4 * Manages VM statistics
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Copyright (C) 2006 Silicon Graphics, Inc.,
9 * Christoph Lameter <christoph@lameter.com>
10 * Copyright (C) 2008-2014 Christoph Lameter
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>
33 #ifdef CONFIG_VM_EVENT_COUNTERS
34 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
35 EXPORT_PER_CPU_SYMBOL(vm_event_states);
37 static void sum_vm_events(unsigned long *ret)
42 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
44 for_each_online_cpu(cpu) {
45 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
47 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
48 ret[i] += this->event[i];
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.
57 void all_vm_events(unsigned long *ret)
63 EXPORT_SYMBOL_GPL(all_vm_events);
66 * Fold the foreign cpu events into our own.
68 * This is adding to the events on one processor
69 * but keeps the global counts constant.
71 void vm_events_fold_cpu(int cpu)
73 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
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;
82 #endif /* CONFIG_VM_EVENT_COUNTERS */
85 * Manage combined zone based / global counters
87 * vm_stat contains the global counters
89 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
90 EXPORT_SYMBOL(vm_stat);
94 int calculate_pressure_threshold(struct zone *zone)
97 int watermark_distance;
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
107 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
108 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
111 * Maximum threshold is 125
113 threshold = min(125, threshold);
118 int calculate_normal_threshold(struct zone *zone)
121 int mem; /* memory in 128 MB units */
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.
129 * Some sample thresholds:
131 * Threshold Processors (fls) Zonesize fls(mem+1)
132 * ------------------------------------------------------------------
149 * 125 1024 10 8-16 GB 8
150 * 125 1024 10 16-32 GB 9
153 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
155 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
158 * Maximum threshold is 125
160 threshold = min(125, threshold);
166 * Refresh the thresholds for each zone.
168 void refresh_zone_stat_thresholds(void)
174 for_each_populated_zone(zone) {
175 unsigned long max_drift, tolerate_drift;
177 threshold = calculate_normal_threshold(zone);
179 for_each_online_cpu(cpu)
180 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
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
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) +
196 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
197 int (*calculate_pressure)(struct zone *))
204 for (i = 0; i < pgdat->nr_zones; i++) {
205 zone = &pgdat->node_zones[i];
206 if (!zone->percpu_drift_mark)
209 threshold = (*calculate_pressure)(zone);
210 for_each_online_cpu(cpu)
211 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
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.
221 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
224 struct per_cpu_pageset __percpu *pcp = zone->pageset;
225 s8 __percpu *p = pcp->vm_stat_diff + item;
229 x = delta + __this_cpu_read(*p);
231 t = __this_cpu_read(pcp->stat_threshold);
233 if (unlikely(x > t || x < -t)) {
234 zone_page_state_add(x, zone, item);
237 __this_cpu_write(*p, x);
239 EXPORT_SYMBOL(__mod_zone_page_state);
242 * Optimized increment and decrement functions.
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.
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
254 * NOTE: These functions are very performance sensitive. Change only
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.
264 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
266 struct per_cpu_pageset __percpu *pcp = zone->pageset;
267 s8 __percpu *p = pcp->vm_stat_diff + item;
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;
275 zone_page_state_add(v + overstep, zone, item);
276 __this_cpu_write(*p, -overstep);
280 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
282 __inc_zone_state(page_zone(page), item);
284 EXPORT_SYMBOL(__inc_zone_page_state);
286 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
288 struct per_cpu_pageset __percpu *pcp = zone->pageset;
289 s8 __percpu *p = pcp->vm_stat_diff + item;
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;
297 zone_page_state_add(v - overstep, zone, item);
298 __this_cpu_write(*p, overstep);
302 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
304 __dec_zone_state(page_zone(page), item);
306 EXPORT_SYMBOL(__dec_zone_page_state);
308 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
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.
313 * mod_state() modifies the zone counter state through atomic per cpu
316 * Overstep mode specifies how overstep should handled:
318 * 1 Overstepping half of threshold
319 * -1 Overstepping minus half of threshold
321 static inline void mod_state(struct zone *zone, enum zone_stat_item item,
322 long delta, int overstep_mode)
324 struct per_cpu_pageset __percpu *pcp = zone->pageset;
325 s8 __percpu *p = pcp->vm_stat_diff + item;
329 z = 0; /* overflow to zone counters */
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.
338 * Most of the time the thresholds are the same anyways
339 * for all cpus in a zone.
341 t = this_cpu_read(pcp->stat_threshold);
343 o = this_cpu_read(*p);
346 if (n > t || n < -t) {
347 int os = overstep_mode * (t >> 1) ;
349 /* Overflow must be added to zone counters */
353 } while (this_cpu_cmpxchg(*p, o, n) != o);
356 zone_page_state_add(z, zone, item);
359 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
362 mod_state(zone, item, delta, 0);
364 EXPORT_SYMBOL(mod_zone_page_state);
366 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
368 mod_state(zone, item, 1, 1);
371 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
373 mod_state(page_zone(page), item, 1, 1);
375 EXPORT_SYMBOL(inc_zone_page_state);
377 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
379 mod_state(page_zone(page), item, -1, -1);
381 EXPORT_SYMBOL(dec_zone_page_state);
384 * Use interrupt disable to serialize counter updates
386 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
391 local_irq_save(flags);
392 __mod_zone_page_state(zone, item, delta);
393 local_irq_restore(flags);
395 EXPORT_SYMBOL(mod_zone_page_state);
397 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
401 local_irq_save(flags);
402 __inc_zone_state(zone, item);
403 local_irq_restore(flags);
406 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
411 zone = page_zone(page);
412 local_irq_save(flags);
413 __inc_zone_state(zone, item);
414 local_irq_restore(flags);
416 EXPORT_SYMBOL(inc_zone_page_state);
418 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
422 local_irq_save(flags);
423 __dec_zone_page_state(page, item);
424 local_irq_restore(flags);
426 EXPORT_SYMBOL(dec_zone_page_state);
431 * Fold a differential into the global counters.
432 * Returns the number of counters updated.
434 static int fold_diff(int *diff)
439 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
441 atomic_long_add(diff[i], &vm_stat[i]);
448 * Update the zone counters for the current cpu.
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
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.
461 * The function returns the number of global counters updated.
463 static int refresh_cpu_vm_stats(bool do_pagesets)
467 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
470 for_each_populated_zone(zone) {
471 struct per_cpu_pageset __percpu *p = zone->pageset;
473 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
476 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
479 atomic_long_add(v, &zone->vm_stat[i]);
482 /* 3 seconds idle till flush */
483 __this_cpu_write(p->expire, 3);
491 * Deal with draining the remote pageset of this
494 * Check if there are pages remaining in this pageset
495 * if not then there is nothing to expire.
497 if (!__this_cpu_read(p->expire) ||
498 !__this_cpu_read(p->pcp.count))
502 * We never drain zones local to this processor.
504 if (zone_to_nid(zone) == numa_node_id()) {
505 __this_cpu_write(p->expire, 0);
509 if (__this_cpu_dec_return(p->expire))
512 if (__this_cpu_read(p->pcp.count)) {
513 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
519 changes += fold_diff(global_diff);
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.
528 void cpu_vm_stats_fold(int cpu)
532 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
534 for_each_populated_zone(zone) {
535 struct per_cpu_pageset *p;
537 p = per_cpu_ptr(zone->pageset, cpu);
539 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
540 if (p->vm_stat_diff[i]) {
543 v = p->vm_stat_diff[i];
544 p->vm_stat_diff[i] = 0;
545 atomic_long_add(v, &zone->vm_stat[i]);
550 fold_diff(global_diff);
554 * this is only called if !populated_zone(zone), which implies no other users of
555 * pset->vm_stat_diff[] exsist.
557 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
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]);
573 * Determine the per node value of a stat item.
575 unsigned long node_page_state(int node, enum zone_stat_item item)
577 struct zone *zones = NODE_DATA(node)->node_zones;
579 unsigned long count = 0;
581 for (i = 0; i < MAX_NR_ZONES; i++)
582 count += zone_page_state(zones + i, item);
589 #ifdef CONFIG_COMPACTION
591 struct contig_page_info {
592 unsigned long free_pages;
593 unsigned long free_blocks_total;
594 unsigned long free_blocks_suitable;
598 * Calculate the number of free pages in a zone, how many contiguous
599 * pages are free and how many are large enough to satisfy an allocation of
600 * the target size. Note that this function makes no attempt to estimate
601 * how many suitable free blocks there *might* be if MOVABLE pages were
602 * migrated. Calculating that is possible, but expensive and can be
603 * figured out from userspace
605 static void fill_contig_page_info(struct zone *zone,
606 unsigned int suitable_order,
607 struct contig_page_info *info)
611 info->free_pages = 0;
612 info->free_blocks_total = 0;
613 info->free_blocks_suitable = 0;
615 for (order = 0; order < MAX_ORDER; order++) {
616 unsigned long blocks;
618 /* Count number of free blocks */
619 blocks = zone->free_area[order].nr_free;
620 info->free_blocks_total += blocks;
622 /* Count free base pages */
623 info->free_pages += blocks << order;
625 /* Count the suitable free blocks */
626 if (order >= suitable_order)
627 info->free_blocks_suitable += blocks <<
628 (order - suitable_order);
633 * A fragmentation index only makes sense if an allocation of a requested
634 * size would fail. If that is true, the fragmentation index indicates
635 * whether external fragmentation or a lack of memory was the problem.
636 * The value can be used to determine if page reclaim or compaction
639 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
641 unsigned long requested = 1UL << order;
643 if (!info->free_blocks_total)
646 /* Fragmentation index only makes sense when a request would fail */
647 if (info->free_blocks_suitable)
651 * Index is between 0 and 1 so return within 3 decimal places
653 * 0 => allocation would fail due to lack of memory
654 * 1 => allocation would fail due to fragmentation
656 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
659 /* Same as __fragmentation index but allocs contig_page_info on stack */
660 int fragmentation_index(struct zone *zone, unsigned int order)
662 struct contig_page_info info;
664 fill_contig_page_info(zone, order, &info);
665 return __fragmentation_index(order, &info);
669 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
670 #ifdef CONFIG_ZONE_DMA
671 #define TEXT_FOR_DMA(xx) xx "_dma",
673 #define TEXT_FOR_DMA(xx)
676 #ifdef CONFIG_ZONE_DMA32
677 #define TEXT_FOR_DMA32(xx) xx "_dma32",
679 #define TEXT_FOR_DMA32(xx)
682 #ifdef CONFIG_HIGHMEM
683 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
685 #define TEXT_FOR_HIGHMEM(xx)
688 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
689 TEXT_FOR_HIGHMEM(xx) xx "_movable",
691 const char * const vmstat_text[] = {
692 /* enum zone_stat_item countes */
706 "nr_slab_reclaimable",
707 "nr_slab_unreclaimable",
708 "nr_page_table_pages",
713 "nr_vmscan_immediate_reclaim",
721 #if IS_ENABLED(CONFIG_ZSMALLOC)
732 "workingset_refault",
733 "workingset_activate",
734 "workingset_nodereclaim",
735 "nr_anon_transparent_hugepages",
738 /* enum writeback_stat_item counters */
739 "nr_dirty_threshold",
740 "nr_dirty_background_threshold",
742 #ifdef CONFIG_VM_EVENT_COUNTERS
743 /* enum vm_event_item counters */
749 TEXTS_FOR_ZONES("pgalloc")
759 TEXTS_FOR_ZONES("pgrefill")
760 TEXTS_FOR_ZONES("pgsteal_kswapd")
761 TEXTS_FOR_ZONES("pgsteal_direct")
762 TEXTS_FOR_ZONES("pgscan_kswapd")
763 TEXTS_FOR_ZONES("pgscan_direct")
764 "pgscan_direct_throttle",
767 "zone_reclaim_failed",
772 "kswapd_low_wmark_hit_quickly",
773 "kswapd_high_wmark_hit_quickly",
782 #ifdef CONFIG_NUMA_BALANCING
784 "numa_huge_pte_updates",
786 "numa_hint_faults_local",
787 "numa_pages_migrated",
789 #ifdef CONFIG_MIGRATION
793 #ifdef CONFIG_COMPACTION
794 "compact_migrate_scanned",
795 "compact_free_scanned",
800 "compact_daemon_wake",
803 #ifdef CONFIG_HUGETLB_PAGE
804 "htlb_buddy_alloc_success",
805 "htlb_buddy_alloc_fail",
807 "unevictable_pgs_culled",
808 "unevictable_pgs_scanned",
809 "unevictable_pgs_rescued",
810 "unevictable_pgs_mlocked",
811 "unevictable_pgs_munlocked",
812 "unevictable_pgs_cleared",
813 "unevictable_pgs_stranded",
815 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
817 "thp_fault_fallback",
818 "thp_collapse_alloc",
819 "thp_collapse_alloc_failed",
821 "thp_split_page_failed",
822 "thp_deferred_split_page",
824 "thp_zero_page_alloc",
825 "thp_zero_page_alloc_failed",
827 #ifdef CONFIG_MEMORY_BALLOON
830 #ifdef CONFIG_BALLOON_COMPACTION
833 #endif /* CONFIG_MEMORY_BALLOON */
834 #ifdef CONFIG_DEBUG_TLBFLUSH
836 "nr_tlb_remote_flush",
837 "nr_tlb_remote_flush_received",
838 #endif /* CONFIG_SMP */
839 "nr_tlb_local_flush_all",
840 "nr_tlb_local_flush_one",
841 #endif /* CONFIG_DEBUG_TLBFLUSH */
843 #ifdef CONFIG_DEBUG_VM_VMACACHE
844 "vmacache_find_calls",
845 "vmacache_find_hits",
846 "vmacache_full_flushes",
848 #endif /* CONFIG_VM_EVENTS_COUNTERS */
850 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
853 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
854 defined(CONFIG_PROC_FS)
855 static void *frag_start(struct seq_file *m, loff_t *pos)
860 for (pgdat = first_online_pgdat();
862 pgdat = next_online_pgdat(pgdat))
868 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
870 pg_data_t *pgdat = (pg_data_t *)arg;
873 return next_online_pgdat(pgdat);
876 static void frag_stop(struct seq_file *m, void *arg)
880 /* Walk all the zones in a node and print using a callback */
881 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
882 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
885 struct zone *node_zones = pgdat->node_zones;
888 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
889 if (!populated_zone(zone))
892 spin_lock_irqsave(&zone->lock, flags);
893 print(m, pgdat, zone);
894 spin_unlock_irqrestore(&zone->lock, flags);
899 #ifdef CONFIG_PROC_FS
900 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
905 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
906 for (order = 0; order < MAX_ORDER; ++order)
907 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
912 * This walks the free areas for each zone.
914 static int frag_show(struct seq_file *m, void *arg)
916 pg_data_t *pgdat = (pg_data_t *)arg;
917 walk_zones_in_node(m, pgdat, frag_show_print);
921 static void pagetypeinfo_showfree_print(struct seq_file *m,
922 pg_data_t *pgdat, struct zone *zone)
926 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
927 seq_printf(m, "Node %4d, zone %8s, type %12s ",
930 migratetype_names[mtype]);
931 for (order = 0; order < MAX_ORDER; ++order) {
932 unsigned long freecount = 0;
933 struct free_area *area;
934 struct list_head *curr;
936 area = &(zone->free_area[order]);
938 list_for_each(curr, &area->free_list[mtype])
940 seq_printf(m, "%6lu ", freecount);
946 /* Print out the free pages at each order for each migatetype */
947 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
950 pg_data_t *pgdat = (pg_data_t *)arg;
953 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
954 for (order = 0; order < MAX_ORDER; ++order)
955 seq_printf(m, "%6d ", order);
958 walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
963 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
964 pg_data_t *pgdat, struct zone *zone)
968 unsigned long start_pfn = zone->zone_start_pfn;
969 unsigned long end_pfn = zone_end_pfn(zone);
970 unsigned long count[MIGRATE_TYPES] = { 0, };
972 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
978 page = pfn_to_page(pfn);
980 /* Watch for unexpected holes punched in the memmap */
981 if (!memmap_valid_within(pfn, page, zone))
984 if (page_zone(page) != zone)
987 mtype = get_pageblock_migratetype(page);
989 if (mtype < MIGRATE_TYPES)
994 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
995 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
996 seq_printf(m, "%12lu ", count[mtype]);
1000 /* Print out the free pages at each order for each migratetype */
1001 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1004 pg_data_t *pgdat = (pg_data_t *)arg;
1006 seq_printf(m, "\n%-23s", "Number of blocks type ");
1007 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1008 seq_printf(m, "%12s ", migratetype_names[mtype]);
1010 walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1015 #ifdef CONFIG_PAGE_OWNER
1016 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1021 struct page_ext *page_ext;
1022 unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1023 unsigned long end_pfn = pfn + zone->spanned_pages;
1024 unsigned long count[MIGRATE_TYPES] = { 0, };
1025 int pageblock_mt, page_mt;
1028 /* Scan block by block. First and last block may be incomplete */
1029 pfn = zone->zone_start_pfn;
1032 * Walk the zone in pageblock_nr_pages steps. If a page block spans
1033 * a zone boundary, it will be double counted between zones. This does
1034 * not matter as the mixed block count will still be correct
1036 for (; pfn < end_pfn; ) {
1037 if (!pfn_valid(pfn)) {
1038 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1042 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1043 block_end_pfn = min(block_end_pfn, end_pfn);
1045 page = pfn_to_page(pfn);
1046 pageblock_mt = get_pageblock_migratetype(page);
1048 for (; pfn < block_end_pfn; pfn++) {
1049 if (!pfn_valid_within(pfn))
1052 page = pfn_to_page(pfn);
1054 if (page_zone(page) != zone)
1057 if (PageBuddy(page)) {
1058 pfn += (1UL << page_order(page)) - 1;
1062 if (PageReserved(page))
1065 page_ext = lookup_page_ext(page);
1066 if (unlikely(!page_ext))
1069 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1072 page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1073 if (pageblock_mt != page_mt) {
1074 if (is_migrate_cma(pageblock_mt))
1075 count[MIGRATE_MOVABLE]++;
1077 count[pageblock_mt]++;
1079 pfn = block_end_pfn;
1082 pfn += (1UL << page_ext->order) - 1;
1087 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1088 for (i = 0; i < MIGRATE_TYPES; i++)
1089 seq_printf(m, "%12lu ", count[i]);
1092 #endif /* CONFIG_PAGE_OWNER */
1095 * Print out the number of pageblocks for each migratetype that contain pages
1096 * of other types. This gives an indication of how well fallbacks are being
1097 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1098 * to determine what is going on
1100 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1102 #ifdef CONFIG_PAGE_OWNER
1105 if (!static_branch_unlikely(&page_owner_inited))
1108 drain_all_pages(NULL);
1110 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1111 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1112 seq_printf(m, "%12s ", migratetype_names[mtype]);
1115 walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1116 #endif /* CONFIG_PAGE_OWNER */
1120 * This prints out statistics in relation to grouping pages by mobility.
1121 * It is expensive to collect so do not constantly read the file.
1123 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1125 pg_data_t *pgdat = (pg_data_t *)arg;
1127 /* check memoryless node */
1128 if (!node_state(pgdat->node_id, N_MEMORY))
1131 seq_printf(m, "Page block order: %d\n", pageblock_order);
1132 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1134 pagetypeinfo_showfree(m, pgdat);
1135 pagetypeinfo_showblockcount(m, pgdat);
1136 pagetypeinfo_showmixedcount(m, pgdat);
1141 static const struct seq_operations fragmentation_op = {
1142 .start = frag_start,
1148 static int fragmentation_open(struct inode *inode, struct file *file)
1150 return seq_open(file, &fragmentation_op);
1153 static const struct file_operations fragmentation_file_operations = {
1154 .open = fragmentation_open,
1156 .llseek = seq_lseek,
1157 .release = seq_release,
1160 static const struct seq_operations pagetypeinfo_op = {
1161 .start = frag_start,
1164 .show = pagetypeinfo_show,
1167 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1169 return seq_open(file, &pagetypeinfo_op);
1172 static const struct file_operations pagetypeinfo_file_ops = {
1173 .open = pagetypeinfo_open,
1175 .llseek = seq_lseek,
1176 .release = seq_release,
1179 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1183 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1193 zone_page_state(zone, NR_FREE_PAGES),
1194 min_wmark_pages(zone),
1195 low_wmark_pages(zone),
1196 high_wmark_pages(zone),
1197 zone_page_state(zone, NR_PAGES_SCANNED),
1198 zone->spanned_pages,
1199 zone->present_pages,
1200 zone->managed_pages);
1202 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1203 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1204 zone_page_state(zone, i));
1207 "\n protection: (%ld",
1208 zone->lowmem_reserve[0]);
1209 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1210 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1214 for_each_online_cpu(i) {
1215 struct per_cpu_pageset *pageset;
1217 pageset = per_cpu_ptr(zone->pageset, i);
1226 pageset->pcp.batch);
1228 seq_printf(m, "\n vm stats threshold: %d",
1229 pageset->stat_threshold);
1233 "\n all_unreclaimable: %u"
1235 "\n inactive_ratio: %u",
1236 !zone_reclaimable(zone),
1237 zone->zone_start_pfn,
1238 zone->inactive_ratio);
1243 * Output information about zones in @pgdat.
1245 static int zoneinfo_show(struct seq_file *m, void *arg)
1247 pg_data_t *pgdat = (pg_data_t *)arg;
1248 walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1252 static const struct seq_operations zoneinfo_op = {
1253 .start = frag_start, /* iterate over all zones. The same as in
1257 .show = zoneinfo_show,
1260 static int zoneinfo_open(struct inode *inode, struct file *file)
1262 return seq_open(file, &zoneinfo_op);
1265 static const struct file_operations proc_zoneinfo_file_operations = {
1266 .open = zoneinfo_open,
1268 .llseek = seq_lseek,
1269 .release = seq_release,
1272 enum writeback_stat_item {
1274 NR_DIRTY_BG_THRESHOLD,
1275 NR_VM_WRITEBACK_STAT_ITEMS,
1278 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1281 int i, stat_items_size;
1283 if (*pos >= ARRAY_SIZE(vmstat_text))
1285 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1286 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1288 #ifdef CONFIG_VM_EVENT_COUNTERS
1289 stat_items_size += sizeof(struct vm_event_state);
1292 v = kmalloc(stat_items_size, GFP_KERNEL);
1295 return ERR_PTR(-ENOMEM);
1296 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1297 v[i] = global_page_state(i);
1298 v += NR_VM_ZONE_STAT_ITEMS;
1300 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1301 v + NR_DIRTY_THRESHOLD);
1302 v += NR_VM_WRITEBACK_STAT_ITEMS;
1304 #ifdef CONFIG_VM_EVENT_COUNTERS
1306 v[PGPGIN] /= 2; /* sectors -> kbytes */
1309 return (unsigned long *)m->private + *pos;
1312 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1315 if (*pos >= ARRAY_SIZE(vmstat_text))
1317 return (unsigned long *)m->private + *pos;
1320 static int vmstat_show(struct seq_file *m, void *arg)
1322 unsigned long *l = arg;
1323 unsigned long off = l - (unsigned long *)m->private;
1325 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1329 static void vmstat_stop(struct seq_file *m, void *arg)
1335 static const struct seq_operations vmstat_op = {
1336 .start = vmstat_start,
1337 .next = vmstat_next,
1338 .stop = vmstat_stop,
1339 .show = vmstat_show,
1342 static int vmstat_open(struct inode *inode, struct file *file)
1344 return seq_open(file, &vmstat_op);
1347 static const struct file_operations proc_vmstat_file_operations = {
1348 .open = vmstat_open,
1350 .llseek = seq_lseek,
1351 .release = seq_release,
1353 #endif /* CONFIG_PROC_FS */
1356 static struct workqueue_struct *vmstat_wq;
1357 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1358 int sysctl_stat_interval __read_mostly = HZ;
1360 #ifdef CONFIG_PROC_FS
1361 static void refresh_vm_stats(struct work_struct *work)
1363 refresh_cpu_vm_stats(true);
1366 int vmstat_refresh(struct ctl_table *table, int write,
1367 void __user *buffer, size_t *lenp, loff_t *ppos)
1374 * The regular update, every sysctl_stat_interval, may come later
1375 * than expected: leaving a significant amount in per_cpu buckets.
1376 * This is particularly misleading when checking a quantity of HUGE
1377 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1378 * which can equally be echo'ed to or cat'ted from (by root),
1379 * can be used to update the stats just before reading them.
1381 * Oh, and since global_page_state() etc. are so careful to hide
1382 * transiently negative values, report an error here if any of
1383 * the stats is negative, so we know to go looking for imbalance.
1385 err = schedule_on_each_cpu(refresh_vm_stats);
1388 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1389 val = atomic_long_read(&vm_stat[i]);
1392 case NR_ALLOC_BATCH:
1393 case NR_PAGES_SCANNED:
1395 * These are often seen to go negative in
1396 * recent kernels, but not to go permanently
1397 * negative. Whilst it would be nicer not to
1398 * have exceptions, rooting them out would be
1399 * another task, of rather low priority.
1403 pr_warn("%s: %s %ld\n",
1404 __func__, vmstat_text[i], val);
1418 #endif /* CONFIG_PROC_FS */
1420 static void vmstat_update(struct work_struct *w)
1422 if (refresh_cpu_vm_stats(true)) {
1424 * Counters were updated so we expect more updates
1425 * to occur in the future. Keep on running the
1426 * update worker thread.
1428 queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1429 this_cpu_ptr(&vmstat_work),
1430 round_jiffies_relative(sysctl_stat_interval));
1435 * Switch off vmstat processing and then fold all the remaining differentials
1436 * until the diffs stay at zero. The function is used by NOHZ and can only be
1437 * invoked when tick processing is not active.
1440 * Check if the diffs for a certain cpu indicate that
1441 * an update is needed.
1443 static bool need_update(int cpu)
1447 for_each_populated_zone(zone) {
1448 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1450 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1452 * The fast way of checking if there are any vmstat diffs.
1453 * This works because the diffs are byte sized items.
1455 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1463 * Switch off vmstat processing and then fold all the remaining differentials
1464 * until the diffs stay at zero. The function is used by NOHZ and can only be
1465 * invoked when tick processing is not active.
1467 void quiet_vmstat(void)
1469 if (system_state != SYSTEM_RUNNING)
1472 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1475 if (!need_update(smp_processor_id()))
1479 * Just refresh counters and do not care about the pending delayed
1480 * vmstat_update. It doesn't fire that often to matter and canceling
1481 * it would be too expensive from this path.
1482 * vmstat_shepherd will take care about that for us.
1484 refresh_cpu_vm_stats(false);
1488 * Shepherd worker thread that checks the
1489 * differentials of processors that have their worker
1490 * threads for vm statistics updates disabled because of
1493 static void vmstat_shepherd(struct work_struct *w);
1495 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1497 static void vmstat_shepherd(struct work_struct *w)
1502 /* Check processors whose vmstat worker threads have been disabled */
1503 for_each_online_cpu(cpu) {
1504 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1506 if (!delayed_work_pending(dw) && need_update(cpu))
1507 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1511 schedule_delayed_work(&shepherd,
1512 round_jiffies_relative(sysctl_stat_interval));
1515 static void __init start_shepherd_timer(void)
1519 for_each_possible_cpu(cpu)
1520 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1523 vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1524 schedule_delayed_work(&shepherd,
1525 round_jiffies_relative(sysctl_stat_interval));
1528 static void vmstat_cpu_dead(int node)
1533 for_each_online_cpu(cpu)
1534 if (cpu_to_node(cpu) == node)
1537 node_clear_state(node, N_CPU);
1543 * Use the cpu notifier to insure that the thresholds are recalculated
1546 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1547 unsigned long action,
1550 long cpu = (long)hcpu;
1554 case CPU_ONLINE_FROZEN:
1555 refresh_zone_stat_thresholds();
1556 node_set_state(cpu_to_node(cpu), N_CPU);
1558 case CPU_DOWN_PREPARE:
1559 case CPU_DOWN_PREPARE_FROZEN:
1560 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1562 case CPU_DOWN_FAILED:
1563 case CPU_DOWN_FAILED_FROZEN:
1566 case CPU_DEAD_FROZEN:
1567 refresh_zone_stat_thresholds();
1568 vmstat_cpu_dead(cpu_to_node(cpu));
1576 static struct notifier_block vmstat_notifier =
1577 { &vmstat_cpuup_callback, NULL, 0 };
1580 static int __init setup_vmstat(void)
1583 cpu_notifier_register_begin();
1584 __register_cpu_notifier(&vmstat_notifier);
1586 start_shepherd_timer();
1587 cpu_notifier_register_done();
1589 #ifdef CONFIG_PROC_FS
1590 proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1591 proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1592 proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1593 proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1597 module_init(setup_vmstat)
1599 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1602 * Return an index indicating how much of the available free memory is
1603 * unusable for an allocation of the requested size.
1605 static int unusable_free_index(unsigned int order,
1606 struct contig_page_info *info)
1608 /* No free memory is interpreted as all free memory is unusable */
1609 if (info->free_pages == 0)
1613 * Index should be a value between 0 and 1. Return a value to 3
1616 * 0 => no fragmentation
1617 * 1 => high fragmentation
1619 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1623 static void unusable_show_print(struct seq_file *m,
1624 pg_data_t *pgdat, struct zone *zone)
1628 struct contig_page_info info;
1630 seq_printf(m, "Node %d, zone %8s ",
1633 for (order = 0; order < MAX_ORDER; ++order) {
1634 fill_contig_page_info(zone, order, &info);
1635 index = unusable_free_index(order, &info);
1636 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1643 * Display unusable free space index
1645 * The unusable free space index measures how much of the available free
1646 * memory cannot be used to satisfy an allocation of a given size and is a
1647 * value between 0 and 1. The higher the value, the more of free memory is
1648 * unusable and by implication, the worse the external fragmentation is. This
1649 * can be expressed as a percentage by multiplying by 100.
1651 static int unusable_show(struct seq_file *m, void *arg)
1653 pg_data_t *pgdat = (pg_data_t *)arg;
1655 /* check memoryless node */
1656 if (!node_state(pgdat->node_id, N_MEMORY))
1659 walk_zones_in_node(m, pgdat, unusable_show_print);
1664 static const struct seq_operations unusable_op = {
1665 .start = frag_start,
1668 .show = unusable_show,
1671 static int unusable_open(struct inode *inode, struct file *file)
1673 return seq_open(file, &unusable_op);
1676 static const struct file_operations unusable_file_ops = {
1677 .open = unusable_open,
1679 .llseek = seq_lseek,
1680 .release = seq_release,
1683 static void extfrag_show_print(struct seq_file *m,
1684 pg_data_t *pgdat, struct zone *zone)
1689 /* Alloc on stack as interrupts are disabled for zone walk */
1690 struct contig_page_info info;
1692 seq_printf(m, "Node %d, zone %8s ",
1695 for (order = 0; order < MAX_ORDER; ++order) {
1696 fill_contig_page_info(zone, order, &info);
1697 index = __fragmentation_index(order, &info);
1698 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1705 * Display fragmentation index for orders that allocations would fail for
1707 static int extfrag_show(struct seq_file *m, void *arg)
1709 pg_data_t *pgdat = (pg_data_t *)arg;
1711 walk_zones_in_node(m, pgdat, extfrag_show_print);
1716 static const struct seq_operations extfrag_op = {
1717 .start = frag_start,
1720 .show = extfrag_show,
1723 static int extfrag_open(struct inode *inode, struct file *file)
1725 return seq_open(file, &extfrag_op);
1728 static const struct file_operations extfrag_file_ops = {
1729 .open = extfrag_open,
1731 .llseek = seq_lseek,
1732 .release = seq_release,
1735 static int __init extfrag_debug_init(void)
1737 struct dentry *extfrag_debug_root;
1739 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1740 if (!extfrag_debug_root)
1743 if (!debugfs_create_file("unusable_index", 0444,
1744 extfrag_debug_root, NULL, &unusable_file_ops))
1747 if (!debugfs_create_file("extfrag_index", 0444,
1748 extfrag_debug_root, NULL, &extfrag_file_ops))
1753 debugfs_remove_recursive(extfrag_debug_root);
1757 module_init(extfrag_debug_init);