks
;
1071 /* Count free base pages */
1072 info
->free_pages
+= blocks
<< order
;
1074 /* Count the suitable free blocks */
1075 if (order
>= suitable_order
)
1076 info
->free_blocks_suitable
+= blocks
<<
1077 (order
- suitable_order
);
1082 * A fragmentation index only makes sense if an allocation of a requested
1083 * size would fail. If that is true, the fragmentation index indicates
1084 * whether external fragmentation or a lack of memory was the problem.
1085 * The value can be used to determine if page reclaim or compaction
1088 static int __fragmentation_index(unsigned int order
, struct contig_page_info
*info
)
1090 unsigned long requested
= 1UL << order
;
1092 if (WARN_ON_ONCE(order
>= MAX_ORDER
))
1095 if (!info
->free_blocks_total
)
1098 /* Fragmentation index only makes sense when a request would fail */
1099 if (info
->free_blocks_suitable
)
1103 * Index is between 0 and 1 so return within 3 decimal places
1105 * 0 => allocation would fail due to lack of memory
1106 * 1 => allocation would fail due to fragmentation
1108 return 1000 - div_u64( (1000+(div_u64(info
->free_pages
* 1000ULL, requested
))), info
->free_blocks_total
);
1112 * Calculates external fragmentation within a zone wrt the given order.
1113 * It is defined as the percentage of pages found in blocks of size
1114 * less than 1 << order. It returns values in range [0, 100].
1116 unsigned int extfrag_for_order(struct zone
*zone
, unsigned int order
)
1118 struct contig_page_info info
;
1120 fill_contig_page_info(zone
, order
, &info
);
1121 if (info
.free_pages
== 0)
1124 return div_u64((info
.free_pages
-
1125 (info
.free_blocks_suitable
<< order
)) * 100,
1129 /* Same as __fragmentation index but allocs contig_page_info on stack */
1130 int fragmentation_index(struct zone
*zone
, unsigned int order
)
1132 struct contig_page_info info
;
1134 fill_contig_page_info(zone
, order
, &info
);
1135 return __fragmentation_index(order
, &info
);
1139 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1140 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1141 #ifdef CONFIG_ZONE_DMA
1142 #define TEXT_FOR_DMA(xx) xx "_dma",
1144 #define TEXT_FOR_DMA(xx)
1147 #ifdef CONFIG_ZONE_DMA32
1148 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1150 #define TEXT_FOR_DMA32(xx)
1153 #ifdef CONFIG_HIGHMEM
1154 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1156 #define TEXT_FOR_HIGHMEM(xx)
1159 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1160 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1162 const char * const vmstat_text
[] = {
1163 /* enum zone_stat_item counters */
1165 "nr_zone_inactive_anon",
1166 "nr_zone_active_anon",
1167 "nr_zone_inactive_file",
1168 "nr_zone_active_file",
1169 "nr_zone_unevictable",
1170 "nr_zone_write_pending",
1173 #if IS_ENABLED(CONFIG_ZSMALLOC)
1178 /* enum numa_stat_item counters */
1188 /* enum node_stat_item counters */
1194 "nr_slab_reclaimable",
1195 "nr_slab_unreclaimable",
1199 "workingset_refault_anon",
1200 "workingset_refault_file",
1201 "workingset_activate_anon",
1202 "workingset_activate_file",
1203 "workingset_restore_anon",
1204 "workingset_restore_file",
1205 "workingset_nodereclaim",
1211 "nr_writeback_temp",
1213 "nr_shmem_hugepages",
1214 "nr_shmem_pmdmapped",
1215 "nr_file_hugepages",
1216 "nr_file_pmdmapped",
1217 "nr_anon_transparent_hugepages",
1219 "nr_vmscan_immediate_reclaim",
1222 "nr_kernel_misc_reclaimable",
1223 "nr_foll_pin_acquired",
1224 "nr_foll_pin_released",
1226 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1227 "nr_shadow_call_stack",
1229 "nr_page_table_pages",
1234 /* enum writeback_stat_item counters */
1235 "nr_dirty_threshold",
1236 "nr_dirty_background_threshold",
1238 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1239 /* enum vm_event_item counters */
1245 TEXTS_FOR_ZONES("pgalloc")
1246 TEXTS_FOR_ZONES("allocstall")
1247 TEXTS_FOR_ZONES("pgskip")
1264 "pgscan_direct_throttle",
1271 "zone_reclaim_failed",
1275 "kswapd_inodesteal",
1276 "kswapd_low_wmark_hit_quickly",
1277 "kswapd_high_wmark_hit_quickly",
1286 #ifdef CONFIG_NUMA_BALANCING
1288 "numa_huge_pte_updates",
1290 "numa_hint_faults_local",
1291 "numa_pages_migrated",
1293 #ifdef CONFIG_MIGRATION
1294 "pgmigrate_success",
1296 "thp_migration_success",
1297 "thp_migration_fail",
1298 "thp_migration_split",
1300 #ifdef CONFIG_COMPACTION
1301 "compact_migrate_scanned",
1302 "compact_free_scanned",
1307 "compact_daemon_wake",
1308 "compact_daemon_migrate_scanned",
1309 "compact_daemon_free_scanned",
1312 #ifdef CONFIG_HUGETLB_PAGE
1313 "htlb_buddy_alloc_success",
1314 "htlb_buddy_alloc_fail",
1316 "unevictable_pgs_culled",
1317 "unevictable_pgs_scanned",
1318 "unevictable_pgs_rescued",
1319 "unevictable_pgs_mlocked",
1320 "unevictable_pgs_munlocked",
1321 "unevictable_pgs_cleared",
1322 "unevictable_pgs_stranded",
1324 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1326 "thp_fault_fallback",
1327 "thp_fault_fallback_charge",
1328 "thp_collapse_alloc",
1329 "thp_collapse_alloc_failed",
1331 "thp_file_fallback",
1332 "thp_file_fallback_charge",
1335 "thp_split_page_failed",
1336 "thp_deferred_split_page",
1338 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1341 "thp_zero_page_alloc",
1342 "thp_zero_page_alloc_failed",
1344 "thp_swpout_fallback",
1346 #ifdef CONFIG_MEMORY_BALLOON
1349 #ifdef CONFIG_BALLOON_COMPACTION
1352 #endif /* CONFIG_MEMORY_BALLOON */
1353 #ifdef CONFIG_DEBUG_TLBFLUSH
1354 "nr_tlb_remote_flush",
1355 "nr_tlb_remote_flush_received",
1356 "nr_tlb_local_flush_all",
1357 "nr_tlb_local_flush_one",
1358 #endif /* CONFIG_DEBUG_TLBFLUSH */
1360 #ifdef CONFIG_DEBUG_VM_VMACACHE
1361 "vmacache_find_calls",
1362 "vmacache_find_hits",
1368 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1370 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1372 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1373 defined(CONFIG_PROC_FS)
1374 static void *frag_start(struct seq_file
*m
, loff_t
*pos
)
1379 for (pgdat
= first_online_pgdat();
1381 pgdat
= next_online_pgdat(pgdat
))
1387 static void *frag_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1389 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1392 return next_online_pgdat(pgdat
);
1395 static void frag_stop(struct seq_file
*m
, void *arg
)
1400 * Walk zones in a node and print using a callback.
1401 * If @assert_populated is true, only use callback for zones that are populated.
1403 static void walk_zones_in_node(struct seq_file
*m
, pg_data_t
*pgdat
,
1404 bool assert_populated
, bool nolock
,
1405 void (*print
)(struct seq_file
*m
, pg_data_t
*, struct zone
*))
1408 struct zone
*node_zones
= pgdat
->node_zones
;
1409 unsigned long flags
;
1411 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
1412 if (assert_populated
&& !populated_zone(zone
))
1416 spin_lock_irqsave(&zone
->lock
, flags
);
1417 print(m
, pgdat
, zone
);
1419 spin_unlock_irqrestore(&zone
->lock
, flags
);
1424 #ifdef CONFIG_PROC_FS
1425 static void frag_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1430 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1431 for (order
= 0; order
< MAX_ORDER
; ++order
)
1432 seq_printf(m
, "%6lu ", zone
->free_area
[order
].nr_free
);
1437 * This walks the free areas for each zone.
1439 static int frag_show(struct seq_file
*m
, void *arg
)
1441 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1442 walk_zones_in_node(m
, pgdat
, true, false, frag_show_print
);
1446 static void pagetypeinfo_showfree_print(struct seq_file
*m
,
1447 pg_data_t
*pgdat
, struct zone
*zone
)
1451 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++) {
1452 seq_printf(m
, "Node %4d, zone %8s, type %12s ",
1455 migratetype_names
[mtype
]);
1456 for (order
= 0; order
< MAX_ORDER
; ++order
) {
1457 unsigned long freecount
= 0;
1458 struct free_area
*area
;
1459 struct list_head
*curr
;
1460 bool overflow
= false;
1462 area
= &(zone
->free_area
[order
]);
1464 list_for_each(curr
, &area
->free_list
[mtype
]) {
1466 * Cap the free_list iteration because it might
1467 * be really large and we are under a spinlock
1468 * so a long time spent here could trigger a
1469 * hard lockup detector. Anyway this is a
1470 * debugging tool so knowing there is a handful
1471 * of pages of this order should be more than
1474 if (++freecount
>= 100000) {
1479 seq_printf(m
, "%s%6lu ", overflow
? ">" : "", freecount
);
1480 spin_unlock_irq(&zone
->lock
);
1482 spin_lock_irq(&zone
->lock
);
1488 /* Print out the free pages at each order for each migatetype */
1489 static int pagetypeinfo_showfree(struct seq_file
*m
, void *arg
)
1492 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1495 seq_printf(m
, "%-43s ", "Free pages count per migrate type at order");
1496 for (order
= 0; order
< MAX_ORDER
; ++order
)
1497 seq_printf(m
, "%6d ", order
);
1500 walk_zones_in_node(m
, pgdat
, true, false, pagetypeinfo_showfree_print
);
1505 static void pagetypeinfo_showblockcount_print(struct seq_file
*m
,
1506 pg_data_t
*pgdat
, struct zone
*zone
)
1510 unsigned long start_pfn
= zone
->zone_start_pfn
;
1511 unsigned long end_pfn
= zone_end_pfn(zone
);
1512 unsigned long count
[MIGRATE_TYPES
] = { 0, };
1514 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
1517 page
= pfn_to_online_page(pfn
);
1521 if (page_zone(page
) != zone
)
1524 mtype
= get_pageblock_migratetype(page
);
1526 if (mtype
< MIGRATE_TYPES
)
1531 seq_printf(m
, "Node %d, zone %8s ", pgdat
->node_id
, zone
->name
);
1532 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1533 seq_printf(m
, "%12lu ", count
[mtype
]);
1537 /* Print out the number of pageblocks for each migratetype */
1538 static int pagetypeinfo_showblockcount(struct seq_file
*m
, void *arg
)
1541 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1543 seq_printf(m
, "\n%-23s", "Number of blocks type ");
1544 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1545 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1547 walk_zones_in_node(m
, pgdat
, true, false,
1548 pagetypeinfo_showblockcount_print
);
1554 * Print out the number of pageblocks for each migratetype that contain pages
1555 * of other types. This gives an indication of how well fallbacks are being
1556 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1557 * to determine what is going on
1559 static void pagetypeinfo_showmixedcount(struct seq_file
*m
, pg_data_t
*pgdat
)
1561 #ifdef CONFIG_PAGE_OWNER
1564 if (!static_branch_unlikely(&page_owner_inited
))
1567 drain_all_pages(NULL
);
1569 seq_printf(m
, "\n%-23s", "Number of mixed blocks ");
1570 for (mtype
= 0; mtype
< MIGRATE_TYPES
; mtype
++)
1571 seq_printf(m
, "%12s ", migratetype_names
[mtype
]);
1574 walk_zones_in_node(m
, pgdat
, true, true,
1575 pagetypeinfo_showmixedcount_print
);
1576 #endif /* CONFIG_PAGE_OWNER */
1580 * This prints out statistics in relation to grouping pages by mobility.
1581 * It is expensive to collect so do not constantly read the file.
1583 static int pagetypeinfo_show(struct seq_file
*m
, void *arg
)
1585 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1587 /* check memoryless node */
1588 if (!node_state(pgdat
->node_id
, N_MEMORY
))
1591 seq_printf(m
, "Page block order: %d\n", pageblock_order
);
1592 seq_printf(m
, "Pages per block: %lu\n", pageblock_nr_pages
);
1594 pagetypeinfo_showfree(m
, pgdat
);
1595 pagetypeinfo_showblockcount(m
, pgdat
);
1596 pagetypeinfo_showmixedcount(m
, pgdat
);
1601 static const struct seq_operations fragmentation_op
= {
1602 .start
= frag_start
,
1608 static const struct seq_operations pagetypeinfo_op
= {
1609 .start
= frag_start
,
1612 .show
= pagetypeinfo_show
,
1615 static bool is_zone_first_populated(pg_data_t
*pgdat
, struct zone
*zone
)
1619 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1620 struct zone
*compare
= &pgdat
->node_zones
[zid
];
1622 if (populated_zone(compare
))
1623 return zone
== compare
;
1629 static void zoneinfo_show_print(struct seq_file
*m
, pg_data_t
*pgdat
,
1633 seq_printf(m
, "Node %d, zone %8s", pgdat
->node_id
, zone
->name
);
1634 if (is_zone_first_populated(pgdat
, zone
)) {
1635 seq_printf(m
, "\n per-node stats");
1636 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
1637 unsigned long pages
= node_page_state_pages(pgdat
, i
);
1639 if (vmstat_item_print_in_thp(i
))
1640 pages
/= HPAGE_PMD_NR
;
1641 seq_printf(m
, "\n %-12s %lu", node_stat_name(i
),
1654 zone_page_state(zone
, NR_FREE_PAGES
),
1655 min_wmark_pages(zone
),
1656 low_wmark_pages(zone
),
1657 high_wmark_pages(zone
),
1658 zone
->spanned_pages
,
1659 zone
->present_pages
,
1660 zone_managed_pages(zone
),
1661 zone_cma_pages(zone
));
1664 "\n protection: (%ld",
1665 zone
->lowmem_reserve
[0]);
1666 for (i
= 1; i
< ARRAY_SIZE(zone
->lowmem_reserve
); i
++)
1667 seq_printf(m
, ", %ld", zone
->lowmem_reserve
[i
]);
1670 /* If unpopulated, no other information is useful */
1671 if (!populated_zone(zone
)) {
1676 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1677 seq_printf(m
, "\n %-12s %lu", zone_stat_name(i
),
1678 zone_page_state(zone
, i
));
1681 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1682 seq_printf(m
, "\n %-12s %lu", numa_stat_name(i
),
1683 zone_numa_state_snapshot(zone
, i
));
1686 seq_printf(m
, "\n pagesets");
1687 for_each_online_cpu(i
) {
1688 struct per_cpu_pageset
*pageset
;
1690 pageset
= per_cpu_ptr(zone
->pageset
, i
);
1699 pageset
->pcp
.batch
);
1701 seq_printf(m
, "\n vm stats threshold: %d",
1702 pageset
->stat_threshold
);
1706 "\n node_unreclaimable: %u"
1707 "\n start_pfn: %lu",
1708 pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
,
1709 zone
->zone_start_pfn
);
1714 * Output information about zones in @pgdat. All zones are printed regardless
1715 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1716 * set of all zones and userspace would not be aware of such zones if they are
1717 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1719 static int zoneinfo_show(struct seq_file
*m
, void *arg
)
1721 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
1722 walk_zones_in_node(m
, pgdat
, false, false, zoneinfo_show_print
);
1726 static const struct seq_operations zoneinfo_op
= {
1727 .start
= frag_start
, /* iterate over all zones. The same as in
1731 .show
= zoneinfo_show
,
1734 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1735 NR_VM_NUMA_STAT_ITEMS + \
1736 NR_VM_NODE_STAT_ITEMS + \
1737 NR_VM_WRITEBACK_STAT_ITEMS + \
1738 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1739 NR_VM_EVENT_ITEMS : 0))
1741 static void *vmstat_start(struct seq_file
*m
, loff_t
*pos
)
1746 if (*pos
>= NR_VMSTAT_ITEMS
)
1749 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text
) < NR_VMSTAT_ITEMS
);
1750 v
= kmalloc_array(NR_VMSTAT_ITEMS
, sizeof(unsigned long), GFP_KERNEL
);
1753 return ERR_PTR(-ENOMEM
);
1754 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++)
1755 v
[i
] = global_zone_page_state(i
);
1756 v
+= NR_VM_ZONE_STAT_ITEMS
;
1759 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++)
1760 v
[i
] = global_numa_state(i
);
1761 v
+= NR_VM_NUMA_STAT_ITEMS
;
1764 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
1765 v
[i
] = global_node_page_state_pages(i
);
1766 if (vmstat_item_print_in_thp(i
))
1767 v
[i
] /= HPAGE_PMD_NR
;
1769 v
+= NR_VM_NODE_STAT_ITEMS
;
1771 global_dirty_limits(v
+ NR_DIRTY_BG_THRESHOLD
,
1772 v
+ NR_DIRTY_THRESHOLD
);
1773 v
+= NR_VM_WRITEBACK_STAT_ITEMS
;
1775 #ifdef CONFIG_VM_EVENT_COUNTERS
1777 v
[PGPGIN
] /= 2; /* sectors -> kbytes */
1780 return (unsigned long *)m
->private + *pos
;
1783 static void *vmstat_next(struct seq_file
*m
, void *arg
, loff_t
*pos
)
1786 if (*pos
>= NR_VMSTAT_ITEMS
)
1788 return (unsigned long *)m
->private + *pos
;
1791 static int vmstat_show(struct seq_file
*m
, void *arg
)
1793 unsigned long *l
= arg
;
1794 unsigned long off
= l
- (unsigned long *)m
->private;
1796 seq_puts(m
, vmstat_text
[off
]);
1797 seq_put_decimal_ull(m
, " ", *l
);
1800 if (off
== NR_VMSTAT_ITEMS
- 1) {
1802 * We've come to the end - add any deprecated counters to avoid
1803 * breaking userspace which might depend on them being present.
1805 seq_puts(m
, "nr_unstable 0\n");
1810 static void vmstat_stop(struct seq_file
*m
, void *arg
)
1816 static const struct seq_operations vmstat_op
= {
1817 .start
= vmstat_start
,
1818 .next
= vmstat_next
,
1819 .stop
= vmstat_stop
,
1820 .show
= vmstat_show
,
1822 #endif /* CONFIG_PROC_FS */
1825 static DEFINE_PER_CPU(struct delayed_work
, vmstat_work
);
1826 int sysctl_stat_interval __read_mostly
= HZ
;
1828 #ifdef CONFIG_PROC_FS
1829 static void refresh_vm_stats(struct work_struct
*work
)
1831 refresh_cpu_vm_stats(true);
1834 int vmstat_refresh(struct ctl_table
*table
, int write
,
1835 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1842 * The regular update, every sysctl_stat_interval, may come later
1843 * than expected: leaving a significant amount in per_cpu buckets.
1844 * This is particularly misleading when checking a quantity of HUGE
1845 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1846 * which can equally be echo'ed to or cat'ted from (by root),
1847 * can be used to update the stats just before reading them.
1849 * Oh, and since global_zone_page_state() etc. are so careful to hide
1850 * transiently negative values, report an error here if any of
1851 * the stats is negative, so we know to go looking for imbalance.
1853 err
= schedule_on_each_cpu(refresh_vm_stats
);
1856 for (i
= 0; i
< NR_VM_ZONE_STAT_ITEMS
; i
++) {
1857 val
= atomic_long_read(&vm_zone_stat
[i
]);
1859 pr_warn("%s: %s %ld\n",
1860 __func__
, zone_stat_name(i
), val
);
1865 for (i
= 0; i
< NR_VM_NUMA_STAT_ITEMS
; i
++) {
1866 val
= atomic_long_read(&vm_numa_stat
[i
]);
1868 pr_warn("%s: %s %ld\n",
1869 __func__
, numa_stat_name(i
), val
);
1882 #endif /* CONFIG_PROC_FS */
1884 static void vmstat_update(struct work_struct
*w
)
1886 if (refresh_cpu_vm_stats(true)) {
1888 * Counters were updated so we expect more updates
1889 * to occur in the future. Keep on running the
1890 * update worker thread.
1892 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq
,
1893 this_cpu_ptr(&vmstat_work
),
1894 round_jiffies_relative(sysctl_stat_interval
));
1899 * Switch off vmstat processing and then fold all the remaining differentials
1900 * until the diffs stay at zero. The function is used by NOHZ and can only be
1901 * invoked when tick processing is not active.
1904 * Check if the diffs for a certain cpu indicate that
1905 * an update is needed.
1907 static bool need_update(int cpu
)
1909 pg_data_t
*last_pgdat
= NULL
;
1912 for_each_populated_zone(zone
) {
1913 struct per_cpu_pageset
*p
= per_cpu_ptr(zone
->pageset
, cpu
);
1914 struct per_cpu_nodestat
*n
;
1916 * The fast way of checking if there are any vmstat diffs.
1918 if (memchr_inv(p
->vm_stat_diff
, 0, NR_VM_ZONE_STAT_ITEMS
*
1919 sizeof(p
->vm_stat_diff
[0])))
1922 if (memchr_inv(p
->vm_numa_stat_diff
, 0, NR_VM_NUMA_STAT_ITEMS
*
1923 sizeof(p
->vm_numa_stat_diff
[0])))
1926 if (last_pgdat
== zone
->zone_pgdat
)
1928 last_pgdat
= zone
->zone_pgdat
;
1929 n
= per_cpu_ptr(zone
->zone_pgdat
->per_cpu_nodestats
, cpu
);
1930 if (memchr_inv(n
->vm_node_stat_diff
, 0, NR_VM_NODE_STAT_ITEMS
*
1931 sizeof(n
->vm_node_stat_diff
[0])))
1938 * Switch off vmstat processing and then fold all the remaining differentials
1939 * until the diffs stay at zero. The function is used by NOHZ and can only be
1940 * invoked when tick processing is not active.
1942 void quiet_vmstat(void)
1944 if (system_state
!= SYSTEM_RUNNING
)
1947 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work
)))
1950 if (!need_update(smp_processor_id()))
1954 * Just refresh counters and do not care about the pending delayed
1955 * vmstat_update. It doesn't fire that often to matter and canceling
1956 * it would be too expensive from this path.
1957 * vmstat_shepherd will take care about that for us.
1959 refresh_cpu_vm_stats(false);
1963 * Shepherd worker thread that checks the
1964 * differentials of processors that have their worker
1965 * threads for vm statistics updates disabled because of
1968 static void vmstat_shepherd(struct work_struct
*w
);
1970 static DECLARE_DEFERRABLE_WORK(shepherd
, vmstat_shepherd
);
1972 static void vmstat_shepherd(struct work_struct
*w
)
1977 /* Check processors whose vmstat worker threads have been disabled */
1978 for_each_online_cpu(cpu
) {
1979 struct delayed_work
*dw
= &per_cpu(vmstat_work
, cpu
);
1981 if (!delayed_work_pending(dw
) && need_update(cpu
))
1982 queue_delayed_work_on(cpu
, mm_percpu_wq
, dw
, 0);
1988 schedule_delayed_work(&shepherd
,
1989 round_jiffies_relative(sysctl_stat_interval
));
1992 static void __init
start_shepherd_timer(void)
1996 for_each_possible_cpu(cpu
)
1997 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work
, cpu
),
2000 schedule_delayed_work(&shepherd
,
2001 round_jiffies_relative(sysctl_stat_interval
));
2004 static void __init
init_cpu_node_state(void)
2008 for_each_online_node(node
) {
2009 if (cpumask_weight(cpumask_of_node(node
)) > 0)
2010 node_set_state(node
, N_CPU
);
2014 static int vmstat_cpu_online(unsigned int cpu
)
2016 refresh_zone_stat_thresholds();
2017 node_set_state(cpu_to_node(cpu
), N_CPU
);
2021 static int vmstat_cpu_down_prep(unsigned int cpu
)
2023 cancel_delayed_work_sync(&per_cpu(vmstat_work
, cpu
));
2027 static int vmstat_cpu_dead(unsigned int cpu
)
2029 const struct cpumask
*node_cpus
;
2032 node
= cpu_to_node(cpu
);
2034 refresh_zone_stat_thresholds();
2035 node_cpus
= cpumask_of_node(node
);
2036 if (cpumask_weight(node_cpus
) > 0)
2039 node_clear_state(node
, N_CPU
);
2045 struct workqueue_struct
*mm_percpu_wq
;
2047 void __init
init_mm_internals(void)
2049 int ret __maybe_unused
;
2051 mm_percpu_wq
= alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM
, 0);
2054 ret
= cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD
, "mm/vmstat:dead",
2055 NULL
, vmstat_cpu_dead
);
2057 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2059 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
, "mm/vmstat:online",
2061 vmstat_cpu_down_prep
);
2063 pr_err("vmstat: failed to register 'online' hotplug state\n");
2066 init_cpu_node_state();
2069 start_shepherd_timer();
2071 #ifdef CONFIG_PROC_FS
2072 proc_create_seq("buddyinfo", 0444, NULL
, &fragmentation_op
);
2073 proc_create_seq("pagetypeinfo", 0400, NULL
, &pagetypeinfo_op
);
2074 proc_create_seq("vmstat", 0444, NULL
, &vmstat_op
);
2075 proc_create_seq("zoneinfo", 0444, NULL
, &zoneinfo_op
);
2079 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2082 * Return an index indicating how much of the available free memory is
2083 * unusable for an allocation of the requested size.
2085 static int unusable_free_index(unsigned int order
,
2086 struct contig_page_info
*info
)
2088 /* No free memory is interpreted as all free memory is unusable */
2089 if (info
->free_pages
== 0)
2093 * Index should be a value between 0 and 1. Return a value to 3
2096 * 0 => no fragmentation
2097 * 1 => high fragmentation
2099 return div_u64((info
->free_pages
- (info
->free_blocks_suitable
<< order
)) * 1000ULL, info
->free_pages
);
2103 static void unusable_show_print(struct seq_file
*m
,
2104 pg_data_t
*pgdat
, struct zone
*zone
)
2108 struct contig_page_info info
;
2110 seq_printf(m
, "Node %d, zone %8s ",
2113 for (order
= 0; order
< MAX_ORDER
; ++order
) {
2114 fill_contig_page_info(zone
, order
, &info
);
2115 index
= unusable_free_index(order
, &info
);
2116 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2123 * Display unusable free space index
2125 * The unusable free space index measures how much of the available free
2126 * memory cannot be used to satisfy an allocation of a given size and is a
2127 * value between 0 and 1. The higher the value, the more of free memory is
2128 * unusable and by implication, the worse the external fragmentation is. This
2129 * can be expressed as a percentage by multiplying by 100.
2131 static int unusable_show(struct seq_file
*m
, void *arg
)
2133 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2135 /* check memoryless node */
2136 if (!node_state(pgdat
->node_id
, N_MEMORY
))
2139 walk_zones_in_node(m
, pgdat
, true, false, unusable_show_print
);
2144 static const struct seq_operations unusable_sops
= {
2145 .start
= frag_start
,
2148 .show
= unusable_show
,
2151 DEFINE_SEQ_ATTRIBUTE(unusable
);
2153 static void extfrag_show_print(struct seq_file
*m
,
2154 pg_data_t
*pgdat
, struct zone
*zone
)
2159 /* Alloc on stack as interrupts are disabled for zone walk */
2160 struct contig_page_info info
;
2162 seq_printf(m
, "Node %d, zone %8s ",
2165 for (order
= 0; order
< MAX_ORDER
; ++order
) {
2166 fill_contig_page_info(zone
, order
, &info
);
2167 index
= __fragmentation_index(order
, &info
);
2168 seq_printf(m
, "%d.%03d ", index
/ 1000, index
% 1000);
2175 * Display fragmentation index for orders that allocations would fail for
2177 static int extfrag_show(struct seq_file
*m
, void *arg
)
2179 pg_data_t
*pgdat
= (pg_data_t
*)arg
;
2181 walk_zones_in_node(m
, pgdat
, true, false, extfrag_show_print
);
2186 static const struct seq_operations extfrag_sops
= {
2187 .start
= frag_start
,
2190 .show
= extfrag_show
,
2193 DEFINE_SEQ_ATTRIBUTE(extfrag
);
2195 static int __init
extfrag_debug_init(void)
2197 struct dentry
*extfrag_debug_root
;
2199 extfrag_debug_root
= debugfs_create_dir("extfrag", NULL
);
2201 debugfs_create_file("unusable_index", 0444, extfrag_debug_root
, NULL
,
2204 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root
, NULL
,
2210 module_init(extfrag_debug_init
);