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_owner.h>
30#include <linux/sched/isolation.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(const 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 cpus_read_lock();
133 sum_vm_events(ret);
134 cpus_read_unlock();
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_NUMA
169static void fold_vm_zone_numa_events(struct zone *zone)
170{
171 unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
172 int cpu;
173 enum numa_stat_item item;
174
175 for_each_online_cpu(cpu) {
176 struct per_cpu_zonestat *pzstats;
177
178 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
179 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
180 zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
181 }
182
183 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
184 zone_numa_event_add(zone_numa_events[item], zone, item);
185}
186
187void fold_vm_numa_events(void)
188{
189 struct zone *zone;
190
191 for_each_populated_zone(zone)
192 fold_vm_zone_numa_events(zone);
193}
194#endif
195
196#ifdef CONFIG_SMP
197
198int calculate_pressure_threshold(struct zone *zone)
199{
200 int threshold;
201 int watermark_distance;
202
203 /*
204 * As vmstats are not up to date, there is drift between the estimated
205 * and real values. For high thresholds and a high number of CPUs, it
206 * is possible for the min watermark to be breached while the estimated
207 * value looks fine. The pressure threshold is a reduced value such
208 * that even the maximum amount of drift will not accidentally breach
209 * the min watermark
210 */
211 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
212 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
213
214 /*
215 * Maximum threshold is 125
216 */
217 threshold = min(125, threshold);
218
219 return threshold;
220}
221
222int calculate_normal_threshold(struct zone *zone)
223{
224 int threshold;
225 int mem; /* memory in 128 MB units */
226
227 /*
228 * The threshold scales with the number of processors and the amount
229 * of memory per zone. More memory means that we can defer updates for
230 * longer, more processors could lead to more contention.
231 * fls() is used to have a cheap way of logarithmic scaling.
232 *
233 * Some sample thresholds:
234 *
235 * Threshold Processors (fls) Zonesize fls(mem)+1
236 * ------------------------------------------------------------------
237 * 8 1 1 0.9-1 GB 4
238 * 16 2 2 0.9-1 GB 4
239 * 20 2 2 1-2 GB 5
240 * 24 2 2 2-4 GB 6
241 * 28 2 2 4-8 GB 7
242 * 32 2 2 8-16 GB 8
243 * 4 2 2 <128M 1
244 * 30 4 3 2-4 GB 5
245 * 48 4 3 8-16 GB 8
246 * 32 8 4 1-2 GB 4
247 * 32 8 4 0.9-1GB 4
248 * 10 16 5 <128M 1
249 * 40 16 5 900M 4
250 * 70 64 7 2-4 GB 5
251 * 84 64 7 4-8 GB 6
252 * 108 512 9 4-8 GB 6
253 * 125 1024 10 8-16 GB 8
254 * 125 1024 10 16-32 GB 9
255 */
256
257 mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
258
259 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
260
261 /*
262 * Maximum threshold is 125
263 */
264 threshold = min(125, threshold);
265
266 return threshold;
267}
268
269/*
270 * Refresh the thresholds for each zone.
271 */
272void refresh_zone_stat_thresholds(void)
273{
274 struct pglist_data *pgdat;
275 struct zone *zone;
276 int cpu;
277 int threshold;
278
279 /* Zero current pgdat thresholds */
280 for_each_online_pgdat(pgdat) {
281 for_each_online_cpu(cpu) {
282 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
283 }
284 }
285
286 for_each_populated_zone(zone) {
287 struct pglist_data *pgdat = zone->zone_pgdat;
288 unsigned long max_drift, tolerate_drift;
289
290 threshold = calculate_normal_threshold(zone);
291
292 for_each_online_cpu(cpu) {
293 int pgdat_threshold;
294
295 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
296 = threshold;
297
298 /* Base nodestat threshold on the largest populated zone. */
299 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
300 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
301 = max(threshold, pgdat_threshold);
302 }
303
304 /*
305 * Only set percpu_drift_mark if there is a danger that
306 * NR_FREE_PAGES reports the low watermark is ok when in fact
307 * the min watermark could be breached by an allocation
308 */
309 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
310 max_drift = num_online_cpus() * threshold;
311 if (max_drift > tolerate_drift)
312 zone->percpu_drift_mark = high_wmark_pages(zone) +
313 max_drift;
314 }
315}
316
317void set_pgdat_percpu_threshold(pg_data_t *pgdat,
318 int (*calculate_pressure)(struct zone *))
319{
320 struct zone *zone;
321 int cpu;
322 int threshold;
323 int i;
324
325 for (i = 0; i < pgdat->nr_zones; i++) {
326 zone = &pgdat->node_zones[i];
327 if (!zone->percpu_drift_mark)
328 continue;
329
330 threshold = (*calculate_pressure)(zone);
331 for_each_online_cpu(cpu)
332 per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
333 = threshold;
334 }
335}
336
337/*
338 * For use when we know that interrupts are disabled,
339 * or when we know that preemption is disabled and that
340 * particular counter cannot be updated from interrupt context.
341 */
342void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
343 long delta)
344{
345 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
346 s8 __percpu *p = pcp->vm_stat_diff + item;
347 long x;
348 long t;
349
350 /*
351 * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
352 * atomicity is provided by IRQs being disabled -- either explicitly
353 * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
354 * CPU migrations and preemption potentially corrupts a counter so
355 * disable preemption.
356 */
357 preempt_disable_nested();
358
359 x = delta + __this_cpu_read(*p);
360
361 t = __this_cpu_read(pcp->stat_threshold);
362
363 if (unlikely(abs(x) > t)) {
364 zone_page_state_add(x, zone, item);
365 x = 0;
366 }
367 __this_cpu_write(*p, x);
368
369 preempt_enable_nested();
370}
371EXPORT_SYMBOL(__mod_zone_page_state);
372
373void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
374 long delta)
375{
376 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
377 s8 __percpu *p = pcp->vm_node_stat_diff + item;
378 long x;
379 long t;
380
381 if (vmstat_item_in_bytes(item)) {
382 /*
383 * Only cgroups use subpage accounting right now; at
384 * the global level, these items still change in
385 * multiples of whole pages. Store them as pages
386 * internally to keep the per-cpu counters compact.
387 */
388 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
389 delta >>= PAGE_SHIFT;
390 }
391
392 /* See __mod_node_page_state */
393 preempt_disable_nested();
394
395 x = delta + __this_cpu_read(*p);
396
397 t = __this_cpu_read(pcp->stat_threshold);
398
399 if (unlikely(abs(x) > t)) {
400 node_page_state_add(x, pgdat, item);
401 x = 0;
402 }
403 __this_cpu_write(*p, x);
404
405 preempt_enable_nested();
406}
407EXPORT_SYMBOL(__mod_node_page_state);
408
409/*
410 * Optimized increment and decrement functions.
411 *
412 * These are only for a single page and therefore can take a struct page *
413 * argument instead of struct zone *. This allows the inclusion of the code
414 * generated for page_zone(page) into the optimized functions.
415 *
416 * No overflow check is necessary and therefore the differential can be
417 * incremented or decremented in place which may allow the compilers to
418 * generate better code.
419 * The increment or decrement is known and therefore one boundary check can
420 * be omitted.
421 *
422 * NOTE: These functions are very performance sensitive. Change only
423 * with care.
424 *
425 * Some processors have inc/dec instructions that are atomic vs an interrupt.
426 * However, the code must first determine the differential location in a zone
427 * based on the processor number and then inc/dec the counter. There is no
428 * guarantee without disabling preemption that the processor will not change
429 * in between and therefore the atomicity vs. interrupt cannot be exploited
430 * in a useful way here.
431 */
432void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
433{
434 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
435 s8 __percpu *p = pcp->vm_stat_diff + item;
436 s8 v, t;
437
438 /* See __mod_node_page_state */
439 preempt_disable_nested();
440
441 v = __this_cpu_inc_return(*p);
442 t = __this_cpu_read(pcp->stat_threshold);
443 if (unlikely(v > t)) {
444 s8 overstep = t >> 1;
445
446 zone_page_state_add(v + overstep, zone, item);
447 __this_cpu_write(*p, -overstep);
448 }
449
450 preempt_enable_nested();
451}
452
453void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
454{
455 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
456 s8 __percpu *p = pcp->vm_node_stat_diff + item;
457 s8 v, t;
458
459 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
460
461 /* See __mod_node_page_state */
462 preempt_disable_nested();
463
464 v = __this_cpu_inc_return(*p);
465 t = __this_cpu_read(pcp->stat_threshold);
466 if (unlikely(v > t)) {
467 s8 overstep = t >> 1;
468
469 node_page_state_add(v + overstep, pgdat, item);
470 __this_cpu_write(*p, -overstep);
471 }
472
473 preempt_enable_nested();
474}
475
476void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
477{
478 __inc_zone_state(page_zone(page), item);
479}
480EXPORT_SYMBOL(__inc_zone_page_state);
481
482void __inc_node_page_state(struct page *page, enum node_stat_item item)
483{
484 __inc_node_state(page_pgdat(page), item);
485}
486EXPORT_SYMBOL(__inc_node_page_state);
487
488void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
489{
490 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
491 s8 __percpu *p = pcp->vm_stat_diff + item;
492 s8 v, t;
493
494 /* See __mod_node_page_state */
495 preempt_disable_nested();
496
497 v = __this_cpu_dec_return(*p);
498 t = __this_cpu_read(pcp->stat_threshold);
499 if (unlikely(v < - t)) {
500 s8 overstep = t >> 1;
501
502 zone_page_state_add(v - overstep, zone, item);
503 __this_cpu_write(*p, overstep);
504 }
505
506 preempt_enable_nested();
507}
508
509void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
510{
511 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
512 s8 __percpu *p = pcp->vm_node_stat_diff + item;
513 s8 v, t;
514
515 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
516
517 /* See __mod_node_page_state */
518 preempt_disable_nested();
519
520 v = __this_cpu_dec_return(*p);
521 t = __this_cpu_read(pcp->stat_threshold);
522 if (unlikely(v < - t)) {
523 s8 overstep = t >> 1;
524
525 node_page_state_add(v - overstep, pgdat, item);
526 __this_cpu_write(*p, overstep);
527 }
528
529 preempt_enable_nested();
530}
531
532void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
533{
534 __dec_zone_state(page_zone(page), item);
535}
536EXPORT_SYMBOL(__dec_zone_page_state);
537
538void __dec_node_page_state(struct page *page, enum node_stat_item item)
539{
540 __dec_node_state(page_pgdat(page), item);
541}
542EXPORT_SYMBOL(__dec_node_page_state);
543
544#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
545/*
546 * If we have cmpxchg_local support then we do not need to incur the overhead
547 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
548 *
549 * mod_state() modifies the zone counter state through atomic per cpu
550 * operations.
551 *
552 * Overstep mode specifies how overstep should handled:
553 * 0 No overstepping
554 * 1 Overstepping half of threshold
555 * -1 Overstepping minus half of threshold
556*/
557static inline void mod_zone_state(struct zone *zone,
558 enum zone_stat_item item, long delta, int overstep_mode)
559{
560 struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
561 s8 __percpu *p = pcp->vm_stat_diff + item;
562 long n, t, z;
563 s8 o;
564
565 o = this_cpu_read(*p);
566 do {
567 z = 0; /* overflow to zone counters */
568
569 /*
570 * The fetching of the stat_threshold is racy. We may apply
571 * a counter threshold to the wrong the cpu if we get
572 * rescheduled while executing here. However, the next
573 * counter update will apply the threshold again and
574 * therefore bring the counter under the threshold again.
575 *
576 * Most of the time the thresholds are the same anyways
577 * for all cpus in a zone.
578 */
579 t = this_cpu_read(pcp->stat_threshold);
580
581 n = delta + (long)o;
582
583 if (abs(n) > t) {
584 int os = overstep_mode * (t >> 1) ;
585
586 /* Overflow must be added to zone counters */
587 z = n + os;
588 n = -os;
589 }
590 } while (!this_cpu_try_cmpxchg(*p, &o, n));
591
592 if (z)
593 zone_page_state_add(z, zone, item);
594}
595
596void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
597 long delta)
598{
599 mod_zone_state(zone, item, delta, 0);
600}
601EXPORT_SYMBOL(mod_zone_page_state);
602
603void inc_zone_page_state(struct page *page, enum zone_stat_item item)
604{
605 mod_zone_state(page_zone(page), item, 1, 1);
606}
607EXPORT_SYMBOL(inc_zone_page_state);
608
609void dec_zone_page_state(struct page *page, enum zone_stat_item item)
610{
611 mod_zone_state(page_zone(page), item, -1, -1);
612}
613EXPORT_SYMBOL(dec_zone_page_state);
614
615static inline void mod_node_state(struct pglist_data *pgdat,
616 enum node_stat_item item, int delta, int overstep_mode)
617{
618 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
619 s8 __percpu *p = pcp->vm_node_stat_diff + item;
620 long n, t, z;
621 s8 o;
622
623 if (vmstat_item_in_bytes(item)) {
624 /*
625 * Only cgroups use subpage accounting right now; at
626 * the global level, these items still change in
627 * multiples of whole pages. Store them as pages
628 * internally to keep the per-cpu counters compact.
629 */
630 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
631 delta >>= PAGE_SHIFT;
632 }
633
634 o = this_cpu_read(*p);
635 do {
636 z = 0; /* overflow to node counters */
637
638 /*
639 * The fetching of the stat_threshold is racy. We may apply
640 * a counter threshold to the wrong the cpu if we get
641 * rescheduled while executing here. However, the next
642 * counter update will apply the threshold again and
643 * therefore bring the counter under the threshold again.
644 *
645 * Most of the time the thresholds are the same anyways
646 * for all cpus in a node.
647 */
648 t = this_cpu_read(pcp->stat_threshold);
649
650 n = delta + (long)o;
651
652 if (abs(n) > t) {
653 int os = overstep_mode * (t >> 1) ;
654
655 /* Overflow must be added to node counters */
656 z = n + os;
657 n = -os;
658 }
659 } while (!this_cpu_try_cmpxchg(*p, &o, n));
660
661 if (z)
662 node_page_state_add(z, pgdat, item);
663}
664
665void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
666 long delta)
667{
668 mod_node_state(pgdat, item, delta, 0);
669}
670EXPORT_SYMBOL(mod_node_page_state);
671
672void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
673{
674 mod_node_state(pgdat, item, 1, 1);
675}
676
677void inc_node_page_state(struct page *page, enum node_stat_item item)
678{
679 mod_node_state(page_pgdat(page), item, 1, 1);
680}
681EXPORT_SYMBOL(inc_node_page_state);
682
683void dec_node_page_state(struct page *page, enum node_stat_item item)
684{
685 mod_node_state(page_pgdat(page), item, -1, -1);
686}
687EXPORT_SYMBOL(dec_node_page_state);
688#else
689/*
690 * Use interrupt disable to serialize counter updates
691 */
692void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
693 long delta)
694{
695 unsigned long flags;
696
697 local_irq_save(flags);
698 __mod_zone_page_state(zone, item, delta);
699 local_irq_restore(flags);
700}
701EXPORT_SYMBOL(mod_zone_page_state);
702
703void inc_zone_page_state(struct page *page, enum zone_stat_item item)
704{
705 unsigned long flags;
706 struct zone *zone;
707
708 zone = page_zone(page);
709 local_irq_save(flags);
710 __inc_zone_state(zone, item);
711 local_irq_restore(flags);
712}
713EXPORT_SYMBOL(inc_zone_page_state);
714
715void dec_zone_page_state(struct page *page, enum zone_stat_item item)
716{
717 unsigned long flags;
718
719 local_irq_save(flags);
720 __dec_zone_page_state(page, item);
721 local_irq_restore(flags);
722}
723EXPORT_SYMBOL(dec_zone_page_state);
724
725void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
726{
727 unsigned long flags;
728
729 local_irq_save(flags);
730 __inc_node_state(pgdat, item);
731 local_irq_restore(flags);
732}
733EXPORT_SYMBOL(inc_node_state);
734
735void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
736 long delta)
737{
738 unsigned long flags;
739
740 local_irq_save(flags);
741 __mod_node_page_state(pgdat, item, delta);
742 local_irq_restore(flags);
743}
744EXPORT_SYMBOL(mod_node_page_state);
745
746void inc_node_page_state(struct page *page, enum node_stat_item item)
747{
748 unsigned long flags;
749 struct pglist_data *pgdat;
750
751 pgdat = page_pgdat(page);
752 local_irq_save(flags);
753 __inc_node_state(pgdat, item);
754 local_irq_restore(flags);
755}
756EXPORT_SYMBOL(inc_node_page_state);
757
758void dec_node_page_state(struct page *page, enum node_stat_item item)
759{
760 unsigned long flags;
761
762 local_irq_save(flags);
763 __dec_node_page_state(page, item);
764 local_irq_restore(flags);
765}
766EXPORT_SYMBOL(dec_node_page_state);
767#endif
768
769/*
770 * Fold a differential into the global counters.
771 * Returns the number of counters updated.
772 */
773static int fold_diff(int *zone_diff, int *node_diff)
774{
775 int i;
776 int changes = 0;
777
778 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
779 if (zone_diff[i]) {
780 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
781 changes++;
782 }
783
784 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
785 if (node_diff[i]) {
786 atomic_long_add(node_diff[i], &vm_node_stat[i]);
787 changes++;
788 }
789 return changes;
790}
791
792/*
793 * Update the zone counters for the current cpu.
794 *
795 * Note that refresh_cpu_vm_stats strives to only access
796 * node local memory. The per cpu pagesets on remote zones are placed
797 * in the memory local to the processor using that pageset. So the
798 * loop over all zones will access a series of cachelines local to
799 * the processor.
800 *
801 * The call to zone_page_state_add updates the cachelines with the
802 * statistics in the remote zone struct as well as the global cachelines
803 * with the global counters. These could cause remote node cache line
804 * bouncing and will have to be only done when necessary.
805 *
806 * The function returns the number of global counters updated.
807 */
808static int refresh_cpu_vm_stats(bool do_pagesets)
809{
810 struct pglist_data *pgdat;
811 struct zone *zone;
812 int i;
813 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
814 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
815 int changes = 0;
816
817 for_each_populated_zone(zone) {
818 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
819 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
820
821 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
822 int v;
823
824 v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
825 if (v) {
826
827 atomic_long_add(v, &zone->vm_stat[i]);
828 global_zone_diff[i] += v;
829#ifdef CONFIG_NUMA
830 /* 3 seconds idle till flush */
831 __this_cpu_write(pcp->expire, 3);
832#endif
833 }
834 }
835
836 if (do_pagesets) {
837 cond_resched();
838
839 changes += decay_pcp_high(zone, this_cpu_ptr(pcp));
840#ifdef CONFIG_NUMA
841 /*
842 * Deal with draining the remote pageset of this
843 * processor
844 *
845 * Check if there are pages remaining in this pageset
846 * if not then there is nothing to expire.
847 */
848 if (!__this_cpu_read(pcp->expire) ||
849 !__this_cpu_read(pcp->count))
850 continue;
851
852 /*
853 * We never drain zones local to this processor.
854 */
855 if (zone_to_nid(zone) == numa_node_id()) {
856 __this_cpu_write(pcp->expire, 0);
857 continue;
858 }
859
860 if (__this_cpu_dec_return(pcp->expire)) {
861 changes++;
862 continue;
863 }
864
865 if (__this_cpu_read(pcp->count)) {
866 drain_zone_pages(zone, this_cpu_ptr(pcp));
867 changes++;
868 }
869#endif
870 }
871 }
872
873 for_each_online_pgdat(pgdat) {
874 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
875
876 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
877 int v;
878
879 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
880 if (v) {
881 atomic_long_add(v, &pgdat->vm_stat[i]);
882 global_node_diff[i] += v;
883 }
884 }
885 }
886
887 changes += fold_diff(global_zone_diff, global_node_diff);
888 return changes;
889}
890
891/*
892 * Fold the data for an offline cpu into the global array.
893 * There cannot be any access by the offline cpu and therefore
894 * synchronization is simplified.
895 */
896void cpu_vm_stats_fold(int cpu)
897{
898 struct pglist_data *pgdat;
899 struct zone *zone;
900 int i;
901 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
902 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
903
904 for_each_populated_zone(zone) {
905 struct per_cpu_zonestat *pzstats;
906
907 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
908
909 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
910 if (pzstats->vm_stat_diff[i]) {
911 int v;
912
913 v = pzstats->vm_stat_diff[i];
914 pzstats->vm_stat_diff[i] = 0;
915 atomic_long_add(v, &zone->vm_stat[i]);
916 global_zone_diff[i] += v;
917 }
918 }
919#ifdef CONFIG_NUMA
920 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
921 if (pzstats->vm_numa_event[i]) {
922 unsigned long v;
923
924 v = pzstats->vm_numa_event[i];
925 pzstats->vm_numa_event[i] = 0;
926 zone_numa_event_add(v, zone, i);
927 }
928 }
929#endif
930 }
931
932 for_each_online_pgdat(pgdat) {
933 struct per_cpu_nodestat *p;
934
935 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
936
937 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
938 if (p->vm_node_stat_diff[i]) {
939 int v;
940
941 v = p->vm_node_stat_diff[i];
942 p->vm_node_stat_diff[i] = 0;
943 atomic_long_add(v, &pgdat->vm_stat[i]);
944 global_node_diff[i] += v;
945 }
946 }
947
948 fold_diff(global_zone_diff, global_node_diff);
949}
950
951/*
952 * this is only called if !populated_zone(zone), which implies no other users of
953 * pset->vm_stat_diff[] exist.
954 */
955void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
956{
957 unsigned long v;
958 int i;
959
960 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
961 if (pzstats->vm_stat_diff[i]) {
962 v = pzstats->vm_stat_diff[i];
963 pzstats->vm_stat_diff[i] = 0;
964 zone_page_state_add(v, zone, i);
965 }
966 }
967
968#ifdef CONFIG_NUMA
969 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
970 if (pzstats->vm_numa_event[i]) {
971 v = pzstats->vm_numa_event[i];
972 pzstats->vm_numa_event[i] = 0;
973 zone_numa_event_add(v, zone, i);
974 }
975 }
976#endif
977}
978#endif
979
980#ifdef CONFIG_NUMA
981/*
982 * Determine the per node value of a stat item. This function
983 * is called frequently in a NUMA machine, so try to be as
984 * frugal as possible.
985 */
986unsigned long sum_zone_node_page_state(int node,
987 enum zone_stat_item item)
988{
989 struct zone *zones = NODE_DATA(node)->node_zones;
990 int i;
991 unsigned long count = 0;
992
993 for (i = 0; i < MAX_NR_ZONES; i++)
994 count += zone_page_state(zones + i, item);
995
996 return count;
997}
998
999/* Determine the per node value of a numa stat item. */
1000unsigned long sum_zone_numa_event_state(int node,
1001 enum numa_stat_item item)
1002{
1003 struct zone *zones = NODE_DATA(node)->node_zones;
1004 unsigned long count = 0;
1005 int i;
1006
1007 for (i = 0; i < MAX_NR_ZONES; i++)
1008 count += zone_numa_event_state(zones + i, item);
1009
1010 return count;
1011}
1012
1013/*
1014 * Determine the per node value of a stat item.
1015 */
1016unsigned long node_page_state_pages(struct pglist_data *pgdat,
1017 enum node_stat_item item)
1018{
1019 long x = atomic_long_read(&pgdat->vm_stat[item]);
1020#ifdef CONFIG_SMP
1021 if (x < 0)
1022 x = 0;
1023#endif
1024 return x;
1025}
1026
1027unsigned long node_page_state(struct pglist_data *pgdat,
1028 enum node_stat_item item)
1029{
1030 VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1031
1032 return node_page_state_pages(pgdat, item);
1033}
1034#endif
1035
1036/*
1037 * Count number of pages "struct page" and "struct page_ext" consume.
1038 * nr_memmap_boot_pages: # of pages allocated by boot allocator
1039 * nr_memmap_pages: # of pages that were allocated by buddy allocator
1040 */
1041static atomic_long_t nr_memmap_boot_pages = ATOMIC_LONG_INIT(0);
1042static atomic_long_t nr_memmap_pages = ATOMIC_LONG_INIT(0);
1043
1044void memmap_boot_pages_add(long delta)
1045{
1046 atomic_long_add(delta, &nr_memmap_boot_pages);
1047}
1048
1049void memmap_pages_add(long delta)
1050{
1051 atomic_long_add(delta, &nr_memmap_pages);
1052}
1053
1054#ifdef CONFIG_COMPACTION
1055
1056struct contig_page_info {
1057 unsigned long free_pages;
1058 unsigned long free_blocks_total;
1059 unsigned long free_blocks_suitable;
1060};
1061
1062/*
1063 * Calculate the number of free pages in a zone, how many contiguous
1064 * pages are free and how many are large enough to satisfy an allocation of
1065 * the target size. Note that this function makes no attempt to estimate
1066 * how many suitable free blocks there *might* be if MOVABLE pages were
1067 * migrated. Calculating that is possible, but expensive and can be
1068 * figured out from userspace
1069 */
1070static void fill_contig_page_info(struct zone *zone,
1071 unsigned int suitable_order,
1072 struct contig_page_info *info)
1073{
1074 unsigned int order;
1075
1076 info->free_pages = 0;
1077 info->free_blocks_total = 0;
1078 info->free_blocks_suitable = 0;
1079
1080 for (order = 0; order < NR_PAGE_ORDERS; order++) {
1081 unsigned long blocks;
1082
1083 /*
1084 * Count number of free blocks.
1085 *
1086 * Access to nr_free is lockless as nr_free is used only for
1087 * diagnostic purposes. Use data_race to avoid KCSAN warning.
1088 */
1089 blocks = data_race(zone->free_area[order].nr_free);
1090 info->free_blocks_total += blocks;
1091
1092 /* Count free base pages */
1093 info->free_pages += blocks << order;
1094
1095 /* Count the suitable free blocks */
1096 if (order >= suitable_order)
1097 info->free_blocks_suitable += blocks <<
1098 (order - suitable_order);
1099 }
1100}
1101
1102/*
1103 * A fragmentation index only makes sense if an allocation of a requested
1104 * size would fail. If that is true, the fragmentation index indicates
1105 * whether external fragmentation or a lack of memory was the problem.
1106 * The value can be used to determine if page reclaim or compaction
1107 * should be used
1108 */
1109static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1110{
1111 unsigned long requested = 1UL << order;
1112
1113 if (WARN_ON_ONCE(order > MAX_PAGE_ORDER))
1114 return 0;
1115
1116 if (!info->free_blocks_total)
1117 return 0;
1118
1119 /* Fragmentation index only makes sense when a request would fail */
1120 if (info->free_blocks_suitable)
1121 return -1000;
1122
1123 /*
1124 * Index is between 0 and 1 so return within 3 decimal places
1125 *
1126 * 0 => allocation would fail due to lack of memory
1127 * 1 => allocation would fail due to fragmentation
1128 */
1129 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1130}
1131
1132/*
1133 * Calculates external fragmentation within a zone wrt the given order.
1134 * It is defined as the percentage of pages found in blocks of size
1135 * less than 1 << order. It returns values in range [0, 100].
1136 */
1137unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1138{
1139 struct contig_page_info info;
1140
1141 fill_contig_page_info(zone, order, &info);
1142 if (info.free_pages == 0)
1143 return 0;
1144
1145 return div_u64((info.free_pages -
1146 (info.free_blocks_suitable << order)) * 100,
1147 info.free_pages);
1148}
1149
1150/* Same as __fragmentation index but allocs contig_page_info on stack */
1151int fragmentation_index(struct zone *zone, unsigned int order)
1152{
1153 struct contig_page_info info;
1154
1155 fill_contig_page_info(zone, order, &info);
1156 return __fragmentation_index(order, &info);
1157}
1158#endif
1159
1160#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1161 defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1162#ifdef CONFIG_ZONE_DMA
1163#define TEXT_FOR_DMA(xx) xx "_dma",
1164#else
1165#define TEXT_FOR_DMA(xx)
1166#endif
1167
1168#ifdef CONFIG_ZONE_DMA32
1169#define TEXT_FOR_DMA32(xx) xx "_dma32",
1170#else
1171#define TEXT_FOR_DMA32(xx)
1172#endif
1173
1174#ifdef CONFIG_HIGHMEM
1175#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1176#else
1177#define TEXT_FOR_HIGHMEM(xx)
1178#endif
1179
1180#ifdef CONFIG_ZONE_DEVICE
1181#define TEXT_FOR_DEVICE(xx) xx "_device",
1182#else
1183#define TEXT_FOR_DEVICE(xx)
1184#endif
1185
1186#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1187 TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1188 TEXT_FOR_DEVICE(xx)
1189
1190const char * const vmstat_text[] = {
1191 /* enum zone_stat_item counters */
1192 "nr_free_pages",
1193 "nr_zone_inactive_anon",
1194 "nr_zone_active_anon",
1195 "nr_zone_inactive_file",
1196 "nr_zone_active_file",
1197 "nr_zone_unevictable",
1198 "nr_zone_write_pending",
1199 "nr_mlock",
1200 "nr_bounce",
1201#if IS_ENABLED(CONFIG_ZSMALLOC)
1202 "nr_zspages",
1203#endif
1204 "nr_free_cma",
1205#ifdef CONFIG_UNACCEPTED_MEMORY
1206 "nr_unaccepted",
1207#endif
1208
1209 /* enum numa_stat_item counters */
1210#ifdef CONFIG_NUMA
1211 "numa_hit",
1212 "numa_miss",
1213 "numa_foreign",
1214 "numa_interleave",
1215 "numa_local",
1216 "numa_other",
1217#endif
1218
1219 /* enum node_stat_item counters */
1220 "nr_inactive_anon",
1221 "nr_active_anon",
1222 "nr_inactive_file",
1223 "nr_active_file",
1224 "nr_unevictable",
1225 "nr_slab_reclaimable",
1226 "nr_slab_unreclaimable",
1227 "nr_isolated_anon",
1228 "nr_isolated_file",
1229 "workingset_nodes",
1230 "workingset_refault_anon",
1231 "workingset_refault_file",
1232 "workingset_activate_anon",
1233 "workingset_activate_file",
1234 "workingset_restore_anon",
1235 "workingset_restore_file",
1236 "workingset_nodereclaim",
1237 "nr_anon_pages",
1238 "nr_mapped",
1239 "nr_file_pages",
1240 "nr_dirty",
1241 "nr_writeback",
1242 "nr_writeback_temp",
1243 "nr_shmem",
1244 "nr_shmem_hugepages",
1245 "nr_shmem_pmdmapped",
1246 "nr_file_hugepages",
1247 "nr_file_pmdmapped",
1248 "nr_anon_transparent_hugepages",
1249 "nr_vmscan_write",
1250 "nr_vmscan_immediate_reclaim",
1251 "nr_dirtied",
1252 "nr_written",
1253 "nr_throttled_written",
1254 "nr_kernel_misc_reclaimable",
1255 "nr_foll_pin_acquired",
1256 "nr_foll_pin_released",
1257 "nr_kernel_stack",
1258#if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1259 "nr_shadow_call_stack",
1260#endif
1261 "nr_page_table_pages",
1262 "nr_sec_page_table_pages",
1263#ifdef CONFIG_IOMMU_SUPPORT
1264 "nr_iommu_pages",
1265#endif
1266#ifdef CONFIG_SWAP
1267 "nr_swapcached",
1268#endif
1269#ifdef CONFIG_NUMA_BALANCING
1270 "pgpromote_success",
1271 "pgpromote_candidate",
1272#endif
1273 "pgdemote_kswapd",
1274 "pgdemote_direct",
1275 "pgdemote_khugepaged",
1276#ifdef CONFIG_HUGETLB_PAGE
1277 "nr_hugetlb",
1278#endif
1279 /* system-wide enum vm_stat_item counters */
1280 "nr_dirty_threshold",
1281 "nr_dirty_background_threshold",
1282 "nr_memmap_pages",
1283 "nr_memmap_boot_pages",
1284
1285#if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1286 /* enum vm_event_item counters */
1287 "pgpgin",
1288 "pgpgout",
1289 "pswpin",
1290 "pswpout",
1291
1292 TEXTS_FOR_ZONES("pgalloc")
1293 TEXTS_FOR_ZONES("allocstall")
1294 TEXTS_FOR_ZONES("pgskip")
1295
1296 "pgfree",
1297 "pgactivate",
1298 "pgdeactivate",
1299 "pglazyfree",
1300
1301 "pgfault",
1302 "pgmajfault",
1303 "pglazyfreed",
1304
1305 "pgrefill",
1306 "pgreuse",
1307 "pgsteal_kswapd",
1308 "pgsteal_direct",
1309 "pgsteal_khugepaged",
1310 "pgscan_kswapd",
1311 "pgscan_direct",
1312 "pgscan_khugepaged",
1313 "pgscan_direct_throttle",
1314 "pgscan_anon",
1315 "pgscan_file",
1316 "pgsteal_anon",
1317 "pgsteal_file",
1318
1319#ifdef CONFIG_NUMA
1320 "zone_reclaim_success",
1321 "zone_reclaim_failed",
1322#endif
1323 "pginodesteal",
1324 "slabs_scanned",
1325 "kswapd_inodesteal",
1326 "kswapd_low_wmark_hit_quickly",
1327 "kswapd_high_wmark_hit_quickly",
1328 "pageoutrun",
1329
1330 "pgrotated",
1331
1332 "drop_pagecache",
1333 "drop_slab",
1334 "oom_kill",
1335
1336#ifdef CONFIG_NUMA_BALANCING
1337 "numa_pte_updates",
1338 "numa_huge_pte_updates",
1339 "numa_hint_faults",
1340 "numa_hint_faults_local",
1341 "numa_pages_migrated",
1342#endif
1343#ifdef CONFIG_MIGRATION
1344 "pgmigrate_success",
1345 "pgmigrate_fail",
1346 "thp_migration_success",
1347 "thp_migration_fail",
1348 "thp_migration_split",
1349#endif
1350#ifdef CONFIG_COMPACTION
1351 "compact_migrate_scanned",
1352 "compact_free_scanned",
1353 "compact_isolated",
1354 "compact_stall",
1355 "compact_fail",
1356 "compact_success",
1357 "compact_daemon_wake",
1358 "compact_daemon_migrate_scanned",
1359 "compact_daemon_free_scanned",
1360#endif
1361
1362#ifdef CONFIG_HUGETLB_PAGE
1363 "htlb_buddy_alloc_success",
1364 "htlb_buddy_alloc_fail",
1365#endif
1366#ifdef CONFIG_CMA
1367 "cma_alloc_success",
1368 "cma_alloc_fail",
1369#endif
1370 "unevictable_pgs_culled",
1371 "unevictable_pgs_scanned",
1372 "unevictable_pgs_rescued",
1373 "unevictable_pgs_mlocked",
1374 "unevictable_pgs_munlocked",
1375 "unevictable_pgs_cleared",
1376 "unevictable_pgs_stranded",
1377
1378#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1379 "thp_fault_alloc",
1380 "thp_fault_fallback",
1381 "thp_fault_fallback_charge",
1382 "thp_collapse_alloc",
1383 "thp_collapse_alloc_failed",
1384 "thp_file_alloc",
1385 "thp_file_fallback",
1386 "thp_file_fallback_charge",
1387 "thp_file_mapped",
1388 "thp_split_page",
1389 "thp_split_page_failed",
1390 "thp_deferred_split_page",
1391 "thp_underused_split_page",
1392 "thp_split_pmd",
1393 "thp_scan_exceed_none_pte",
1394 "thp_scan_exceed_swap_pte",
1395 "thp_scan_exceed_share_pte",
1396#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1397 "thp_split_pud",
1398#endif
1399 "thp_zero_page_alloc",
1400 "thp_zero_page_alloc_failed",
1401 "thp_swpout",
1402 "thp_swpout_fallback",
1403#endif
1404#ifdef CONFIG_MEMORY_BALLOON
1405 "balloon_inflate",
1406 "balloon_deflate",
1407#ifdef CONFIG_BALLOON_COMPACTION
1408 "balloon_migrate",
1409#endif
1410#endif /* CONFIG_MEMORY_BALLOON */
1411#ifdef CONFIG_DEBUG_TLBFLUSH
1412 "nr_tlb_remote_flush",
1413 "nr_tlb_remote_flush_received",
1414 "nr_tlb_local_flush_all",
1415 "nr_tlb_local_flush_one",
1416#endif /* CONFIG_DEBUG_TLBFLUSH */
1417
1418#ifdef CONFIG_SWAP
1419 "swap_ra",
1420 "swap_ra_hit",
1421 "swpin_zero",
1422 "swpout_zero",
1423#ifdef CONFIG_KSM
1424 "ksm_swpin_copy",
1425#endif
1426#endif
1427#ifdef CONFIG_KSM
1428 "cow_ksm",
1429#endif
1430#ifdef CONFIG_ZSWAP
1431 "zswpin",
1432 "zswpout",
1433 "zswpwb",
1434#endif
1435#ifdef CONFIG_X86
1436 "direct_map_level2_splits",
1437 "direct_map_level3_splits",
1438#endif
1439#ifdef CONFIG_PER_VMA_LOCK_STATS
1440 "vma_lock_success",
1441 "vma_lock_abort",
1442 "vma_lock_retry",
1443 "vma_lock_miss",
1444#endif
1445#ifdef CONFIG_DEBUG_STACK_USAGE
1446 "kstack_1k",
1447#if THREAD_SIZE > 1024
1448 "kstack_2k",
1449#endif
1450#if THREAD_SIZE > 2048
1451 "kstack_4k",
1452#endif
1453#if THREAD_SIZE > 4096
1454 "kstack_8k",
1455#endif
1456#if THREAD_SIZE > 8192
1457 "kstack_16k",
1458#endif
1459#if THREAD_SIZE > 16384
1460 "kstack_32k",
1461#endif
1462#if THREAD_SIZE > 32768
1463 "kstack_64k",
1464#endif
1465#if THREAD_SIZE > 65536
1466 "kstack_rest",
1467#endif
1468#endif
1469#endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1470};
1471#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1472
1473#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1474 defined(CONFIG_PROC_FS)
1475static void *frag_start(struct seq_file *m, loff_t *pos)
1476{
1477 pg_data_t *pgdat;
1478 loff_t node = *pos;
1479
1480 for (pgdat = first_online_pgdat();
1481 pgdat && node;
1482 pgdat = next_online_pgdat(pgdat))
1483 --node;
1484
1485 return pgdat;
1486}
1487
1488static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1489{
1490 pg_data_t *pgdat = (pg_data_t *)arg;
1491
1492 (*pos)++;
1493 return next_online_pgdat(pgdat);
1494}
1495
1496static void frag_stop(struct seq_file *m, void *arg)
1497{
1498}
1499
1500/*
1501 * Walk zones in a node and print using a callback.
1502 * If @assert_populated is true, only use callback for zones that are populated.
1503 */
1504static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1505 bool assert_populated, bool nolock,
1506 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1507{
1508 struct zone *zone;
1509 struct zone *node_zones = pgdat->node_zones;
1510 unsigned long flags;
1511
1512 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1513 if (assert_populated && !populated_zone(zone))
1514 continue;
1515
1516 if (!nolock)
1517 spin_lock_irqsave(&zone->lock, flags);
1518 print(m, pgdat, zone);
1519 if (!nolock)
1520 spin_unlock_irqrestore(&zone->lock, flags);
1521 }
1522}
1523#endif
1524
1525#ifdef CONFIG_PROC_FS
1526static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1527 struct zone *zone)
1528{
1529 int order;
1530
1531 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1532 for (order = 0; order < NR_PAGE_ORDERS; ++order)
1533 /*
1534 * Access to nr_free is lockless as nr_free is used only for
1535 * printing purposes. Use data_race to avoid KCSAN warning.
1536 */
1537 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1538 seq_putc(m, '\n');
1539}
1540
1541/*
1542 * This walks the free areas for each zone.
1543 */
1544static int frag_show(struct seq_file *m, void *arg)
1545{
1546 pg_data_t *pgdat = (pg_data_t *)arg;
1547 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1548 return 0;
1549}
1550
1551static void pagetypeinfo_showfree_print(struct seq_file *m,
1552 pg_data_t *pgdat, struct zone *zone)
1553{
1554 int order, mtype;
1555
1556 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1557 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1558 pgdat->node_id,
1559 zone->name,
1560 migratetype_names[mtype]);
1561 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1562 unsigned long freecount = 0;
1563 struct free_area *area;
1564 struct list_head *curr;
1565 bool overflow = false;
1566
1567 area = &(zone->free_area[order]);
1568
1569 list_for_each(curr, &area->free_list[mtype]) {
1570 /*
1571 * Cap the free_list iteration because it might
1572 * be really large and we are under a spinlock
1573 * so a long time spent here could trigger a
1574 * hard lockup detector. Anyway this is a
1575 * debugging tool so knowing there is a handful
1576 * of pages of this order should be more than
1577 * sufficient.
1578 */
1579 if (++freecount >= 100000) {
1580 overflow = true;
1581 break;
1582 }
1583 }
1584 seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1585 spin_unlock_irq(&zone->lock);
1586 cond_resched();
1587 spin_lock_irq(&zone->lock);
1588 }
1589 seq_putc(m, '\n');
1590 }
1591}
1592
1593/* Print out the free pages at each order for each migatetype */
1594static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1595{
1596 int order;
1597 pg_data_t *pgdat = (pg_data_t *)arg;
1598
1599 /* Print header */
1600 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1601 for (order = 0; order < NR_PAGE_ORDERS; ++order)
1602 seq_printf(m, "%6d ", order);
1603 seq_putc(m, '\n');
1604
1605 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1606}
1607
1608static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1609 pg_data_t *pgdat, struct zone *zone)
1610{
1611 int mtype;
1612 unsigned long pfn;
1613 unsigned long start_pfn = zone->zone_start_pfn;
1614 unsigned long end_pfn = zone_end_pfn(zone);
1615 unsigned long count[MIGRATE_TYPES] = { 0, };
1616
1617 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1618 struct page *page;
1619
1620 page = pfn_to_online_page(pfn);
1621 if (!page)
1622 continue;
1623
1624 if (page_zone(page) != zone)
1625 continue;
1626
1627 mtype = get_pageblock_migratetype(page);
1628
1629 if (mtype < MIGRATE_TYPES)
1630 count[mtype]++;
1631 }
1632
1633 /* Print counts */
1634 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1635 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1636 seq_printf(m, "%12lu ", count[mtype]);
1637 seq_putc(m, '\n');
1638}
1639
1640/* Print out the number of pageblocks for each migratetype */
1641static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1642{
1643 int mtype;
1644 pg_data_t *pgdat = (pg_data_t *)arg;
1645
1646 seq_printf(m, "\n%-23s", "Number of blocks type ");
1647 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1648 seq_printf(m, "%12s ", migratetype_names[mtype]);
1649 seq_putc(m, '\n');
1650 walk_zones_in_node(m, pgdat, true, false,
1651 pagetypeinfo_showblockcount_print);
1652}
1653
1654/*
1655 * Print out the number of pageblocks for each migratetype that contain pages
1656 * of other types. This gives an indication of how well fallbacks are being
1657 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1658 * to determine what is going on
1659 */
1660static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1661{
1662#ifdef CONFIG_PAGE_OWNER
1663 int mtype;
1664
1665 if (!static_branch_unlikely(&page_owner_inited))
1666 return;
1667
1668 drain_all_pages(NULL);
1669
1670 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1671 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1672 seq_printf(m, "%12s ", migratetype_names[mtype]);
1673 seq_putc(m, '\n');
1674
1675 walk_zones_in_node(m, pgdat, true, true,
1676 pagetypeinfo_showmixedcount_print);
1677#endif /* CONFIG_PAGE_OWNER */
1678}
1679
1680/*
1681 * This prints out statistics in relation to grouping pages by mobility.
1682 * It is expensive to collect so do not constantly read the file.
1683 */
1684static int pagetypeinfo_show(struct seq_file *m, void *arg)
1685{
1686 pg_data_t *pgdat = (pg_data_t *)arg;
1687
1688 /* check memoryless node */
1689 if (!node_state(pgdat->node_id, N_MEMORY))
1690 return 0;
1691
1692 seq_printf(m, "Page block order: %d\n", pageblock_order);
1693 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1694 seq_putc(m, '\n');
1695 pagetypeinfo_showfree(m, pgdat);
1696 pagetypeinfo_showblockcount(m, pgdat);
1697 pagetypeinfo_showmixedcount(m, pgdat);
1698
1699 return 0;
1700}
1701
1702static const struct seq_operations fragmentation_op = {
1703 .start = frag_start,
1704 .next = frag_next,
1705 .stop = frag_stop,
1706 .show = frag_show,
1707};
1708
1709static const struct seq_operations pagetypeinfo_op = {
1710 .start = frag_start,
1711 .next = frag_next,
1712 .stop = frag_stop,
1713 .show = pagetypeinfo_show,
1714};
1715
1716static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1717{
1718 int zid;
1719
1720 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1721 struct zone *compare = &pgdat->node_zones[zid];
1722
1723 if (populated_zone(compare))
1724 return zone == compare;
1725 }
1726
1727 return false;
1728}
1729
1730static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1731 struct zone *zone)
1732{
1733 int i;
1734 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1735 if (is_zone_first_populated(pgdat, zone)) {
1736 seq_printf(m, "\n per-node stats");
1737 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1738 unsigned long pages = node_page_state_pages(pgdat, i);
1739
1740 if (vmstat_item_print_in_thp(i))
1741 pages /= HPAGE_PMD_NR;
1742 seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1743 pages);
1744 }
1745 }
1746 seq_printf(m,
1747 "\n pages free %lu"
1748 "\n boost %lu"
1749 "\n min %lu"
1750 "\n low %lu"
1751 "\n high %lu"
1752 "\n promo %lu"
1753 "\n spanned %lu"
1754 "\n present %lu"
1755 "\n managed %lu"
1756 "\n cma %lu",
1757 zone_page_state(zone, NR_FREE_PAGES),
1758 zone->watermark_boost,
1759 min_wmark_pages(zone),
1760 low_wmark_pages(zone),
1761 high_wmark_pages(zone),
1762 promo_wmark_pages(zone),
1763 zone->spanned_pages,
1764 zone->present_pages,
1765 zone_managed_pages(zone),
1766 zone_cma_pages(zone));
1767
1768 seq_printf(m,
1769 "\n protection: (%ld",
1770 zone->lowmem_reserve[0]);
1771 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1772 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1773 seq_putc(m, ')');
1774
1775 /* If unpopulated, no other information is useful */
1776 if (!populated_zone(zone)) {
1777 seq_putc(m, '\n');
1778 return;
1779 }
1780
1781 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1782 seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1783 zone_page_state(zone, i));
1784
1785#ifdef CONFIG_NUMA
1786 fold_vm_zone_numa_events(zone);
1787 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1788 seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1789 zone_numa_event_state(zone, i));
1790#endif
1791
1792 seq_printf(m, "\n pagesets");
1793 for_each_online_cpu(i) {
1794 struct per_cpu_pages *pcp;
1795 struct per_cpu_zonestat __maybe_unused *pzstats;
1796
1797 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1798 seq_printf(m,
1799 "\n cpu: %i"
1800 "\n count: %i"
1801 "\n high: %i"
1802 "\n batch: %i"
1803 "\n high_min: %i"
1804 "\n high_max: %i",
1805 i,
1806 pcp->count,
1807 pcp->high,
1808 pcp->batch,
1809 pcp->high_min,
1810 pcp->high_max);
1811#ifdef CONFIG_SMP
1812 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1813 seq_printf(m, "\n vm stats threshold: %d",
1814 pzstats->stat_threshold);
1815#endif
1816 }
1817 seq_printf(m,
1818 "\n node_unreclaimable: %u"
1819 "\n start_pfn: %lu",
1820 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1821 zone->zone_start_pfn);
1822 seq_putc(m, '\n');
1823}
1824
1825/*
1826 * Output information about zones in @pgdat. All zones are printed regardless
1827 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1828 * set of all zones and userspace would not be aware of such zones if they are
1829 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1830 */
1831static int zoneinfo_show(struct seq_file *m, void *arg)
1832{
1833 pg_data_t *pgdat = (pg_data_t *)arg;
1834 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1835 return 0;
1836}
1837
1838static const struct seq_operations zoneinfo_op = {
1839 .start = frag_start, /* iterate over all zones. The same as in
1840 * fragmentation. */
1841 .next = frag_next,
1842 .stop = frag_stop,
1843 .show = zoneinfo_show,
1844};
1845
1846#define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1847 NR_VM_NUMA_EVENT_ITEMS + \
1848 NR_VM_NODE_STAT_ITEMS + \
1849 NR_VM_STAT_ITEMS + \
1850 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1851 NR_VM_EVENT_ITEMS : 0))
1852
1853static void *vmstat_start(struct seq_file *m, loff_t *pos)
1854{
1855 unsigned long *v;
1856 int i;
1857
1858 if (*pos >= NR_VMSTAT_ITEMS)
1859 return NULL;
1860
1861 BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1862 fold_vm_numa_events();
1863 v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1864 m->private = v;
1865 if (!v)
1866 return ERR_PTR(-ENOMEM);
1867 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1868 v[i] = global_zone_page_state(i);
1869 v += NR_VM_ZONE_STAT_ITEMS;
1870
1871#ifdef CONFIG_NUMA
1872 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1873 v[i] = global_numa_event_state(i);
1874 v += NR_VM_NUMA_EVENT_ITEMS;
1875#endif
1876
1877 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1878 v[i] = global_node_page_state_pages(i);
1879 if (vmstat_item_print_in_thp(i))
1880 v[i] /= HPAGE_PMD_NR;
1881 }
1882 v += NR_VM_NODE_STAT_ITEMS;
1883
1884 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1885 v + NR_DIRTY_THRESHOLD);
1886 v[NR_MEMMAP_PAGES] = atomic_long_read(&nr_memmap_pages);
1887 v[NR_MEMMAP_BOOT_PAGES] = atomic_long_read(&nr_memmap_boot_pages);
1888 v += NR_VM_STAT_ITEMS;
1889
1890#ifdef CONFIG_VM_EVENT_COUNTERS
1891 all_vm_events(v);
1892 v[PGPGIN] /= 2; /* sectors -> kbytes */
1893 v[PGPGOUT] /= 2;
1894#endif
1895 return (unsigned long *)m->private + *pos;
1896}
1897
1898static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1899{
1900 (*pos)++;
1901 if (*pos >= NR_VMSTAT_ITEMS)
1902 return NULL;
1903 return (unsigned long *)m->private + *pos;
1904}
1905
1906static int vmstat_show(struct seq_file *m, void *arg)
1907{
1908 unsigned long *l = arg;
1909 unsigned long off = l - (unsigned long *)m->private;
1910
1911 seq_puts(m, vmstat_text[off]);
1912 seq_put_decimal_ull(m, " ", *l);
1913 seq_putc(m, '\n');
1914
1915 if (off == NR_VMSTAT_ITEMS - 1) {
1916 /*
1917 * We've come to the end - add any deprecated counters to avoid
1918 * breaking userspace which might depend on them being present.
1919 */
1920 seq_puts(m, "nr_unstable 0\n");
1921 }
1922 return 0;
1923}
1924
1925static void vmstat_stop(struct seq_file *m, void *arg)
1926{
1927 kfree(m->private);
1928 m->private = NULL;
1929}
1930
1931static const struct seq_operations vmstat_op = {
1932 .start = vmstat_start,
1933 .next = vmstat_next,
1934 .stop = vmstat_stop,
1935 .show = vmstat_show,
1936};
1937#endif /* CONFIG_PROC_FS */
1938
1939#ifdef CONFIG_SMP
1940static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1941int sysctl_stat_interval __read_mostly = HZ;
1942static int vmstat_late_init_done;
1943
1944#ifdef CONFIG_PROC_FS
1945static void refresh_vm_stats(struct work_struct *work)
1946{
1947 refresh_cpu_vm_stats(true);
1948}
1949
1950int vmstat_refresh(const struct ctl_table *table, int write,
1951 void *buffer, size_t *lenp, loff_t *ppos)
1952{
1953 long val;
1954 int err;
1955 int i;
1956
1957 /*
1958 * The regular update, every sysctl_stat_interval, may come later
1959 * than expected: leaving a significant amount in per_cpu buckets.
1960 * This is particularly misleading when checking a quantity of HUGE
1961 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1962 * which can equally be echo'ed to or cat'ted from (by root),
1963 * can be used to update the stats just before reading them.
1964 *
1965 * Oh, and since global_zone_page_state() etc. are so careful to hide
1966 * transiently negative values, report an error here if any of
1967 * the stats is negative, so we know to go looking for imbalance.
1968 */
1969 err = schedule_on_each_cpu(refresh_vm_stats);
1970 if (err)
1971 return err;
1972 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1973 /*
1974 * Skip checking stats known to go negative occasionally.
1975 */
1976 switch (i) {
1977 case NR_ZONE_WRITE_PENDING:
1978 case NR_FREE_CMA_PAGES:
1979 continue;
1980 }
1981 val = atomic_long_read(&vm_zone_stat[i]);
1982 if (val < 0) {
1983 pr_warn("%s: %s %ld\n",
1984 __func__, zone_stat_name(i), val);
1985 }
1986 }
1987 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1988 /*
1989 * Skip checking stats known to go negative occasionally.
1990 */
1991 switch (i) {
1992 case NR_WRITEBACK:
1993 continue;
1994 }
1995 val = atomic_long_read(&vm_node_stat[i]);
1996 if (val < 0) {
1997 pr_warn("%s: %s %ld\n",
1998 __func__, node_stat_name(i), val);
1999 }
2000 }
2001 if (write)
2002 *ppos += *lenp;
2003 else
2004 *lenp = 0;
2005 return 0;
2006}
2007#endif /* CONFIG_PROC_FS */
2008
2009static void vmstat_update(struct work_struct *w)
2010{
2011 if (refresh_cpu_vm_stats(true)) {
2012 /*
2013 * Counters were updated so we expect more updates
2014 * to occur in the future. Keep on running the
2015 * update worker thread.
2016 */
2017 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
2018 this_cpu_ptr(&vmstat_work),
2019 round_jiffies_relative(sysctl_stat_interval));
2020 }
2021}
2022
2023/*
2024 * Check if the diffs for a certain cpu indicate that
2025 * an update is needed.
2026 */
2027static bool need_update(int cpu)
2028{
2029 pg_data_t *last_pgdat = NULL;
2030 struct zone *zone;
2031
2032 for_each_populated_zone(zone) {
2033 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
2034 struct per_cpu_nodestat *n;
2035
2036 /*
2037 * The fast way of checking if there are any vmstat diffs.
2038 */
2039 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
2040 return true;
2041
2042 if (last_pgdat == zone->zone_pgdat)
2043 continue;
2044 last_pgdat = zone->zone_pgdat;
2045 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
2046 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
2047 return true;
2048 }
2049 return false;
2050}
2051
2052/*
2053 * Switch off vmstat processing and then fold all the remaining differentials
2054 * until the diffs stay at zero. The function is used by NOHZ and can only be
2055 * invoked when tick processing is not active.
2056 */
2057void quiet_vmstat(void)
2058{
2059 if (system_state != SYSTEM_RUNNING)
2060 return;
2061
2062 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
2063 return;
2064
2065 if (!need_update(smp_processor_id()))
2066 return;
2067
2068 /*
2069 * Just refresh counters and do not care about the pending delayed
2070 * vmstat_update. It doesn't fire that often to matter and canceling
2071 * it would be too expensive from this path.
2072 * vmstat_shepherd will take care about that for us.
2073 */
2074 refresh_cpu_vm_stats(false);
2075}
2076
2077/*
2078 * Shepherd worker thread that checks the
2079 * differentials of processors that have their worker
2080 * threads for vm statistics updates disabled because of
2081 * inactivity.
2082 */
2083static void vmstat_shepherd(struct work_struct *w);
2084
2085static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2086
2087static void vmstat_shepherd(struct work_struct *w)
2088{
2089 int cpu;
2090
2091 cpus_read_lock();
2092 /* Check processors whose vmstat worker threads have been disabled */
2093 for_each_online_cpu(cpu) {
2094 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2095
2096 /*
2097 * In kernel users of vmstat counters either require the precise value and
2098 * they are using zone_page_state_snapshot interface or they can live with
2099 * an imprecision as the regular flushing can happen at arbitrary time and
2100 * cumulative error can grow (see calculate_normal_threshold).
2101 *
2102 * From that POV the regular flushing can be postponed for CPUs that have
2103 * been isolated from the kernel interference without critical
2104 * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2105 * for all isolated CPUs to avoid interference with the isolated workload.
2106 */
2107 if (cpu_is_isolated(cpu))
2108 continue;
2109
2110 if (!delayed_work_pending(dw) && need_update(cpu))
2111 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2112
2113 cond_resched();
2114 }
2115 cpus_read_unlock();
2116
2117 schedule_delayed_work(&shepherd,
2118 round_jiffies_relative(sysctl_stat_interval));
2119}
2120
2121static void __init start_shepherd_timer(void)
2122{
2123 int cpu;
2124
2125 for_each_possible_cpu(cpu) {
2126 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2127 vmstat_update);
2128
2129 /*
2130 * For secondary CPUs during CPU hotplug scenarios,
2131 * vmstat_cpu_online() will enable the work.
2132 * mm/vmstat:online enables and disables vmstat_work
2133 * symmetrically during CPU hotplug events.
2134 */
2135 if (!cpu_online(cpu))
2136 disable_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2137 }
2138
2139 schedule_delayed_work(&shepherd,
2140 round_jiffies_relative(sysctl_stat_interval));
2141}
2142
2143static void __init init_cpu_node_state(void)
2144{
2145 int node;
2146
2147 for_each_online_node(node) {
2148 if (!cpumask_empty(cpumask_of_node(node)))
2149 node_set_state(node, N_CPU);
2150 }
2151}
2152
2153static int vmstat_cpu_online(unsigned int cpu)
2154{
2155 if (vmstat_late_init_done)
2156 refresh_zone_stat_thresholds();
2157
2158 if (!node_state(cpu_to_node(cpu), N_CPU)) {
2159 node_set_state(cpu_to_node(cpu), N_CPU);
2160 }
2161 enable_delayed_work(&per_cpu(vmstat_work, cpu));
2162
2163 return 0;
2164}
2165
2166static int vmstat_cpu_down_prep(unsigned int cpu)
2167{
2168 disable_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2169 return 0;
2170}
2171
2172static int vmstat_cpu_dead(unsigned int cpu)
2173{
2174 const struct cpumask *node_cpus;
2175 int node;
2176
2177 node = cpu_to_node(cpu);
2178
2179 refresh_zone_stat_thresholds();
2180 node_cpus = cpumask_of_node(node);
2181 if (!cpumask_empty(node_cpus))
2182 return 0;
2183
2184 node_clear_state(node, N_CPU);
2185
2186 return 0;
2187}
2188
2189static int __init vmstat_late_init(void)
2190{
2191 refresh_zone_stat_thresholds();
2192 vmstat_late_init_done = 1;
2193
2194 return 0;
2195}
2196late_initcall(vmstat_late_init);
2197#endif
2198
2199struct workqueue_struct *mm_percpu_wq;
2200
2201void __init init_mm_internals(void)
2202{
2203 int ret __maybe_unused;
2204
2205 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2206
2207#ifdef CONFIG_SMP
2208 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2209 NULL, vmstat_cpu_dead);
2210 if (ret < 0)
2211 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2212
2213 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2214 vmstat_cpu_online,
2215 vmstat_cpu_down_prep);
2216 if (ret < 0)
2217 pr_err("vmstat: failed to register 'online' hotplug state\n");
2218
2219 cpus_read_lock();
2220 init_cpu_node_state();
2221 cpus_read_unlock();
2222
2223 start_shepherd_timer();
2224#endif
2225#ifdef CONFIG_PROC_FS
2226 proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2227 proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2228 proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2229 proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2230#endif
2231}
2232
2233#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2234
2235/*
2236 * Return an index indicating how much of the available free memory is
2237 * unusable for an allocation of the requested size.
2238 */
2239static int unusable_free_index(unsigned int order,
2240 struct contig_page_info *info)
2241{
2242 /* No free memory is interpreted as all free memory is unusable */
2243 if (info->free_pages == 0)
2244 return 1000;
2245
2246 /*
2247 * Index should be a value between 0 and 1. Return a value to 3
2248 * decimal places.
2249 *
2250 * 0 => no fragmentation
2251 * 1 => high fragmentation
2252 */
2253 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2254
2255}
2256
2257static void unusable_show_print(struct seq_file *m,
2258 pg_data_t *pgdat, struct zone *zone)
2259{
2260 unsigned int order;
2261 int index;
2262 struct contig_page_info info;
2263
2264 seq_printf(m, "Node %d, zone %8s ",
2265 pgdat->node_id,
2266 zone->name);
2267 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2268 fill_contig_page_info(zone, order, &info);
2269 index = unusable_free_index(order, &info);
2270 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2271 }
2272
2273 seq_putc(m, '\n');
2274}
2275
2276/*
2277 * Display unusable free space index
2278 *
2279 * The unusable free space index measures how much of the available free
2280 * memory cannot be used to satisfy an allocation of a given size and is a
2281 * value between 0 and 1. The higher the value, the more of free memory is
2282 * unusable and by implication, the worse the external fragmentation is. This
2283 * can be expressed as a percentage by multiplying by 100.
2284 */
2285static int unusable_show(struct seq_file *m, void *arg)
2286{
2287 pg_data_t *pgdat = (pg_data_t *)arg;
2288
2289 /* check memoryless node */
2290 if (!node_state(pgdat->node_id, N_MEMORY))
2291 return 0;
2292
2293 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2294
2295 return 0;
2296}
2297
2298static const struct seq_operations unusable_sops = {
2299 .start = frag_start,
2300 .next = frag_next,
2301 .stop = frag_stop,
2302 .show = unusable_show,
2303};
2304
2305DEFINE_SEQ_ATTRIBUTE(unusable);
2306
2307static void extfrag_show_print(struct seq_file *m,
2308 pg_data_t *pgdat, struct zone *zone)
2309{
2310 unsigned int order;
2311 int index;
2312
2313 /* Alloc on stack as interrupts are disabled for zone walk */
2314 struct contig_page_info info;
2315
2316 seq_printf(m, "Node %d, zone %8s ",
2317 pgdat->node_id,
2318 zone->name);
2319 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2320 fill_contig_page_info(zone, order, &info);
2321 index = __fragmentation_index(order, &info);
2322 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2323 }
2324
2325 seq_putc(m, '\n');
2326}
2327
2328/*
2329 * Display fragmentation index for orders that allocations would fail for
2330 */
2331static int extfrag_show(struct seq_file *m, void *arg)
2332{
2333 pg_data_t *pgdat = (pg_data_t *)arg;
2334
2335 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2336
2337 return 0;
2338}
2339
2340static const struct seq_operations extfrag_sops = {
2341 .start = frag_start,
2342 .next = frag_next,
2343 .stop = frag_stop,
2344 .show = extfrag_show,
2345};
2346
2347DEFINE_SEQ_ATTRIBUTE(extfrag);
2348
2349static int __init extfrag_debug_init(void)
2350{
2351 struct dentry *extfrag_debug_root;
2352
2353 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2354
2355 debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2356 &unusable_fops);
2357
2358 debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2359 &extfrag_fops);
2360
2361 return 0;
2362}
2363
2364module_init(extfrag_debug_init);
2365
2366#endif