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_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
90EXPORT_SYMBOL(vm_stat);
91
92#ifdef CONFIG_SMP
93
94int calculate_pressure_threshold(struct zone *zone)
95{
96 int threshold;
97 int watermark_distance;
98
99 /*
100 * As vmstats are not up to date, there is drift between the estimated
101 * and real values. For high thresholds and a high number of CPUs, it
102 * is possible for the min watermark to be breached while the estimated
103 * value looks fine. The pressure threshold is a reduced value such
104 * that even the maximum amount of drift will not accidentally breach
105 * the min watermark
106 */
107 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
108 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
109
110 /*
111 * Maximum threshold is 125
112 */
113 threshold = min(125, threshold);
114
115 return threshold;
116}
117
118int calculate_normal_threshold(struct zone *zone)
119{
120 int threshold;
121 int mem; /* memory in 128 MB units */
122
123 /*
124 * The threshold scales with the number of processors and the amount
125 * of memory per zone. More memory means that we can defer updates for
126 * longer, more processors could lead to more contention.
127 * fls() is used to have a cheap way of logarithmic scaling.
128 *
129 * Some sample thresholds:
130 *
131 * Threshold Processors (fls) Zonesize fls(mem+1)
132 * ------------------------------------------------------------------
133 * 8 1 1 0.9-1 GB 4
134 * 16 2 2 0.9-1 GB 4
135 * 20 2 2 1-2 GB 5
136 * 24 2 2 2-4 GB 6
137 * 28 2 2 4-8 GB 7
138 * 32 2 2 8-16 GB 8
139 * 4 2 2 <128M 1
140 * 30 4 3 2-4 GB 5
141 * 48 4 3 8-16 GB 8
142 * 32 8 4 1-2 GB 4
143 * 32 8 4 0.9-1GB 4
144 * 10 16 5 <128M 1
145 * 40 16 5 900M 4
146 * 70 64 7 2-4 GB 5
147 * 84 64 7 4-8 GB 6
148 * 108 512 9 4-8 GB 6
149 * 125 1024 10 8-16 GB 8
150 * 125 1024 10 16-32 GB 9
151 */
152
153 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
154
155 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
156
157 /*
158 * Maximum threshold is 125
159 */
160 threshold = min(125, threshold);
161
162 return threshold;
163}
164
165/*
166 * Refresh the thresholds for each zone.
167 */
168void refresh_zone_stat_thresholds(void)
169{
170 struct zone *zone;
171 int cpu;
172 int threshold;
173
174 for_each_populated_zone(zone) {
175 unsigned long max_drift, tolerate_drift;
176
177 threshold = calculate_normal_threshold(zone);
178
179 for_each_online_cpu(cpu)
180 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
181 = threshold;
182
183 /*
184 * Only set percpu_drift_mark if there is a danger that
185 * NR_FREE_PAGES reports the low watermark is ok when in fact
186 * the min watermark could be breached by an allocation
187 */
188 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
189 max_drift = num_online_cpus() * threshold;
190 if (max_drift > tolerate_drift)
191 zone->percpu_drift_mark = high_wmark_pages(zone) +
192 max_drift;
193 }
194}
195
196void set_pgdat_percpu_threshold(pg_data_t *pgdat,
197 int (*calculate_pressure)(struct zone *))
198{
199 struct zone *zone;
200 int cpu;
201 int threshold;
202 int i;
203
204 for (i = 0; i < pgdat->nr_zones; i++) {
205 zone = &pgdat->node_zones[i];
206 if (!zone->percpu_drift_mark)
207 continue;
208
209 threshold = (*calculate_pressure)(zone);
210 for_each_online_cpu(cpu)
211 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
212 = threshold;
213 }
214}
215
216/*
217 * For use when we know that interrupts are disabled,
218 * or when we know that preemption is disabled and that
219 * particular counter cannot be updated from interrupt context.
220 */
221void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
222 long delta)
223{
224 struct per_cpu_pageset __percpu *pcp = zone->pageset;
225 s8 __percpu *p = pcp->vm_stat_diff + item;
226 long x;
227 long t;
228
229 x = delta + __this_cpu_read(*p);
230
231 t = __this_cpu_read(pcp->stat_threshold);
232
233 if (unlikely(x > t || x < -t)) {
234 zone_page_state_add(x, zone, item);
235 x = 0;
236 }
237 __this_cpu_write(*p, x);
238}
239EXPORT_SYMBOL(__mod_zone_page_state);
240
241/*
242 * Optimized increment and decrement functions.
243 *
244 * These are only for a single page and therefore can take a struct page *
245 * argument instead of struct zone *. This allows the inclusion of the code
246 * generated for page_zone(page) into the optimized functions.
247 *
248 * No overflow check is necessary and therefore the differential can be
249 * incremented or decremented in place which may allow the compilers to
250 * generate better code.
251 * The increment or decrement is known and therefore one boundary check can
252 * be omitted.
253 *
254 * NOTE: These functions are very performance sensitive. Change only
255 * with care.
256 *
257 * Some processors have inc/dec instructions that are atomic vs an interrupt.
258 * However, the code must first determine the differential location in a zone
259 * based on the processor number and then inc/dec the counter. There is no
260 * guarantee without disabling preemption that the processor will not change
261 * in between and therefore the atomicity vs. interrupt cannot be exploited
262 * in a useful way here.
263 */
264void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
265{
266 struct per_cpu_pageset __percpu *pcp = zone->pageset;
267 s8 __percpu *p = pcp->vm_stat_diff + item;
268 s8 v, t;
269
270 v = __this_cpu_inc_return(*p);
271 t = __this_cpu_read(pcp->stat_threshold);
272 if (unlikely(v > t)) {
273 s8 overstep = t >> 1;
274
275 zone_page_state_add(v + overstep, zone, item);
276 __this_cpu_write(*p, -overstep);
277 }
278}
279
280void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
281{
282 __inc_zone_state(page_zone(page), item);
283}
284EXPORT_SYMBOL(__inc_zone_page_state);
285
286void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
287{
288 struct per_cpu_pageset __percpu *pcp = zone->pageset;
289 s8 __percpu *p = pcp->vm_stat_diff + item;
290 s8 v, t;
291
292 v = __this_cpu_dec_return(*p);
293 t = __this_cpu_read(pcp->stat_threshold);
294 if (unlikely(v < - t)) {
295 s8 overstep = t >> 1;
296
297 zone_page_state_add(v - overstep, zone, item);
298 __this_cpu_write(*p, overstep);
299 }
300}
301
302void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
303{
304 __dec_zone_state(page_zone(page), item);
305}
306EXPORT_SYMBOL(__dec_zone_page_state);
307
308#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
309/*
310 * If we have cmpxchg_local support then we do not need to incur the overhead
311 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
312 *
313 * mod_state() modifies the zone counter state through atomic per cpu
314 * operations.
315 *
316 * Overstep mode specifies how overstep should handled:
317 * 0 No overstepping
318 * 1 Overstepping half of threshold
319 * -1 Overstepping minus half of threshold
320*/
321static inline void mod_state(struct zone *zone, enum zone_stat_item item,
322 long delta, int overstep_mode)
323{
324 struct per_cpu_pageset __percpu *pcp = zone->pageset;
325 s8 __percpu *p = pcp->vm_stat_diff + item;
326 long o, n, t, z;
327
328 do {
329 z = 0; /* overflow to zone counters */
330
331 /*
332 * The fetching of the stat_threshold is racy. We may apply
333 * a counter threshold to the wrong the cpu if we get
334 * rescheduled while executing here. However, the next
335 * counter update will apply the threshold again and
336 * therefore bring the counter under the threshold again.
337 *
338 * Most of the time the thresholds are the same anyways
339 * for all cpus in a zone.
340 */
341 t = this_cpu_read(pcp->stat_threshold);
342
343 o = this_cpu_read(*p);
344 n = delta + o;
345
346 if (n > t || n < -t) {
347 int os = overstep_mode * (t >> 1) ;
348
349 /* Overflow must be added to zone counters */
350 z = n + os;
351 n = -os;
352 }
353 } while (this_cpu_cmpxchg(*p, o, n) != o);
354
355 if (z)
356 zone_page_state_add(z, zone, item);
357}
358
359void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
360 long delta)
361{
362 mod_state(zone, item, delta, 0);
363}
364EXPORT_SYMBOL(mod_zone_page_state);
365
366void inc_zone_state(struct zone *zone, enum zone_stat_item item)
367{
368 mod_state(zone, item, 1, 1);
369}
370
371void inc_zone_page_state(struct page *page, enum zone_stat_item item)
372{
373 mod_state(page_zone(page), item, 1, 1);
374}
375EXPORT_SYMBOL(inc_zone_page_state);
376
377void dec_zone_page_state(struct page *page, enum zone_stat_item item)
378{
379 mod_state(page_zone(page), item, -1, -1);
380}
381EXPORT_SYMBOL(dec_zone_page_state);
382#else
383/*
384 * Use interrupt disable to serialize counter updates
385 */
386void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
387 long delta)
388{
389 unsigned long flags;
390
391 local_irq_save(flags);
392 __mod_zone_page_state(zone, item, delta);
393 local_irq_restore(flags);
394}
395EXPORT_SYMBOL(mod_zone_page_state);
396
397void inc_zone_state(struct zone *zone, enum zone_stat_item item)
398{
399 unsigned long flags;
400
401 local_irq_save(flags);
402 __inc_zone_state(zone, item);
403 local_irq_restore(flags);
404}
405
406void inc_zone_page_state(struct page *page, enum zone_stat_item item)
407{
408 unsigned long flags;
409 struct zone *zone;
410
411 zone = page_zone(page);
412 local_irq_save(flags);
413 __inc_zone_state(zone, item);
414 local_irq_restore(flags);
415}
416EXPORT_SYMBOL(inc_zone_page_state);
417
418void dec_zone_page_state(struct page *page, enum zone_stat_item item)
419{
420 unsigned long flags;
421
422 local_irq_save(flags);
423 __dec_zone_page_state(page, item);
424 local_irq_restore(flags);
425}
426EXPORT_SYMBOL(dec_zone_page_state);
427#endif
428
429
430/*
431 * Fold a differential into the global counters.
432 * Returns the number of counters updated.
433 */
434static int fold_diff(int *diff)
435{
436 int i;
437 int changes = 0;
438
439 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
440 if (diff[i]) {
441 atomic_long_add(diff[i], &vm_stat[i]);
442 changes++;
443 }
444 return changes;
445}
446
447/*
448 * Update the zone counters for the current cpu.
449 *
450 * Note that refresh_cpu_vm_stats strives to only access
451 * node local memory. The per cpu pagesets on remote zones are placed
452 * in the memory local to the processor using that pageset. So the
453 * loop over all zones will access a series of cachelines local to
454 * the processor.
455 *
456 * The call to zone_page_state_add updates the cachelines with the
457 * statistics in the remote zone struct as well as the global cachelines
458 * with the global counters. These could cause remote node cache line
459 * bouncing and will have to be only done when necessary.
460 *
461 * The function returns the number of global counters updated.
462 */
463static int refresh_cpu_vm_stats(bool do_pagesets)
464{
465 struct zone *zone;
466 int i;
467 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
468 int changes = 0;
469
470 for_each_populated_zone(zone) {
471 struct per_cpu_pageset __percpu *p = zone->pageset;
472
473 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
474 int v;
475
476 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
477 if (v) {
478
479 atomic_long_add(v, &zone->vm_stat[i]);
480 global_diff[i] += v;
481#ifdef CONFIG_NUMA
482 /* 3 seconds idle till flush */
483 __this_cpu_write(p->expire, 3);
484#endif
485 }
486 }
487#ifdef CONFIG_NUMA
488 if (do_pagesets) {
489 cond_resched();
490 /*
491 * Deal with draining the remote pageset of this
492 * processor
493 *
494 * Check if there are pages remaining in this pageset
495 * if not then there is nothing to expire.
496 */
497 if (!__this_cpu_read(p->expire) ||
498 !__this_cpu_read(p->pcp.count))
499 continue;
500
501 /*
502 * We never drain zones local to this processor.
503 */
504 if (zone_to_nid(zone) == numa_node_id()) {
505 __this_cpu_write(p->expire, 0);
506 continue;
507 }
508
509 if (__this_cpu_dec_return(p->expire))
510 continue;
511
512 if (__this_cpu_read(p->pcp.count)) {
513 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
514 changes++;
515 }
516 }
517#endif
518 }
519 changes += fold_diff(global_diff);
520 return changes;
521}
522
523/*
524 * Fold the data for an offline cpu into the global array.
525 * There cannot be any access by the offline cpu and therefore
526 * synchronization is simplified.
527 */
528void cpu_vm_stats_fold(int cpu)
529{
530 struct zone *zone;
531 int i;
532 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
533
534 for_each_populated_zone(zone) {
535 struct per_cpu_pageset *p;
536
537 p = per_cpu_ptr(zone->pageset, cpu);
538
539 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
540 if (p->vm_stat_diff[i]) {
541 int v;
542
543 v = p->vm_stat_diff[i];
544 p->vm_stat_diff[i] = 0;
545 atomic_long_add(v, &zone->vm_stat[i]);
546 global_diff[i] += v;
547 }
548 }
549
550 fold_diff(global_diff);
551}
552
553/*
554 * this is only called if !populated_zone(zone), which implies no other users of
555 * pset->vm_stat_diff[] exsist.
556 */
557void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
558{
559 int i;
560
561 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
562 if (pset->vm_stat_diff[i]) {
563 int v = pset->vm_stat_diff[i];
564 pset->vm_stat_diff[i] = 0;
565 atomic_long_add(v, &zone->vm_stat[i]);
566 atomic_long_add(v, &vm_stat[i]);
567 }
568}
569#endif
570
571#ifdef CONFIG_NUMA
572/*
573 * zonelist = the list of zones passed to the allocator
574 * z = the zone from which the allocation occurred.
575 *
576 * Must be called with interrupts disabled.
577 *
578 * When __GFP_OTHER_NODE is set assume the node of the preferred
579 * zone is the local node. This is useful for daemons who allocate
580 * memory on behalf of other processes.
581 */
582void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
583{
584 if (z->zone_pgdat == preferred_zone->zone_pgdat) {
585 __inc_zone_state(z, NUMA_HIT);
586 } else {
587 __inc_zone_state(z, NUMA_MISS);
588 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
589 }
590 if (z->node == ((flags & __GFP_OTHER_NODE) ?
591 preferred_zone->node : numa_node_id()))
592 __inc_zone_state(z, NUMA_LOCAL);
593 else
594 __inc_zone_state(z, NUMA_OTHER);
595}
596
597/*
598 * Determine the per node value of a stat item.
599 */
600unsigned long node_page_state(int node, enum zone_stat_item item)
601{
602 struct zone *zones = NODE_DATA(node)->node_zones;
603
604 return
605#ifdef CONFIG_ZONE_DMA
606 zone_page_state(&zones[ZONE_DMA], item) +
607#endif
608#ifdef CONFIG_ZONE_DMA32
609 zone_page_state(&zones[ZONE_DMA32], item) +
610#endif
611#ifdef CONFIG_HIGHMEM
612 zone_page_state(&zones[ZONE_HIGHMEM], item) +
613#endif
614 zone_page_state(&zones[ZONE_NORMAL], item) +
615 zone_page_state(&zones[ZONE_MOVABLE], item);
616}
617
618#endif
619
620#ifdef CONFIG_COMPACTION
621
622struct contig_page_info {
623 unsigned long free_pages;
624 unsigned long free_blocks_total;
625 unsigned long free_blocks_suitable;
626};
627
628/*
629 * Calculate the number of free pages in a zone, how many contiguous
630 * pages are free and how many are large enough to satisfy an allocation of
631 * the target size. Note that this function makes no attempt to estimate
632 * how many suitable free blocks there *might* be if MOVABLE pages were
633 * migrated. Calculating that is possible, but expensive and can be
634 * figured out from userspace
635 */
636static void fill_contig_page_info(struct zone *zone,
637 unsigned int suitable_order,
638 struct contig_page_info *info)
639{
640 unsigned int order;
641
642 info->free_pages = 0;
643 info->free_blocks_total = 0;
644 info->free_blocks_suitable = 0;
645
646 for (order = 0; order < MAX_ORDER; order++) {
647 unsigned long blocks;
648
649 /* Count number of free blocks */
650 blocks = zone->free_area[order].nr_free;
651 info->free_blocks_total += blocks;
652
653 /* Count free base pages */
654 info->free_pages += blocks << order;
655
656 /* Count the suitable free blocks */
657 if (order >= suitable_order)
658 info->free_blocks_suitable += blocks <<
659 (order - suitable_order);
660 }
661}
662
663/*
664 * A fragmentation index only makes sense if an allocation of a requested
665 * size would fail. If that is true, the fragmentation index indicates
666 * whether external fragmentation or a lack of memory was the problem.
667 * The value can be used to determine if page reclaim or compaction
668 * should be used
669 */
670static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
671{
672 unsigned long requested = 1UL << order;
673
674 if (!info->free_blocks_total)
675 return 0;
676
677 /* Fragmentation index only makes sense when a request would fail */
678 if (info->free_blocks_suitable)
679 return -1000;
680
681 /*
682 * Index is between 0 and 1 so return within 3 decimal places
683 *
684 * 0 => allocation would fail due to lack of memory
685 * 1 => allocation would fail due to fragmentation
686 */
687 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
688}
689
690/* Same as __fragmentation index but allocs contig_page_info on stack */
691int fragmentation_index(struct zone *zone, unsigned int order)
692{
693 struct contig_page_info info;
694
695 fill_contig_page_info(zone, order, &info);
696 return __fragmentation_index(order, &info);
697}
698#endif
699
700#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
701#ifdef CONFIG_ZONE_DMA
702#define TEXT_FOR_DMA(xx) xx "_dma",
703#else
704#define TEXT_FOR_DMA(xx)
705#endif
706
707#ifdef CONFIG_ZONE_DMA32
708#define TEXT_FOR_DMA32(xx) xx "_dma32",
709#else
710#define TEXT_FOR_DMA32(xx)
711#endif
712
713#ifdef CONFIG_HIGHMEM
714#define TEXT_FOR_HIGHMEM(xx) xx "_high",
715#else
716#define TEXT_FOR_HIGHMEM(xx)
717#endif
718
719#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
720 TEXT_FOR_HIGHMEM(xx) xx "_movable",
721
722const char * const vmstat_text[] = {
723 /* enum zone_stat_item countes */
724 "nr_free_pages",
725 "nr_alloc_batch",
726 "nr_inactive_anon",
727 "nr_active_anon",
728 "nr_inactive_file",
729 "nr_active_file",
730 "nr_unevictable",
731 "nr_mlock",
732 "nr_anon_pages",
733 "nr_mapped",
734 "nr_file_pages",
735 "nr_dirty",
736 "nr_writeback",
737 "nr_slab_reclaimable",
738 "nr_slab_unreclaimable",
739 "nr_page_table_pages",
740 "nr_kernel_stack",
741 "nr_unstable",
742 "nr_bounce",
743 "nr_vmscan_write",
744 "nr_vmscan_immediate_reclaim",
745 "nr_writeback_temp",
746 "nr_isolated_anon",
747 "nr_isolated_file",
748 "nr_shmem",
749 "nr_dirtied",
750 "nr_written",
751 "nr_pages_scanned",
752
753#ifdef CONFIG_NUMA
754 "numa_hit",
755 "numa_miss",
756 "numa_foreign",
757 "numa_interleave",
758 "numa_local",
759 "numa_other",
760#endif
761 "workingset_refault",
762 "workingset_activate",
763 "workingset_nodereclaim",
764 "nr_anon_transparent_hugepages",
765 "nr_free_cma",
766
767 /* enum writeback_stat_item counters */
768 "nr_dirty_threshold",
769 "nr_dirty_background_threshold",
770
771#ifdef CONFIG_VM_EVENT_COUNTERS
772 /* enum vm_event_item counters */
773 "pgpgin",
774 "pgpgout",
775 "pswpin",
776 "pswpout",
777
778 TEXTS_FOR_ZONES("pgalloc")
779
780 "pgfree",
781 "pgactivate",
782 "pgdeactivate",
783
784 "pgfault",
785 "pgmajfault",
786 "pglazyfreed",
787
788 TEXTS_FOR_ZONES("pgrefill")
789 TEXTS_FOR_ZONES("pgsteal_kswapd")
790 TEXTS_FOR_ZONES("pgsteal_direct")
791 TEXTS_FOR_ZONES("pgscan_kswapd")
792 TEXTS_FOR_ZONES("pgscan_direct")
793 "pgscan_direct_throttle",
794
795#ifdef CONFIG_NUMA
796 "zone_reclaim_failed",
797#endif
798 "pginodesteal",
799 "slabs_scanned",
800 "kswapd_inodesteal",
801 "kswapd_low_wmark_hit_quickly",
802 "kswapd_high_wmark_hit_quickly",
803 "pageoutrun",
804 "allocstall",
805
806 "pgrotated",
807
808 "drop_pagecache",
809 "drop_slab",
810
811#ifdef CONFIG_NUMA_BALANCING
812 "numa_pte_updates",
813 "numa_huge_pte_updates",
814 "numa_hint_faults",
815 "numa_hint_faults_local",
816 "numa_pages_migrated",
817#endif
818#ifdef CONFIG_MIGRATION
819 "pgmigrate_success",
820 "pgmigrate_fail",
821#endif
822#ifdef CONFIG_COMPACTION
823 "compact_migrate_scanned",
824 "compact_free_scanned",
825 "compact_isolated",
826 "compact_stall",
827 "compact_fail",
828 "compact_success",
829 "compact_daemon_wake",
830#endif
831
832#ifdef CONFIG_HUGETLB_PAGE
833 "htlb_buddy_alloc_success",
834 "htlb_buddy_alloc_fail",
835#endif
836 "unevictable_pgs_culled",
837 "unevictable_pgs_scanned",
838 "unevictable_pgs_rescued",
839 "unevictable_pgs_mlocked",
840 "unevictable_pgs_munlocked",
841 "unevictable_pgs_cleared",
842 "unevictable_pgs_stranded",
843
844#ifdef CONFIG_TRANSPARENT_HUGEPAGE
845 "thp_fault_alloc",
846 "thp_fault_fallback",
847 "thp_collapse_alloc",
848 "thp_collapse_alloc_failed",
849 "thp_split_page",
850 "thp_split_page_failed",
851 "thp_deferred_split_page",
852 "thp_split_pmd",
853 "thp_zero_page_alloc",
854 "thp_zero_page_alloc_failed",
855#endif
856#ifdef CONFIG_MEMORY_BALLOON
857 "balloon_inflate",
858 "balloon_deflate",
859#ifdef CONFIG_BALLOON_COMPACTION
860 "balloon_migrate",
861#endif
862#endif /* CONFIG_MEMORY_BALLOON */
863#ifdef CONFIG_DEBUG_TLBFLUSH
864#ifdef CONFIG_SMP
865 "nr_tlb_remote_flush",
866 "nr_tlb_remote_flush_received",
867#endif /* CONFIG_SMP */
868 "nr_tlb_local_flush_all",
869 "nr_tlb_local_flush_one",
870#endif /* CONFIG_DEBUG_TLBFLUSH */
871
872#ifdef CONFIG_DEBUG_VM_VMACACHE
873 "vmacache_find_calls",
874 "vmacache_find_hits",
875 "vmacache_full_flushes",
876#endif
877#endif /* CONFIG_VM_EVENTS_COUNTERS */
878};
879#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
880
881
882#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
883 defined(CONFIG_PROC_FS)
884static void *frag_start(struct seq_file *m, loff_t *pos)
885{
886 pg_data_t *pgdat;
887 loff_t node = *pos;
888
889 for (pgdat = first_online_pgdat();
890 pgdat && node;
891 pgdat = next_online_pgdat(pgdat))
892 --node;
893
894 return pgdat;
895}
896
897static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
898{
899 pg_data_t *pgdat = (pg_data_t *)arg;
900
901 (*pos)++;
902 return next_online_pgdat(pgdat);
903}
904
905static void frag_stop(struct seq_file *m, void *arg)
906{
907}
908
909/* Walk all the zones in a node and print using a callback */
910static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
911 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
912{
913 struct zone *zone;
914 struct zone *node_zones = pgdat->node_zones;
915 unsigned long flags;
916
917 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
918 if (!populated_zone(zone))
919 continue;
920
921 spin_lock_irqsave(&zone->lock, flags);
922 print(m, pgdat, zone);
923 spin_unlock_irqrestore(&zone->lock, flags);
924 }
925}
926#endif
927
928#ifdef CONFIG_PROC_FS
929static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
930 struct zone *zone)
931{
932 int order;
933
934 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
935 for (order = 0; order < MAX_ORDER; ++order)
936 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
937 seq_putc(m, '\n');
938}
939
940/*
941 * This walks the free areas for each zone.
942 */
943static int frag_show(struct seq_file *m, void *arg)
944{
945 pg_data_t *pgdat = (pg_data_t *)arg;
946 walk_zones_in_node(m, pgdat, frag_show_print);
947 return 0;
948}
949
950static void pagetypeinfo_showfree_print(struct seq_file *m,
951 pg_data_t *pgdat, struct zone *zone)
952{
953 int order, mtype;
954
955 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
956 seq_printf(m, "Node %4d, zone %8s, type %12s ",
957 pgdat->node_id,
958 zone->name,
959 migratetype_names[mtype]);
960 for (order = 0; order < MAX_ORDER; ++order) {
961 unsigned long freecount = 0;
962 struct free_area *area;
963 struct list_head *curr;
964
965 area = &(zone->free_area[order]);
966
967 list_for_each(curr, &area->free_list[mtype])
968 freecount++;
969 seq_printf(m, "%6lu ", freecount);
970 }
971 seq_putc(m, '\n');
972 }
973}
974
975/* Print out the free pages at each order for each migatetype */
976static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
977{
978 int order;
979 pg_data_t *pgdat = (pg_data_t *)arg;
980
981 /* Print header */
982 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
983 for (order = 0; order < MAX_ORDER; ++order)
984 seq_printf(m, "%6d ", order);
985 seq_putc(m, '\n');
986
987 walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
988
989 return 0;
990}
991
992static void pagetypeinfo_showblockcount_print(struct seq_file *m,
993 pg_data_t *pgdat, struct zone *zone)
994{
995 int mtype;
996 unsigned long pfn;
997 unsigned long start_pfn = zone->zone_start_pfn;
998 unsigned long end_pfn = zone_end_pfn(zone);
999 unsigned long count[MIGRATE_TYPES] = { 0, };
1000
1001 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1002 struct page *page;
1003
1004 if (!pfn_valid(pfn))
1005 continue;
1006
1007 page = pfn_to_page(pfn);
1008
1009 /* Watch for unexpected holes punched in the memmap */
1010 if (!memmap_valid_within(pfn, page, zone))
1011 continue;
1012
1013 mtype = get_pageblock_migratetype(page);
1014
1015 if (mtype < MIGRATE_TYPES)
1016 count[mtype]++;
1017 }
1018
1019 /* Print counts */
1020 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1021 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1022 seq_printf(m, "%12lu ", count[mtype]);
1023 seq_putc(m, '\n');
1024}
1025
1026/* Print out the free pages at each order for each migratetype */
1027static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1028{
1029 int mtype;
1030 pg_data_t *pgdat = (pg_data_t *)arg;
1031
1032 seq_printf(m, "\n%-23s", "Number of blocks type ");
1033 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1034 seq_printf(m, "%12s ", migratetype_names[mtype]);
1035 seq_putc(m, '\n');
1036 walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1037
1038 return 0;
1039}
1040
1041#ifdef CONFIG_PAGE_OWNER
1042static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1043 pg_data_t *pgdat,
1044 struct zone *zone)
1045{
1046 struct page *page;
1047 struct page_ext *page_ext;
1048 unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1049 unsigned long end_pfn = pfn + zone->spanned_pages;
1050 unsigned long count[MIGRATE_TYPES] = { 0, };
1051 int pageblock_mt, page_mt;
1052 int i;
1053
1054 /* Scan block by block. First and last block may be incomplete */
1055 pfn = zone->zone_start_pfn;
1056
1057 /*
1058 * Walk the zone in pageblock_nr_pages steps. If a page block spans
1059 * a zone boundary, it will be double counted between zones. This does
1060 * not matter as the mixed block count will still be correct
1061 */
1062 for (; pfn < end_pfn; ) {
1063 if (!pfn_valid(pfn)) {
1064 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1065 continue;
1066 }
1067
1068 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1069 block_end_pfn = min(block_end_pfn, end_pfn);
1070
1071 page = pfn_to_page(pfn);
1072 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1073
1074 for (; pfn < block_end_pfn; pfn++) {
1075 if (!pfn_valid_within(pfn))
1076 continue;
1077
1078 page = pfn_to_page(pfn);
1079 if (PageBuddy(page)) {
1080 pfn += (1UL << page_order(page)) - 1;
1081 continue;
1082 }
1083
1084 if (PageReserved(page))
1085 continue;
1086
1087 page_ext = lookup_page_ext(page);
1088
1089 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1090 continue;
1091
1092 page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1093 if (pageblock_mt != page_mt) {
1094 if (is_migrate_cma(pageblock_mt))
1095 count[MIGRATE_MOVABLE]++;
1096 else
1097 count[pageblock_mt]++;
1098
1099 pfn = block_end_pfn;
1100 break;
1101 }
1102 pfn += (1UL << page_ext->order) - 1;
1103 }
1104 }
1105
1106 /* Print counts */
1107 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1108 for (i = 0; i < MIGRATE_TYPES; i++)
1109 seq_printf(m, "%12lu ", count[i]);
1110 seq_putc(m, '\n');
1111}
1112#endif /* CONFIG_PAGE_OWNER */
1113
1114/*
1115 * Print out the number of pageblocks for each migratetype that contain pages
1116 * of other types. This gives an indication of how well fallbacks are being
1117 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1118 * to determine what is going on
1119 */
1120static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1121{
1122#ifdef CONFIG_PAGE_OWNER
1123 int mtype;
1124
1125 if (!static_branch_unlikely(&page_owner_inited))
1126 return;
1127
1128 drain_all_pages(NULL);
1129
1130 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1131 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1132 seq_printf(m, "%12s ", migratetype_names[mtype]);
1133 seq_putc(m, '\n');
1134
1135 walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1136#endif /* CONFIG_PAGE_OWNER */
1137}
1138
1139/*
1140 * This prints out statistics in relation to grouping pages by mobility.
1141 * It is expensive to collect so do not constantly read the file.
1142 */
1143static int pagetypeinfo_show(struct seq_file *m, void *arg)
1144{
1145 pg_data_t *pgdat = (pg_data_t *)arg;
1146
1147 /* check memoryless node */
1148 if (!node_state(pgdat->node_id, N_MEMORY))
1149 return 0;
1150
1151 seq_printf(m, "Page block order: %d\n", pageblock_order);
1152 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1153 seq_putc(m, '\n');
1154 pagetypeinfo_showfree(m, pgdat);
1155 pagetypeinfo_showblockcount(m, pgdat);
1156 pagetypeinfo_showmixedcount(m, pgdat);
1157
1158 return 0;
1159}
1160
1161static const struct seq_operations fragmentation_op = {
1162 .start = frag_start,
1163 .next = frag_next,
1164 .stop = frag_stop,
1165 .show = frag_show,
1166};
1167
1168static int fragmentation_open(struct inode *inode, struct file *file)
1169{
1170 return seq_open(file, &fragmentation_op);
1171}
1172
1173static const struct file_operations fragmentation_file_operations = {
1174 .open = fragmentation_open,
1175 .read = seq_read,
1176 .llseek = seq_lseek,
1177 .release = seq_release,
1178};
1179
1180static const struct seq_operations pagetypeinfo_op = {
1181 .start = frag_start,
1182 .next = frag_next,
1183 .stop = frag_stop,
1184 .show = pagetypeinfo_show,
1185};
1186
1187static int pagetypeinfo_open(struct inode *inode, struct file *file)
1188{
1189 return seq_open(file, &pagetypeinfo_op);
1190}
1191
1192static const struct file_operations pagetypeinfo_file_ops = {
1193 .open = pagetypeinfo_open,
1194 .read = seq_read,
1195 .llseek = seq_lseek,
1196 .release = seq_release,
1197};
1198
1199static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1200 struct zone *zone)
1201{
1202 int i;
1203 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1204 seq_printf(m,
1205 "\n pages free %lu"
1206 "\n min %lu"
1207 "\n low %lu"
1208 "\n high %lu"
1209 "\n scanned %lu"
1210 "\n spanned %lu"
1211 "\n present %lu"
1212 "\n managed %lu",
1213 zone_page_state(zone, NR_FREE_PAGES),
1214 min_wmark_pages(zone),
1215 low_wmark_pages(zone),
1216 high_wmark_pages(zone),
1217 zone_page_state(zone, NR_PAGES_SCANNED),
1218 zone->spanned_pages,
1219 zone->present_pages,
1220 zone->managed_pages);
1221
1222 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1223 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1224 zone_page_state(zone, i));
1225
1226 seq_printf(m,
1227 "\n protection: (%ld",
1228 zone->lowmem_reserve[0]);
1229 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1230 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1231 seq_printf(m,
1232 ")"
1233 "\n pagesets");
1234 for_each_online_cpu(i) {
1235 struct per_cpu_pageset *pageset;
1236
1237 pageset = per_cpu_ptr(zone->pageset, i);
1238 seq_printf(m,
1239 "\n cpu: %i"
1240 "\n count: %i"
1241 "\n high: %i"
1242 "\n batch: %i",
1243 i,
1244 pageset->pcp.count,
1245 pageset->pcp.high,
1246 pageset->pcp.batch);
1247#ifdef CONFIG_SMP
1248 seq_printf(m, "\n vm stats threshold: %d",
1249 pageset->stat_threshold);
1250#endif
1251 }
1252 seq_printf(m,
1253 "\n all_unreclaimable: %u"
1254 "\n start_pfn: %lu"
1255 "\n inactive_ratio: %u",
1256 !zone_reclaimable(zone),
1257 zone->zone_start_pfn,
1258 zone->inactive_ratio);
1259 seq_putc(m, '\n');
1260}
1261
1262/*
1263 * Output information about zones in @pgdat.
1264 */
1265static int zoneinfo_show(struct seq_file *m, void *arg)
1266{
1267 pg_data_t *pgdat = (pg_data_t *)arg;
1268 walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1269 return 0;
1270}
1271
1272static const struct seq_operations zoneinfo_op = {
1273 .start = frag_start, /* iterate over all zones. The same as in
1274 * fragmentation. */
1275 .next = frag_next,
1276 .stop = frag_stop,
1277 .show = zoneinfo_show,
1278};
1279
1280static int zoneinfo_open(struct inode *inode, struct file *file)
1281{
1282 return seq_open(file, &zoneinfo_op);
1283}
1284
1285static const struct file_operations proc_zoneinfo_file_operations = {
1286 .open = zoneinfo_open,
1287 .read = seq_read,
1288 .llseek = seq_lseek,
1289 .release = seq_release,
1290};
1291
1292enum writeback_stat_item {
1293 NR_DIRTY_THRESHOLD,
1294 NR_DIRTY_BG_THRESHOLD,
1295 NR_VM_WRITEBACK_STAT_ITEMS,
1296};
1297
1298static void *vmstat_start(struct seq_file *m, loff_t *pos)
1299{
1300 unsigned long *v;
1301 int i, stat_items_size;
1302
1303 if (*pos >= ARRAY_SIZE(vmstat_text))
1304 return NULL;
1305 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1306 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1307
1308#ifdef CONFIG_VM_EVENT_COUNTERS
1309 stat_items_size += sizeof(struct vm_event_state);
1310#endif
1311
1312 v = kmalloc(stat_items_size, GFP_KERNEL);
1313 m->private = v;
1314 if (!v)
1315 return ERR_PTR(-ENOMEM);
1316 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1317 v[i] = global_page_state(i);
1318 v += NR_VM_ZONE_STAT_ITEMS;
1319
1320 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1321 v + NR_DIRTY_THRESHOLD);
1322 v += NR_VM_WRITEBACK_STAT_ITEMS;
1323
1324#ifdef CONFIG_VM_EVENT_COUNTERS
1325 all_vm_events(v);
1326 v[PGPGIN] /= 2; /* sectors -> kbytes */
1327 v[PGPGOUT] /= 2;
1328#endif
1329 return (unsigned long *)m->private + *pos;
1330}
1331
1332static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1333{
1334 (*pos)++;
1335 if (*pos >= ARRAY_SIZE(vmstat_text))
1336 return NULL;
1337 return (unsigned long *)m->private + *pos;
1338}
1339
1340static int vmstat_show(struct seq_file *m, void *arg)
1341{
1342 unsigned long *l = arg;
1343 unsigned long off = l - (unsigned long *)m->private;
1344
1345 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1346 return 0;
1347}
1348
1349static void vmstat_stop(struct seq_file *m, void *arg)
1350{
1351 kfree(m->private);
1352 m->private = NULL;
1353}
1354
1355static const struct seq_operations vmstat_op = {
1356 .start = vmstat_start,
1357 .next = vmstat_next,
1358 .stop = vmstat_stop,
1359 .show = vmstat_show,
1360};
1361
1362static int vmstat_open(struct inode *inode, struct file *file)
1363{
1364 return seq_open(file, &vmstat_op);
1365}
1366
1367static const struct file_operations proc_vmstat_file_operations = {
1368 .open = vmstat_open,
1369 .read = seq_read,
1370 .llseek = seq_lseek,
1371 .release = seq_release,
1372};
1373#endif /* CONFIG_PROC_FS */
1374
1375#ifdef CONFIG_SMP
1376static struct workqueue_struct *vmstat_wq;
1377static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1378int sysctl_stat_interval __read_mostly = HZ;
1379static cpumask_var_t cpu_stat_off;
1380
1381static void vmstat_update(struct work_struct *w)
1382{
1383 if (refresh_cpu_vm_stats(true)) {
1384 /*
1385 * Counters were updated so we expect more updates
1386 * to occur in the future. Keep on running the
1387 * update worker thread.
1388 * If we were marked on cpu_stat_off clear the flag
1389 * so that vmstat_shepherd doesn't schedule us again.
1390 */
1391 if (!cpumask_test_and_clear_cpu(smp_processor_id(),
1392 cpu_stat_off)) {
1393 queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1394 this_cpu_ptr(&vmstat_work),
1395 round_jiffies_relative(sysctl_stat_interval));
1396 }
1397 } else {
1398 /*
1399 * We did not update any counters so the app may be in
1400 * a mode where it does not cause counter updates.
1401 * We may be uselessly running vmstat_update.
1402 * Defer the checking for differentials to the
1403 * shepherd thread on a different processor.
1404 */
1405 cpumask_set_cpu(smp_processor_id(), cpu_stat_off);
1406 }
1407}
1408
1409/*
1410 * Switch off vmstat processing and then fold all the remaining differentials
1411 * until the diffs stay at zero. The function is used by NOHZ and can only be
1412 * invoked when tick processing is not active.
1413 */
1414/*
1415 * Check if the diffs for a certain cpu indicate that
1416 * an update is needed.
1417 */
1418static bool need_update(int cpu)
1419{
1420 struct zone *zone;
1421
1422 for_each_populated_zone(zone) {
1423 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1424
1425 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1426 /*
1427 * The fast way of checking if there are any vmstat diffs.
1428 * This works because the diffs are byte sized items.
1429 */
1430 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1431 return true;
1432
1433 }
1434 return false;
1435}
1436
1437void quiet_vmstat(void)
1438{
1439 if (system_state != SYSTEM_RUNNING)
1440 return;
1441
1442 /*
1443 * If we are already in hands of the shepherd then there
1444 * is nothing for us to do here.
1445 */
1446 if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
1447 return;
1448
1449 if (!need_update(smp_processor_id()))
1450 return;
1451
1452 /*
1453 * Just refresh counters and do not care about the pending delayed
1454 * vmstat_update. It doesn't fire that often to matter and canceling
1455 * it would be too expensive from this path.
1456 * vmstat_shepherd will take care about that for us.
1457 */
1458 refresh_cpu_vm_stats(false);
1459}
1460
1461
1462/*
1463 * Shepherd worker thread that checks the
1464 * differentials of processors that have their worker
1465 * threads for vm statistics updates disabled because of
1466 * inactivity.
1467 */
1468static void vmstat_shepherd(struct work_struct *w);
1469
1470static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1471
1472static void vmstat_shepherd(struct work_struct *w)
1473{
1474 int cpu;
1475
1476 get_online_cpus();
1477 /* Check processors whose vmstat worker threads have been disabled */
1478 for_each_cpu(cpu, cpu_stat_off) {
1479 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1480
1481 if (need_update(cpu)) {
1482 if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1483 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1484 } else {
1485 /*
1486 * Cancel the work if quiet_vmstat has put this
1487 * cpu on cpu_stat_off because the work item might
1488 * be still scheduled
1489 */
1490 cancel_delayed_work(dw);
1491 }
1492 }
1493 put_online_cpus();
1494
1495 schedule_delayed_work(&shepherd,
1496 round_jiffies_relative(sysctl_stat_interval));
1497}
1498
1499static void __init start_shepherd_timer(void)
1500{
1501 int cpu;
1502
1503 for_each_possible_cpu(cpu)
1504 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1505 vmstat_update);
1506
1507 if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1508 BUG();
1509 cpumask_copy(cpu_stat_off, cpu_online_mask);
1510
1511 vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1512 schedule_delayed_work(&shepherd,
1513 round_jiffies_relative(sysctl_stat_interval));
1514}
1515
1516static void vmstat_cpu_dead(int node)
1517{
1518 int cpu;
1519
1520 get_online_cpus();
1521 for_each_online_cpu(cpu)
1522 if (cpu_to_node(cpu) == node)
1523 goto end;
1524
1525 node_clear_state(node, N_CPU);
1526end:
1527 put_online_cpus();
1528}
1529
1530/*
1531 * Use the cpu notifier to insure that the thresholds are recalculated
1532 * when necessary.
1533 */
1534static int vmstat_cpuup_callback(struct notifier_block *nfb,
1535 unsigned long action,
1536 void *hcpu)
1537{
1538 long cpu = (long)hcpu;
1539
1540 switch (action) {
1541 case CPU_ONLINE:
1542 case CPU_ONLINE_FROZEN:
1543 refresh_zone_stat_thresholds();
1544 node_set_state(cpu_to_node(cpu), N_CPU);
1545 cpumask_set_cpu(cpu, cpu_stat_off);
1546 break;
1547 case CPU_DOWN_PREPARE:
1548 case CPU_DOWN_PREPARE_FROZEN:
1549 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1550 cpumask_clear_cpu(cpu, cpu_stat_off);
1551 break;
1552 case CPU_DOWN_FAILED:
1553 case CPU_DOWN_FAILED_FROZEN:
1554 cpumask_set_cpu(cpu, cpu_stat_off);
1555 break;
1556 case CPU_DEAD:
1557 case CPU_DEAD_FROZEN:
1558 refresh_zone_stat_thresholds();
1559 vmstat_cpu_dead(cpu_to_node(cpu));
1560 break;
1561 default:
1562 break;
1563 }
1564 return NOTIFY_OK;
1565}
1566
1567static struct notifier_block vmstat_notifier =
1568 { &vmstat_cpuup_callback, NULL, 0 };
1569#endif
1570
1571static int __init setup_vmstat(void)
1572{
1573#ifdef CONFIG_SMP
1574 cpu_notifier_register_begin();
1575 __register_cpu_notifier(&vmstat_notifier);
1576
1577 start_shepherd_timer();
1578 cpu_notifier_register_done();
1579#endif
1580#ifdef CONFIG_PROC_FS
1581 proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1582 proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1583 proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1584 proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1585#endif
1586 return 0;
1587}
1588module_init(setup_vmstat)
1589
1590#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1591
1592/*
1593 * Return an index indicating how much of the available free memory is
1594 * unusable for an allocation of the requested size.
1595 */
1596static int unusable_free_index(unsigned int order,
1597 struct contig_page_info *info)
1598{
1599 /* No free memory is interpreted as all free memory is unusable */
1600 if (info->free_pages == 0)
1601 return 1000;
1602
1603 /*
1604 * Index should be a value between 0 and 1. Return a value to 3
1605 * decimal places.
1606 *
1607 * 0 => no fragmentation
1608 * 1 => high fragmentation
1609 */
1610 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1611
1612}
1613
1614static void unusable_show_print(struct seq_file *m,
1615 pg_data_t *pgdat, struct zone *zone)
1616{
1617 unsigned int order;
1618 int index;
1619 struct contig_page_info info;
1620
1621 seq_printf(m, "Node %d, zone %8s ",
1622 pgdat->node_id,
1623 zone->name);
1624 for (order = 0; order < MAX_ORDER; ++order) {
1625 fill_contig_page_info(zone, order, &info);
1626 index = unusable_free_index(order, &info);
1627 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1628 }
1629
1630 seq_putc(m, '\n');
1631}
1632
1633/*
1634 * Display unusable free space index
1635 *
1636 * The unusable free space index measures how much of the available free
1637 * memory cannot be used to satisfy an allocation of a given size and is a
1638 * value between 0 and 1. The higher the value, the more of free memory is
1639 * unusable and by implication, the worse the external fragmentation is. This
1640 * can be expressed as a percentage by multiplying by 100.
1641 */
1642static int unusable_show(struct seq_file *m, void *arg)
1643{
1644 pg_data_t *pgdat = (pg_data_t *)arg;
1645
1646 /* check memoryless node */
1647 if (!node_state(pgdat->node_id, N_MEMORY))
1648 return 0;
1649
1650 walk_zones_in_node(m, pgdat, unusable_show_print);
1651
1652 return 0;
1653}
1654
1655static const struct seq_operations unusable_op = {
1656 .start = frag_start,
1657 .next = frag_next,
1658 .stop = frag_stop,
1659 .show = unusable_show,
1660};
1661
1662static int unusable_open(struct inode *inode, struct file *file)
1663{
1664 return seq_open(file, &unusable_op);
1665}
1666
1667static const struct file_operations unusable_file_ops = {
1668 .open = unusable_open,
1669 .read = seq_read,
1670 .llseek = seq_lseek,
1671 .release = seq_release,
1672};
1673
1674static void extfrag_show_print(struct seq_file *m,
1675 pg_data_t *pgdat, struct zone *zone)
1676{
1677 unsigned int order;
1678 int index;
1679
1680 /* Alloc on stack as interrupts are disabled for zone walk */
1681 struct contig_page_info info;
1682
1683 seq_printf(m, "Node %d, zone %8s ",
1684 pgdat->node_id,
1685 zone->name);
1686 for (order = 0; order < MAX_ORDER; ++order) {
1687 fill_contig_page_info(zone, order, &info);
1688 index = __fragmentation_index(order, &info);
1689 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1690 }
1691
1692 seq_putc(m, '\n');
1693}
1694
1695/*
1696 * Display fragmentation index for orders that allocations would fail for
1697 */
1698static int extfrag_show(struct seq_file *m, void *arg)
1699{
1700 pg_data_t *pgdat = (pg_data_t *)arg;
1701
1702 walk_zones_in_node(m, pgdat, extfrag_show_print);
1703
1704 return 0;
1705}
1706
1707static const struct seq_operations extfrag_op = {
1708 .start = frag_start,
1709 .next = frag_next,
1710 .stop = frag_stop,
1711 .show = extfrag_show,
1712};
1713
1714static int extfrag_open(struct inode *inode, struct file *file)
1715{
1716 return seq_open(file, &extfrag_op);
1717}
1718
1719static const struct file_operations extfrag_file_ops = {
1720 .open = extfrag_open,
1721 .read = seq_read,
1722 .llseek = seq_lseek,
1723 .release = seq_release,
1724};
1725
1726static int __init extfrag_debug_init(void)
1727{
1728 struct dentry *extfrag_debug_root;
1729
1730 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1731 if (!extfrag_debug_root)
1732 return -ENOMEM;
1733
1734 if (!debugfs_create_file("unusable_index", 0444,
1735 extfrag_debug_root, NULL, &unusable_file_ops))
1736 goto fail;
1737
1738 if (!debugfs_create_file("extfrag_index", 0444,
1739 extfrag_debug_root, NULL, &extfrag_file_ops))
1740 goto fail;
1741
1742 return 0;
1743fail:
1744 debugfs_remove_recursive(extfrag_debug_root);
1745 return -ENOMEM;
1746}
1747
1748module_init(extfrag_debug_init);
1749#endif