Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmstat.c
4 *
5 * Manages VM statistics
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 *
8 * zoned VM statistics
9 * Copyright (C) 2006 Silicon Graphics, Inc.,
10 * Christoph Lameter <christoph@lameter.com>
11 * Copyright (C) 2008-2014 Christoph Lameter
12 */
13#include <linux/fs.h>
14#include <linux/mm.h>
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>
31
32#include "internal.h"
33
34#define NUMA_STATS_THRESHOLD (U16_MAX - 2)
35
36#ifdef CONFIG_NUMA
37int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
38
39/* zero numa counters within a zone */
40static void zero_zone_numa_counters(struct zone *zone)
41{
42 int item, cpu;
43
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]
48 = 0;
49 }
50}
51
52/* zero numa counters of all the populated zones */
53static void zero_zones_numa_counters(void)
54{
55 struct zone *zone;
56
57 for_each_populated_zone(zone)
58 zero_zone_numa_counters(zone);
59}
60
61/* zero global numa counters */
62static void zero_global_numa_counters(void)
63{
64 int item;
65
66 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
67 atomic_long_set(&vm_numa_stat[item], 0);
68}
69
70static void invalid_numa_statistics(void)
71{
72 zero_zones_numa_counters();
73 zero_global_numa_counters();
74}
75
76static DEFINE_MUTEX(vm_numa_stat_lock);
77
78int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
79 void *buffer, size_t *length, loff_t *ppos)
80{
81 int ret, oldval;
82
83 mutex_lock(&vm_numa_stat_lock);
84 if (write)
85 oldval = sysctl_vm_numa_stat;
86 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
87 if (ret || !write)
88 goto out;
89
90 if (oldval == sysctl_vm_numa_stat)
91 goto out;
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");
95 } else {
96 static_branch_disable(&vm_numa_stat_key);
97 invalid_numa_statistics();
98 pr_info("disable numa statistics, and clear numa counters\n");
99 }
100
101out:
102 mutex_unlock(&vm_numa_stat_lock);
103 return ret;
104}
105#endif
106
107#ifdef CONFIG_VM_EVENT_COUNTERS
108DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
109EXPORT_PER_CPU_SYMBOL(vm_event_states);
110
111static void sum_vm_events(unsigned long *ret)
112{
113 int cpu;
114 int i;
115
116 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
117
118 for_each_online_cpu(cpu) {
119 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
120
121 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
122 ret[i] += this->event[i];
123 }
124}
125
126/*
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.
130*/
131void all_vm_events(unsigned long *ret)
132{
133 get_online_cpus();
134 sum_vm_events(ret);
135 put_online_cpus();
136}
137EXPORT_SYMBOL_GPL(all_vm_events);
138
139/*
140 * Fold the foreign cpu events into our own.
141 *
142 * This is adding to the events on one processor
143 * but keeps the global counts constant.
144 */
145void vm_events_fold_cpu(int cpu)
146{
147 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
148 int i;
149
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;
153 }
154}
155
156#endif /* CONFIG_VM_EVENT_COUNTERS */
157
158/*
159 * Manage combined zone based / global counters
160 *
161 * vm_stat contains the global counters
162 */
163atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
165atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
166EXPORT_SYMBOL(vm_zone_stat);
167EXPORT_SYMBOL(vm_numa_stat);
168EXPORT_SYMBOL(vm_node_stat);
169
170#ifdef CONFIG_SMP
171
172int calculate_pressure_threshold(struct zone *zone)
173{
174 int threshold;
175 int watermark_distance;
176
177 /*
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
183 * the min watermark
184 */
185 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
186 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
187
188 /*
189 * Maximum threshold is 125
190 */
191 threshold = min(125, threshold);
192
193 return threshold;
194}
195
196int calculate_normal_threshold(struct zone *zone)
197{
198 int threshold;
199 int mem; /* memory in 128 MB units */
200
201 /*
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.
206 *
207 * Some sample thresholds:
208 *
209 * Threshold Processors (fls) Zonesize fls(mem+1)
210 * ------------------------------------------------------------------
211 * 8 1 1 0.9-1 GB 4
212 * 16 2 2 0.9-1 GB 4
213 * 20 2 2 1-2 GB 5
214 * 24 2 2 2-4 GB 6
215 * 28 2 2 4-8 GB 7
216 * 32 2 2 8-16 GB 8
217 * 4 2 2 <128M 1
218 * 30 4 3 2-4 GB 5
219 * 48 4 3 8-16 GB 8
220 * 32 8 4 1-2 GB 4
221 * 32 8 4 0.9-1GB 4
222 * 10 16 5 <128M 1
223 * 40 16 5 900M 4
224 * 70 64 7 2-4 GB 5
225 * 84 64 7 4-8 GB 6
226 * 108 512 9 4-8 GB 6
227 * 125 1024 10 8-16 GB 8
228 * 125 1024 10 16-32 GB 9
229 */
230
231 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
232
233 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
234
235 /*
236 * Maximum threshold is 125
237 */
238 threshold = min(125, threshold);
239
240 return threshold;
241}
242
243/*
244 * Refresh the thresholds for each zone.
245 */
246void refresh_zone_stat_thresholds(void)
247{
248 struct pglist_data *pgdat;
249 struct zone *zone;
250 int cpu;
251 int threshold;
252
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;
257 }
258 }
259
260 for_each_populated_zone(zone) {
261 struct pglist_data *pgdat = zone->zone_pgdat;
262 unsigned long max_drift, tolerate_drift;
263
264 threshold = calculate_normal_threshold(zone);
265
266 for_each_online_cpu(cpu) {
267 int pgdat_threshold;
268
269 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
270 = threshold;
271
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);
276 }
277
278 /*
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
282 */
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) +
287 max_drift;
288 }
289}
290
291void set_pgdat_percpu_threshold(pg_data_t *pgdat,
292 int (*calculate_pressure)(struct zone *))
293{
294 struct zone *zone;
295 int cpu;
296 int threshold;
297 int i;
298
299 for (i = 0; i < pgdat->nr_zones; i++) {
300 zone = &pgdat->node_zones[i];
301 if (!zone->percpu_drift_mark)
302 continue;
303
304 threshold = (*calculate_pressure)(zone);
305 for_each_online_cpu(cpu)
306 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
307 = threshold;
308 }
309}
310
311/*
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.
315 */
316void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
317 long delta)
318{
319 struct per_cpu_pageset __percpu *pcp = zone->pageset;
320 s8 __percpu *p = pcp->vm_stat_diff + item;
321 long x;
322 long t;
323
324 x = delta + __this_cpu_read(*p);
325
326 t = __this_cpu_read(pcp->stat_threshold);
327
328 if (unlikely(x > t || x < -t)) {
329 zone_page_state_add(x, zone, item);
330 x = 0;
331 }
332 __this_cpu_write(*p, x);
333}
334EXPORT_SYMBOL(__mod_zone_page_state);
335
336void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
337 long delta)
338{
339 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
340 s8 __percpu *p = pcp->vm_node_stat_diff + item;
341 long x;
342 long t;
343
344 if (vmstat_item_in_bytes(item)) {
345 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
346 delta >>= PAGE_SHIFT;
347 }
348
349 x = delta + __this_cpu_read(*p);
350
351 t = __this_cpu_read(pcp->stat_threshold);
352
353 if (unlikely(x > t || x < -t)) {
354 node_page_state_add(x, pgdat, item);
355 x = 0;
356 }
357 __this_cpu_write(*p, x);
358}
359EXPORT_SYMBOL(__mod_node_page_state);
360
361/*
362 * Optimized increment and decrement functions.
363 *
364 * These are only for a single page and therefore can take a struct page *
365 * argument instead of struct zone *. This allows the inclusion of the code
366 * generated for page_zone(page) into the optimized functions.
367 *
368 * No overflow check is necessary and therefore the differential can be
369 * incremented or decremented in place which may allow the compilers to
370 * generate better code.
371 * The increment or decrement is known and therefore one boundary check can
372 * be omitted.
373 *
374 * NOTE: These functions are very performance sensitive. Change only
375 * with care.
376 *
377 * Some processors have inc/dec instructions that are atomic vs an interrupt.
378 * However, the code must first determine the differential location in a zone
379 * based on the processor number and then inc/dec the counter. There is no
380 * guarantee without disabling preemption that the processor will not change
381 * in between and therefore the atomicity vs. interrupt cannot be exploited
382 * in a useful way here.
383 */
384void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
385{
386 struct per_cpu_pageset __percpu *pcp = zone->pageset;
387 s8 __percpu *p = pcp->vm_stat_diff + item;
388 s8 v, t;
389
390 v = __this_cpu_inc_return(*p);
391 t = __this_cpu_read(pcp->stat_threshold);
392 if (unlikely(v > t)) {
393 s8 overstep = t >> 1;
394
395 zone_page_state_add(v + overstep, zone, item);
396 __this_cpu_write(*p, -overstep);
397 }
398}
399
400void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
401{
402 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
403 s8 __percpu *p = pcp->vm_node_stat_diff + item;
404 s8 v, t;
405
406 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
407
408 v = __this_cpu_inc_return(*p);
409 t = __this_cpu_read(pcp->stat_threshold);
410 if (unlikely(v > t)) {
411 s8 overstep = t >> 1;
412
413 node_page_state_add(v + overstep, pgdat, item);
414 __this_cpu_write(*p, -overstep);
415 }
416}
417
418void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
419{
420 __inc_zone_state(page_zone(page), item);
421}
422EXPORT_SYMBOL(__inc_zone_page_state);
423
424void __inc_node_page_state(struct page *page, enum node_stat_item item)
425{
426 __inc_node_state(page_pgdat(page), item);
427}
428EXPORT_SYMBOL(__inc_node_page_state);
429
430void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
431{
432 struct per_cpu_pageset __percpu *pcp = zone->pageset;
433 s8 __percpu *p = pcp->vm_stat_diff + item;
434 s8 v, t;
435
436 v = __this_cpu_dec_return(*p);
437 t = __this_cpu_read(pcp->stat_threshold);
438 if (unlikely(v < - t)) {
439 s8 overstep = t >> 1;
440
441 zone_page_state_add(v - overstep, zone, item);
442 __this_cpu_write(*p, overstep);
443 }
444}
445
446void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
447{
448 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
449 s8 __percpu *p = pcp->vm_node_stat_diff + item;
450 s8 v, t;
451
452 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
453
454 v = __this_cpu_dec_return(*p);
455 t = __this_cpu_read(pcp->stat_threshold);
456 if (unlikely(v < - t)) {
457 s8 overstep = t >> 1;
458
459 node_page_state_add(v - overstep, pgdat, item);
460 __this_cpu_write(*p, overstep);
461 }
462}
463
464void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
465{
466 __dec_zone_state(page_zone(page), item);
467}
468EXPORT_SYMBOL(__dec_zone_page_state);
469
470void __dec_node_page_state(struct page *page, enum node_stat_item item)
471{
472 __dec_node_state(page_pgdat(page), item);
473}
474EXPORT_SYMBOL(__dec_node_page_state);
475
476#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
477/*
478 * If we have cmpxchg_local support then we do not need to incur the overhead
479 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
480 *
481 * mod_state() modifies the zone counter state through atomic per cpu
482 * operations.
483 *
484 * Overstep mode specifies how overstep should handled:
485 * 0 No overstepping
486 * 1 Overstepping half of threshold
487 * -1 Overstepping minus half of threshold
488*/
489static inline void mod_zone_state(struct zone *zone,
490 enum zone_stat_item item, long delta, int overstep_mode)
491{
492 struct per_cpu_pageset __percpu *pcp = zone->pageset;
493 s8 __percpu *p = pcp->vm_stat_diff + item;
494 long o, n, t, z;
495
496 do {
497 z = 0; /* overflow to zone counters */
498
499 /*
500 * The fetching of the stat_threshold is racy. We may apply
501 * a counter threshold to the wrong the cpu if we get
502 * rescheduled while executing here. However, the next
503 * counter update will apply the threshold again and
504 * therefore bring the counter under the threshold again.
505 *
506 * Most of the time the thresholds are the same anyways
507 * for all cpus in a zone.
508 */
509 t = this_cpu_read(pcp->stat_threshold);
510
511 o = this_cpu_read(*p);
512 n = delta + o;
513
514 if (n > t || n < -t) {
515 int os = overstep_mode * (t >> 1) ;
516
517 /* Overflow must be added to zone counters */
518 z = n + os;
519 n = -os;
520 }
521 } while (this_cpu_cmpxchg(*p, o, n) != o);
522
523 if (z)
524 zone_page_state_add(z, zone, item);
525}
526
527void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
528 long delta)
529{
530 mod_zone_state(zone, item, delta, 0);
531}
532EXPORT_SYMBOL(mod_zone_page_state);
533
534void inc_zone_page_state(struct page *page, enum zone_stat_item item)
535{
536 mod_zone_state(page_zone(page), item, 1, 1);
537}
538EXPORT_SYMBOL(inc_zone_page_state);
539
540void dec_zone_page_state(struct page *page, enum zone_stat_item item)
541{
542 mod_zone_state(page_zone(page), item, -1, -1);
543}
544EXPORT_SYMBOL(dec_zone_page_state);
545
546static inline void mod_node_state(struct pglist_data *pgdat,
547 enum node_stat_item item, int delta, int overstep_mode)
548{
549 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
550 s8 __percpu *p = pcp->vm_node_stat_diff + item;
551 long o, n, t, z;
552
553 if (vmstat_item_in_bytes(item)) {
554 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
555 delta >>= PAGE_SHIFT;
556 }
557
558 do {
559 z = 0; /* overflow to node counters */
560
561 /*
562 * The fetching of the stat_threshold is racy. We may apply
563 * a counter threshold to the wrong the cpu if we get
564 * rescheduled while executing here. However, the next
565 * counter update will apply the threshold again and
566 * therefore bring the counter under the threshold again.
567 *
568 * Most of the time the thresholds are the same anyways
569 * for all cpus in a node.
570 */
571 t = this_cpu_read(pcp->stat_threshold);
572
573 o = this_cpu_read(*p);
574 n = delta + o;
575
576 if (n > t || n < -t) {
577 int os = overstep_mode * (t >> 1) ;
578
579 /* Overflow must be added to node counters */
580 z = n + os;
581 n = -os;
582 }
583 } while (this_cpu_cmpxchg(*p, o, n) != o);
584
585 if (z)
586 node_page_state_add(z, pgdat, item);
587}
588
589void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
590 long delta)
591{
592 mod_node_state(pgdat, item, delta, 0);
593}
594EXPORT_SYMBOL(mod_node_page_state);
595
596void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
597{
598 mod_node_state(pgdat, item, 1, 1);
599}
600
601void inc_node_page_state(struct page *page, enum node_stat_item item)
602{
603 mod_node_state(page_pgdat(page), item, 1, 1);
604}
605EXPORT_SYMBOL(inc_node_page_state);
606
607void dec_node_page_state(struct page *page, enum node_stat_item item)
608{
609 mod_node_state(page_pgdat(page), item, -1, -1);
610}
611EXPORT_SYMBOL(dec_node_page_state);
612#else
613/*
614 * Use interrupt disable to serialize counter updates
615 */
616void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
617 long delta)
618{
619 unsigned long flags;
620
621 local_irq_save(flags);
622 __mod_zone_page_state(zone, item, delta);
623 local_irq_restore(flags);
624}
625EXPORT_SYMBOL(mod_zone_page_state);
626
627void inc_zone_page_state(struct page *page, enum zone_stat_item item)
628{
629 unsigned long flags;
630 struct zone *zone;
631
632 zone = page_zone(page);
633 local_irq_save(flags);
634 __inc_zone_state(zone, item);
635 local_irq_restore(flags);
636}
637EXPORT_SYMBOL(inc_zone_page_state);
638
639void dec_zone_page_state(struct page *page, enum zone_stat_item item)
640{
641 unsigned long flags;
642
643 local_irq_save(flags);
644 __dec_zone_page_state(page, item);
645 local_irq_restore(flags);
646}
647EXPORT_SYMBOL(dec_zone_page_state);
648
649void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
650{
651 unsigned long flags;
652
653 local_irq_save(flags);
654 __inc_node_state(pgdat, item);
655 local_irq_restore(flags);
656}
657EXPORT_SYMBOL(inc_node_state);
658
659void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
660 long delta)
661{
662 unsigned long flags;
663
664 local_irq_save(flags);
665 __mod_node_page_state(pgdat, item, delta);
666 local_irq_restore(flags);
667}
668EXPORT_SYMBOL(mod_node_page_state);
669
670void inc_node_page_state(struct page *page, enum node_stat_item item)
671{
672 unsigned long flags;
673 struct pglist_data *pgdat;
674
675 pgdat = page_pgdat(page);
676 local_irq_save(flags);
677 __inc_node_state(pgdat, item);
678 local_irq_restore(flags);
679}
680EXPORT_SYMBOL(inc_node_page_state);
681
682void dec_node_page_state(struct page *page, enum node_stat_item item)
683{
684 unsigned long flags;
685
686 local_irq_save(flags);
687 __dec_node_page_state(page, item);
688 local_irq_restore(flags);
689}
690EXPORT_SYMBOL(dec_node_page_state);
691#endif
692
693/*
694 * Fold a differential into the global counters.
695 * Returns the number of counters updated.
696 */
697#ifdef CONFIG_NUMA
698static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
699{
700 int i;
701 int changes = 0;
702
703 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
704 if (zone_diff[i]) {
705 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
706 changes++;
707 }
708
709 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
710 if (numa_diff[i]) {
711 atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
712 changes++;
713 }
714
715 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
716 if (node_diff[i]) {
717 atomic_long_add(node_diff[i], &vm_node_stat[i]);
718 changes++;
719 }
720 return changes;
721}
722#else
723static int fold_diff(int *zone_diff, int *node_diff)
724{
725 int i;
726 int changes = 0;
727
728 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
729 if (zone_diff[i]) {
730 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
731 changes++;
732 }
733
734 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
735 if (node_diff[i]) {
736 atomic_long_add(node_diff[i], &vm_node_stat[i]);
737 changes++;
738 }
739 return changes;
740}
741#endif /* CONFIG_NUMA */
742
743/*
744 * Update the zone counters for the current cpu.
745 *
746 * Note that refresh_cpu_vm_stats strives to only access
747 * node local memory. The per cpu pagesets on remote zones are placed
748 * in the memory local to the processor using that pageset. So the
749 * loop over all zones will access a series of cachelines local to
750 * the processor.
751 *
752 * The call to zone_page_state_add updates the cachelines with the
753 * statistics in the remote zone struct as well as the global cachelines
754 * with the global counters. These could cause remote node cache line
755 * bouncing and will have to be only done when necessary.
756 *
757 * The function returns the number of global counters updated.
758 */
759static int refresh_cpu_vm_stats(bool do_pagesets)
760{
761 struct pglist_data *pgdat;
762 struct zone *zone;
763 int i;
764 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
765#ifdef CONFIG_NUMA
766 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
767#endif
768 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
769 int changes = 0;
770
771 for_each_populated_zone(zone) {
772 struct per_cpu_pageset __percpu *p = zone->pageset;
773
774 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
775 int v;
776
777 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
778 if (v) {
779
780 atomic_long_add(v, &zone->vm_stat[i]);
781 global_zone_diff[i] += v;
782#ifdef CONFIG_NUMA
783 /* 3 seconds idle till flush */
784 __this_cpu_write(p->expire, 3);
785#endif
786 }
787 }
788#ifdef CONFIG_NUMA
789 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
790 int v;
791
792 v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
793 if (v) {
794
795 atomic_long_add(v, &zone->vm_numa_stat[i]);
796 global_numa_diff[i] += v;
797 __this_cpu_write(p->expire, 3);
798 }
799 }
800
801 if (do_pagesets) {
802 cond_resched();
803 /*
804 * Deal with draining the remote pageset of this
805 * processor
806 *
807 * Check if there are pages remaining in this pageset
808 * if not then there is nothing to expire.
809 */
810 if (!__this_cpu_read(p->expire) ||
811 !__this_cpu_read(p->pcp.count))
812 continue;
813
814 /*
815 * We never drain zones local to this processor.
816 */
817 if (zone_to_nid(zone) == numa_node_id()) {
818 __this_cpu_write(p->expire, 0);
819 continue;
820 }
821
822 if (__this_cpu_dec_return(p->expire))
823 continue;
824
825 if (__this_cpu_read(p->pcp.count)) {
826 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
827 changes++;
828 }
829 }
830#endif
831 }
832
833 for_each_online_pgdat(pgdat) {
834 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
835
836 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
837 int v;
838
839 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
840 if (v) {
841 atomic_long_add(v, &pgdat->vm_stat[i]);
842 global_node_diff[i] += v;
843 }
844 }
845 }
846
847#ifdef CONFIG_NUMA
848 changes += fold_diff(global_zone_diff, global_numa_diff,
849 global_node_diff);
850#else
851 changes += fold_diff(global_zone_diff, global_node_diff);
852#endif
853 return changes;
854}
855
856/*
857 * Fold the data for an offline cpu into the global array.
858 * There cannot be any access by the offline cpu and therefore
859 * synchronization is simplified.
860 */
861void cpu_vm_stats_fold(int cpu)
862{
863 struct pglist_data *pgdat;
864 struct zone *zone;
865 int i;
866 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
867#ifdef CONFIG_NUMA
868 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
869#endif
870 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
871
872 for_each_populated_zone(zone) {
873 struct per_cpu_pageset *p;
874
875 p = per_cpu_ptr(zone->pageset, cpu);
876
877 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
878 if (p->vm_stat_diff[i]) {
879 int v;
880
881 v = p->vm_stat_diff[i];
882 p->vm_stat_diff[i] = 0;
883 atomic_long_add(v, &zone->vm_stat[i]);
884 global_zone_diff[i] += v;
885 }
886
887#ifdef CONFIG_NUMA
888 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
889 if (p->vm_numa_stat_diff[i]) {
890 int v;
891
892 v = p->vm_numa_stat_diff[i];
893 p->vm_numa_stat_diff[i] = 0;
894 atomic_long_add(v, &zone->vm_numa_stat[i]);
895 global_numa_diff[i] += v;
896 }
897#endif
898 }
899
900 for_each_online_pgdat(pgdat) {
901 struct per_cpu_nodestat *p;
902
903 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
904
905 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
906 if (p->vm_node_stat_diff[i]) {
907 int v;
908
909 v = p->vm_node_stat_diff[i];
910 p->vm_node_stat_diff[i] = 0;
911 atomic_long_add(v, &pgdat->vm_stat[i]);
912 global_node_diff[i] += v;
913 }
914 }
915
916#ifdef CONFIG_NUMA
917 fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
918#else
919 fold_diff(global_zone_diff, global_node_diff);
920#endif
921}
922
923/*
924 * this is only called if !populated_zone(zone), which implies no other users of
925 * pset->vm_stat_diff[] exsist.
926 */
927void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
928{
929 int i;
930
931 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
932 if (pset->vm_stat_diff[i]) {
933 int v = pset->vm_stat_diff[i];
934 pset->vm_stat_diff[i] = 0;
935 atomic_long_add(v, &zone->vm_stat[i]);
936 atomic_long_add(v, &vm_zone_stat[i]);
937 }
938
939#ifdef CONFIG_NUMA
940 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
941 if (pset->vm_numa_stat_diff[i]) {
942 int v = pset->vm_numa_stat_diff[i];
943
944 pset->vm_numa_stat_diff[i] = 0;
945 atomic_long_add(v, &zone->vm_numa_stat[i]);
946 atomic_long_add(v, &vm_numa_stat[i]);
947 }
948#endif
949}
950#endif
951
952#ifdef CONFIG_NUMA
953void __inc_numa_state(struct zone *zone,
954 enum numa_stat_item item)
955{
956 struct per_cpu_pageset __percpu *pcp = zone->pageset;
957 u16 __percpu *p = pcp->vm_numa_stat_diff + item;
958 u16 v;
959
960 v = __this_cpu_inc_return(*p);
961
962 if (unlikely(v > NUMA_STATS_THRESHOLD)) {
963 zone_numa_state_add(v, zone, item);
964 __this_cpu_write(*p, 0);
965 }
966}
967
968/*
969 * Determine the per node value of a stat item. This function
970 * is called frequently in a NUMA machine, so try to be as
971 * frugal as possible.
972 */
973unsigned long sum_zone_node_page_state(int node,
974 enum zone_stat_item item)
975{
976 struct zone *zones = NODE_DATA(node)->node_zones;
977 int i;
978 unsigned long count = 0;
979
980 for (i = 0; i < MAX_NR_ZONES; i++)
981 count += zone_page_state(zones + i, item);
982
983 return count;
984}
985
986/*
987 * Determine the per node value of a numa stat item. To avoid deviation,
988 * the per cpu stat number in vm_numa_stat_diff[] is also included.
989 */
990unsigned long sum_zone_numa_state(int node,
991 enum numa_stat_item item)
992{
993 struct zone *zones = NODE_DATA(node)->node_zones;
994 int i;
995 unsigned long count = 0;
996
997 for (i = 0; i < MAX_NR_ZONES; i++)
998 count += zone_numa_state_snapshot(zones + i, item);
999
1000 return count;
1001}
1002
1003/*
1004 * Determine the per node value of a stat item.
1005 */
1006unsigned long node_page_state_pages(struct pglist_data *pgdat,
1007 enum node_stat_item item)
1008{
1009 long x = atomic_long_read(&pgdat->vm_stat[item]);
1010#ifdef CONFIG_SMP
1011 if (x < 0)
1012 x = 0;
1013#endif
1014 return x;
1015}
1016
1017unsigned long node_page_state(struct pglist_data *pgdat,
1018 enum node_stat_item item)
1019{
1020 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1021
1022 return node_page_state_pages(pgdat, item);
1023}
1024#endif
1025
1026#ifdef CONFIG_COMPACTION
1027
1028struct contig_page_info {
1029 unsigned long free_pages;
1030 unsigned long free_blocks_total;
1031 unsigned long free_blocks_suitable;
1032};
1033
1034/*
1035 * Calculate the number of free pages in a zone, how many contiguous
1036 * pages are free and how many are large enough to satisfy an allocation of
1037 * the target size. Note that this function makes no attempt to estimate
1038 * how many suitable free blocks there *might* be if MOVABLE pages were
1039 * migrated. Calculating that is possible, but expensive and can be
1040 * figured out from userspace
1041 */
1042static void fill_contig_page_info(struct zone *zone,
1043 unsigned int suitable_order,
1044 struct contig_page_info *info)
1045{
1046 unsigned int order;
1047
1048 info->free_pages = 0;
1049 info->free_blocks_total = 0;
1050 info->free_blocks_suitable = 0;
1051
1052 for (order = 0; order < MAX_ORDER; order++) {
1053 unsigned long blocks;
1054
1055 /* Count number of free blocks */
1056 blocks = zone->free_area[order].nr_free;
1057 info->free_blocks_total += blocks;
1058
1059 /* Count free base pages */
1060 info->free_pages += blocks << order;
1061
1062 /* Count the suitable free blocks */
1063 if (order >= suitable_order)
1064 info->free_blocks_suitable += blocks <<
1065 (order - suitable_order);
1066 }
1067}
1068
1069/*
1070 * A fragmentation index only makes sense if an allocation of a requested
1071 * size would fail. If that is true, the fragmentation index indicates
1072 * whether external fragmentation or a lack of memory was the problem.
1073 * The value can be used to determine if page reclaim or compaction
1074 * should be used
1075 */
1076static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1077{
1078 unsigned long requested = 1UL << order;
1079
1080 if (WARN_ON_ONCE(order >= MAX_ORDER))
1081 return 0;
1082
1083 if (!info->free_blocks_total)
1084 return 0;
1085
1086 /* Fragmentation index only makes sense when a request would fail */
1087 if (info->free_blocks_suitable)
1088 return -1000;
1089
1090 /*
1091 * Index is between 0 and 1 so return within 3 decimal places
1092 *
1093 * 0 => allocation would fail due to lack of memory
1094 * 1 => allocation would fail due to fragmentation
1095 */
1096 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1097}
1098
1099/*
1100 * Calculates external fragmentation within a zone wrt the given order.
1101 * It is defined as the percentage of pages found in blocks of size
1102 * less than 1 << order. It returns values in range [0, 100].
1103 */
1104unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1105{
1106 struct contig_page_info info;
1107
1108 fill_contig_page_info(zone, order, &info);
1109 if (info.free_pages == 0)
1110 return 0;
1111
1112 return div_u64((info.free_pages -
1113 (info.free_blocks_suitable << order)) * 100,
1114 info.free_pages);
1115}
1116
1117/* Same as __fragmentation index but allocs contig_page_info on stack */
1118int fragmentation_index(struct zone *zone, unsigned int order)
1119{
1120 struct contig_page_info info;
1121
1122 fill_contig_page_info(zone, order, &info);
1123 return __fragmentation_index(order, &info);
1124}
1125#endif
1126
1127#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1128 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1129#ifdef CONFIG_ZONE_DMA
1130#define TEXT_FOR_DMA(xx) xx "_dma",
1131#else
1132#define TEXT_FOR_DMA(xx)
1133#endif
1134
1135#ifdef CONFIG_ZONE_DMA32
1136#define TEXT_FOR_DMA32(xx) xx "_dma32",
1137#else
1138#define TEXT_FOR_DMA32(xx)
1139#endif
1140
1141#ifdef CONFIG_HIGHMEM
1142#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1143#else
1144#define TEXT_FOR_HIGHMEM(xx)
1145#endif
1146
1147#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1148 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1149
1150const char * const vmstat_text[] = {
1151 /* enum zone_stat_item counters */
1152 "nr_free_pages",
1153 "nr_zone_inactive_anon",
1154 "nr_zone_active_anon",
1155 "nr_zone_inactive_file",
1156 "nr_zone_active_file",
1157 "nr_zone_unevictable",
1158 "nr_zone_write_pending",
1159 "nr_mlock",
1160 "nr_page_table_pages",
1161 "nr_bounce",
1162#if IS_ENABLED(CONFIG_ZSMALLOC)
1163 "nr_zspages",
1164#endif
1165 "nr_free_cma",
1166
1167 /* enum numa_stat_item counters */
1168#ifdef CONFIG_NUMA
1169 "numa_hit",
1170 "numa_miss",
1171 "numa_foreign",
1172 "numa_interleave",
1173 "numa_local",
1174 "numa_other",
1175#endif
1176
1177 /* enum node_stat_item counters */
1178 "nr_inactive_anon",
1179 "nr_active_anon",
1180 "nr_inactive_file",
1181 "nr_active_file",
1182 "nr_unevictable",
1183 "nr_slab_reclaimable",
1184 "nr_slab_unreclaimable",
1185 "nr_isolated_anon",
1186 "nr_isolated_file",
1187 "workingset_nodes",
1188 "workingset_refault_anon",
1189 "workingset_refault_file",
1190 "workingset_activate_anon",
1191 "workingset_activate_file",
1192 "workingset_restore_anon",
1193 "workingset_restore_file",
1194 "workingset_nodereclaim",
1195 "nr_anon_pages",
1196 "nr_mapped",
1197 "nr_file_pages",
1198 "nr_dirty",
1199 "nr_writeback",
1200 "nr_writeback_temp",
1201 "nr_shmem",
1202 "nr_shmem_hugepages",
1203 "nr_shmem_pmdmapped",
1204 "nr_file_hugepages",
1205 "nr_file_pmdmapped",
1206 "nr_anon_transparent_hugepages",
1207 "nr_vmscan_write",
1208 "nr_vmscan_immediate_reclaim",
1209 "nr_dirtied",
1210 "nr_written",
1211 "nr_kernel_misc_reclaimable",
1212 "nr_foll_pin_acquired",
1213 "nr_foll_pin_released",
1214 "nr_kernel_stack",
1215#if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1216 "nr_shadow_call_stack",
1217#endif
1218
1219 /* enum writeback_stat_item counters */
1220 "nr_dirty_threshold",
1221 "nr_dirty_background_threshold",
1222
1223#if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1224 /* enum vm_event_item counters */
1225 "pgpgin",
1226 "pgpgout",
1227 "pswpin",
1228 "pswpout",
1229
1230 TEXTS_FOR_ZONES("pgalloc")
1231 TEXTS_FOR_ZONES("allocstall")
1232 TEXTS_FOR_ZONES("pgskip")
1233
1234 "pgfree",
1235 "pgactivate",
1236 "pgdeactivate",
1237 "pglazyfree",
1238
1239 "pgfault",
1240 "pgmajfault",
1241 "pglazyfreed",
1242
1243 "pgrefill",
1244 "pgreuse",
1245 "pgsteal_kswapd",
1246 "pgsteal_direct",
1247 "pgscan_kswapd",
1248 "pgscan_direct",
1249 "pgscan_direct_throttle",
1250 "pgscan_anon",
1251 "pgscan_file",
1252 "pgsteal_anon",
1253 "pgsteal_file",
1254
1255#ifdef CONFIG_NUMA
1256 "zone_reclaim_failed",
1257#endif
1258 "pginodesteal",
1259 "slabs_scanned",
1260 "kswapd_inodesteal",
1261 "kswapd_low_wmark_hit_quickly",
1262 "kswapd_high_wmark_hit_quickly",
1263 "pageoutrun",
1264
1265 "pgrotated",
1266
1267 "drop_pagecache",
1268 "drop_slab",
1269 "oom_kill",
1270
1271#ifdef CONFIG_NUMA_BALANCING
1272 "numa_pte_updates",
1273 "numa_huge_pte_updates",
1274 "numa_hint_faults",
1275 "numa_hint_faults_local",
1276 "numa_pages_migrated",
1277#endif
1278#ifdef CONFIG_MIGRATION
1279 "pgmigrate_success",
1280 "pgmigrate_fail",
1281 "thp_migration_success",
1282 "thp_migration_fail",
1283 "thp_migration_split",
1284#endif
1285#ifdef CONFIG_COMPACTION
1286 "compact_migrate_scanned",
1287 "compact_free_scanned",
1288 "compact_isolated",
1289 "compact_stall",
1290 "compact_fail",
1291 "compact_success",
1292 "compact_daemon_wake",
1293 "compact_daemon_migrate_scanned",
1294 "compact_daemon_free_scanned",
1295#endif
1296
1297#ifdef CONFIG_HUGETLB_PAGE
1298 "htlb_buddy_alloc_success",
1299 "htlb_buddy_alloc_fail",
1300#endif
1301 "unevictable_pgs_culled",
1302 "unevictable_pgs_scanned",
1303 "unevictable_pgs_rescued",
1304 "unevictable_pgs_mlocked",
1305 "unevictable_pgs_munlocked",
1306 "unevictable_pgs_cleared",
1307 "unevictable_pgs_stranded",
1308
1309#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1310 "thp_fault_alloc",
1311 "thp_fault_fallback",
1312 "thp_fault_fallback_charge",
1313 "thp_collapse_alloc",
1314 "thp_collapse_alloc_failed",
1315 "thp_file_alloc",
1316 "thp_file_fallback",
1317 "thp_file_fallback_charge",
1318 "thp_file_mapped",
1319 "thp_split_page",
1320 "thp_split_page_failed",
1321 "thp_deferred_split_page",
1322 "thp_split_pmd",
1323#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1324 "thp_split_pud",
1325#endif
1326 "thp_zero_page_alloc",
1327 "thp_zero_page_alloc_failed",
1328 "thp_swpout",
1329 "thp_swpout_fallback",
1330#endif
1331#ifdef CONFIG_MEMORY_BALLOON
1332 "balloon_inflate",
1333 "balloon_deflate",
1334#ifdef CONFIG_BALLOON_COMPACTION
1335 "balloon_migrate",
1336#endif
1337#endif /* CONFIG_MEMORY_BALLOON */
1338#ifdef CONFIG_DEBUG_TLBFLUSH
1339 "nr_tlb_remote_flush",
1340 "nr_tlb_remote_flush_received",
1341 "nr_tlb_local_flush_all",
1342 "nr_tlb_local_flush_one",
1343#endif /* CONFIG_DEBUG_TLBFLUSH */
1344
1345#ifdef CONFIG_DEBUG_VM_VMACACHE
1346 "vmacache_find_calls",
1347 "vmacache_find_hits",
1348#endif
1349#ifdef CONFIG_SWAP
1350 "swap_ra",
1351 "swap_ra_hit",
1352#endif
1353#endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1354};
1355#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1356
1357#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1358 defined(CONFIG_PROC_FS)
1359static void *frag_start(struct seq_file *m, loff_t *pos)
1360{
1361 pg_data_t *pgdat;
1362 loff_t node = *pos;
1363
1364 for (pgdat = first_online_pgdat();
1365 pgdat && node;
1366 pgdat = next_online_pgdat(pgdat))
1367 --node;
1368
1369 return pgdat;
1370}
1371
1372static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1373{
1374 pg_data_t *pgdat = (pg_data_t *)arg;
1375
1376 (*pos)++;
1377 return next_online_pgdat(pgdat);
1378}
1379
1380static void frag_stop(struct seq_file *m, void *arg)
1381{
1382}
1383
1384/*
1385 * Walk zones in a node and print using a callback.
1386 * If @assert_populated is true, only use callback for zones that are populated.
1387 */
1388static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1389 bool assert_populated, bool nolock,
1390 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1391{
1392 struct zone *zone;
1393 struct zone *node_zones = pgdat->node_zones;
1394 unsigned long flags;
1395
1396 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1397 if (assert_populated && !populated_zone(zone))
1398 continue;
1399
1400 if (!nolock)
1401 spin_lock_irqsave(&zone->lock, flags);
1402 print(m, pgdat, zone);
1403 if (!nolock)
1404 spin_unlock_irqrestore(&zone->lock, flags);
1405 }
1406}
1407#endif
1408
1409#ifdef CONFIG_PROC_FS
1410static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1411 struct zone *zone)
1412{
1413 int order;
1414
1415 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1416 for (order = 0; order < MAX_ORDER; ++order)
1417 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1418 seq_putc(m, '\n');
1419}
1420
1421/*
1422 * This walks the free areas for each zone.
1423 */
1424static int frag_show(struct seq_file *m, void *arg)
1425{
1426 pg_data_t *pgdat = (pg_data_t *)arg;
1427 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1428 return 0;
1429}
1430
1431static void pagetypeinfo_showfree_print(struct seq_file *m,
1432 pg_data_t *pgdat, struct zone *zone)
1433{
1434 int order, mtype;
1435
1436 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1437 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1438 pgdat->node_id,
1439 zone->name,
1440 migratetype_names[mtype]);
1441 for (order = 0; order < MAX_ORDER; ++order) {
1442 unsigned long freecount = 0;
1443 struct free_area *area;
1444 struct list_head *curr;
1445 bool overflow = false;
1446
1447 area = &(zone->free_area[order]);
1448
1449 list_for_each(curr, &area->free_list[mtype]) {
1450 /*
1451 * Cap the free_list iteration because it might
1452 * be really large and we are under a spinlock
1453 * so a long time spent here could trigger a
1454 * hard lockup detector. Anyway this is a
1455 * debugging tool so knowing there is a handful
1456 * of pages of this order should be more than
1457 * sufficient.
1458 */
1459 if (++freecount >= 100000) {
1460 overflow = true;
1461 break;
1462 }
1463 }
1464 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1465 spin_unlock_irq(&zone->lock);
1466 cond_resched();
1467 spin_lock_irq(&zone->lock);
1468 }
1469 seq_putc(m, '\n');
1470 }
1471}
1472
1473/* Print out the free pages at each order for each migatetype */
1474static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1475{
1476 int order;
1477 pg_data_t *pgdat = (pg_data_t *)arg;
1478
1479 /* Print header */
1480 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1481 for (order = 0; order < MAX_ORDER; ++order)
1482 seq_printf(m, "%6d ", order);
1483 seq_putc(m, '\n');
1484
1485 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1486
1487 return 0;
1488}
1489
1490static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1491 pg_data_t *pgdat, struct zone *zone)
1492{
1493 int mtype;
1494 unsigned long pfn;
1495 unsigned long start_pfn = zone->zone_start_pfn;
1496 unsigned long end_pfn = zone_end_pfn(zone);
1497 unsigned long count[MIGRATE_TYPES] = { 0, };
1498
1499 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1500 struct page *page;
1501
1502 page = pfn_to_online_page(pfn);
1503 if (!page)
1504 continue;
1505
1506 /* Watch for unexpected holes punched in the memmap */
1507 if (!memmap_valid_within(pfn, page, zone))
1508 continue;
1509
1510 if (page_zone(page) != zone)
1511 continue;
1512
1513 mtype = get_pageblock_migratetype(page);
1514
1515 if (mtype < MIGRATE_TYPES)
1516 count[mtype]++;
1517 }
1518
1519 /* Print counts */
1520 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1521 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1522 seq_printf(m, "%12lu ", count[mtype]);
1523 seq_putc(m, '\n');
1524}
1525
1526/* Print out the number of pageblocks for each migratetype */
1527static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1528{
1529 int mtype;
1530 pg_data_t *pgdat = (pg_data_t *)arg;
1531
1532 seq_printf(m, "\n%-23s", "Number of blocks type ");
1533 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1534 seq_printf(m, "%12s ", migratetype_names[mtype]);
1535 seq_putc(m, '\n');
1536 walk_zones_in_node(m, pgdat, true, false,
1537 pagetypeinfo_showblockcount_print);
1538
1539 return 0;
1540}
1541
1542/*
1543 * Print out the number of pageblocks for each migratetype that contain pages
1544 * of other types. This gives an indication of how well fallbacks are being
1545 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1546 * to determine what is going on
1547 */
1548static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1549{
1550#ifdef CONFIG_PAGE_OWNER
1551 int mtype;
1552
1553 if (!static_branch_unlikely(&page_owner_inited))
1554 return;
1555
1556 drain_all_pages(NULL);
1557
1558 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1559 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1560 seq_printf(m, "%12s ", migratetype_names[mtype]);
1561 seq_putc(m, '\n');
1562
1563 walk_zones_in_node(m, pgdat, true, true,
1564 pagetypeinfo_showmixedcount_print);
1565#endif /* CONFIG_PAGE_OWNER */
1566}
1567
1568/*
1569 * This prints out statistics in relation to grouping pages by mobility.
1570 * It is expensive to collect so do not constantly read the file.
1571 */
1572static int pagetypeinfo_show(struct seq_file *m, void *arg)
1573{
1574 pg_data_t *pgdat = (pg_data_t *)arg;
1575
1576 /* check memoryless node */
1577 if (!node_state(pgdat->node_id, N_MEMORY))
1578 return 0;
1579
1580 seq_printf(m, "Page block order: %d\n", pageblock_order);
1581 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1582 seq_putc(m, '\n');
1583 pagetypeinfo_showfree(m, pgdat);
1584 pagetypeinfo_showblockcount(m, pgdat);
1585 pagetypeinfo_showmixedcount(m, pgdat);
1586
1587 return 0;
1588}
1589
1590static const struct seq_operations fragmentation_op = {
1591 .start = frag_start,
1592 .next = frag_next,
1593 .stop = frag_stop,
1594 .show = frag_show,
1595};
1596
1597static const struct seq_operations pagetypeinfo_op = {
1598 .start = frag_start,
1599 .next = frag_next,
1600 .stop = frag_stop,
1601 .show = pagetypeinfo_show,
1602};
1603
1604static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1605{
1606 int zid;
1607
1608 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1609 struct zone *compare = &pgdat->node_zones[zid];
1610
1611 if (populated_zone(compare))
1612 return zone == compare;
1613 }
1614
1615 return false;
1616}
1617
1618static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1619 struct zone *zone)
1620{
1621 int i;
1622 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1623 if (is_zone_first_populated(pgdat, zone)) {
1624 seq_printf(m, "\n per-node stats");
1625 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1626 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1627 node_page_state_pages(pgdat, i));
1628 }
1629 }
1630 seq_printf(m,
1631 "\n pages free %lu"
1632 "\n min %lu"
1633 "\n low %lu"
1634 "\n high %lu"
1635 "\n spanned %lu"
1636 "\n present %lu"
1637 "\n managed %lu",
1638 zone_page_state(zone, NR_FREE_PAGES),
1639 min_wmark_pages(zone),
1640 low_wmark_pages(zone),
1641 high_wmark_pages(zone),
1642 zone->spanned_pages,
1643 zone->present_pages,
1644 zone_managed_pages(zone));
1645
1646 seq_printf(m,
1647 "\n protection: (%ld",
1648 zone->lowmem_reserve[0]);
1649 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1650 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1651 seq_putc(m, ')');
1652
1653 /* If unpopulated, no other information is useful */
1654 if (!populated_zone(zone)) {
1655 seq_putc(m, '\n');
1656 return;
1657 }
1658
1659 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1660 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1661 zone_page_state(zone, i));
1662
1663#ifdef CONFIG_NUMA
1664 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1665 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1666 zone_numa_state_snapshot(zone, i));
1667#endif
1668
1669 seq_printf(m, "\n pagesets");
1670 for_each_online_cpu(i) {
1671 struct per_cpu_pageset *pageset;
1672
1673 pageset = per_cpu_ptr(zone->pageset, i);
1674 seq_printf(m,
1675 "\n cpu: %i"
1676 "\n count: %i"
1677 "\n high: %i"
1678 "\n batch: %i",
1679 i,
1680 pageset->pcp.count,
1681 pageset->pcp.high,
1682 pageset->pcp.batch);
1683#ifdef CONFIG_SMP
1684 seq_printf(m, "\n vm stats threshold: %d",
1685 pageset->stat_threshold);
1686#endif
1687 }
1688 seq_printf(m,
1689 "\n node_unreclaimable: %u"
1690 "\n start_pfn: %lu",
1691 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1692 zone->zone_start_pfn);
1693 seq_putc(m, '\n');
1694}
1695
1696/*
1697 * Output information about zones in @pgdat. All zones are printed regardless
1698 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1699 * set of all zones and userspace would not be aware of such zones if they are
1700 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1701 */
1702static int zoneinfo_show(struct seq_file *m, void *arg)
1703{
1704 pg_data_t *pgdat = (pg_data_t *)arg;
1705 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1706 return 0;
1707}
1708
1709static const struct seq_operations zoneinfo_op = {
1710 .start = frag_start, /* iterate over all zones. The same as in
1711 * fragmentation. */
1712 .next = frag_next,
1713 .stop = frag_stop,
1714 .show = zoneinfo_show,
1715};
1716
1717#define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1718 NR_VM_NUMA_STAT_ITEMS + \
1719 NR_VM_NODE_STAT_ITEMS + \
1720 NR_VM_WRITEBACK_STAT_ITEMS + \
1721 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1722 NR_VM_EVENT_ITEMS : 0))
1723
1724static void *vmstat_start(struct seq_file *m, loff_t *pos)
1725{
1726 unsigned long *v;
1727 int i;
1728
1729 if (*pos >= NR_VMSTAT_ITEMS)
1730 return NULL;
1731
1732 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1733 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1734 m->private = v;
1735 if (!v)
1736 return ERR_PTR(-ENOMEM);
1737 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1738 v[i] = global_zone_page_state(i);
1739 v += NR_VM_ZONE_STAT_ITEMS;
1740
1741#ifdef CONFIG_NUMA
1742 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1743 v[i] = global_numa_state(i);
1744 v += NR_VM_NUMA_STAT_ITEMS;
1745#endif
1746
1747 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1748 v[i] = global_node_page_state_pages(i);
1749 v += NR_VM_NODE_STAT_ITEMS;
1750
1751 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1752 v + NR_DIRTY_THRESHOLD);
1753 v += NR_VM_WRITEBACK_STAT_ITEMS;
1754
1755#ifdef CONFIG_VM_EVENT_COUNTERS
1756 all_vm_events(v);
1757 v[PGPGIN] /= 2; /* sectors -> kbytes */
1758 v[PGPGOUT] /= 2;
1759#endif
1760 return (unsigned long *)m->private + *pos;
1761}
1762
1763static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1764{
1765 (*pos)++;
1766 if (*pos >= NR_VMSTAT_ITEMS)
1767 return NULL;
1768 return (unsigned long *)m->private + *pos;
1769}
1770
1771static int vmstat_show(struct seq_file *m, void *arg)
1772{
1773 unsigned long *l = arg;
1774 unsigned long off = l - (unsigned long *)m->private;
1775
1776 seq_puts(m, vmstat_text[off]);
1777 seq_put_decimal_ull(m, " ", *l);
1778 seq_putc(m, '\n');
1779
1780 if (off == NR_VMSTAT_ITEMS - 1) {
1781 /*
1782 * We've come to the end - add any deprecated counters to avoid
1783 * breaking userspace which might depend on them being present.
1784 */
1785 seq_puts(m, "nr_unstable 0\n");
1786 }
1787 return 0;
1788}
1789
1790static void vmstat_stop(struct seq_file *m, void *arg)
1791{
1792 kfree(m->private);
1793 m->private = NULL;
1794}
1795
1796static const struct seq_operations vmstat_op = {
1797 .start = vmstat_start,
1798 .next = vmstat_next,
1799 .stop = vmstat_stop,
1800 .show = vmstat_show,
1801};
1802#endif /* CONFIG_PROC_FS */
1803
1804#ifdef CONFIG_SMP
1805static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1806int sysctl_stat_interval __read_mostly = HZ;
1807
1808#ifdef CONFIG_PROC_FS
1809static void refresh_vm_stats(struct work_struct *work)
1810{
1811 refresh_cpu_vm_stats(true);
1812}
1813
1814int vmstat_refresh(struct ctl_table *table, int write,
1815 void *buffer, size_t *lenp, loff_t *ppos)
1816{
1817 long val;
1818 int err;
1819 int i;
1820
1821 /*
1822 * The regular update, every sysctl_stat_interval, may come later
1823 * than expected: leaving a significant amount in per_cpu buckets.
1824 * This is particularly misleading when checking a quantity of HUGE
1825 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1826 * which can equally be echo'ed to or cat'ted from (by root),
1827 * can be used to update the stats just before reading them.
1828 *
1829 * Oh, and since global_zone_page_state() etc. are so careful to hide
1830 * transiently negative values, report an error here if any of
1831 * the stats is negative, so we know to go looking for imbalance.
1832 */
1833 err = schedule_on_each_cpu(refresh_vm_stats);
1834 if (err)
1835 return err;
1836 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1837 val = atomic_long_read(&vm_zone_stat[i]);
1838 if (val < 0) {
1839 pr_warn("%s: %s %ld\n",
1840 __func__, zone_stat_name(i), val);
1841 err = -EINVAL;
1842 }
1843 }
1844#ifdef CONFIG_NUMA
1845 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1846 val = atomic_long_read(&vm_numa_stat[i]);
1847 if (val < 0) {
1848 pr_warn("%s: %s %ld\n",
1849 __func__, numa_stat_name(i), val);
1850 err = -EINVAL;
1851 }
1852 }
1853#endif
1854 if (err)
1855 return err;
1856 if (write)
1857 *ppos += *lenp;
1858 else
1859 *lenp = 0;
1860 return 0;
1861}
1862#endif /* CONFIG_PROC_FS */
1863
1864static void vmstat_update(struct work_struct *w)
1865{
1866 if (refresh_cpu_vm_stats(true)) {
1867 /*
1868 * Counters were updated so we expect more updates
1869 * to occur in the future. Keep on running the
1870 * update worker thread.
1871 */
1872 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1873 this_cpu_ptr(&vmstat_work),
1874 round_jiffies_relative(sysctl_stat_interval));
1875 }
1876}
1877
1878/*
1879 * Switch off vmstat processing and then fold all the remaining differentials
1880 * until the diffs stay at zero. The function is used by NOHZ and can only be
1881 * invoked when tick processing is not active.
1882 */
1883/*
1884 * Check if the diffs for a certain cpu indicate that
1885 * an update is needed.
1886 */
1887static bool need_update(int cpu)
1888{
1889 struct zone *zone;
1890
1891 for_each_populated_zone(zone) {
1892 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1893
1894 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1895#ifdef CONFIG_NUMA
1896 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1897#endif
1898
1899 /*
1900 * The fast way of checking if there are any vmstat diffs.
1901 */
1902 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1903 sizeof(p->vm_stat_diff[0])))
1904 return true;
1905#ifdef CONFIG_NUMA
1906 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1907 sizeof(p->vm_numa_stat_diff[0])))
1908 return true;
1909#endif
1910 }
1911 return false;
1912}
1913
1914/*
1915 * Switch off vmstat processing and then fold all the remaining differentials
1916 * until the diffs stay at zero. The function is used by NOHZ and can only be
1917 * invoked when tick processing is not active.
1918 */
1919void quiet_vmstat(void)
1920{
1921 if (system_state != SYSTEM_RUNNING)
1922 return;
1923
1924 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1925 return;
1926
1927 if (!need_update(smp_processor_id()))
1928 return;
1929
1930 /*
1931 * Just refresh counters and do not care about the pending delayed
1932 * vmstat_update. It doesn't fire that often to matter and canceling
1933 * it would be too expensive from this path.
1934 * vmstat_shepherd will take care about that for us.
1935 */
1936 refresh_cpu_vm_stats(false);
1937}
1938
1939/*
1940 * Shepherd worker thread that checks the
1941 * differentials of processors that have their worker
1942 * threads for vm statistics updates disabled because of
1943 * inactivity.
1944 */
1945static void vmstat_shepherd(struct work_struct *w);
1946
1947static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1948
1949static void vmstat_shepherd(struct work_struct *w)
1950{
1951 int cpu;
1952
1953 get_online_cpus();
1954 /* Check processors whose vmstat worker threads have been disabled */
1955 for_each_online_cpu(cpu) {
1956 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1957
1958 if (!delayed_work_pending(dw) && need_update(cpu))
1959 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1960 }
1961 put_online_cpus();
1962
1963 schedule_delayed_work(&shepherd,
1964 round_jiffies_relative(sysctl_stat_interval));
1965}
1966
1967static void __init start_shepherd_timer(void)
1968{
1969 int cpu;
1970
1971 for_each_possible_cpu(cpu)
1972 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1973 vmstat_update);
1974
1975 schedule_delayed_work(&shepherd,
1976 round_jiffies_relative(sysctl_stat_interval));
1977}
1978
1979static void __init init_cpu_node_state(void)
1980{
1981 int node;
1982
1983 for_each_online_node(node) {
1984 if (cpumask_weight(cpumask_of_node(node)) > 0)
1985 node_set_state(node, N_CPU);
1986 }
1987}
1988
1989static int vmstat_cpu_online(unsigned int cpu)
1990{
1991 refresh_zone_stat_thresholds();
1992 node_set_state(cpu_to_node(cpu), N_CPU);
1993 return 0;
1994}
1995
1996static int vmstat_cpu_down_prep(unsigned int cpu)
1997{
1998 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1999 return 0;
2000}
2001
2002static int vmstat_cpu_dead(unsigned int cpu)
2003{
2004 const struct cpumask *node_cpus;
2005 int node;
2006
2007 node = cpu_to_node(cpu);
2008
2009 refresh_zone_stat_thresholds();
2010 node_cpus = cpumask_of_node(node);
2011 if (cpumask_weight(node_cpus) > 0)
2012 return 0;
2013
2014 node_clear_state(node, N_CPU);
2015 return 0;
2016}
2017
2018#endif
2019
2020struct workqueue_struct *mm_percpu_wq;
2021
2022void __init init_mm_internals(void)
2023{
2024 int ret __maybe_unused;
2025
2026 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2027
2028#ifdef CONFIG_SMP
2029 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2030 NULL, vmstat_cpu_dead);
2031 if (ret < 0)
2032 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2033
2034 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2035 vmstat_cpu_online,
2036 vmstat_cpu_down_prep);
2037 if (ret < 0)
2038 pr_err("vmstat: failed to register 'online' hotplug state\n");
2039
2040 get_online_cpus();
2041 init_cpu_node_state();
2042 put_online_cpus();
2043
2044 start_shepherd_timer();
2045#endif
2046#ifdef CONFIG_PROC_FS
2047 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2048 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2049 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2050 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2051#endif
2052}
2053
2054#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2055
2056/*
2057 * Return an index indicating how much of the available free memory is
2058 * unusable for an allocation of the requested size.
2059 */
2060static int unusable_free_index(unsigned int order,
2061 struct contig_page_info *info)
2062{
2063 /* No free memory is interpreted as all free memory is unusable */
2064 if (info->free_pages == 0)
2065 return 1000;
2066
2067 /*
2068 * Index should be a value between 0 and 1. Return a value to 3
2069 * decimal places.
2070 *
2071 * 0 => no fragmentation
2072 * 1 => high fragmentation
2073 */
2074 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2075
2076}
2077
2078static void unusable_show_print(struct seq_file *m,
2079 pg_data_t *pgdat, struct zone *zone)
2080{
2081 unsigned int order;
2082 int index;
2083 struct contig_page_info info;
2084
2085 seq_printf(m, "Node %d, zone %8s ",
2086 pgdat->node_id,
2087 zone->name);
2088 for (order = 0; order < MAX_ORDER; ++order) {
2089 fill_contig_page_info(zone, order, &info);
2090 index = unusable_free_index(order, &info);
2091 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2092 }
2093
2094 seq_putc(m, '\n');
2095}
2096
2097/*
2098 * Display unusable free space index
2099 *
2100 * The unusable free space index measures how much of the available free
2101 * memory cannot be used to satisfy an allocation of a given size and is a
2102 * value between 0 and 1. The higher the value, the more of free memory is
2103 * unusable and by implication, the worse the external fragmentation is. This
2104 * can be expressed as a percentage by multiplying by 100.
2105 */
2106static int unusable_show(struct seq_file *m, void *arg)
2107{
2108 pg_data_t *pgdat = (pg_data_t *)arg;
2109
2110 /* check memoryless node */
2111 if (!node_state(pgdat->node_id, N_MEMORY))
2112 return 0;
2113
2114 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2115
2116 return 0;
2117}
2118
2119static const struct seq_operations unusable_sops = {
2120 .start = frag_start,
2121 .next = frag_next,
2122 .stop = frag_stop,
2123 .show = unusable_show,
2124};
2125
2126DEFINE_SEQ_ATTRIBUTE(unusable);
2127
2128static void extfrag_show_print(struct seq_file *m,
2129 pg_data_t *pgdat, struct zone *zone)
2130{
2131 unsigned int order;
2132 int index;
2133
2134 /* Alloc on stack as interrupts are disabled for zone walk */
2135 struct contig_page_info info;
2136
2137 seq_printf(m, "Node %d, zone %8s ",
2138 pgdat->node_id,
2139 zone->name);
2140 for (order = 0; order < MAX_ORDER; ++order) {
2141 fill_contig_page_info(zone, order, &info);
2142 index = __fragmentation_index(order, &info);
2143 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2144 }
2145
2146 seq_putc(m, '\n');
2147}
2148
2149/*
2150 * Display fragmentation index for orders that allocations would fail for
2151 */
2152static int extfrag_show(struct seq_file *m, void *arg)
2153{
2154 pg_data_t *pgdat = (pg_data_t *)arg;
2155
2156 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2157
2158 return 0;
2159}
2160
2161static const struct seq_operations extfrag_sops = {
2162 .start = frag_start,
2163 .next = frag_next,
2164 .stop = frag_stop,
2165 .show = extfrag_show,
2166};
2167
2168DEFINE_SEQ_ATTRIBUTE(extfrag);
2169
2170static int __init extfrag_debug_init(void)
2171{
2172 struct dentry *extfrag_debug_root;
2173
2174 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2175
2176 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2177 &unusable_fops);
2178
2179 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2180 &extfrag_fops);
2181
2182 return 0;
2183}
2184
2185module_init(extfrag_debug_init);
2186#endif
1/*
2 * linux/mm/vmstat.c
3 *
4 * Manages VM statistics
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * zoned VM statistics
8 * Copyright (C) 2006 Silicon Graphics, Inc.,
9 * Christoph Lameter <christoph@lameter.com>
10 */
11#include <linux/fs.h>
12#include <linux/mm.h>
13#include <linux/err.h>
14#include <linux/module.h>
15#include <linux/slab.h>
16#include <linux/cpu.h>
17#include <linux/vmstat.h>
18#include <linux/sched.h>
19#include <linux/math64.h>
20#include <linux/writeback.h>
21#include <linux/compaction.h>
22
23#ifdef CONFIG_VM_EVENT_COUNTERS
24DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
25EXPORT_PER_CPU_SYMBOL(vm_event_states);
26
27static void sum_vm_events(unsigned long *ret)
28{
29 int cpu;
30 int i;
31
32 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
33
34 for_each_online_cpu(cpu) {
35 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
36
37 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
38 ret[i] += this->event[i];
39 }
40}
41
42/*
43 * Accumulate the vm event counters across all CPUs.
44 * The result is unavoidably approximate - it can change
45 * during and after execution of this function.
46*/
47void all_vm_events(unsigned long *ret)
48{
49 get_online_cpus();
50 sum_vm_events(ret);
51 put_online_cpus();
52}
53EXPORT_SYMBOL_GPL(all_vm_events);
54
55#ifdef CONFIG_HOTPLUG
56/*
57 * Fold the foreign cpu events into our own.
58 *
59 * This is adding to the events on one processor
60 * but keeps the global counts constant.
61 */
62void vm_events_fold_cpu(int cpu)
63{
64 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
65 int i;
66
67 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
68 count_vm_events(i, fold_state->event[i]);
69 fold_state->event[i] = 0;
70 }
71}
72#endif /* CONFIG_HOTPLUG */
73
74#endif /* CONFIG_VM_EVENT_COUNTERS */
75
76/*
77 * Manage combined zone based / global counters
78 *
79 * vm_stat contains the global counters
80 */
81atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
82EXPORT_SYMBOL(vm_stat);
83
84#ifdef CONFIG_SMP
85
86int calculate_pressure_threshold(struct zone *zone)
87{
88 int threshold;
89 int watermark_distance;
90
91 /*
92 * As vmstats are not up to date, there is drift between the estimated
93 * and real values. For high thresholds and a high number of CPUs, it
94 * is possible for the min watermark to be breached while the estimated
95 * value looks fine. The pressure threshold is a reduced value such
96 * that even the maximum amount of drift will not accidentally breach
97 * the min watermark
98 */
99 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
100 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
101
102 /*
103 * Maximum threshold is 125
104 */
105 threshold = min(125, threshold);
106
107 return threshold;
108}
109
110int calculate_normal_threshold(struct zone *zone)
111{
112 int threshold;
113 int mem; /* memory in 128 MB units */
114
115 /*
116 * The threshold scales with the number of processors and the amount
117 * of memory per zone. More memory means that we can defer updates for
118 * longer, more processors could lead to more contention.
119 * fls() is used to have a cheap way of logarithmic scaling.
120 *
121 * Some sample thresholds:
122 *
123 * Threshold Processors (fls) Zonesize fls(mem+1)
124 * ------------------------------------------------------------------
125 * 8 1 1 0.9-1 GB 4
126 * 16 2 2 0.9-1 GB 4
127 * 20 2 2 1-2 GB 5
128 * 24 2 2 2-4 GB 6
129 * 28 2 2 4-8 GB 7
130 * 32 2 2 8-16 GB 8
131 * 4 2 2 <128M 1
132 * 30 4 3 2-4 GB 5
133 * 48 4 3 8-16 GB 8
134 * 32 8 4 1-2 GB 4
135 * 32 8 4 0.9-1GB 4
136 * 10 16 5 <128M 1
137 * 40 16 5 900M 4
138 * 70 64 7 2-4 GB 5
139 * 84 64 7 4-8 GB 6
140 * 108 512 9 4-8 GB 6
141 * 125 1024 10 8-16 GB 8
142 * 125 1024 10 16-32 GB 9
143 */
144
145 mem = zone->present_pages >> (27 - PAGE_SHIFT);
146
147 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
148
149 /*
150 * Maximum threshold is 125
151 */
152 threshold = min(125, threshold);
153
154 return threshold;
155}
156
157/*
158 * Refresh the thresholds for each zone.
159 */
160void refresh_zone_stat_thresholds(void)
161{
162 struct zone *zone;
163 int cpu;
164 int threshold;
165
166 for_each_populated_zone(zone) {
167 unsigned long max_drift, tolerate_drift;
168
169 threshold = calculate_normal_threshold(zone);
170
171 for_each_online_cpu(cpu)
172 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
173 = threshold;
174
175 /*
176 * Only set percpu_drift_mark if there is a danger that
177 * NR_FREE_PAGES reports the low watermark is ok when in fact
178 * the min watermark could be breached by an allocation
179 */
180 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
181 max_drift = num_online_cpus() * threshold;
182 if (max_drift > tolerate_drift)
183 zone->percpu_drift_mark = high_wmark_pages(zone) +
184 max_drift;
185 }
186}
187
188void set_pgdat_percpu_threshold(pg_data_t *pgdat,
189 int (*calculate_pressure)(struct zone *))
190{
191 struct zone *zone;
192 int cpu;
193 int threshold;
194 int i;
195
196 for (i = 0; i < pgdat->nr_zones; i++) {
197 zone = &pgdat->node_zones[i];
198 if (!zone->percpu_drift_mark)
199 continue;
200
201 threshold = (*calculate_pressure)(zone);
202 for_each_possible_cpu(cpu)
203 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
204 = threshold;
205 }
206}
207
208/*
209 * For use when we know that interrupts are disabled.
210 */
211void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
212 int delta)
213{
214 struct per_cpu_pageset __percpu *pcp = zone->pageset;
215 s8 __percpu *p = pcp->vm_stat_diff + item;
216 long x;
217 long t;
218
219 x = delta + __this_cpu_read(*p);
220
221 t = __this_cpu_read(pcp->stat_threshold);
222
223 if (unlikely(x > t || x < -t)) {
224 zone_page_state_add(x, zone, item);
225 x = 0;
226 }
227 __this_cpu_write(*p, x);
228}
229EXPORT_SYMBOL(__mod_zone_page_state);
230
231/*
232 * Optimized increment and decrement functions.
233 *
234 * These are only for a single page and therefore can take a struct page *
235 * argument instead of struct zone *. This allows the inclusion of the code
236 * generated for page_zone(page) into the optimized functions.
237 *
238 * No overflow check is necessary and therefore the differential can be
239 * incremented or decremented in place which may allow the compilers to
240 * generate better code.
241 * The increment or decrement is known and therefore one boundary check can
242 * be omitted.
243 *
244 * NOTE: These functions are very performance sensitive. Change only
245 * with care.
246 *
247 * Some processors have inc/dec instructions that are atomic vs an interrupt.
248 * However, the code must first determine the differential location in a zone
249 * based on the processor number and then inc/dec the counter. There is no
250 * guarantee without disabling preemption that the processor will not change
251 * in between and therefore the atomicity vs. interrupt cannot be exploited
252 * in a useful way here.
253 */
254void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
255{
256 struct per_cpu_pageset __percpu *pcp = zone->pageset;
257 s8 __percpu *p = pcp->vm_stat_diff + item;
258 s8 v, t;
259
260 v = __this_cpu_inc_return(*p);
261 t = __this_cpu_read(pcp->stat_threshold);
262 if (unlikely(v > t)) {
263 s8 overstep = t >> 1;
264
265 zone_page_state_add(v + overstep, zone, item);
266 __this_cpu_write(*p, -overstep);
267 }
268}
269
270void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
271{
272 __inc_zone_state(page_zone(page), item);
273}
274EXPORT_SYMBOL(__inc_zone_page_state);
275
276void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
277{
278 struct per_cpu_pageset __percpu *pcp = zone->pageset;
279 s8 __percpu *p = pcp->vm_stat_diff + item;
280 s8 v, t;
281
282 v = __this_cpu_dec_return(*p);
283 t = __this_cpu_read(pcp->stat_threshold);
284 if (unlikely(v < - t)) {
285 s8 overstep = t >> 1;
286
287 zone_page_state_add(v - overstep, zone, item);
288 __this_cpu_write(*p, overstep);
289 }
290}
291
292void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
293{
294 __dec_zone_state(page_zone(page), item);
295}
296EXPORT_SYMBOL(__dec_zone_page_state);
297
298#ifdef CONFIG_CMPXCHG_LOCAL
299/*
300 * If we have cmpxchg_local support then we do not need to incur the overhead
301 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
302 *
303 * mod_state() modifies the zone counter state through atomic per cpu
304 * operations.
305 *
306 * Overstep mode specifies how overstep should handled:
307 * 0 No overstepping
308 * 1 Overstepping half of threshold
309 * -1 Overstepping minus half of threshold
310*/
311static inline void mod_state(struct zone *zone,
312 enum zone_stat_item item, int delta, int overstep_mode)
313{
314 struct per_cpu_pageset __percpu *pcp = zone->pageset;
315 s8 __percpu *p = pcp->vm_stat_diff + item;
316 long o, n, t, z;
317
318 do {
319 z = 0; /* overflow to zone counters */
320
321 /*
322 * The fetching of the stat_threshold is racy. We may apply
323 * a counter threshold to the wrong the cpu if we get
324 * rescheduled while executing here. However, the next
325 * counter update will apply the threshold again and
326 * therefore bring the counter under the threshold again.
327 *
328 * Most of the time the thresholds are the same anyways
329 * for all cpus in a zone.
330 */
331 t = this_cpu_read(pcp->stat_threshold);
332
333 o = this_cpu_read(*p);
334 n = delta + o;
335
336 if (n > t || n < -t) {
337 int os = overstep_mode * (t >> 1) ;
338
339 /* Overflow must be added to zone counters */
340 z = n + os;
341 n = -os;
342 }
343 } while (this_cpu_cmpxchg(*p, o, n) != o);
344
345 if (z)
346 zone_page_state_add(z, zone, item);
347}
348
349void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
350 int delta)
351{
352 mod_state(zone, item, delta, 0);
353}
354EXPORT_SYMBOL(mod_zone_page_state);
355
356void inc_zone_state(struct zone *zone, enum zone_stat_item item)
357{
358 mod_state(zone, item, 1, 1);
359}
360
361void inc_zone_page_state(struct page *page, enum zone_stat_item item)
362{
363 mod_state(page_zone(page), item, 1, 1);
364}
365EXPORT_SYMBOL(inc_zone_page_state);
366
367void dec_zone_page_state(struct page *page, enum zone_stat_item item)
368{
369 mod_state(page_zone(page), item, -1, -1);
370}
371EXPORT_SYMBOL(dec_zone_page_state);
372#else
373/*
374 * Use interrupt disable to serialize counter updates
375 */
376void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
377 int delta)
378{
379 unsigned long flags;
380
381 local_irq_save(flags);
382 __mod_zone_page_state(zone, item, delta);
383 local_irq_restore(flags);
384}
385EXPORT_SYMBOL(mod_zone_page_state);
386
387void inc_zone_state(struct zone *zone, enum zone_stat_item item)
388{
389 unsigned long flags;
390
391 local_irq_save(flags);
392 __inc_zone_state(zone, item);
393 local_irq_restore(flags);
394}
395
396void inc_zone_page_state(struct page *page, enum zone_stat_item item)
397{
398 unsigned long flags;
399 struct zone *zone;
400
401 zone = page_zone(page);
402 local_irq_save(flags);
403 __inc_zone_state(zone, item);
404 local_irq_restore(flags);
405}
406EXPORT_SYMBOL(inc_zone_page_state);
407
408void dec_zone_page_state(struct page *page, enum zone_stat_item item)
409{
410 unsigned long flags;
411
412 local_irq_save(flags);
413 __dec_zone_page_state(page, item);
414 local_irq_restore(flags);
415}
416EXPORT_SYMBOL(dec_zone_page_state);
417#endif
418
419/*
420 * Update the zone counters for one cpu.
421 *
422 * The cpu specified must be either the current cpu or a processor that
423 * is not online. If it is the current cpu then the execution thread must
424 * be pinned to the current cpu.
425 *
426 * Note that refresh_cpu_vm_stats strives to only access
427 * node local memory. The per cpu pagesets on remote zones are placed
428 * in the memory local to the processor using that pageset. So the
429 * loop over all zones will access a series of cachelines local to
430 * the processor.
431 *
432 * The call to zone_page_state_add updates the cachelines with the
433 * statistics in the remote zone struct as well as the global cachelines
434 * with the global counters. These could cause remote node cache line
435 * bouncing and will have to be only done when necessary.
436 */
437void refresh_cpu_vm_stats(int cpu)
438{
439 struct zone *zone;
440 int i;
441 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
442
443 for_each_populated_zone(zone) {
444 struct per_cpu_pageset *p;
445
446 p = per_cpu_ptr(zone->pageset, cpu);
447
448 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
449 if (p->vm_stat_diff[i]) {
450 unsigned long flags;
451 int v;
452
453 local_irq_save(flags);
454 v = p->vm_stat_diff[i];
455 p->vm_stat_diff[i] = 0;
456 local_irq_restore(flags);
457 atomic_long_add(v, &zone->vm_stat[i]);
458 global_diff[i] += v;
459#ifdef CONFIG_NUMA
460 /* 3 seconds idle till flush */
461 p->expire = 3;
462#endif
463 }
464 cond_resched();
465#ifdef CONFIG_NUMA
466 /*
467 * Deal with draining the remote pageset of this
468 * processor
469 *
470 * Check if there are pages remaining in this pageset
471 * if not then there is nothing to expire.
472 */
473 if (!p->expire || !p->pcp.count)
474 continue;
475
476 /*
477 * We never drain zones local to this processor.
478 */
479 if (zone_to_nid(zone) == numa_node_id()) {
480 p->expire = 0;
481 continue;
482 }
483
484 p->expire--;
485 if (p->expire)
486 continue;
487
488 if (p->pcp.count)
489 drain_zone_pages(zone, &p->pcp);
490#endif
491 }
492
493 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
494 if (global_diff[i])
495 atomic_long_add(global_diff[i], &vm_stat[i]);
496}
497
498#endif
499
500#ifdef CONFIG_NUMA
501/*
502 * zonelist = the list of zones passed to the allocator
503 * z = the zone from which the allocation occurred.
504 *
505 * Must be called with interrupts disabled.
506 *
507 * When __GFP_OTHER_NODE is set assume the node of the preferred
508 * zone is the local node. This is useful for daemons who allocate
509 * memory on behalf of other processes.
510 */
511void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
512{
513 if (z->zone_pgdat == preferred_zone->zone_pgdat) {
514 __inc_zone_state(z, NUMA_HIT);
515 } else {
516 __inc_zone_state(z, NUMA_MISS);
517 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
518 }
519 if (z->node == ((flags & __GFP_OTHER_NODE) ?
520 preferred_zone->node : numa_node_id()))
521 __inc_zone_state(z, NUMA_LOCAL);
522 else
523 __inc_zone_state(z, NUMA_OTHER);
524}
525#endif
526
527#ifdef CONFIG_COMPACTION
528
529struct contig_page_info {
530 unsigned long free_pages;
531 unsigned long free_blocks_total;
532 unsigned long free_blocks_suitable;
533};
534
535/*
536 * Calculate the number of free pages in a zone, how many contiguous
537 * pages are free and how many are large enough to satisfy an allocation of
538 * the target size. Note that this function makes no attempt to estimate
539 * how many suitable free blocks there *might* be if MOVABLE pages were
540 * migrated. Calculating that is possible, but expensive and can be
541 * figured out from userspace
542 */
543static void fill_contig_page_info(struct zone *zone,
544 unsigned int suitable_order,
545 struct contig_page_info *info)
546{
547 unsigned int order;
548
549 info->free_pages = 0;
550 info->free_blocks_total = 0;
551 info->free_blocks_suitable = 0;
552
553 for (order = 0; order < MAX_ORDER; order++) {
554 unsigned long blocks;
555
556 /* Count number of free blocks */
557 blocks = zone->free_area[order].nr_free;
558 info->free_blocks_total += blocks;
559
560 /* Count free base pages */
561 info->free_pages += blocks << order;
562
563 /* Count the suitable free blocks */
564 if (order >= suitable_order)
565 info->free_blocks_suitable += blocks <<
566 (order - suitable_order);
567 }
568}
569
570/*
571 * A fragmentation index only makes sense if an allocation of a requested
572 * size would fail. If that is true, the fragmentation index indicates
573 * whether external fragmentation or a lack of memory was the problem.
574 * The value can be used to determine if page reclaim or compaction
575 * should be used
576 */
577static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
578{
579 unsigned long requested = 1UL << order;
580
581 if (!info->free_blocks_total)
582 return 0;
583
584 /* Fragmentation index only makes sense when a request would fail */
585 if (info->free_blocks_suitable)
586 return -1000;
587
588 /*
589 * Index is between 0 and 1 so return within 3 decimal places
590 *
591 * 0 => allocation would fail due to lack of memory
592 * 1 => allocation would fail due to fragmentation
593 */
594 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
595}
596
597/* Same as __fragmentation index but allocs contig_page_info on stack */
598int fragmentation_index(struct zone *zone, unsigned int order)
599{
600 struct contig_page_info info;
601
602 fill_contig_page_info(zone, order, &info);
603 return __fragmentation_index(order, &info);
604}
605#endif
606
607#if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
608#include <linux/proc_fs.h>
609#include <linux/seq_file.h>
610
611static char * const migratetype_names[MIGRATE_TYPES] = {
612 "Unmovable",
613 "Reclaimable",
614 "Movable",
615 "Reserve",
616 "Isolate",
617};
618
619static void *frag_start(struct seq_file *m, loff_t *pos)
620{
621 pg_data_t *pgdat;
622 loff_t node = *pos;
623 for (pgdat = first_online_pgdat();
624 pgdat && node;
625 pgdat = next_online_pgdat(pgdat))
626 --node;
627
628 return pgdat;
629}
630
631static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
632{
633 pg_data_t *pgdat = (pg_data_t *)arg;
634
635 (*pos)++;
636 return next_online_pgdat(pgdat);
637}
638
639static void frag_stop(struct seq_file *m, void *arg)
640{
641}
642
643/* Walk all the zones in a node and print using a callback */
644static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
645 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
646{
647 struct zone *zone;
648 struct zone *node_zones = pgdat->node_zones;
649 unsigned long flags;
650
651 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
652 if (!populated_zone(zone))
653 continue;
654
655 spin_lock_irqsave(&zone->lock, flags);
656 print(m, pgdat, zone);
657 spin_unlock_irqrestore(&zone->lock, flags);
658 }
659}
660#endif
661
662#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
663#ifdef CONFIG_ZONE_DMA
664#define TEXT_FOR_DMA(xx) xx "_dma",
665#else
666#define TEXT_FOR_DMA(xx)
667#endif
668
669#ifdef CONFIG_ZONE_DMA32
670#define TEXT_FOR_DMA32(xx) xx "_dma32",
671#else
672#define TEXT_FOR_DMA32(xx)
673#endif
674
675#ifdef CONFIG_HIGHMEM
676#define TEXT_FOR_HIGHMEM(xx) xx "_high",
677#else
678#define TEXT_FOR_HIGHMEM(xx)
679#endif
680
681#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
682 TEXT_FOR_HIGHMEM(xx) xx "_movable",
683
684const char * const vmstat_text[] = {
685 /* Zoned VM counters */
686 "nr_free_pages",
687 "nr_inactive_anon",
688 "nr_active_anon",
689 "nr_inactive_file",
690 "nr_active_file",
691 "nr_unevictable",
692 "nr_mlock",
693 "nr_anon_pages",
694 "nr_mapped",
695 "nr_file_pages",
696 "nr_dirty",
697 "nr_writeback",
698 "nr_slab_reclaimable",
699 "nr_slab_unreclaimable",
700 "nr_page_table_pages",
701 "nr_kernel_stack",
702 "nr_unstable",
703 "nr_bounce",
704 "nr_vmscan_write",
705 "nr_writeback_temp",
706 "nr_isolated_anon",
707 "nr_isolated_file",
708 "nr_shmem",
709 "nr_dirtied",
710 "nr_written",
711
712#ifdef CONFIG_NUMA
713 "numa_hit",
714 "numa_miss",
715 "numa_foreign",
716 "numa_interleave",
717 "numa_local",
718 "numa_other",
719#endif
720 "nr_anon_transparent_hugepages",
721 "nr_dirty_threshold",
722 "nr_dirty_background_threshold",
723
724#ifdef CONFIG_VM_EVENT_COUNTERS
725 "pgpgin",
726 "pgpgout",
727 "pswpin",
728 "pswpout",
729
730 TEXTS_FOR_ZONES("pgalloc")
731
732 "pgfree",
733 "pgactivate",
734 "pgdeactivate",
735
736 "pgfault",
737 "pgmajfault",
738
739 TEXTS_FOR_ZONES("pgrefill")
740 TEXTS_FOR_ZONES("pgsteal")
741 TEXTS_FOR_ZONES("pgscan_kswapd")
742 TEXTS_FOR_ZONES("pgscan_direct")
743
744#ifdef CONFIG_NUMA
745 "zone_reclaim_failed",
746#endif
747 "pginodesteal",
748 "slabs_scanned",
749 "kswapd_steal",
750 "kswapd_inodesteal",
751 "kswapd_low_wmark_hit_quickly",
752 "kswapd_high_wmark_hit_quickly",
753 "kswapd_skip_congestion_wait",
754 "pageoutrun",
755 "allocstall",
756
757 "pgrotated",
758
759#ifdef CONFIG_COMPACTION
760 "compact_blocks_moved",
761 "compact_pages_moved",
762 "compact_pagemigrate_failed",
763 "compact_stall",
764 "compact_fail",
765 "compact_success",
766#endif
767
768#ifdef CONFIG_HUGETLB_PAGE
769 "htlb_buddy_alloc_success",
770 "htlb_buddy_alloc_fail",
771#endif
772 "unevictable_pgs_culled",
773 "unevictable_pgs_scanned",
774 "unevictable_pgs_rescued",
775 "unevictable_pgs_mlocked",
776 "unevictable_pgs_munlocked",
777 "unevictable_pgs_cleared",
778 "unevictable_pgs_stranded",
779 "unevictable_pgs_mlockfreed",
780
781#ifdef CONFIG_TRANSPARENT_HUGEPAGE
782 "thp_fault_alloc",
783 "thp_fault_fallback",
784 "thp_collapse_alloc",
785 "thp_collapse_alloc_failed",
786 "thp_split",
787#endif
788
789#endif /* CONFIG_VM_EVENTS_COUNTERS */
790};
791#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
792
793
794#ifdef CONFIG_PROC_FS
795static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
796 struct zone *zone)
797{
798 int order;
799
800 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
801 for (order = 0; order < MAX_ORDER; ++order)
802 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
803 seq_putc(m, '\n');
804}
805
806/*
807 * This walks the free areas for each zone.
808 */
809static int frag_show(struct seq_file *m, void *arg)
810{
811 pg_data_t *pgdat = (pg_data_t *)arg;
812 walk_zones_in_node(m, pgdat, frag_show_print);
813 return 0;
814}
815
816static void pagetypeinfo_showfree_print(struct seq_file *m,
817 pg_data_t *pgdat, struct zone *zone)
818{
819 int order, mtype;
820
821 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
822 seq_printf(m, "Node %4d, zone %8s, type %12s ",
823 pgdat->node_id,
824 zone->name,
825 migratetype_names[mtype]);
826 for (order = 0; order < MAX_ORDER; ++order) {
827 unsigned long freecount = 0;
828 struct free_area *area;
829 struct list_head *curr;
830
831 area = &(zone->free_area[order]);
832
833 list_for_each(curr, &area->free_list[mtype])
834 freecount++;
835 seq_printf(m, "%6lu ", freecount);
836 }
837 seq_putc(m, '\n');
838 }
839}
840
841/* Print out the free pages at each order for each migatetype */
842static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
843{
844 int order;
845 pg_data_t *pgdat = (pg_data_t *)arg;
846
847 /* Print header */
848 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
849 for (order = 0; order < MAX_ORDER; ++order)
850 seq_printf(m, "%6d ", order);
851 seq_putc(m, '\n');
852
853 walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
854
855 return 0;
856}
857
858static void pagetypeinfo_showblockcount_print(struct seq_file *m,
859 pg_data_t *pgdat, struct zone *zone)
860{
861 int mtype;
862 unsigned long pfn;
863 unsigned long start_pfn = zone->zone_start_pfn;
864 unsigned long end_pfn = start_pfn + zone->spanned_pages;
865 unsigned long count[MIGRATE_TYPES] = { 0, };
866
867 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
868 struct page *page;
869
870 if (!pfn_valid(pfn))
871 continue;
872
873 page = pfn_to_page(pfn);
874
875 /* Watch for unexpected holes punched in the memmap */
876 if (!memmap_valid_within(pfn, page, zone))
877 continue;
878
879 mtype = get_pageblock_migratetype(page);
880
881 if (mtype < MIGRATE_TYPES)
882 count[mtype]++;
883 }
884
885 /* Print counts */
886 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
887 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
888 seq_printf(m, "%12lu ", count[mtype]);
889 seq_putc(m, '\n');
890}
891
892/* Print out the free pages at each order for each migratetype */
893static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
894{
895 int mtype;
896 pg_data_t *pgdat = (pg_data_t *)arg;
897
898 seq_printf(m, "\n%-23s", "Number of blocks type ");
899 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
900 seq_printf(m, "%12s ", migratetype_names[mtype]);
901 seq_putc(m, '\n');
902 walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
903
904 return 0;
905}
906
907/*
908 * This prints out statistics in relation to grouping pages by mobility.
909 * It is expensive to collect so do not constantly read the file.
910 */
911static int pagetypeinfo_show(struct seq_file *m, void *arg)
912{
913 pg_data_t *pgdat = (pg_data_t *)arg;
914
915 /* check memoryless node */
916 if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
917 return 0;
918
919 seq_printf(m, "Page block order: %d\n", pageblock_order);
920 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
921 seq_putc(m, '\n');
922 pagetypeinfo_showfree(m, pgdat);
923 pagetypeinfo_showblockcount(m, pgdat);
924
925 return 0;
926}
927
928static const struct seq_operations fragmentation_op = {
929 .start = frag_start,
930 .next = frag_next,
931 .stop = frag_stop,
932 .show = frag_show,
933};
934
935static int fragmentation_open(struct inode *inode, struct file *file)
936{
937 return seq_open(file, &fragmentation_op);
938}
939
940static const struct file_operations fragmentation_file_operations = {
941 .open = fragmentation_open,
942 .read = seq_read,
943 .llseek = seq_lseek,
944 .release = seq_release,
945};
946
947static const struct seq_operations pagetypeinfo_op = {
948 .start = frag_start,
949 .next = frag_next,
950 .stop = frag_stop,
951 .show = pagetypeinfo_show,
952};
953
954static int pagetypeinfo_open(struct inode *inode, struct file *file)
955{
956 return seq_open(file, &pagetypeinfo_op);
957}
958
959static const struct file_operations pagetypeinfo_file_ops = {
960 .open = pagetypeinfo_open,
961 .read = seq_read,
962 .llseek = seq_lseek,
963 .release = seq_release,
964};
965
966static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
967 struct zone *zone)
968{
969 int i;
970 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
971 seq_printf(m,
972 "\n pages free %lu"
973 "\n min %lu"
974 "\n low %lu"
975 "\n high %lu"
976 "\n scanned %lu"
977 "\n spanned %lu"
978 "\n present %lu",
979 zone_page_state(zone, NR_FREE_PAGES),
980 min_wmark_pages(zone),
981 low_wmark_pages(zone),
982 high_wmark_pages(zone),
983 zone->pages_scanned,
984 zone->spanned_pages,
985 zone->present_pages);
986
987 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
988 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
989 zone_page_state(zone, i));
990
991 seq_printf(m,
992 "\n protection: (%lu",
993 zone->lowmem_reserve[0]);
994 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
995 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
996 seq_printf(m,
997 ")"
998 "\n pagesets");
999 for_each_online_cpu(i) {
1000 struct per_cpu_pageset *pageset;
1001
1002 pageset = per_cpu_ptr(zone->pageset, i);
1003 seq_printf(m,
1004 "\n cpu: %i"
1005 "\n count: %i"
1006 "\n high: %i"
1007 "\n batch: %i",
1008 i,
1009 pageset->pcp.count,
1010 pageset->pcp.high,
1011 pageset->pcp.batch);
1012#ifdef CONFIG_SMP
1013 seq_printf(m, "\n vm stats threshold: %d",
1014 pageset->stat_threshold);
1015#endif
1016 }
1017 seq_printf(m,
1018 "\n all_unreclaimable: %u"
1019 "\n start_pfn: %lu"
1020 "\n inactive_ratio: %u",
1021 zone->all_unreclaimable,
1022 zone->zone_start_pfn,
1023 zone->inactive_ratio);
1024 seq_putc(m, '\n');
1025}
1026
1027/*
1028 * Output information about zones in @pgdat.
1029 */
1030static int zoneinfo_show(struct seq_file *m, void *arg)
1031{
1032 pg_data_t *pgdat = (pg_data_t *)arg;
1033 walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1034 return 0;
1035}
1036
1037static const struct seq_operations zoneinfo_op = {
1038 .start = frag_start, /* iterate over all zones. The same as in
1039 * fragmentation. */
1040 .next = frag_next,
1041 .stop = frag_stop,
1042 .show = zoneinfo_show,
1043};
1044
1045static int zoneinfo_open(struct inode *inode, struct file *file)
1046{
1047 return seq_open(file, &zoneinfo_op);
1048}
1049
1050static const struct file_operations proc_zoneinfo_file_operations = {
1051 .open = zoneinfo_open,
1052 .read = seq_read,
1053 .llseek = seq_lseek,
1054 .release = seq_release,
1055};
1056
1057enum writeback_stat_item {
1058 NR_DIRTY_THRESHOLD,
1059 NR_DIRTY_BG_THRESHOLD,
1060 NR_VM_WRITEBACK_STAT_ITEMS,
1061};
1062
1063static void *vmstat_start(struct seq_file *m, loff_t *pos)
1064{
1065 unsigned long *v;
1066 int i, stat_items_size;
1067
1068 if (*pos >= ARRAY_SIZE(vmstat_text))
1069 return NULL;
1070 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1071 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1072
1073#ifdef CONFIG_VM_EVENT_COUNTERS
1074 stat_items_size += sizeof(struct vm_event_state);
1075#endif
1076
1077 v = kmalloc(stat_items_size, GFP_KERNEL);
1078 m->private = v;
1079 if (!v)
1080 return ERR_PTR(-ENOMEM);
1081 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1082 v[i] = global_page_state(i);
1083 v += NR_VM_ZONE_STAT_ITEMS;
1084
1085 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1086 v + NR_DIRTY_THRESHOLD);
1087 v += NR_VM_WRITEBACK_STAT_ITEMS;
1088
1089#ifdef CONFIG_VM_EVENT_COUNTERS
1090 all_vm_events(v);
1091 v[PGPGIN] /= 2; /* sectors -> kbytes */
1092 v[PGPGOUT] /= 2;
1093#endif
1094 return (unsigned long *)m->private + *pos;
1095}
1096
1097static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1098{
1099 (*pos)++;
1100 if (*pos >= ARRAY_SIZE(vmstat_text))
1101 return NULL;
1102 return (unsigned long *)m->private + *pos;
1103}
1104
1105static int vmstat_show(struct seq_file *m, void *arg)
1106{
1107 unsigned long *l = arg;
1108 unsigned long off = l - (unsigned long *)m->private;
1109
1110 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1111 return 0;
1112}
1113
1114static void vmstat_stop(struct seq_file *m, void *arg)
1115{
1116 kfree(m->private);
1117 m->private = NULL;
1118}
1119
1120static const struct seq_operations vmstat_op = {
1121 .start = vmstat_start,
1122 .next = vmstat_next,
1123 .stop = vmstat_stop,
1124 .show = vmstat_show,
1125};
1126
1127static int vmstat_open(struct inode *inode, struct file *file)
1128{
1129 return seq_open(file, &vmstat_op);
1130}
1131
1132static const struct file_operations proc_vmstat_file_operations = {
1133 .open = vmstat_open,
1134 .read = seq_read,
1135 .llseek = seq_lseek,
1136 .release = seq_release,
1137};
1138#endif /* CONFIG_PROC_FS */
1139
1140#ifdef CONFIG_SMP
1141static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1142int sysctl_stat_interval __read_mostly = HZ;
1143
1144static void vmstat_update(struct work_struct *w)
1145{
1146 refresh_cpu_vm_stats(smp_processor_id());
1147 schedule_delayed_work(&__get_cpu_var(vmstat_work),
1148 round_jiffies_relative(sysctl_stat_interval));
1149}
1150
1151static void __cpuinit start_cpu_timer(int cpu)
1152{
1153 struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1154
1155 INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
1156 schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1157}
1158
1159/*
1160 * Use the cpu notifier to insure that the thresholds are recalculated
1161 * when necessary.
1162 */
1163static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1164 unsigned long action,
1165 void *hcpu)
1166{
1167 long cpu = (long)hcpu;
1168
1169 switch (action) {
1170 case CPU_ONLINE:
1171 case CPU_ONLINE_FROZEN:
1172 refresh_zone_stat_thresholds();
1173 start_cpu_timer(cpu);
1174 node_set_state(cpu_to_node(cpu), N_CPU);
1175 break;
1176 case CPU_DOWN_PREPARE:
1177 case CPU_DOWN_PREPARE_FROZEN:
1178 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1179 per_cpu(vmstat_work, cpu).work.func = NULL;
1180 break;
1181 case CPU_DOWN_FAILED:
1182 case CPU_DOWN_FAILED_FROZEN:
1183 start_cpu_timer(cpu);
1184 break;
1185 case CPU_DEAD:
1186 case CPU_DEAD_FROZEN:
1187 refresh_zone_stat_thresholds();
1188 break;
1189 default:
1190 break;
1191 }
1192 return NOTIFY_OK;
1193}
1194
1195static struct notifier_block __cpuinitdata vmstat_notifier =
1196 { &vmstat_cpuup_callback, NULL, 0 };
1197#endif
1198
1199static int __init setup_vmstat(void)
1200{
1201#ifdef CONFIG_SMP
1202 int cpu;
1203
1204 register_cpu_notifier(&vmstat_notifier);
1205
1206 for_each_online_cpu(cpu)
1207 start_cpu_timer(cpu);
1208#endif
1209#ifdef CONFIG_PROC_FS
1210 proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1211 proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1212 proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1213 proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1214#endif
1215 return 0;
1216}
1217module_init(setup_vmstat)
1218
1219#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1220#include <linux/debugfs.h>
1221
1222static struct dentry *extfrag_debug_root;
1223
1224/*
1225 * Return an index indicating how much of the available free memory is
1226 * unusable for an allocation of the requested size.
1227 */
1228static int unusable_free_index(unsigned int order,
1229 struct contig_page_info *info)
1230{
1231 /* No free memory is interpreted as all free memory is unusable */
1232 if (info->free_pages == 0)
1233 return 1000;
1234
1235 /*
1236 * Index should be a value between 0 and 1. Return a value to 3
1237 * decimal places.
1238 *
1239 * 0 => no fragmentation
1240 * 1 => high fragmentation
1241 */
1242 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1243
1244}
1245
1246static void unusable_show_print(struct seq_file *m,
1247 pg_data_t *pgdat, struct zone *zone)
1248{
1249 unsigned int order;
1250 int index;
1251 struct contig_page_info info;
1252
1253 seq_printf(m, "Node %d, zone %8s ",
1254 pgdat->node_id,
1255 zone->name);
1256 for (order = 0; order < MAX_ORDER; ++order) {
1257 fill_contig_page_info(zone, order, &info);
1258 index = unusable_free_index(order, &info);
1259 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1260 }
1261
1262 seq_putc(m, '\n');
1263}
1264
1265/*
1266 * Display unusable free space index
1267 *
1268 * The unusable free space index measures how much of the available free
1269 * memory cannot be used to satisfy an allocation of a given size and is a
1270 * value between 0 and 1. The higher the value, the more of free memory is
1271 * unusable and by implication, the worse the external fragmentation is. This
1272 * can be expressed as a percentage by multiplying by 100.
1273 */
1274static int unusable_show(struct seq_file *m, void *arg)
1275{
1276 pg_data_t *pgdat = (pg_data_t *)arg;
1277
1278 /* check memoryless node */
1279 if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1280 return 0;
1281
1282 walk_zones_in_node(m, pgdat, unusable_show_print);
1283
1284 return 0;
1285}
1286
1287static const struct seq_operations unusable_op = {
1288 .start = frag_start,
1289 .next = frag_next,
1290 .stop = frag_stop,
1291 .show = unusable_show,
1292};
1293
1294static int unusable_open(struct inode *inode, struct file *file)
1295{
1296 return seq_open(file, &unusable_op);
1297}
1298
1299static const struct file_operations unusable_file_ops = {
1300 .open = unusable_open,
1301 .read = seq_read,
1302 .llseek = seq_lseek,
1303 .release = seq_release,
1304};
1305
1306static void extfrag_show_print(struct seq_file *m,
1307 pg_data_t *pgdat, struct zone *zone)
1308{
1309 unsigned int order;
1310 int index;
1311
1312 /* Alloc on stack as interrupts are disabled for zone walk */
1313 struct contig_page_info info;
1314
1315 seq_printf(m, "Node %d, zone %8s ",
1316 pgdat->node_id,
1317 zone->name);
1318 for (order = 0; order < MAX_ORDER; ++order) {
1319 fill_contig_page_info(zone, order, &info);
1320 index = __fragmentation_index(order, &info);
1321 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1322 }
1323
1324 seq_putc(m, '\n');
1325}
1326
1327/*
1328 * Display fragmentation index for orders that allocations would fail for
1329 */
1330static int extfrag_show(struct seq_file *m, void *arg)
1331{
1332 pg_data_t *pgdat = (pg_data_t *)arg;
1333
1334 walk_zones_in_node(m, pgdat, extfrag_show_print);
1335
1336 return 0;
1337}
1338
1339static const struct seq_operations extfrag_op = {
1340 .start = frag_start,
1341 .next = frag_next,
1342 .stop = frag_stop,
1343 .show = extfrag_show,
1344};
1345
1346static int extfrag_open(struct inode *inode, struct file *file)
1347{
1348 return seq_open(file, &extfrag_op);
1349}
1350
1351static const struct file_operations extfrag_file_ops = {
1352 .open = extfrag_open,
1353 .read = seq_read,
1354 .llseek = seq_lseek,
1355 .release = seq_release,
1356};
1357
1358static int __init extfrag_debug_init(void)
1359{
1360 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1361 if (!extfrag_debug_root)
1362 return -ENOMEM;
1363
1364 if (!debugfs_create_file("unusable_index", 0444,
1365 extfrag_debug_root, NULL, &unusable_file_ops))
1366 return -ENOMEM;
1367
1368 if (!debugfs_create_file("extfrag_index", 0444,
1369 extfrag_debug_root, NULL, &extfrag_file_ops))
1370 return -ENOMEM;
1371
1372 return 0;
1373}
1374
1375module_init(extfrag_debug_init);
1376#endif