c9d1f68120cddebe03ad99ee16c30d4e1c758197
[~shefty/rdma-dev.git] / mm / vmstat.c
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
24 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
25 EXPORT_PER_CPU_SYMBOL(vm_event_states);
26
27 static 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 */
47 void all_vm_events(unsigned long *ret)
48 {
49         get_online_cpus();
50         sum_vm_events(ret);
51         put_online_cpus();
52 }
53 EXPORT_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  */
62 void 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  */
81 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
82 EXPORT_SYMBOL(vm_stat);
83
84 #ifdef CONFIG_SMP
85
86 int 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
110 int 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->managed_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  */
160 void 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
188 void 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  */
211 void __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 }
229 EXPORT_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  */
254 void __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
270 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
271 {
272         __inc_zone_state(page_zone(page), item);
273 }
274 EXPORT_SYMBOL(__inc_zone_page_state);
275
276 void __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
292 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
293 {
294         __dec_zone_state(page_zone(page), item);
295 }
296 EXPORT_SYMBOL(__dec_zone_page_state);
297
298 #ifdef CONFIG_HAVE_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 */
311 static 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
349 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
350                                         int delta)
351 {
352         mod_state(zone, item, delta, 0);
353 }
354 EXPORT_SYMBOL(mod_zone_page_state);
355
356 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
357 {
358         mod_state(zone, item, 1, 1);
359 }
360
361 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
362 {
363         mod_state(page_zone(page), item, 1, 1);
364 }
365 EXPORT_SYMBOL(inc_zone_page_state);
366
367 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
368 {
369         mod_state(page_zone(page), item, -1, -1);
370 }
371 EXPORT_SYMBOL(dec_zone_page_state);
372 #else
373 /*
374  * Use interrupt disable to serialize counter updates
375  */
376 void 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 }
385 EXPORT_SYMBOL(mod_zone_page_state);
386
387 void 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
396 void 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 }
406 EXPORT_SYMBOL(inc_zone_page_state);
407
408 void 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 }
416 EXPORT_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  */
437 void 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 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
499 {
500         int i;
501
502         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
503                 if (pset->vm_stat_diff[i]) {
504                         int v = pset->vm_stat_diff[i];
505                         pset->vm_stat_diff[i] = 0;
506                         atomic_long_add(v, &zone->vm_stat[i]);
507                         atomic_long_add(v, &vm_stat[i]);
508                 }
509 }
510 #endif
511
512 #ifdef CONFIG_NUMA
513 /*
514  * zonelist = the list of zones passed to the allocator
515  * z        = the zone from which the allocation occurred.
516  *
517  * Must be called with interrupts disabled.
518  *
519  * When __GFP_OTHER_NODE is set assume the node of the preferred
520  * zone is the local node. This is useful for daemons who allocate
521  * memory on behalf of other processes.
522  */
523 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
524 {
525         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
526                 __inc_zone_state(z, NUMA_HIT);
527         } else {
528                 __inc_zone_state(z, NUMA_MISS);
529                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
530         }
531         if (z->node == ((flags & __GFP_OTHER_NODE) ?
532                         preferred_zone->node : numa_node_id()))
533                 __inc_zone_state(z, NUMA_LOCAL);
534         else
535                 __inc_zone_state(z, NUMA_OTHER);
536 }
537 #endif
538
539 #ifdef CONFIG_COMPACTION
540
541 struct contig_page_info {
542         unsigned long free_pages;
543         unsigned long free_blocks_total;
544         unsigned long free_blocks_suitable;
545 };
546
547 /*
548  * Calculate the number of free pages in a zone, how many contiguous
549  * pages are free and how many are large enough to satisfy an allocation of
550  * the target size. Note that this function makes no attempt to estimate
551  * how many suitable free blocks there *might* be if MOVABLE pages were
552  * migrated. Calculating that is possible, but expensive and can be
553  * figured out from userspace
554  */
555 static void fill_contig_page_info(struct zone *zone,
556                                 unsigned int suitable_order,
557                                 struct contig_page_info *info)
558 {
559         unsigned int order;
560
561         info->free_pages = 0;
562         info->free_blocks_total = 0;
563         info->free_blocks_suitable = 0;
564
565         for (order = 0; order < MAX_ORDER; order++) {
566                 unsigned long blocks;
567
568                 /* Count number of free blocks */
569                 blocks = zone->free_area[order].nr_free;
570                 info->free_blocks_total += blocks;
571
572                 /* Count free base pages */
573                 info->free_pages += blocks << order;
574
575                 /* Count the suitable free blocks */
576                 if (order >= suitable_order)
577                         info->free_blocks_suitable += blocks <<
578                                                 (order - suitable_order);
579         }
580 }
581
582 /*
583  * A fragmentation index only makes sense if an allocation of a requested
584  * size would fail. If that is true, the fragmentation index indicates
585  * whether external fragmentation or a lack of memory was the problem.
586  * The value can be used to determine if page reclaim or compaction
587  * should be used
588  */
589 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
590 {
591         unsigned long requested = 1UL << order;
592
593         if (!info->free_blocks_total)
594                 return 0;
595
596         /* Fragmentation index only makes sense when a request would fail */
597         if (info->free_blocks_suitable)
598                 return -1000;
599
600         /*
601          * Index is between 0 and 1 so return within 3 decimal places
602          *
603          * 0 => allocation would fail due to lack of memory
604          * 1 => allocation would fail due to fragmentation
605          */
606         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
607 }
608
609 /* Same as __fragmentation index but allocs contig_page_info on stack */
610 int fragmentation_index(struct zone *zone, unsigned int order)
611 {
612         struct contig_page_info info;
613
614         fill_contig_page_info(zone, order, &info);
615         return __fragmentation_index(order, &info);
616 }
617 #endif
618
619 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
620 #include <linux/proc_fs.h>
621 #include <linux/seq_file.h>
622
623 static char * const migratetype_names[MIGRATE_TYPES] = {
624         "Unmovable",
625         "Reclaimable",
626         "Movable",
627         "Reserve",
628 #ifdef CONFIG_CMA
629         "CMA",
630 #endif
631 #ifdef CONFIG_MEMORY_ISOLATION
632         "Isolate",
633 #endif
634 };
635
636 static void *frag_start(struct seq_file *m, loff_t *pos)
637 {
638         pg_data_t *pgdat;
639         loff_t node = *pos;
640         for (pgdat = first_online_pgdat();
641              pgdat && node;
642              pgdat = next_online_pgdat(pgdat))
643                 --node;
644
645         return pgdat;
646 }
647
648 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
649 {
650         pg_data_t *pgdat = (pg_data_t *)arg;
651
652         (*pos)++;
653         return next_online_pgdat(pgdat);
654 }
655
656 static void frag_stop(struct seq_file *m, void *arg)
657 {
658 }
659
660 /* Walk all the zones in a node and print using a callback */
661 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
662                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
663 {
664         struct zone *zone;
665         struct zone *node_zones = pgdat->node_zones;
666         unsigned long flags;
667
668         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
669                 if (!populated_zone(zone))
670                         continue;
671
672                 spin_lock_irqsave(&zone->lock, flags);
673                 print(m, pgdat, zone);
674                 spin_unlock_irqrestore(&zone->lock, flags);
675         }
676 }
677 #endif
678
679 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
680 #ifdef CONFIG_ZONE_DMA
681 #define TEXT_FOR_DMA(xx) xx "_dma",
682 #else
683 #define TEXT_FOR_DMA(xx)
684 #endif
685
686 #ifdef CONFIG_ZONE_DMA32
687 #define TEXT_FOR_DMA32(xx) xx "_dma32",
688 #else
689 #define TEXT_FOR_DMA32(xx)
690 #endif
691
692 #ifdef CONFIG_HIGHMEM
693 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
694 #else
695 #define TEXT_FOR_HIGHMEM(xx)
696 #endif
697
698 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
699                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
700
701 const char * const vmstat_text[] = {
702         /* Zoned VM counters */
703         "nr_free_pages",
704         "nr_inactive_anon",
705         "nr_active_anon",
706         "nr_inactive_file",
707         "nr_active_file",
708         "nr_unevictable",
709         "nr_mlock",
710         "nr_anon_pages",
711         "nr_mapped",
712         "nr_file_pages",
713         "nr_dirty",
714         "nr_writeback",
715         "nr_slab_reclaimable",
716         "nr_slab_unreclaimable",
717         "nr_page_table_pages",
718         "nr_kernel_stack",
719         "nr_unstable",
720         "nr_bounce",
721         "nr_vmscan_write",
722         "nr_vmscan_immediate_reclaim",
723         "nr_writeback_temp",
724         "nr_isolated_anon",
725         "nr_isolated_file",
726         "nr_shmem",
727         "nr_dirtied",
728         "nr_written",
729
730 #ifdef CONFIG_NUMA
731         "numa_hit",
732         "numa_miss",
733         "numa_foreign",
734         "numa_interleave",
735         "numa_local",
736         "numa_other",
737 #endif
738         "nr_anon_transparent_hugepages",
739         "nr_free_cma",
740         "nr_dirty_threshold",
741         "nr_dirty_background_threshold",
742
743 #ifdef CONFIG_VM_EVENT_COUNTERS
744         "pgpgin",
745         "pgpgout",
746         "pswpin",
747         "pswpout",
748
749         TEXTS_FOR_ZONES("pgalloc")
750
751         "pgfree",
752         "pgactivate",
753         "pgdeactivate",
754
755         "pgfault",
756         "pgmajfault",
757
758         TEXTS_FOR_ZONES("pgrefill")
759         TEXTS_FOR_ZONES("pgsteal_kswapd")
760         TEXTS_FOR_ZONES("pgsteal_direct")
761         TEXTS_FOR_ZONES("pgscan_kswapd")
762         TEXTS_FOR_ZONES("pgscan_direct")
763         "pgscan_direct_throttle",
764
765 #ifdef CONFIG_NUMA
766         "zone_reclaim_failed",
767 #endif
768         "pginodesteal",
769         "slabs_scanned",
770         "kswapd_inodesteal",
771         "kswapd_low_wmark_hit_quickly",
772         "kswapd_high_wmark_hit_quickly",
773         "kswapd_skip_congestion_wait",
774         "pageoutrun",
775         "allocstall",
776
777         "pgrotated",
778
779 #ifdef CONFIG_NUMA_BALANCING
780         "numa_pte_updates",
781         "numa_hint_faults",
782         "numa_hint_faults_local",
783         "numa_pages_migrated",
784 #endif
785 #ifdef CONFIG_MIGRATION
786         "pgmigrate_success",
787         "pgmigrate_fail",
788 #endif
789 #ifdef CONFIG_COMPACTION
790         "compact_migrate_scanned",
791         "compact_free_scanned",
792         "compact_isolated",
793         "compact_stall",
794         "compact_fail",
795         "compact_success",
796 #endif
797
798 #ifdef CONFIG_HUGETLB_PAGE
799         "htlb_buddy_alloc_success",
800         "htlb_buddy_alloc_fail",
801 #endif
802         "unevictable_pgs_culled",
803         "unevictable_pgs_scanned",
804         "unevictable_pgs_rescued",
805         "unevictable_pgs_mlocked",
806         "unevictable_pgs_munlocked",
807         "unevictable_pgs_cleared",
808         "unevictable_pgs_stranded",
809
810 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
811         "thp_fault_alloc",
812         "thp_fault_fallback",
813         "thp_collapse_alloc",
814         "thp_collapse_alloc_failed",
815         "thp_split",
816         "thp_zero_page_alloc",
817         "thp_zero_page_alloc_failed",
818 #endif
819
820 #endif /* CONFIG_VM_EVENTS_COUNTERS */
821 };
822 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
823
824
825 #ifdef CONFIG_PROC_FS
826 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
827                                                 struct zone *zone)
828 {
829         int order;
830
831         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
832         for (order = 0; order < MAX_ORDER; ++order)
833                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
834         seq_putc(m, '\n');
835 }
836
837 /*
838  * This walks the free areas for each zone.
839  */
840 static int frag_show(struct seq_file *m, void *arg)
841 {
842         pg_data_t *pgdat = (pg_data_t *)arg;
843         walk_zones_in_node(m, pgdat, frag_show_print);
844         return 0;
845 }
846
847 static void pagetypeinfo_showfree_print(struct seq_file *m,
848                                         pg_data_t *pgdat, struct zone *zone)
849 {
850         int order, mtype;
851
852         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
853                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
854                                         pgdat->node_id,
855                                         zone->name,
856                                         migratetype_names[mtype]);
857                 for (order = 0; order < MAX_ORDER; ++order) {
858                         unsigned long freecount = 0;
859                         struct free_area *area;
860                         struct list_head *curr;
861
862                         area = &(zone->free_area[order]);
863
864                         list_for_each(curr, &area->free_list[mtype])
865                                 freecount++;
866                         seq_printf(m, "%6lu ", freecount);
867                 }
868                 seq_putc(m, '\n');
869         }
870 }
871
872 /* Print out the free pages at each order for each migatetype */
873 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
874 {
875         int order;
876         pg_data_t *pgdat = (pg_data_t *)arg;
877
878         /* Print header */
879         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
880         for (order = 0; order < MAX_ORDER; ++order)
881                 seq_printf(m, "%6d ", order);
882         seq_putc(m, '\n');
883
884         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
885
886         return 0;
887 }
888
889 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
890                                         pg_data_t *pgdat, struct zone *zone)
891 {
892         int mtype;
893         unsigned long pfn;
894         unsigned long start_pfn = zone->zone_start_pfn;
895         unsigned long end_pfn = start_pfn + zone->spanned_pages;
896         unsigned long count[MIGRATE_TYPES] = { 0, };
897
898         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
899                 struct page *page;
900
901                 if (!pfn_valid(pfn))
902                         continue;
903
904                 page = pfn_to_page(pfn);
905
906                 /* Watch for unexpected holes punched in the memmap */
907                 if (!memmap_valid_within(pfn, page, zone))
908                         continue;
909
910                 mtype = get_pageblock_migratetype(page);
911
912                 if (mtype < MIGRATE_TYPES)
913                         count[mtype]++;
914         }
915
916         /* Print counts */
917         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
918         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
919                 seq_printf(m, "%12lu ", count[mtype]);
920         seq_putc(m, '\n');
921 }
922
923 /* Print out the free pages at each order for each migratetype */
924 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
925 {
926         int mtype;
927         pg_data_t *pgdat = (pg_data_t *)arg;
928
929         seq_printf(m, "\n%-23s", "Number of blocks type ");
930         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
931                 seq_printf(m, "%12s ", migratetype_names[mtype]);
932         seq_putc(m, '\n');
933         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
934
935         return 0;
936 }
937
938 /*
939  * This prints out statistics in relation to grouping pages by mobility.
940  * It is expensive to collect so do not constantly read the file.
941  */
942 static int pagetypeinfo_show(struct seq_file *m, void *arg)
943 {
944         pg_data_t *pgdat = (pg_data_t *)arg;
945
946         /* check memoryless node */
947         if (!node_state(pgdat->node_id, N_MEMORY))
948                 return 0;
949
950         seq_printf(m, "Page block order: %d\n", pageblock_order);
951         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
952         seq_putc(m, '\n');
953         pagetypeinfo_showfree(m, pgdat);
954         pagetypeinfo_showblockcount(m, pgdat);
955
956         return 0;
957 }
958
959 static const struct seq_operations fragmentation_op = {
960         .start  = frag_start,
961         .next   = frag_next,
962         .stop   = frag_stop,
963         .show   = frag_show,
964 };
965
966 static int fragmentation_open(struct inode *inode, struct file *file)
967 {
968         return seq_open(file, &fragmentation_op);
969 }
970
971 static const struct file_operations fragmentation_file_operations = {
972         .open           = fragmentation_open,
973         .read           = seq_read,
974         .llseek         = seq_lseek,
975         .release        = seq_release,
976 };
977
978 static const struct seq_operations pagetypeinfo_op = {
979         .start  = frag_start,
980         .next   = frag_next,
981         .stop   = frag_stop,
982         .show   = pagetypeinfo_show,
983 };
984
985 static int pagetypeinfo_open(struct inode *inode, struct file *file)
986 {
987         return seq_open(file, &pagetypeinfo_op);
988 }
989
990 static const struct file_operations pagetypeinfo_file_ops = {
991         .open           = pagetypeinfo_open,
992         .read           = seq_read,
993         .llseek         = seq_lseek,
994         .release        = seq_release,
995 };
996
997 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
998                                                         struct zone *zone)
999 {
1000         int i;
1001         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1002         seq_printf(m,
1003                    "\n  pages free     %lu"
1004                    "\n        min      %lu"
1005                    "\n        low      %lu"
1006                    "\n        high     %lu"
1007                    "\n        scanned  %lu"
1008                    "\n        spanned  %lu"
1009                    "\n        present  %lu"
1010                    "\n        managed  %lu",
1011                    zone_page_state(zone, NR_FREE_PAGES),
1012                    min_wmark_pages(zone),
1013                    low_wmark_pages(zone),
1014                    high_wmark_pages(zone),
1015                    zone->pages_scanned,
1016                    zone->spanned_pages,
1017                    zone->present_pages,
1018                    zone->managed_pages);
1019
1020         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1021                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1022                                 zone_page_state(zone, i));
1023
1024         seq_printf(m,
1025                    "\n        protection: (%lu",
1026                    zone->lowmem_reserve[0]);
1027         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1028                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1029         seq_printf(m,
1030                    ")"
1031                    "\n  pagesets");
1032         for_each_online_cpu(i) {
1033                 struct per_cpu_pageset *pageset;
1034
1035                 pageset = per_cpu_ptr(zone->pageset, i);
1036                 seq_printf(m,
1037                            "\n    cpu: %i"
1038                            "\n              count: %i"
1039                            "\n              high:  %i"
1040                            "\n              batch: %i",
1041                            i,
1042                            pageset->pcp.count,
1043                            pageset->pcp.high,
1044                            pageset->pcp.batch);
1045 #ifdef CONFIG_SMP
1046                 seq_printf(m, "\n  vm stats threshold: %d",
1047                                 pageset->stat_threshold);
1048 #endif
1049         }
1050         seq_printf(m,
1051                    "\n  all_unreclaimable: %u"
1052                    "\n  start_pfn:         %lu"
1053                    "\n  inactive_ratio:    %u",
1054                    zone->all_unreclaimable,
1055                    zone->zone_start_pfn,
1056                    zone->inactive_ratio);
1057         seq_putc(m, '\n');
1058 }
1059
1060 /*
1061  * Output information about zones in @pgdat.
1062  */
1063 static int zoneinfo_show(struct seq_file *m, void *arg)
1064 {
1065         pg_data_t *pgdat = (pg_data_t *)arg;
1066         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1067         return 0;
1068 }
1069
1070 static const struct seq_operations zoneinfo_op = {
1071         .start  = frag_start, /* iterate over all zones. The same as in
1072                                * fragmentation. */
1073         .next   = frag_next,
1074         .stop   = frag_stop,
1075         .show   = zoneinfo_show,
1076 };
1077
1078 static int zoneinfo_open(struct inode *inode, struct file *file)
1079 {
1080         return seq_open(file, &zoneinfo_op);
1081 }
1082
1083 static const struct file_operations proc_zoneinfo_file_operations = {
1084         .open           = zoneinfo_open,
1085         .read           = seq_read,
1086         .llseek         = seq_lseek,
1087         .release        = seq_release,
1088 };
1089
1090 enum writeback_stat_item {
1091         NR_DIRTY_THRESHOLD,
1092         NR_DIRTY_BG_THRESHOLD,
1093         NR_VM_WRITEBACK_STAT_ITEMS,
1094 };
1095
1096 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1097 {
1098         unsigned long *v;
1099         int i, stat_items_size;
1100
1101         if (*pos >= ARRAY_SIZE(vmstat_text))
1102                 return NULL;
1103         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1104                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1105
1106 #ifdef CONFIG_VM_EVENT_COUNTERS
1107         stat_items_size += sizeof(struct vm_event_state);
1108 #endif
1109
1110         v = kmalloc(stat_items_size, GFP_KERNEL);
1111         m->private = v;
1112         if (!v)
1113                 return ERR_PTR(-ENOMEM);
1114         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1115                 v[i] = global_page_state(i);
1116         v += NR_VM_ZONE_STAT_ITEMS;
1117
1118         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1119                             v + NR_DIRTY_THRESHOLD);
1120         v += NR_VM_WRITEBACK_STAT_ITEMS;
1121
1122 #ifdef CONFIG_VM_EVENT_COUNTERS
1123         all_vm_events(v);
1124         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1125         v[PGPGOUT] /= 2;
1126 #endif
1127         return (unsigned long *)m->private + *pos;
1128 }
1129
1130 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1131 {
1132         (*pos)++;
1133         if (*pos >= ARRAY_SIZE(vmstat_text))
1134                 return NULL;
1135         return (unsigned long *)m->private + *pos;
1136 }
1137
1138 static int vmstat_show(struct seq_file *m, void *arg)
1139 {
1140         unsigned long *l = arg;
1141         unsigned long off = l - (unsigned long *)m->private;
1142
1143         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1144         return 0;
1145 }
1146
1147 static void vmstat_stop(struct seq_file *m, void *arg)
1148 {
1149         kfree(m->private);
1150         m->private = NULL;
1151 }
1152
1153 static const struct seq_operations vmstat_op = {
1154         .start  = vmstat_start,
1155         .next   = vmstat_next,
1156         .stop   = vmstat_stop,
1157         .show   = vmstat_show,
1158 };
1159
1160 static int vmstat_open(struct inode *inode, struct file *file)
1161 {
1162         return seq_open(file, &vmstat_op);
1163 }
1164
1165 static const struct file_operations proc_vmstat_file_operations = {
1166         .open           = vmstat_open,
1167         .read           = seq_read,
1168         .llseek         = seq_lseek,
1169         .release        = seq_release,
1170 };
1171 #endif /* CONFIG_PROC_FS */
1172
1173 #ifdef CONFIG_SMP
1174 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1175 int sysctl_stat_interval __read_mostly = HZ;
1176
1177 static void vmstat_update(struct work_struct *w)
1178 {
1179         refresh_cpu_vm_stats(smp_processor_id());
1180         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1181                 round_jiffies_relative(sysctl_stat_interval));
1182 }
1183
1184 static void __cpuinit start_cpu_timer(int cpu)
1185 {
1186         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1187
1188         INIT_DEFERRABLE_WORK(work, vmstat_update);
1189         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1190 }
1191
1192 /*
1193  * Use the cpu notifier to insure that the thresholds are recalculated
1194  * when necessary.
1195  */
1196 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1197                 unsigned long action,
1198                 void *hcpu)
1199 {
1200         long cpu = (long)hcpu;
1201
1202         switch (action) {
1203         case CPU_ONLINE:
1204         case CPU_ONLINE_FROZEN:
1205                 refresh_zone_stat_thresholds();
1206                 start_cpu_timer(cpu);
1207                 node_set_state(cpu_to_node(cpu), N_CPU);
1208                 break;
1209         case CPU_DOWN_PREPARE:
1210         case CPU_DOWN_PREPARE_FROZEN:
1211                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1212                 per_cpu(vmstat_work, cpu).work.func = NULL;
1213                 break;
1214         case CPU_DOWN_FAILED:
1215         case CPU_DOWN_FAILED_FROZEN:
1216                 start_cpu_timer(cpu);
1217                 break;
1218         case CPU_DEAD:
1219         case CPU_DEAD_FROZEN:
1220                 refresh_zone_stat_thresholds();
1221                 break;
1222         default:
1223                 break;
1224         }
1225         return NOTIFY_OK;
1226 }
1227
1228 static struct notifier_block __cpuinitdata vmstat_notifier =
1229         { &vmstat_cpuup_callback, NULL, 0 };
1230 #endif
1231
1232 static int __init setup_vmstat(void)
1233 {
1234 #ifdef CONFIG_SMP
1235         int cpu;
1236
1237         register_cpu_notifier(&vmstat_notifier);
1238
1239         for_each_online_cpu(cpu)
1240                 start_cpu_timer(cpu);
1241 #endif
1242 #ifdef CONFIG_PROC_FS
1243         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1244         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1245         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1246         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1247 #endif
1248         return 0;
1249 }
1250 module_init(setup_vmstat)
1251
1252 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1253 #include <linux/debugfs.h>
1254
1255
1256 /*
1257  * Return an index indicating how much of the available free memory is
1258  * unusable for an allocation of the requested size.
1259  */
1260 static int unusable_free_index(unsigned int order,
1261                                 struct contig_page_info *info)
1262 {
1263         /* No free memory is interpreted as all free memory is unusable */
1264         if (info->free_pages == 0)
1265                 return 1000;
1266
1267         /*
1268          * Index should be a value between 0 and 1. Return a value to 3
1269          * decimal places.
1270          *
1271          * 0 => no fragmentation
1272          * 1 => high fragmentation
1273          */
1274         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1275
1276 }
1277
1278 static void unusable_show_print(struct seq_file *m,
1279                                         pg_data_t *pgdat, struct zone *zone)
1280 {
1281         unsigned int order;
1282         int index;
1283         struct contig_page_info info;
1284
1285         seq_printf(m, "Node %d, zone %8s ",
1286                                 pgdat->node_id,
1287                                 zone->name);
1288         for (order = 0; order < MAX_ORDER; ++order) {
1289                 fill_contig_page_info(zone, order, &info);
1290                 index = unusable_free_index(order, &info);
1291                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1292         }
1293
1294         seq_putc(m, '\n');
1295 }
1296
1297 /*
1298  * Display unusable free space index
1299  *
1300  * The unusable free space index measures how much of the available free
1301  * memory cannot be used to satisfy an allocation of a given size and is a
1302  * value between 0 and 1. The higher the value, the more of free memory is
1303  * unusable and by implication, the worse the external fragmentation is. This
1304  * can be expressed as a percentage by multiplying by 100.
1305  */
1306 static int unusable_show(struct seq_file *m, void *arg)
1307 {
1308         pg_data_t *pgdat = (pg_data_t *)arg;
1309
1310         /* check memoryless node */
1311         if (!node_state(pgdat->node_id, N_MEMORY))
1312                 return 0;
1313
1314         walk_zones_in_node(m, pgdat, unusable_show_print);
1315
1316         return 0;
1317 }
1318
1319 static const struct seq_operations unusable_op = {
1320         .start  = frag_start,
1321         .next   = frag_next,
1322         .stop   = frag_stop,
1323         .show   = unusable_show,
1324 };
1325
1326 static int unusable_open(struct inode *inode, struct file *file)
1327 {
1328         return seq_open(file, &unusable_op);
1329 }
1330
1331 static const struct file_operations unusable_file_ops = {
1332         .open           = unusable_open,
1333         .read           = seq_read,
1334         .llseek         = seq_lseek,
1335         .release        = seq_release,
1336 };
1337
1338 static void extfrag_show_print(struct seq_file *m,
1339                                         pg_data_t *pgdat, struct zone *zone)
1340 {
1341         unsigned int order;
1342         int index;
1343
1344         /* Alloc on stack as interrupts are disabled for zone walk */
1345         struct contig_page_info info;
1346
1347         seq_printf(m, "Node %d, zone %8s ",
1348                                 pgdat->node_id,
1349                                 zone->name);
1350         for (order = 0; order < MAX_ORDER; ++order) {
1351                 fill_contig_page_info(zone, order, &info);
1352                 index = __fragmentation_index(order, &info);
1353                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1354         }
1355
1356         seq_putc(m, '\n');
1357 }
1358
1359 /*
1360  * Display fragmentation index for orders that allocations would fail for
1361  */
1362 static int extfrag_show(struct seq_file *m, void *arg)
1363 {
1364         pg_data_t *pgdat = (pg_data_t *)arg;
1365
1366         walk_zones_in_node(m, pgdat, extfrag_show_print);
1367
1368         return 0;
1369 }
1370
1371 static const struct seq_operations extfrag_op = {
1372         .start  = frag_start,
1373         .next   = frag_next,
1374         .stop   = frag_stop,
1375         .show   = extfrag_show,
1376 };
1377
1378 static int extfrag_open(struct inode *inode, struct file *file)
1379 {
1380         return seq_open(file, &extfrag_op);
1381 }
1382
1383 static const struct file_operations extfrag_file_ops = {
1384         .open           = extfrag_open,
1385         .read           = seq_read,
1386         .llseek         = seq_lseek,
1387         .release        = seq_release,
1388 };
1389
1390 static int __init extfrag_debug_init(void)
1391 {
1392         struct dentry *extfrag_debug_root;
1393
1394         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1395         if (!extfrag_debug_root)
1396                 return -ENOMEM;
1397
1398         if (!debugfs_create_file("unusable_index", 0444,
1399                         extfrag_debug_root, NULL, &unusable_file_ops))
1400                 goto fail;
1401
1402         if (!debugfs_create_file("extfrag_index", 0444,
1403                         extfrag_debug_root, NULL, &extfrag_file_ops))
1404                 goto fail;
1405
1406         return 0;
1407 fail:
1408         debugfs_remove_recursive(extfrag_debug_root);
1409         return -ENOMEM;
1410 }
1411
1412 module_init(extfrag_debug_init);
1413 #endif