Loading...
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
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#ifdef CONFIG_NUMA
35int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
36
37/* zero numa counters within a zone */
38static void zero_zone_numa_counters(struct zone *zone)
39{
40 int item, cpu;
41
42 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
43 atomic_long_set(&zone->vm_numa_event[item], 0);
44 for_each_online_cpu(cpu) {
45 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
46 = 0;
47 }
48 }
49}
50
51/* zero numa counters of all the populated zones */
52static void zero_zones_numa_counters(void)
53{
54 struct zone *zone;
55
56 for_each_populated_zone(zone)
57 zero_zone_numa_counters(zone);
58}
59
60/* zero global numa counters */
61static void zero_global_numa_counters(void)
62{
63 int item;
64
65 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
66 atomic_long_set(&vm_numa_event[item], 0);
67}
68
69static void invalid_numa_statistics(void)
70{
71 zero_zones_numa_counters();
72 zero_global_numa_counters();
73}
74
75static DEFINE_MUTEX(vm_numa_stat_lock);
76
77int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
78 void *buffer, size_t *length, loff_t *ppos)
79{
80 int ret, oldval;
81
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;
88
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 }
99
100out:
101 mutex_unlock(&vm_numa_stat_lock);
102 return ret;
103}
104#endif
105
106#ifdef CONFIG_VM_EVENT_COUNTERS
107DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
108EXPORT_PER_CPU_SYMBOL(vm_event_states);
109
110static void sum_vm_events(unsigned long *ret)
111{
112 int cpu;
113 int i;
114
115 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
116
117 for_each_online_cpu(cpu) {
118 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
119
120 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
121 ret[i] += this->event[i];
122 }
123}
124
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*/
130void all_vm_events(unsigned long *ret)
131{
132 get_online_cpus();
133 sum_vm_events(ret);
134 put_online_cpus();
135}
136EXPORT_SYMBOL_GPL(all_vm_events);
137
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 */
144void vm_events_fold_cpu(int cpu)
145{
146 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
147 int i;
148
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}
154
155#endif /* CONFIG_VM_EVENT_COUNTERS */
156
157/*
158 * Manage combined zone based / global counters
159 *
160 * vm_stat contains the global counters
161 */
162atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
163atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
165EXPORT_SYMBOL(vm_zone_stat);
166EXPORT_SYMBOL(vm_node_stat);
167
168#ifdef CONFIG_SMP
169
170int calculate_pressure_threshold(struct zone *zone)
171{
172 int threshold;
173 int watermark_distance;
174
175 /*
176 * As vmstats are not up to date, there is drift between the estimated
177 * and real values. For high thresholds and a high number of CPUs, it
178 * is possible for the min watermark to be breached while the estimated
179 * value looks fine. The pressure threshold is a reduced value such
180 * that even the maximum amount of drift will not accidentally breach
181 * the min watermark
182 */
183 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
184 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
185
186 /*
187 * Maximum threshold is 125
188 */
189 threshold = min(125, threshold);
190
191 return threshold;
192}
193
194int calculate_normal_threshold(struct zone *zone)
195{
196 int threshold;
197 int mem; /* memory in 128 MB units */
198
199 /*
200 * The threshold scales with the number of processors and the amount
201 * of memory per zone. More memory means that we can defer updates for
202 * longer, more processors could lead to more contention.
203 * fls() is used to have a cheap way of logarithmic scaling.
204 *
205 * Some sample thresholds:
206 *
207 * Threshold Processors (fls) Zonesize fls(mem+1)
208 * ------------------------------------------------------------------
209 * 8 1 1 0.9-1 GB 4
210 * 16 2 2 0.9-1 GB 4
211 * 20 2 2 1-2 GB 5
212 * 24 2 2 2-4 GB 6
213 * 28 2 2 4-8 GB 7
214 * 32 2 2 8-16 GB 8
215 * 4 2 2 <128M 1
216 * 30 4 3 2-4 GB 5
217 * 48 4 3 8-16 GB 8
218 * 32 8 4 1-2 GB 4
219 * 32 8 4 0.9-1GB 4
220 * 10 16 5 <128M 1
221 * 40 16 5 900M 4
222 * 70 64 7 2-4 GB 5
223 * 84 64 7 4-8 GB 6
224 * 108 512 9 4-8 GB 6
225 * 125 1024 10 8-16 GB 8
226 * 125 1024 10 16-32 GB 9
227 */
228
229 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
230
231 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
232
233 /*
234 * Maximum threshold is 125
235 */
236 threshold = min(125, threshold);
237
238 return threshold;
239}
240
241/*
242 * Refresh the thresholds for each zone.
243 */
244void refresh_zone_stat_thresholds(void)
245{
246 struct pglist_data *pgdat;
247 struct zone *zone;
248 int cpu;
249 int threshold;
250
251 /* Zero current pgdat thresholds */
252 for_each_online_pgdat(pgdat) {
253 for_each_online_cpu(cpu) {
254 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
255 }
256 }
257
258 for_each_populated_zone(zone) {
259 struct pglist_data *pgdat = zone->zone_pgdat;
260 unsigned long max_drift, tolerate_drift;
261
262 threshold = calculate_normal_threshold(zone);
263
264 for_each_online_cpu(cpu) {
265 int pgdat_threshold;
266
267 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
268 = threshold;
269
270 /* Base nodestat threshold on the largest populated zone. */
271 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
272 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
273 = max(threshold, pgdat_threshold);
274 }
275
276 /*
277 * Only set percpu_drift_mark if there is a danger that
278 * NR_FREE_PAGES reports the low watermark is ok when in fact
279 * the min watermark could be breached by an allocation
280 */
281 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
282 max_drift = num_online_cpus() * threshold;
283 if (max_drift > tolerate_drift)
284 zone->percpu_drift_mark = high_wmark_pages(zone) +
285 max_drift;
286 }
287}
288
289void set_pgdat_percpu_threshold(pg_data_t *pgdat,
290 int (*calculate_pressure)(struct zone *))
291{
292 struct zone *zone;
293 int cpu;
294 int threshold;
295 int i;
296
297 for (i = 0; i < pgdat->nr_zones; i++) {
298 zone = &pgdat->node_zones[i];
299 if (!zone->percpu_drift_mark)
300 continue;
301
302 threshold = (*calculate_pressure)(zone);
303 for_each_online_cpu(cpu)
304 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
305 = threshold;
306 }
307}
308
309/*
310 * For use when we know that interrupts are disabled,
311 * or when we know that preemption is disabled and that
312 * particular counter cannot be updated from interrupt context.
313 */
314void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
315 long delta)
316{
317 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
318 s8 __percpu *p = pcp->vm_stat_diff + item;
319 long x;
320 long t;
321
322 x = delta + __this_cpu_read(*p);
323
324 t = __this_cpu_read(pcp->stat_threshold);
325
326 if (unlikely(abs(x) > t)) {
327 zone_page_state_add(x, zone, item);
328 x = 0;
329 }
330 __this_cpu_write(*p, x);
331}
332EXPORT_SYMBOL(__mod_zone_page_state);
333
334void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
335 long delta)
336{
337 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
338 s8 __percpu *p = pcp->vm_node_stat_diff + item;
339 long x;
340 long t;
341
342 if (vmstat_item_in_bytes(item)) {
343 /*
344 * Only cgroups use subpage accounting right now; at
345 * the global level, these items still change in
346 * multiples of whole pages. Store them as pages
347 * internally to keep the per-cpu counters compact.
348 */
349 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
350 delta >>= PAGE_SHIFT;
351 }
352
353 x = delta + __this_cpu_read(*p);
354
355 t = __this_cpu_read(pcp->stat_threshold);
356
357 if (unlikely(abs(x) > t)) {
358 node_page_state_add(x, pgdat, item);
359 x = 0;
360 }
361 __this_cpu_write(*p, x);
362}
363EXPORT_SYMBOL(__mod_node_page_state);
364
365/*
366 * Optimized increment and decrement functions.
367 *
368 * These are only for a single page and therefore can take a struct page *
369 * argument instead of struct zone *. This allows the inclusion of the code
370 * generated for page_zone(page) into the optimized functions.
371 *
372 * No overflow check is necessary and therefore the differential can be
373 * incremented or decremented in place which may allow the compilers to
374 * generate better code.
375 * The increment or decrement is known and therefore one boundary check can
376 * be omitted.
377 *
378 * NOTE: These functions are very performance sensitive. Change only
379 * with care.
380 *
381 * Some processors have inc/dec instructions that are atomic vs an interrupt.
382 * However, the code must first determine the differential location in a zone
383 * based on the processor number and then inc/dec the counter. There is no
384 * guarantee without disabling preemption that the processor will not change
385 * in between and therefore the atomicity vs. interrupt cannot be exploited
386 * in a useful way here.
387 */
388void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
389{
390 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
391 s8 __percpu *p = pcp->vm_stat_diff + item;
392 s8 v, t;
393
394 v = __this_cpu_inc_return(*p);
395 t = __this_cpu_read(pcp->stat_threshold);
396 if (unlikely(v > t)) {
397 s8 overstep = t >> 1;
398
399 zone_page_state_add(v + overstep, zone, item);
400 __this_cpu_write(*p, -overstep);
401 }
402}
403
404void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
405{
406 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
407 s8 __percpu *p = pcp->vm_node_stat_diff + item;
408 s8 v, t;
409
410 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
411
412 v = __this_cpu_inc_return(*p);
413 t = __this_cpu_read(pcp->stat_threshold);
414 if (unlikely(v > t)) {
415 s8 overstep = t >> 1;
416
417 node_page_state_add(v + overstep, pgdat, item);
418 __this_cpu_write(*p, -overstep);
419 }
420}
421
422void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
423{
424 __inc_zone_state(page_zone(page), item);
425}
426EXPORT_SYMBOL(__inc_zone_page_state);
427
428void __inc_node_page_state(struct page *page, enum node_stat_item item)
429{
430 __inc_node_state(page_pgdat(page), item);
431}
432EXPORT_SYMBOL(__inc_node_page_state);
433
434void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
435{
436 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
437 s8 __percpu *p = pcp->vm_stat_diff + item;
438 s8 v, t;
439
440 v = __this_cpu_dec_return(*p);
441 t = __this_cpu_read(pcp->stat_threshold);
442 if (unlikely(v < - t)) {
443 s8 overstep = t >> 1;
444
445 zone_page_state_add(v - overstep, zone, item);
446 __this_cpu_write(*p, overstep);
447 }
448}
449
450void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
451{
452 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
453 s8 __percpu *p = pcp->vm_node_stat_diff + item;
454 s8 v, t;
455
456 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
457
458 v = __this_cpu_dec_return(*p);
459 t = __this_cpu_read(pcp->stat_threshold);
460 if (unlikely(v < - t)) {
461 s8 overstep = t >> 1;
462
463 node_page_state_add(v - overstep, pgdat, item);
464 __this_cpu_write(*p, overstep);
465 }
466}
467
468void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
469{
470 __dec_zone_state(page_zone(page), item);
471}
472EXPORT_SYMBOL(__dec_zone_page_state);
473
474void __dec_node_page_state(struct page *page, enum node_stat_item item)
475{
476 __dec_node_state(page_pgdat(page), item);
477}
478EXPORT_SYMBOL(__dec_node_page_state);
479
480#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
481/*
482 * If we have cmpxchg_local support then we do not need to incur the overhead
483 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
484 *
485 * mod_state() modifies the zone counter state through atomic per cpu
486 * operations.
487 *
488 * Overstep mode specifies how overstep should handled:
489 * 0 No overstepping
490 * 1 Overstepping half of threshold
491 * -1 Overstepping minus half of threshold
492*/
493static inline void mod_zone_state(struct zone *zone,
494 enum zone_stat_item item, long delta, int overstep_mode)
495{
496 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
497 s8 __percpu *p = pcp->vm_stat_diff + item;
498 long o, n, t, z;
499
500 do {
501 z = 0; /* overflow to zone counters */
502
503 /*
504 * The fetching of the stat_threshold is racy. We may apply
505 * a counter threshold to the wrong the cpu if we get
506 * rescheduled while executing here. However, the next
507 * counter update will apply the threshold again and
508 * therefore bring the counter under the threshold again.
509 *
510 * Most of the time the thresholds are the same anyways
511 * for all cpus in a zone.
512 */
513 t = this_cpu_read(pcp->stat_threshold);
514
515 o = this_cpu_read(*p);
516 n = delta + o;
517
518 if (abs(n) > t) {
519 int os = overstep_mode * (t >> 1) ;
520
521 /* Overflow must be added to zone counters */
522 z = n + os;
523 n = -os;
524 }
525 } while (this_cpu_cmpxchg(*p, o, n) != o);
526
527 if (z)
528 zone_page_state_add(z, zone, item);
529}
530
531void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
532 long delta)
533{
534 mod_zone_state(zone, item, delta, 0);
535}
536EXPORT_SYMBOL(mod_zone_page_state);
537
538void inc_zone_page_state(struct page *page, enum zone_stat_item item)
539{
540 mod_zone_state(page_zone(page), item, 1, 1);
541}
542EXPORT_SYMBOL(inc_zone_page_state);
543
544void dec_zone_page_state(struct page *page, enum zone_stat_item item)
545{
546 mod_zone_state(page_zone(page), item, -1, -1);
547}
548EXPORT_SYMBOL(dec_zone_page_state);
549
550static inline void mod_node_state(struct pglist_data *pgdat,
551 enum node_stat_item item, int delta, int overstep_mode)
552{
553 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
554 s8 __percpu *p = pcp->vm_node_stat_diff + item;
555 long o, n, t, z;
556
557 if (vmstat_item_in_bytes(item)) {
558 /*
559 * Only cgroups use subpage accounting right now; at
560 * the global level, these items still change in
561 * multiples of whole pages. Store them as pages
562 * internally to keep the per-cpu counters compact.
563 */
564 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
565 delta >>= PAGE_SHIFT;
566 }
567
568 do {
569 z = 0; /* overflow to node counters */
570
571 /*
572 * The fetching of the stat_threshold is racy. We may apply
573 * a counter threshold to the wrong the cpu if we get
574 * rescheduled while executing here. However, the next
575 * counter update will apply the threshold again and
576 * therefore bring the counter under the threshold again.
577 *
578 * Most of the time the thresholds are the same anyways
579 * for all cpus in a node.
580 */
581 t = this_cpu_read(pcp->stat_threshold);
582
583 o = this_cpu_read(*p);
584 n = delta + o;
585
586 if (abs(n) > t) {
587 int os = overstep_mode * (t >> 1) ;
588
589 /* Overflow must be added to node counters */
590 z = n + os;
591 n = -os;
592 }
593 } while (this_cpu_cmpxchg(*p, o, n) != o);
594
595 if (z)
596 node_page_state_add(z, pgdat, item);
597}
598
599void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
600 long delta)
601{
602 mod_node_state(pgdat, item, delta, 0);
603}
604EXPORT_SYMBOL(mod_node_page_state);
605
606void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
607{
608 mod_node_state(pgdat, item, 1, 1);
609}
610
611void inc_node_page_state(struct page *page, enum node_stat_item item)
612{
613 mod_node_state(page_pgdat(page), item, 1, 1);
614}
615EXPORT_SYMBOL(inc_node_page_state);
616
617void dec_node_page_state(struct page *page, enum node_stat_item item)
618{
619 mod_node_state(page_pgdat(page), item, -1, -1);
620}
621EXPORT_SYMBOL(dec_node_page_state);
622#else
623/*
624 * Use interrupt disable to serialize counter updates
625 */
626void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
627 long delta)
628{
629 unsigned long flags;
630
631 local_irq_save(flags);
632 __mod_zone_page_state(zone, item, delta);
633 local_irq_restore(flags);
634}
635EXPORT_SYMBOL(mod_zone_page_state);
636
637void inc_zone_page_state(struct page *page, enum zone_stat_item item)
638{
639 unsigned long flags;
640 struct zone *zone;
641
642 zone = page_zone(page);
643 local_irq_save(flags);
644 __inc_zone_state(zone, item);
645 local_irq_restore(flags);
646}
647EXPORT_SYMBOL(inc_zone_page_state);
648
649void dec_zone_page_state(struct page *page, enum zone_stat_item item)
650{
651 unsigned long flags;
652
653 local_irq_save(flags);
654 __dec_zone_page_state(page, item);
655 local_irq_restore(flags);
656}
657EXPORT_SYMBOL(dec_zone_page_state);
658
659void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
660{
661 unsigned long flags;
662
663 local_irq_save(flags);
664 __inc_node_state(pgdat, item);
665 local_irq_restore(flags);
666}
667EXPORT_SYMBOL(inc_node_state);
668
669void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
670 long delta)
671{
672 unsigned long flags;
673
674 local_irq_save(flags);
675 __mod_node_page_state(pgdat, item, delta);
676 local_irq_restore(flags);
677}
678EXPORT_SYMBOL(mod_node_page_state);
679
680void inc_node_page_state(struct page *page, enum node_stat_item item)
681{
682 unsigned long flags;
683 struct pglist_data *pgdat;
684
685 pgdat = page_pgdat(page);
686 local_irq_save(flags);
687 __inc_node_state(pgdat, item);
688 local_irq_restore(flags);
689}
690EXPORT_SYMBOL(inc_node_page_state);
691
692void dec_node_page_state(struct page *page, enum node_stat_item item)
693{
694 unsigned long flags;
695
696 local_irq_save(flags);
697 __dec_node_page_state(page, item);
698 local_irq_restore(flags);
699}
700EXPORT_SYMBOL(dec_node_page_state);
701#endif
702
703/*
704 * Fold a differential into the global counters.
705 * Returns the number of counters updated.
706 */
707static int fold_diff(int *zone_diff, int *node_diff)
708{
709 int i;
710 int changes = 0;
711
712 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
713 if (zone_diff[i]) {
714 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
715 changes++;
716 }
717
718 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
719 if (node_diff[i]) {
720 atomic_long_add(node_diff[i], &vm_node_stat[i]);
721 changes++;
722 }
723 return changes;
724}
725
726#ifdef CONFIG_NUMA
727static void fold_vm_zone_numa_events(struct zone *zone)
728{
729 unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
730 int cpu;
731 enum numa_stat_item item;
732
733 for_each_online_cpu(cpu) {
734 struct per_cpu_zonestat *pzstats;
735
736 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
737 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
738 zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
739 }
740
741 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
742 zone_numa_event_add(zone_numa_events[item], zone, item);
743}
744
745void fold_vm_numa_events(void)
746{
747 struct zone *zone;
748
749 for_each_populated_zone(zone)
750 fold_vm_zone_numa_events(zone);
751}
752#endif
753
754/*
755 * Update the zone counters for the current cpu.
756 *
757 * Note that refresh_cpu_vm_stats strives to only access
758 * node local memory. The per cpu pagesets on remote zones are placed
759 * in the memory local to the processor using that pageset. So the
760 * loop over all zones will access a series of cachelines local to
761 * the processor.
762 *
763 * The call to zone_page_state_add updates the cachelines with the
764 * statistics in the remote zone struct as well as the global cachelines
765 * with the global counters. These could cause remote node cache line
766 * bouncing and will have to be only done when necessary.
767 *
768 * The function returns the number of global counters updated.
769 */
770static int refresh_cpu_vm_stats(bool do_pagesets)
771{
772 struct pglist_data *pgdat;
773 struct zone *zone;
774 int i;
775 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
776 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
777 int changes = 0;
778
779 for_each_populated_zone(zone) {
780 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
781#ifdef CONFIG_NUMA
782 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
783#endif
784
785 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
786 int v;
787
788 v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
789 if (v) {
790
791 atomic_long_add(v, &zone->vm_stat[i]);
792 global_zone_diff[i] += v;
793#ifdef CONFIG_NUMA
794 /* 3 seconds idle till flush */
795 __this_cpu_write(pcp->expire, 3);
796#endif
797 }
798 }
799#ifdef CONFIG_NUMA
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(pcp->expire) ||
811 !__this_cpu_read(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(pcp->expire, 0);
819 continue;
820 }
821
822 if (__this_cpu_dec_return(pcp->expire))
823 continue;
824
825 if (__this_cpu_read(pcp->count)) {
826 drain_zone_pages(zone, this_cpu_ptr(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 changes += fold_diff(global_zone_diff, global_node_diff);
848 return changes;
849}
850
851/*
852 * Fold the data for an offline cpu into the global array.
853 * There cannot be any access by the offline cpu and therefore
854 * synchronization is simplified.
855 */
856void cpu_vm_stats_fold(int cpu)
857{
858 struct pglist_data *pgdat;
859 struct zone *zone;
860 int i;
861 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
862 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
863
864 for_each_populated_zone(zone) {
865 struct per_cpu_zonestat *pzstats;
866
867 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
868
869 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
870 if (pzstats->vm_stat_diff[i]) {
871 int v;
872
873 v = pzstats->vm_stat_diff[i];
874 pzstats->vm_stat_diff[i] = 0;
875 atomic_long_add(v, &zone->vm_stat[i]);
876 global_zone_diff[i] += v;
877 }
878 }
879#ifdef CONFIG_NUMA
880 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
881 if (pzstats->vm_numa_event[i]) {
882 unsigned long v;
883
884 v = pzstats->vm_numa_event[i];
885 pzstats->vm_numa_event[i] = 0;
886 zone_numa_event_add(v, zone, i);
887 }
888 }
889#endif
890 }
891
892 for_each_online_pgdat(pgdat) {
893 struct per_cpu_nodestat *p;
894
895 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
896
897 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
898 if (p->vm_node_stat_diff[i]) {
899 int v;
900
901 v = p->vm_node_stat_diff[i];
902 p->vm_node_stat_diff[i] = 0;
903 atomic_long_add(v, &pgdat->vm_stat[i]);
904 global_node_diff[i] += v;
905 }
906 }
907
908 fold_diff(global_zone_diff, global_node_diff);
909}
910
911/*
912 * this is only called if !populated_zone(zone), which implies no other users of
913 * pset->vm_stat_diff[] exist.
914 */
915void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
916{
917 unsigned long v;
918 int i;
919
920 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
921 if (pzstats->vm_stat_diff[i]) {
922 v = pzstats->vm_stat_diff[i];
923 pzstats->vm_stat_diff[i] = 0;
924 zone_page_state_add(v, zone, i);
925 }
926 }
927
928#ifdef CONFIG_NUMA
929 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
930 if (pzstats->vm_numa_event[i]) {
931 v = pzstats->vm_numa_event[i];
932 pzstats->vm_numa_event[i] = 0;
933 zone_numa_event_add(v, zone, i);
934 }
935 }
936#endif
937}
938#endif
939
940#ifdef CONFIG_NUMA
941/*
942 * Determine the per node value of a stat item. This function
943 * is called frequently in a NUMA machine, so try to be as
944 * frugal as possible.
945 */
946unsigned long sum_zone_node_page_state(int node,
947 enum zone_stat_item item)
948{
949 struct zone *zones = NODE_DATA(node)->node_zones;
950 int i;
951 unsigned long count = 0;
952
953 for (i = 0; i < MAX_NR_ZONES; i++)
954 count += zone_page_state(zones + i, item);
955
956 return count;
957}
958
959/* Determine the per node value of a numa stat item. */
960unsigned long sum_zone_numa_event_state(int node,
961 enum numa_stat_item item)
962{
963 struct zone *zones = NODE_DATA(node)->node_zones;
964 unsigned long count = 0;
965 int i;
966
967 for (i = 0; i < MAX_NR_ZONES; i++)
968 count += zone_numa_event_state(zones + i, item);
969
970 return count;
971}
972
973/*
974 * Determine the per node value of a stat item.
975 */
976unsigned long node_page_state_pages(struct pglist_data *pgdat,
977 enum node_stat_item item)
978{
979 long x = atomic_long_read(&pgdat->vm_stat[item]);
980#ifdef CONFIG_SMP
981 if (x < 0)
982 x = 0;
983#endif
984 return x;
985}
986
987unsigned long node_page_state(struct pglist_data *pgdat,
988 enum node_stat_item item)
989{
990 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
991
992 return node_page_state_pages(pgdat, item);
993}
994#endif
995
996#ifdef CONFIG_COMPACTION
997
998struct contig_page_info {
999 unsigned long free_pages;
1000 unsigned long free_blocks_total;
1001 unsigned long free_blocks_suitable;
1002};
1003
1004/*
1005 * Calculate the number of free pages in a zone, how many contiguous
1006 * pages are free and how many are large enough to satisfy an allocation of
1007 * the target size. Note that this function makes no attempt to estimate
1008 * how many suitable free blocks there *might* be if MOVABLE pages were
1009 * migrated. Calculating that is possible, but expensive and can be
1010 * figured out from userspace
1011 */
1012static void fill_contig_page_info(struct zone *zone,
1013 unsigned int suitable_order,
1014 struct contig_page_info *info)
1015{
1016 unsigned int order;
1017
1018 info->free_pages = 0;
1019 info->free_blocks_total = 0;
1020 info->free_blocks_suitable = 0;
1021
1022 for (order = 0; order < MAX_ORDER; order++) {
1023 unsigned long blocks;
1024
1025 /* Count number of free blocks */
1026 blocks = zone->free_area[order].nr_free;
1027 info->free_blocks_total += blocks;
1028
1029 /* Count free base pages */
1030 info->free_pages += blocks << order;
1031
1032 /* Count the suitable free blocks */
1033 if (order >= suitable_order)
1034 info->free_blocks_suitable += blocks <<
1035 (order - suitable_order);
1036 }
1037}
1038
1039/*
1040 * A fragmentation index only makes sense if an allocation of a requested
1041 * size would fail. If that is true, the fragmentation index indicates
1042 * whether external fragmentation or a lack of memory was the problem.
1043 * The value can be used to determine if page reclaim or compaction
1044 * should be used
1045 */
1046static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1047{
1048 unsigned long requested = 1UL << order;
1049
1050 if (WARN_ON_ONCE(order >= MAX_ORDER))
1051 return 0;
1052
1053 if (!info->free_blocks_total)
1054 return 0;
1055
1056 /* Fragmentation index only makes sense when a request would fail */
1057 if (info->free_blocks_suitable)
1058 return -1000;
1059
1060 /*
1061 * Index is between 0 and 1 so return within 3 decimal places
1062 *
1063 * 0 => allocation would fail due to lack of memory
1064 * 1 => allocation would fail due to fragmentation
1065 */
1066 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1067}
1068
1069/*
1070 * Calculates external fragmentation within a zone wrt the given order.
1071 * It is defined as the percentage of pages found in blocks of size
1072 * less than 1 << order. It returns values in range [0, 100].
1073 */
1074unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1075{
1076 struct contig_page_info info;
1077
1078 fill_contig_page_info(zone, order, &info);
1079 if (info.free_pages == 0)
1080 return 0;
1081
1082 return div_u64((info.free_pages -
1083 (info.free_blocks_suitable << order)) * 100,
1084 info.free_pages);
1085}
1086
1087/* Same as __fragmentation index but allocs contig_page_info on stack */
1088int fragmentation_index(struct zone *zone, unsigned int order)
1089{
1090 struct contig_page_info info;
1091
1092 fill_contig_page_info(zone, order, &info);
1093 return __fragmentation_index(order, &info);
1094}
1095#endif
1096
1097#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1098 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1099#ifdef CONFIG_ZONE_DMA
1100#define TEXT_FOR_DMA(xx) xx "_dma",
1101#else
1102#define TEXT_FOR_DMA(xx)
1103#endif
1104
1105#ifdef CONFIG_ZONE_DMA32
1106#define TEXT_FOR_DMA32(xx) xx "_dma32",
1107#else
1108#define TEXT_FOR_DMA32(xx)
1109#endif
1110
1111#ifdef CONFIG_HIGHMEM
1112#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1113#else
1114#define TEXT_FOR_HIGHMEM(xx)
1115#endif
1116
1117#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1118 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1119
1120const char * const vmstat_text[] = {
1121 /* enum zone_stat_item counters */
1122 "nr_free_pages",
1123 "nr_zone_inactive_anon",
1124 "nr_zone_active_anon",
1125 "nr_zone_inactive_file",
1126 "nr_zone_active_file",
1127 "nr_zone_unevictable",
1128 "nr_zone_write_pending",
1129 "nr_mlock",
1130 "nr_bounce",
1131#if IS_ENABLED(CONFIG_ZSMALLOC)
1132 "nr_zspages",
1133#endif
1134 "nr_free_cma",
1135
1136 /* enum numa_stat_item counters */
1137#ifdef CONFIG_NUMA
1138 "numa_hit",
1139 "numa_miss",
1140 "numa_foreign",
1141 "numa_interleave",
1142 "numa_local",
1143 "numa_other",
1144#endif
1145
1146 /* enum node_stat_item counters */
1147 "nr_inactive_anon",
1148 "nr_active_anon",
1149 "nr_inactive_file",
1150 "nr_active_file",
1151 "nr_unevictable",
1152 "nr_slab_reclaimable",
1153 "nr_slab_unreclaimable",
1154 "nr_isolated_anon",
1155 "nr_isolated_file",
1156 "workingset_nodes",
1157 "workingset_refault_anon",
1158 "workingset_refault_file",
1159 "workingset_activate_anon",
1160 "workingset_activate_file",
1161 "workingset_restore_anon",
1162 "workingset_restore_file",
1163 "workingset_nodereclaim",
1164 "nr_anon_pages",
1165 "nr_mapped",
1166 "nr_file_pages",
1167 "nr_dirty",
1168 "nr_writeback",
1169 "nr_writeback_temp",
1170 "nr_shmem",
1171 "nr_shmem_hugepages",
1172 "nr_shmem_pmdmapped",
1173 "nr_file_hugepages",
1174 "nr_file_pmdmapped",
1175 "nr_anon_transparent_hugepages",
1176 "nr_vmscan_write",
1177 "nr_vmscan_immediate_reclaim",
1178 "nr_dirtied",
1179 "nr_written",
1180 "nr_kernel_misc_reclaimable",
1181 "nr_foll_pin_acquired",
1182 "nr_foll_pin_released",
1183 "nr_kernel_stack",
1184#if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1185 "nr_shadow_call_stack",
1186#endif
1187 "nr_page_table_pages",
1188#ifdef CONFIG_SWAP
1189 "nr_swapcached",
1190#endif
1191
1192 /* enum writeback_stat_item counters */
1193 "nr_dirty_threshold",
1194 "nr_dirty_background_threshold",
1195
1196#if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1197 /* enum vm_event_item counters */
1198 "pgpgin",
1199 "pgpgout",
1200 "pswpin",
1201 "pswpout",
1202
1203 TEXTS_FOR_ZONES("pgalloc")
1204 TEXTS_FOR_ZONES("allocstall")
1205 TEXTS_FOR_ZONES("pgskip")
1206
1207 "pgfree",
1208 "pgactivate",
1209 "pgdeactivate",
1210 "pglazyfree",
1211
1212 "pgfault",
1213 "pgmajfault",
1214 "pglazyfreed",
1215
1216 "pgrefill",
1217 "pgreuse",
1218 "pgsteal_kswapd",
1219 "pgsteal_direct",
1220 "pgscan_kswapd",
1221 "pgscan_direct",
1222 "pgscan_direct_throttle",
1223 "pgscan_anon",
1224 "pgscan_file",
1225 "pgsteal_anon",
1226 "pgsteal_file",
1227
1228#ifdef CONFIG_NUMA
1229 "zone_reclaim_failed",
1230#endif
1231 "pginodesteal",
1232 "slabs_scanned",
1233 "kswapd_inodesteal",
1234 "kswapd_low_wmark_hit_quickly",
1235 "kswapd_high_wmark_hit_quickly",
1236 "pageoutrun",
1237
1238 "pgrotated",
1239
1240 "drop_pagecache",
1241 "drop_slab",
1242 "oom_kill",
1243
1244#ifdef CONFIG_NUMA_BALANCING
1245 "numa_pte_updates",
1246 "numa_huge_pte_updates",
1247 "numa_hint_faults",
1248 "numa_hint_faults_local",
1249 "numa_pages_migrated",
1250#endif
1251#ifdef CONFIG_MIGRATION
1252 "pgmigrate_success",
1253 "pgmigrate_fail",
1254 "thp_migration_success",
1255 "thp_migration_fail",
1256 "thp_migration_split",
1257#endif
1258#ifdef CONFIG_COMPACTION
1259 "compact_migrate_scanned",
1260 "compact_free_scanned",
1261 "compact_isolated",
1262 "compact_stall",
1263 "compact_fail",
1264 "compact_success",
1265 "compact_daemon_wake",
1266 "compact_daemon_migrate_scanned",
1267 "compact_daemon_free_scanned",
1268#endif
1269
1270#ifdef CONFIG_HUGETLB_PAGE
1271 "htlb_buddy_alloc_success",
1272 "htlb_buddy_alloc_fail",
1273#endif
1274#ifdef CONFIG_CMA
1275 "cma_alloc_success",
1276 "cma_alloc_fail",
1277#endif
1278 "unevictable_pgs_culled",
1279 "unevictable_pgs_scanned",
1280 "unevictable_pgs_rescued",
1281 "unevictable_pgs_mlocked",
1282 "unevictable_pgs_munlocked",
1283 "unevictable_pgs_cleared",
1284 "unevictable_pgs_stranded",
1285
1286#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1287 "thp_fault_alloc",
1288 "thp_fault_fallback",
1289 "thp_fault_fallback_charge",
1290 "thp_collapse_alloc",
1291 "thp_collapse_alloc_failed",
1292 "thp_file_alloc",
1293 "thp_file_fallback",
1294 "thp_file_fallback_charge",
1295 "thp_file_mapped",
1296 "thp_split_page",
1297 "thp_split_page_failed",
1298 "thp_deferred_split_page",
1299 "thp_split_pmd",
1300#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1301 "thp_split_pud",
1302#endif
1303 "thp_zero_page_alloc",
1304 "thp_zero_page_alloc_failed",
1305 "thp_swpout",
1306 "thp_swpout_fallback",
1307#endif
1308#ifdef CONFIG_MEMORY_BALLOON
1309 "balloon_inflate",
1310 "balloon_deflate",
1311#ifdef CONFIG_BALLOON_COMPACTION
1312 "balloon_migrate",
1313#endif
1314#endif /* CONFIG_MEMORY_BALLOON */
1315#ifdef CONFIG_DEBUG_TLBFLUSH
1316 "nr_tlb_remote_flush",
1317 "nr_tlb_remote_flush_received",
1318 "nr_tlb_local_flush_all",
1319 "nr_tlb_local_flush_one",
1320#endif /* CONFIG_DEBUG_TLBFLUSH */
1321
1322#ifdef CONFIG_DEBUG_VM_VMACACHE
1323 "vmacache_find_calls",
1324 "vmacache_find_hits",
1325#endif
1326#ifdef CONFIG_SWAP
1327 "swap_ra",
1328 "swap_ra_hit",
1329#endif
1330#ifdef CONFIG_X86
1331 "direct_map_level2_splits",
1332 "direct_map_level3_splits",
1333#endif
1334#endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1335};
1336#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1337
1338#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1339 defined(CONFIG_PROC_FS)
1340static void *frag_start(struct seq_file *m, loff_t *pos)
1341{
1342 pg_data_t *pgdat;
1343 loff_t node = *pos;
1344
1345 for (pgdat = first_online_pgdat();
1346 pgdat && node;
1347 pgdat = next_online_pgdat(pgdat))
1348 --node;
1349
1350 return pgdat;
1351}
1352
1353static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1354{
1355 pg_data_t *pgdat = (pg_data_t *)arg;
1356
1357 (*pos)++;
1358 return next_online_pgdat(pgdat);
1359}
1360
1361static void frag_stop(struct seq_file *m, void *arg)
1362{
1363}
1364
1365/*
1366 * Walk zones in a node and print using a callback.
1367 * If @assert_populated is true, only use callback for zones that are populated.
1368 */
1369static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1370 bool assert_populated, bool nolock,
1371 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1372{
1373 struct zone *zone;
1374 struct zone *node_zones = pgdat->node_zones;
1375 unsigned long flags;
1376
1377 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1378 if (assert_populated && !populated_zone(zone))
1379 continue;
1380
1381 if (!nolock)
1382 spin_lock_irqsave(&zone->lock, flags);
1383 print(m, pgdat, zone);
1384 if (!nolock)
1385 spin_unlock_irqrestore(&zone->lock, flags);
1386 }
1387}
1388#endif
1389
1390#ifdef CONFIG_PROC_FS
1391static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1392 struct zone *zone)
1393{
1394 int order;
1395
1396 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1397 for (order = 0; order < MAX_ORDER; ++order)
1398 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1399 seq_putc(m, '\n');
1400}
1401
1402/*
1403 * This walks the free areas for each zone.
1404 */
1405static int frag_show(struct seq_file *m, void *arg)
1406{
1407 pg_data_t *pgdat = (pg_data_t *)arg;
1408 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1409 return 0;
1410}
1411
1412static void pagetypeinfo_showfree_print(struct seq_file *m,
1413 pg_data_t *pgdat, struct zone *zone)
1414{
1415 int order, mtype;
1416
1417 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1418 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1419 pgdat->node_id,
1420 zone->name,
1421 migratetype_names[mtype]);
1422 for (order = 0; order < MAX_ORDER; ++order) {
1423 unsigned long freecount = 0;
1424 struct free_area *area;
1425 struct list_head *curr;
1426 bool overflow = false;
1427
1428 area = &(zone->free_area[order]);
1429
1430 list_for_each(curr, &area->free_list[mtype]) {
1431 /*
1432 * Cap the free_list iteration because it might
1433 * be really large and we are under a spinlock
1434 * so a long time spent here could trigger a
1435 * hard lockup detector. Anyway this is a
1436 * debugging tool so knowing there is a handful
1437 * of pages of this order should be more than
1438 * sufficient.
1439 */
1440 if (++freecount >= 100000) {
1441 overflow = true;
1442 break;
1443 }
1444 }
1445 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1446 spin_unlock_irq(&zone->lock);
1447 cond_resched();
1448 spin_lock_irq(&zone->lock);
1449 }
1450 seq_putc(m, '\n');
1451 }
1452}
1453
1454/* Print out the free pages at each order for each migatetype */
1455static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1456{
1457 int order;
1458 pg_data_t *pgdat = (pg_data_t *)arg;
1459
1460 /* Print header */
1461 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1462 for (order = 0; order < MAX_ORDER; ++order)
1463 seq_printf(m, "%6d ", order);
1464 seq_putc(m, '\n');
1465
1466 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1467
1468 return 0;
1469}
1470
1471static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1472 pg_data_t *pgdat, struct zone *zone)
1473{
1474 int mtype;
1475 unsigned long pfn;
1476 unsigned long start_pfn = zone->zone_start_pfn;
1477 unsigned long end_pfn = zone_end_pfn(zone);
1478 unsigned long count[MIGRATE_TYPES] = { 0, };
1479
1480 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1481 struct page *page;
1482
1483 page = pfn_to_online_page(pfn);
1484 if (!page)
1485 continue;
1486
1487 if (page_zone(page) != zone)
1488 continue;
1489
1490 mtype = get_pageblock_migratetype(page);
1491
1492 if (mtype < MIGRATE_TYPES)
1493 count[mtype]++;
1494 }
1495
1496 /* Print counts */
1497 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1498 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1499 seq_printf(m, "%12lu ", count[mtype]);
1500 seq_putc(m, '\n');
1501}
1502
1503/* Print out the number of pageblocks for each migratetype */
1504static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1505{
1506 int mtype;
1507 pg_data_t *pgdat = (pg_data_t *)arg;
1508
1509 seq_printf(m, "\n%-23s", "Number of blocks type ");
1510 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1511 seq_printf(m, "%12s ", migratetype_names[mtype]);
1512 seq_putc(m, '\n');
1513 walk_zones_in_node(m, pgdat, true, false,
1514 pagetypeinfo_showblockcount_print);
1515
1516 return 0;
1517}
1518
1519/*
1520 * Print out the number of pageblocks for each migratetype that contain pages
1521 * of other types. This gives an indication of how well fallbacks are being
1522 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1523 * to determine what is going on
1524 */
1525static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1526{
1527#ifdef CONFIG_PAGE_OWNER
1528 int mtype;
1529
1530 if (!static_branch_unlikely(&page_owner_inited))
1531 return;
1532
1533 drain_all_pages(NULL);
1534
1535 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1536 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1537 seq_printf(m, "%12s ", migratetype_names[mtype]);
1538 seq_putc(m, '\n');
1539
1540 walk_zones_in_node(m, pgdat, true, true,
1541 pagetypeinfo_showmixedcount_print);
1542#endif /* CONFIG_PAGE_OWNER */
1543}
1544
1545/*
1546 * This prints out statistics in relation to grouping pages by mobility.
1547 * It is expensive to collect so do not constantly read the file.
1548 */
1549static int pagetypeinfo_show(struct seq_file *m, void *arg)
1550{
1551 pg_data_t *pgdat = (pg_data_t *)arg;
1552
1553 /* check memoryless node */
1554 if (!node_state(pgdat->node_id, N_MEMORY))
1555 return 0;
1556
1557 seq_printf(m, "Page block order: %d\n", pageblock_order);
1558 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1559 seq_putc(m, '\n');
1560 pagetypeinfo_showfree(m, pgdat);
1561 pagetypeinfo_showblockcount(m, pgdat);
1562 pagetypeinfo_showmixedcount(m, pgdat);
1563
1564 return 0;
1565}
1566
1567static const struct seq_operations fragmentation_op = {
1568 .start = frag_start,
1569 .next = frag_next,
1570 .stop = frag_stop,
1571 .show = frag_show,
1572};
1573
1574static const struct seq_operations pagetypeinfo_op = {
1575 .start = frag_start,
1576 .next = frag_next,
1577 .stop = frag_stop,
1578 .show = pagetypeinfo_show,
1579};
1580
1581static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1582{
1583 int zid;
1584
1585 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1586 struct zone *compare = &pgdat->node_zones[zid];
1587
1588 if (populated_zone(compare))
1589 return zone == compare;
1590 }
1591
1592 return false;
1593}
1594
1595static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1596 struct zone *zone)
1597{
1598 int i;
1599 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1600 if (is_zone_first_populated(pgdat, zone)) {
1601 seq_printf(m, "\n per-node stats");
1602 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1603 unsigned long pages = node_page_state_pages(pgdat, i);
1604
1605 if (vmstat_item_print_in_thp(i))
1606 pages /= HPAGE_PMD_NR;
1607 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1608 pages);
1609 }
1610 }
1611 seq_printf(m,
1612 "\n pages free %lu"
1613 "\n min %lu"
1614 "\n low %lu"
1615 "\n high %lu"
1616 "\n spanned %lu"
1617 "\n present %lu"
1618 "\n managed %lu"
1619 "\n cma %lu",
1620 zone_page_state(zone, NR_FREE_PAGES),
1621 min_wmark_pages(zone),
1622 low_wmark_pages(zone),
1623 high_wmark_pages(zone),
1624 zone->spanned_pages,
1625 zone->present_pages,
1626 zone_managed_pages(zone),
1627 zone_cma_pages(zone));
1628
1629 seq_printf(m,
1630 "\n protection: (%ld",
1631 zone->lowmem_reserve[0]);
1632 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1633 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1634 seq_putc(m, ')');
1635
1636 /* If unpopulated, no other information is useful */
1637 if (!populated_zone(zone)) {
1638 seq_putc(m, '\n');
1639 return;
1640 }
1641
1642 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1643 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1644 zone_page_state(zone, i));
1645
1646#ifdef CONFIG_NUMA
1647 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1648 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1649 zone_numa_event_state(zone, i));
1650#endif
1651
1652 seq_printf(m, "\n pagesets");
1653 for_each_online_cpu(i) {
1654 struct per_cpu_pages *pcp;
1655 struct per_cpu_zonestat __maybe_unused *pzstats;
1656
1657 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1658 seq_printf(m,
1659 "\n cpu: %i"
1660 "\n count: %i"
1661 "\n high: %i"
1662 "\n batch: %i",
1663 i,
1664 pcp->count,
1665 pcp->high,
1666 pcp->batch);
1667#ifdef CONFIG_SMP
1668 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1669 seq_printf(m, "\n vm stats threshold: %d",
1670 pzstats->stat_threshold);
1671#endif
1672 }
1673 seq_printf(m,
1674 "\n node_unreclaimable: %u"
1675 "\n start_pfn: %lu",
1676 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1677 zone->zone_start_pfn);
1678 seq_putc(m, '\n');
1679}
1680
1681/*
1682 * Output information about zones in @pgdat. All zones are printed regardless
1683 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1684 * set of all zones and userspace would not be aware of such zones if they are
1685 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1686 */
1687static int zoneinfo_show(struct seq_file *m, void *arg)
1688{
1689 pg_data_t *pgdat = (pg_data_t *)arg;
1690 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1691 return 0;
1692}
1693
1694static const struct seq_operations zoneinfo_op = {
1695 .start = frag_start, /* iterate over all zones. The same as in
1696 * fragmentation. */
1697 .next = frag_next,
1698 .stop = frag_stop,
1699 .show = zoneinfo_show,
1700};
1701
1702#define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1703 NR_VM_NUMA_EVENT_ITEMS + \
1704 NR_VM_NODE_STAT_ITEMS + \
1705 NR_VM_WRITEBACK_STAT_ITEMS + \
1706 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1707 NR_VM_EVENT_ITEMS : 0))
1708
1709static void *vmstat_start(struct seq_file *m, loff_t *pos)
1710{
1711 unsigned long *v;
1712 int i;
1713
1714 if (*pos >= NR_VMSTAT_ITEMS)
1715 return NULL;
1716
1717 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1718 fold_vm_numa_events();
1719 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1720 m->private = v;
1721 if (!v)
1722 return ERR_PTR(-ENOMEM);
1723 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1724 v[i] = global_zone_page_state(i);
1725 v += NR_VM_ZONE_STAT_ITEMS;
1726
1727#ifdef CONFIG_NUMA
1728 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1729 v[i] = global_numa_event_state(i);
1730 v += NR_VM_NUMA_EVENT_ITEMS;
1731#endif
1732
1733 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1734 v[i] = global_node_page_state_pages(i);
1735 if (vmstat_item_print_in_thp(i))
1736 v[i] /= HPAGE_PMD_NR;
1737 }
1738 v += NR_VM_NODE_STAT_ITEMS;
1739
1740 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1741 v + NR_DIRTY_THRESHOLD);
1742 v += NR_VM_WRITEBACK_STAT_ITEMS;
1743
1744#ifdef CONFIG_VM_EVENT_COUNTERS
1745 all_vm_events(v);
1746 v[PGPGIN] /= 2; /* sectors -> kbytes */
1747 v[PGPGOUT] /= 2;
1748#endif
1749 return (unsigned long *)m->private + *pos;
1750}
1751
1752static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1753{
1754 (*pos)++;
1755 if (*pos >= NR_VMSTAT_ITEMS)
1756 return NULL;
1757 return (unsigned long *)m->private + *pos;
1758}
1759
1760static int vmstat_show(struct seq_file *m, void *arg)
1761{
1762 unsigned long *l = arg;
1763 unsigned long off = l - (unsigned long *)m->private;
1764
1765 seq_puts(m, vmstat_text[off]);
1766 seq_put_decimal_ull(m, " ", *l);
1767 seq_putc(m, '\n');
1768
1769 if (off == NR_VMSTAT_ITEMS - 1) {
1770 /*
1771 * We've come to the end - add any deprecated counters to avoid
1772 * breaking userspace which might depend on them being present.
1773 */
1774 seq_puts(m, "nr_unstable 0\n");
1775 }
1776 return 0;
1777}
1778
1779static void vmstat_stop(struct seq_file *m, void *arg)
1780{
1781 kfree(m->private);
1782 m->private = NULL;
1783}
1784
1785static const struct seq_operations vmstat_op = {
1786 .start = vmstat_start,
1787 .next = vmstat_next,
1788 .stop = vmstat_stop,
1789 .show = vmstat_show,
1790};
1791#endif /* CONFIG_PROC_FS */
1792
1793#ifdef CONFIG_SMP
1794static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1795int sysctl_stat_interval __read_mostly = HZ;
1796
1797#ifdef CONFIG_PROC_FS
1798static void refresh_vm_stats(struct work_struct *work)
1799{
1800 refresh_cpu_vm_stats(true);
1801}
1802
1803int vmstat_refresh(struct ctl_table *table, int write,
1804 void *buffer, size_t *lenp, loff_t *ppos)
1805{
1806 long val;
1807 int err;
1808 int i;
1809
1810 /*
1811 * The regular update, every sysctl_stat_interval, may come later
1812 * than expected: leaving a significant amount in per_cpu buckets.
1813 * This is particularly misleading when checking a quantity of HUGE
1814 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1815 * which can equally be echo'ed to or cat'ted from (by root),
1816 * can be used to update the stats just before reading them.
1817 *
1818 * Oh, and since global_zone_page_state() etc. are so careful to hide
1819 * transiently negative values, report an error here if any of
1820 * the stats is negative, so we know to go looking for imbalance.
1821 */
1822 err = schedule_on_each_cpu(refresh_vm_stats);
1823 if (err)
1824 return err;
1825 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1826 /*
1827 * Skip checking stats known to go negative occasionally.
1828 */
1829 switch (i) {
1830 case NR_ZONE_WRITE_PENDING:
1831 case NR_FREE_CMA_PAGES:
1832 continue;
1833 }
1834 val = atomic_long_read(&vm_zone_stat[i]);
1835 if (val < 0) {
1836 pr_warn("%s: %s %ld\n",
1837 __func__, zone_stat_name(i), val);
1838 }
1839 }
1840 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1841 /*
1842 * Skip checking stats known to go negative occasionally.
1843 */
1844 switch (i) {
1845 case NR_WRITEBACK:
1846 continue;
1847 }
1848 val = atomic_long_read(&vm_node_stat[i]);
1849 if (val < 0) {
1850 pr_warn("%s: %s %ld\n",
1851 __func__, node_stat_name(i), val);
1852 }
1853 }
1854 if (write)
1855 *ppos += *lenp;
1856 else
1857 *lenp = 0;
1858 return 0;
1859}
1860#endif /* CONFIG_PROC_FS */
1861
1862static void vmstat_update(struct work_struct *w)
1863{
1864 if (refresh_cpu_vm_stats(true)) {
1865 /*
1866 * Counters were updated so we expect more updates
1867 * to occur in the future. Keep on running the
1868 * update worker thread.
1869 */
1870 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1871 this_cpu_ptr(&vmstat_work),
1872 round_jiffies_relative(sysctl_stat_interval));
1873 }
1874}
1875
1876/*
1877 * Switch off vmstat processing and then fold all the remaining differentials
1878 * until the diffs stay at zero. The function is used by NOHZ and can only be
1879 * invoked when tick processing is not active.
1880 */
1881/*
1882 * Check if the diffs for a certain cpu indicate that
1883 * an update is needed.
1884 */
1885static bool need_update(int cpu)
1886{
1887 pg_data_t *last_pgdat = NULL;
1888 struct zone *zone;
1889
1890 for_each_populated_zone(zone) {
1891 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1892 struct per_cpu_nodestat *n;
1893
1894 /*
1895 * The fast way of checking if there are any vmstat diffs.
1896 */
1897 if (memchr_inv(pzstats->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1898 sizeof(pzstats->vm_stat_diff[0])))
1899 return true;
1900
1901 if (last_pgdat == zone->zone_pgdat)
1902 continue;
1903 last_pgdat = zone->zone_pgdat;
1904 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1905 if (memchr_inv(n->vm_node_stat_diff, 0, NR_VM_NODE_STAT_ITEMS *
1906 sizeof(n->vm_node_stat_diff[0])))
1907 return true;
1908 }
1909 return false;
1910}
1911
1912/*
1913 * Switch off vmstat processing and then fold all the remaining differentials
1914 * until the diffs stay at zero. The function is used by NOHZ and can only be
1915 * invoked when tick processing is not active.
1916 */
1917void quiet_vmstat(void)
1918{
1919 if (system_state != SYSTEM_RUNNING)
1920 return;
1921
1922 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1923 return;
1924
1925 if (!need_update(smp_processor_id()))
1926 return;
1927
1928 /*
1929 * Just refresh counters and do not care about the pending delayed
1930 * vmstat_update. It doesn't fire that often to matter and canceling
1931 * it would be too expensive from this path.
1932 * vmstat_shepherd will take care about that for us.
1933 */
1934 refresh_cpu_vm_stats(false);
1935}
1936
1937/*
1938 * Shepherd worker thread that checks the
1939 * differentials of processors that have their worker
1940 * threads for vm statistics updates disabled because of
1941 * inactivity.
1942 */
1943static void vmstat_shepherd(struct work_struct *w);
1944
1945static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1946
1947static void vmstat_shepherd(struct work_struct *w)
1948{
1949 int cpu;
1950
1951 get_online_cpus();
1952 /* Check processors whose vmstat worker threads have been disabled */
1953 for_each_online_cpu(cpu) {
1954 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1955
1956 if (!delayed_work_pending(dw) && need_update(cpu))
1957 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1958
1959 cond_resched();
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