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