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