mm: Close races between THP migration and PMD numa clearing
[linux-3.10.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 /*
56  * Fold the foreign cpu events into our own.
57  *
58  * This is adding to the events on one processor
59  * but keeps the global counts constant.
60  */
61 void vm_events_fold_cpu(int cpu)
62 {
63         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
64         int i;
65
66         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
67                 count_vm_events(i, fold_state->event[i]);
68                 fold_state->event[i] = 0;
69         }
70 }
71
72 #endif /* CONFIG_VM_EVENT_COUNTERS */
73
74 /*
75  * Manage combined zone based / global counters
76  *
77  * vm_stat contains the global counters
78  */
79 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
80 EXPORT_SYMBOL(vm_stat);
81
82 #ifdef CONFIG_SMP
83
84 int calculate_pressure_threshold(struct zone *zone)
85 {
86         int threshold;
87         int watermark_distance;
88
89         /*
90          * As vmstats are not up to date, there is drift between the estimated
91          * and real values. For high thresholds and a high number of CPUs, it
92          * is possible for the min watermark to be breached while the estimated
93          * value looks fine. The pressure threshold is a reduced value such
94          * that even the maximum amount of drift will not accidentally breach
95          * the min watermark
96          */
97         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
98         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
99
100         /*
101          * Maximum threshold is 125
102          */
103         threshold = min(125, threshold);
104
105         return threshold;
106 }
107
108 int calculate_normal_threshold(struct zone *zone)
109 {
110         int threshold;
111         int mem;        /* memory in 128 MB units */
112
113         /*
114          * The threshold scales with the number of processors and the amount
115          * of memory per zone. More memory means that we can defer updates for
116          * longer, more processors could lead to more contention.
117          * fls() is used to have a cheap way of logarithmic scaling.
118          *
119          * Some sample thresholds:
120          *
121          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
122          * ------------------------------------------------------------------
123          * 8            1               1       0.9-1 GB        4
124          * 16           2               2       0.9-1 GB        4
125          * 20           2               2       1-2 GB          5
126          * 24           2               2       2-4 GB          6
127          * 28           2               2       4-8 GB          7
128          * 32           2               2       8-16 GB         8
129          * 4            2               2       <128M           1
130          * 30           4               3       2-4 GB          5
131          * 48           4               3       8-16 GB         8
132          * 32           8               4       1-2 GB          4
133          * 32           8               4       0.9-1GB         4
134          * 10           16              5       <128M           1
135          * 40           16              5       900M            4
136          * 70           64              7       2-4 GB          5
137          * 84           64              7       4-8 GB          6
138          * 108          512             9       4-8 GB          6
139          * 125          1024            10      8-16 GB         8
140          * 125          1024            10      16-32 GB        9
141          */
142
143         mem = zone->managed_pages >> (27 - PAGE_SHIFT);
144
145         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
146
147         /*
148          * Maximum threshold is 125
149          */
150         threshold = min(125, threshold);
151
152         return threshold;
153 }
154
155 /*
156  * Refresh the thresholds for each zone.
157  */
158 void refresh_zone_stat_thresholds(void)
159 {
160         struct zone *zone;
161         int cpu;
162         int threshold;
163
164         for_each_populated_zone(zone) {
165                 unsigned long max_drift, tolerate_drift;
166
167                 threshold = calculate_normal_threshold(zone);
168
169                 for_each_online_cpu(cpu)
170                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
171                                                         = threshold;
172
173                 /*
174                  * Only set percpu_drift_mark if there is a danger that
175                  * NR_FREE_PAGES reports the low watermark is ok when in fact
176                  * the min watermark could be breached by an allocation
177                  */
178                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
179                 max_drift = num_online_cpus() * threshold;
180                 if (max_drift > tolerate_drift)
181                         zone->percpu_drift_mark = high_wmark_pages(zone) +
182                                         max_drift;
183         }
184 }
185
186 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
187                                 int (*calculate_pressure)(struct zone *))
188 {
189         struct zone *zone;
190         int cpu;
191         int threshold;
192         int i;
193
194         for (i = 0; i < pgdat->nr_zones; i++) {
195                 zone = &pgdat->node_zones[i];
196                 if (!zone->percpu_drift_mark)
197                         continue;
198
199                 threshold = (*calculate_pressure)(zone);
200                 for_each_possible_cpu(cpu)
201                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
202                                                         = threshold;
203         }
204 }
205
206 /*
207  * For use when we know that interrupts are disabled.
208  */
209 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
210                                 int delta)
211 {
212         struct per_cpu_pageset __percpu *pcp = zone->pageset;
213         s8 __percpu *p = pcp->vm_stat_diff + item;
214         long x;
215         long t;
216
217         x = delta + __this_cpu_read(*p);
218
219         t = __this_cpu_read(pcp->stat_threshold);
220
221         if (unlikely(x > t || x < -t)) {
222                 zone_page_state_add(x, zone, item);
223                 x = 0;
224         }
225         __this_cpu_write(*p, x);
226 }
227 EXPORT_SYMBOL(__mod_zone_page_state);
228
229 /*
230  * Optimized increment and decrement functions.
231  *
232  * These are only for a single page and therefore can take a struct page *
233  * argument instead of struct zone *. This allows the inclusion of the code
234  * generated for page_zone(page) into the optimized functions.
235  *
236  * No overflow check is necessary and therefore the differential can be
237  * incremented or decremented in place which may allow the compilers to
238  * generate better code.
239  * The increment or decrement is known and therefore one boundary check can
240  * be omitted.
241  *
242  * NOTE: These functions are very performance sensitive. Change only
243  * with care.
244  *
245  * Some processors have inc/dec instructions that are atomic vs an interrupt.
246  * However, the code must first determine the differential location in a zone
247  * based on the processor number and then inc/dec the counter. There is no
248  * guarantee without disabling preemption that the processor will not change
249  * in between and therefore the atomicity vs. interrupt cannot be exploited
250  * in a useful way here.
251  */
252 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
253 {
254         struct per_cpu_pageset __percpu *pcp = zone->pageset;
255         s8 __percpu *p = pcp->vm_stat_diff + item;
256         s8 v, t;
257
258         v = __this_cpu_inc_return(*p);
259         t = __this_cpu_read(pcp->stat_threshold);
260         if (unlikely(v > t)) {
261                 s8 overstep = t >> 1;
262
263                 zone_page_state_add(v + overstep, zone, item);
264                 __this_cpu_write(*p, -overstep);
265         }
266 }
267
268 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
269 {
270         __inc_zone_state(page_zone(page), item);
271 }
272 EXPORT_SYMBOL(__inc_zone_page_state);
273
274 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
275 {
276         struct per_cpu_pageset __percpu *pcp = zone->pageset;
277         s8 __percpu *p = pcp->vm_stat_diff + item;
278         s8 v, t;
279
280         v = __this_cpu_dec_return(*p);
281         t = __this_cpu_read(pcp->stat_threshold);
282         if (unlikely(v < - t)) {
283                 s8 overstep = t >> 1;
284
285                 zone_page_state_add(v - overstep, zone, item);
286                 __this_cpu_write(*p, overstep);
287         }
288 }
289
290 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
291 {
292         __dec_zone_state(page_zone(page), item);
293 }
294 EXPORT_SYMBOL(__dec_zone_page_state);
295
296 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
297 /*
298  * If we have cmpxchg_local support then we do not need to incur the overhead
299  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
300  *
301  * mod_state() modifies the zone counter state through atomic per cpu
302  * operations.
303  *
304  * Overstep mode specifies how overstep should handled:
305  *     0       No overstepping
306  *     1       Overstepping half of threshold
307  *     -1      Overstepping minus half of threshold
308 */
309 static inline void mod_state(struct zone *zone,
310        enum zone_stat_item item, int delta, int overstep_mode)
311 {
312         struct per_cpu_pageset __percpu *pcp = zone->pageset;
313         s8 __percpu *p = pcp->vm_stat_diff + item;
314         long o, n, t, z;
315
316         do {
317                 z = 0;  /* overflow to zone counters */
318
319                 /*
320                  * The fetching of the stat_threshold is racy. We may apply
321                  * a counter threshold to the wrong the cpu if we get
322                  * rescheduled while executing here. However, the next
323                  * counter update will apply the threshold again and
324                  * therefore bring the counter under the threshold again.
325                  *
326                  * Most of the time the thresholds are the same anyways
327                  * for all cpus in a zone.
328                  */
329                 t = this_cpu_read(pcp->stat_threshold);
330
331                 o = this_cpu_read(*p);
332                 n = delta + o;
333
334                 if (n > t || n < -t) {
335                         int os = overstep_mode * (t >> 1) ;
336
337                         /* Overflow must be added to zone counters */
338                         z = n + os;
339                         n = -os;
340                 }
341         } while (this_cpu_cmpxchg(*p, o, n) != o);
342
343         if (z)
344                 zone_page_state_add(z, zone, item);
345 }
346
347 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
348                                         int delta)
349 {
350         mod_state(zone, item, delta, 0);
351 }
352 EXPORT_SYMBOL(mod_zone_page_state);
353
354 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
355 {
356         mod_state(zone, item, 1, 1);
357 }
358
359 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
360 {
361         mod_state(page_zone(page), item, 1, 1);
362 }
363 EXPORT_SYMBOL(inc_zone_page_state);
364
365 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
366 {
367         mod_state(page_zone(page), item, -1, -1);
368 }
369 EXPORT_SYMBOL(dec_zone_page_state);
370 #else
371 /*
372  * Use interrupt disable to serialize counter updates
373  */
374 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
375                                         int delta)
376 {
377         unsigned long flags;
378
379         local_irq_save(flags);
380         __mod_zone_page_state(zone, item, delta);
381         local_irq_restore(flags);
382 }
383 EXPORT_SYMBOL(mod_zone_page_state);
384
385 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
386 {
387         unsigned long flags;
388
389         local_irq_save(flags);
390         __inc_zone_state(zone, item);
391         local_irq_restore(flags);
392 }
393
394 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
395 {
396         unsigned long flags;
397         struct zone *zone;
398
399         zone = page_zone(page);
400         local_irq_save(flags);
401         __inc_zone_state(zone, item);
402         local_irq_restore(flags);
403 }
404 EXPORT_SYMBOL(inc_zone_page_state);
405
406 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
407 {
408         unsigned long flags;
409
410         local_irq_save(flags);
411         __dec_zone_page_state(page, item);
412         local_irq_restore(flags);
413 }
414 EXPORT_SYMBOL(dec_zone_page_state);
415 #endif
416
417 /*
418  * Update the zone counters for one cpu.
419  *
420  * The cpu specified must be either the current cpu or a processor that
421  * is not online. If it is the current cpu then the execution thread must
422  * be pinned to the current cpu.
423  *
424  * Note that refresh_cpu_vm_stats strives to only access
425  * node local memory. The per cpu pagesets on remote zones are placed
426  * in the memory local to the processor using that pageset. So the
427  * loop over all zones will access a series of cachelines local to
428  * the processor.
429  *
430  * The call to zone_page_state_add updates the cachelines with the
431  * statistics in the remote zone struct as well as the global cachelines
432  * with the global counters. These could cause remote node cache line
433  * bouncing and will have to be only done when necessary.
434  */
435 void refresh_cpu_vm_stats(int cpu)
436 {
437         struct zone *zone;
438         int i;
439         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
440
441         for_each_populated_zone(zone) {
442                 struct per_cpu_pageset *p;
443
444                 p = per_cpu_ptr(zone->pageset, cpu);
445
446                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
447                         if (p->vm_stat_diff[i]) {
448                                 unsigned long flags;
449                                 int v;
450
451                                 local_irq_save(flags);
452                                 v = p->vm_stat_diff[i];
453                                 p->vm_stat_diff[i] = 0;
454                                 local_irq_restore(flags);
455                                 atomic_long_add(v, &zone->vm_stat[i]);
456                                 global_diff[i] += v;
457 #ifdef CONFIG_NUMA
458                                 /* 3 seconds idle till flush */
459                                 p->expire = 3;
460 #endif
461                         }
462                 cond_resched();
463 #ifdef CONFIG_NUMA
464                 /*
465                  * Deal with draining the remote pageset of this
466                  * processor
467                  *
468                  * Check if there are pages remaining in this pageset
469                  * if not then there is nothing to expire.
470                  */
471                 if (!p->expire || !p->pcp.count)
472                         continue;
473
474                 /*
475                  * We never drain zones local to this processor.
476                  */
477                 if (zone_to_nid(zone) == numa_node_id()) {
478                         p->expire = 0;
479                         continue;
480                 }
481
482                 p->expire--;
483                 if (p->expire)
484                         continue;
485
486                 if (p->pcp.count)
487                         drain_zone_pages(zone, &p->pcp);
488 #endif
489         }
490
491         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
492                 if (global_diff[i])
493                         atomic_long_add(global_diff[i], &vm_stat[i]);
494 }
495
496 /*
497  * this is only called if !populated_zone(zone), which implies no other users of
498  * pset->vm_stat_diff[] exsist.
499  */
500 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
501 {
502         int i;
503
504         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
505                 if (pset->vm_stat_diff[i]) {
506                         int v = pset->vm_stat_diff[i];
507                         pset->vm_stat_diff[i] = 0;
508                         atomic_long_add(v, &zone->vm_stat[i]);
509                         atomic_long_add(v, &vm_stat[i]);
510                 }
511 }
512 #endif
513
514 #ifdef CONFIG_NUMA
515 /*
516  * zonelist = the list of zones passed to the allocator
517  * z        = the zone from which the allocation occurred.
518  *
519  * Must be called with interrupts disabled.
520  *
521  * When __GFP_OTHER_NODE is set assume the node of the preferred
522  * zone is the local node. This is useful for daemons who allocate
523  * memory on behalf of other processes.
524  */
525 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
526 {
527         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
528                 __inc_zone_state(z, NUMA_HIT);
529         } else {
530                 __inc_zone_state(z, NUMA_MISS);
531                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
532         }
533         if (z->node == ((flags & __GFP_OTHER_NODE) ?
534                         preferred_zone->node : numa_node_id()))
535                 __inc_zone_state(z, NUMA_LOCAL);
536         else
537                 __inc_zone_state(z, NUMA_OTHER);
538 }
539 #endif
540
541 #ifdef CONFIG_COMPACTION
542
543 struct contig_page_info {
544         unsigned long free_pages;
545         unsigned long free_blocks_total;
546         unsigned long free_blocks_suitable;
547 };
548
549 /*
550  * Calculate the number of free pages in a zone, how many contiguous
551  * pages are free and how many are large enough to satisfy an allocation of
552  * the target size. Note that this function makes no attempt to estimate
553  * how many suitable free blocks there *might* be if MOVABLE pages were
554  * migrated. Calculating that is possible, but expensive and can be
555  * figured out from userspace
556  */
557 static void fill_contig_page_info(struct zone *zone,
558                                 unsigned int suitable_order,
559                                 struct contig_page_info *info)
560 {
561         unsigned int order;
562
563         info->free_pages = 0;
564         info->free_blocks_total = 0;
565         info->free_blocks_suitable = 0;
566
567         for (order = 0; order < MAX_ORDER; order++) {
568                 unsigned long blocks;
569
570                 /* Count number of free blocks */
571                 blocks = zone->free_area[order].nr_free;
572                 info->free_blocks_total += blocks;
573
574                 /* Count free base pages */
575                 info->free_pages += blocks << order;
576
577                 /* Count the suitable free blocks */
578                 if (order >= suitable_order)
579                         info->free_blocks_suitable += blocks <<
580                                                 (order - suitable_order);
581         }
582 }
583
584 /*
585  * A fragmentation index only makes sense if an allocation of a requested
586  * size would fail. If that is true, the fragmentation index indicates
587  * whether external fragmentation or a lack of memory was the problem.
588  * The value can be used to determine if page reclaim or compaction
589  * should be used
590  */
591 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
592 {
593         unsigned long requested = 1UL << order;
594
595         if (!info->free_blocks_total)
596                 return 0;
597
598         /* Fragmentation index only makes sense when a request would fail */
599         if (info->free_blocks_suitable)
600                 return -1000;
601
602         /*
603          * Index is between 0 and 1 so return within 3 decimal places
604          *
605          * 0 => allocation would fail due to lack of memory
606          * 1 => allocation would fail due to fragmentation
607          */
608         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
609 }
610
611 /* Same as __fragmentation index but allocs contig_page_info on stack */
612 int fragmentation_index(struct zone *zone, unsigned int order)
613 {
614         struct contig_page_info info;
615
616         fill_contig_page_info(zone, order, &info);
617         return __fragmentation_index(order, &info);
618 }
619 #endif
620
621 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
622 #include <linux/proc_fs.h>
623 #include <linux/seq_file.h>
624
625 static char * const migratetype_names[MIGRATE_TYPES] = {
626         "Unmovable",
627         "Reclaimable",
628         "Movable",
629         "Reserve",
630 #ifdef CONFIG_CMA
631         "CMA",
632 #endif
633 #ifdef CONFIG_MEMORY_ISOLATION
634         "Isolate",
635 #endif
636 };
637
638 static void *frag_start(struct seq_file *m, loff_t *pos)
639 {
640         pg_data_t *pgdat;
641         loff_t node = *pos;
642         for (pgdat = first_online_pgdat();
643              pgdat && node;
644              pgdat = next_online_pgdat(pgdat))
645                 --node;
646
647         return pgdat;
648 }
649
650 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
651 {
652         pg_data_t *pgdat = (pg_data_t *)arg;
653
654         (*pos)++;
655         return next_online_pgdat(pgdat);
656 }
657
658 static void frag_stop(struct seq_file *m, void *arg)
659 {
660 }
661
662 /* Walk all the zones in a node and print using a callback */
663 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
664                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
665 {
666         struct zone *zone;
667         struct zone *node_zones = pgdat->node_zones;
668         unsigned long flags;
669
670         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
671                 if (!populated_zone(zone))
672                         continue;
673
674                 spin_lock_irqsave(&zone->lock, flags);
675                 print(m, pgdat, zone);
676                 spin_unlock_irqrestore(&zone->lock, flags);
677         }
678 }
679 #endif
680
681 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
682 #ifdef CONFIG_ZONE_DMA
683 #define TEXT_FOR_DMA(xx) xx "_dma",
684 #else
685 #define TEXT_FOR_DMA(xx)
686 #endif
687
688 #ifdef CONFIG_ZONE_DMA32
689 #define TEXT_FOR_DMA32(xx) xx "_dma32",
690 #else
691 #define TEXT_FOR_DMA32(xx)
692 #endif
693
694 #ifdef CONFIG_HIGHMEM
695 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
696 #else
697 #define TEXT_FOR_HIGHMEM(xx)
698 #endif
699
700 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
701                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
702
703 const char * const vmstat_text[] = {
704         /* Zoned VM counters */
705         "nr_free_pages",
706         "nr_inactive_anon",
707         "nr_active_anon",
708         "nr_inactive_file",
709         "nr_active_file",
710         "nr_unevictable",
711         "nr_mlock",
712         "nr_anon_pages",
713         "nr_mapped",
714         "nr_file_pages",
715         "nr_dirty",
716         "nr_writeback",
717         "nr_slab_reclaimable",
718         "nr_slab_unreclaimable",
719         "nr_page_table_pages",
720         "nr_kernel_stack",
721         "nr_unstable",
722         "nr_bounce",
723         "nr_vmscan_write",
724         "nr_vmscan_immediate_reclaim",
725         "nr_writeback_temp",
726         "nr_isolated_anon",
727         "nr_isolated_file",
728         "nr_shmem",
729         "nr_dirtied",
730         "nr_written",
731
732 #ifdef CONFIG_NUMA
733         "numa_hit",
734         "numa_miss",
735         "numa_foreign",
736         "numa_interleave",
737         "numa_local",
738         "numa_other",
739 #endif
740         "nr_anon_transparent_hugepages",
741         "nr_free_cma",
742         "nr_dirty_threshold",
743         "nr_dirty_background_threshold",
744
745 #ifdef CONFIG_VM_EVENT_COUNTERS
746         "pgpgin",
747         "pgpgout",
748         "pswpin",
749         "pswpout",
750
751         TEXTS_FOR_ZONES("pgalloc")
752
753         "pgfree",
754         "pgactivate",
755         "pgdeactivate",
756
757         "pgfault",
758         "pgmajfault",
759
760         TEXTS_FOR_ZONES("pgrefill")
761         TEXTS_FOR_ZONES("pgsteal_kswapd")
762         TEXTS_FOR_ZONES("pgsteal_direct")
763         TEXTS_FOR_ZONES("pgscan_kswapd")
764         TEXTS_FOR_ZONES("pgscan_direct")
765         "pgscan_direct_throttle",
766
767 #ifdef CONFIG_NUMA
768         "zone_reclaim_failed",
769 #endif
770         "pginodesteal",
771         "slabs_scanned",
772         "kswapd_inodesteal",
773         "kswapd_low_wmark_hit_quickly",
774         "kswapd_high_wmark_hit_quickly",
775         "pageoutrun",
776         "allocstall",
777
778         "pgrotated",
779
780 #ifdef CONFIG_NUMA_BALANCING
781         "numa_pte_updates",
782         "numa_hint_faults",
783         "numa_hint_faults_local",
784         "numa_pages_migrated",
785 #endif
786 #ifdef CONFIG_MIGRATION
787         "pgmigrate_success",
788         "pgmigrate_fail",
789 #endif
790 #ifdef CONFIG_COMPACTION
791         "compact_migrate_scanned",
792         "compact_free_scanned",
793         "compact_isolated",
794         "compact_stall",
795         "compact_fail",
796         "compact_success",
797 #endif
798
799 #ifdef CONFIG_HUGETLB_PAGE
800         "htlb_buddy_alloc_success",
801         "htlb_buddy_alloc_fail",
802 #endif
803         "unevictable_pgs_culled",
804         "unevictable_pgs_scanned",
805         "unevictable_pgs_rescued",
806         "unevictable_pgs_mlocked",
807         "unevictable_pgs_munlocked",
808         "unevictable_pgs_cleared",
809         "unevictable_pgs_stranded",
810
811 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
812         "thp_fault_alloc",
813         "thp_fault_fallback",
814         "thp_collapse_alloc",
815         "thp_collapse_alloc_failed",
816         "thp_split",
817         "thp_zero_page_alloc",
818         "thp_zero_page_alloc_failed",
819 #endif
820
821 #endif /* CONFIG_VM_EVENTS_COUNTERS */
822 };
823 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
824
825
826 #ifdef CONFIG_PROC_FS
827 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
828                                                 struct zone *zone)
829 {
830         int order;
831
832         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
833         for (order = 0; order < MAX_ORDER; ++order)
834                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
835         seq_putc(m, '\n');
836 }
837
838 /*
839  * This walks the free areas for each zone.
840  */
841 static int frag_show(struct seq_file *m, void *arg)
842 {
843         pg_data_t *pgdat = (pg_data_t *)arg;
844         walk_zones_in_node(m, pgdat, frag_show_print);
845         return 0;
846 }
847
848 static void pagetypeinfo_showfree_print(struct seq_file *m,
849                                         pg_data_t *pgdat, struct zone *zone)
850 {
851         int order, mtype;
852
853         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
854                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
855                                         pgdat->node_id,
856                                         zone->name,
857                                         migratetype_names[mtype]);
858                 for (order = 0; order < MAX_ORDER; ++order) {
859                         unsigned long freecount = 0;
860                         struct free_area *area;
861                         struct list_head *curr;
862
863                         area = &(zone->free_area[order]);
864
865                         list_for_each(curr, &area->free_list[mtype])
866                                 freecount++;
867                         seq_printf(m, "%6lu ", freecount);
868                 }
869                 seq_putc(m, '\n');
870         }
871 }
872
873 /* Print out the free pages at each order for each migatetype */
874 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
875 {
876         int order;
877         pg_data_t *pgdat = (pg_data_t *)arg;
878
879         /* Print header */
880         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
881         for (order = 0; order < MAX_ORDER; ++order)
882                 seq_printf(m, "%6d ", order);
883         seq_putc(m, '\n');
884
885         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
886
887         return 0;
888 }
889
890 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
891                                         pg_data_t *pgdat, struct zone *zone)
892 {
893         int mtype;
894         unsigned long pfn;
895         unsigned long start_pfn = zone->zone_start_pfn;
896         unsigned long end_pfn = zone_end_pfn(zone);
897         unsigned long count[MIGRATE_TYPES] = { 0, };
898
899         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
900                 struct page *page;
901
902                 if (!pfn_valid(pfn))
903                         continue;
904
905                 page = pfn_to_page(pfn);
906
907                 /* Watch for unexpected holes punched in the memmap */
908                 if (!memmap_valid_within(pfn, page, zone))
909                         continue;
910
911                 mtype = get_pageblock_migratetype(page);
912
913                 if (mtype < MIGRATE_TYPES)
914                         count[mtype]++;
915         }
916
917         /* Print counts */
918         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
919         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
920                 seq_printf(m, "%12lu ", count[mtype]);
921         seq_putc(m, '\n');
922 }
923
924 /* Print out the free pages at each order for each migratetype */
925 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
926 {
927         int mtype;
928         pg_data_t *pgdat = (pg_data_t *)arg;
929
930         seq_printf(m, "\n%-23s", "Number of blocks type ");
931         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
932                 seq_printf(m, "%12s ", migratetype_names[mtype]);
933         seq_putc(m, '\n');
934         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
935
936         return 0;
937 }
938
939 /*
940  * This prints out statistics in relation to grouping pages by mobility.
941  * It is expensive to collect so do not constantly read the file.
942  */
943 static int pagetypeinfo_show(struct seq_file *m, void *arg)
944 {
945         pg_data_t *pgdat = (pg_data_t *)arg;
946
947         /* check memoryless node */
948         if (!node_state(pgdat->node_id, N_MEMORY))
949                 return 0;
950
951         seq_printf(m, "Page block order: %d\n", pageblock_order);
952         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
953         seq_putc(m, '\n');
954         pagetypeinfo_showfree(m, pgdat);
955         pagetypeinfo_showblockcount(m, pgdat);
956
957         return 0;
958 }
959
960 static const struct seq_operations fragmentation_op = {
961         .start  = frag_start,
962         .next   = frag_next,
963         .stop   = frag_stop,
964         .show   = frag_show,
965 };
966
967 static int fragmentation_open(struct inode *inode, struct file *file)
968 {
969         return seq_open(file, &fragmentation_op);
970 }
971
972 static const struct file_operations fragmentation_file_operations = {
973         .open           = fragmentation_open,
974         .read           = seq_read,
975         .llseek         = seq_lseek,
976         .release        = seq_release,
977 };
978
979 static const struct seq_operations pagetypeinfo_op = {
980         .start  = frag_start,
981         .next   = frag_next,
982         .stop   = frag_stop,
983         .show   = pagetypeinfo_show,
984 };
985
986 static int pagetypeinfo_open(struct inode *inode, struct file *file)
987 {
988         return seq_open(file, &pagetypeinfo_op);
989 }
990
991 static const struct file_operations pagetypeinfo_file_ops = {
992         .open           = pagetypeinfo_open,
993         .read           = seq_read,
994         .llseek         = seq_lseek,
995         .release        = seq_release,
996 };
997
998 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
999                                                         struct zone *zone)
1000 {
1001         int i;
1002         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1003         seq_printf(m,
1004                    "\n  pages free     %lu"
1005                    "\n        min      %lu"
1006                    "\n        low      %lu"
1007                    "\n        high     %lu"
1008                    "\n        scanned  %lu"
1009                    "\n        spanned  %lu"
1010                    "\n        present  %lu"
1011                    "\n        managed  %lu",
1012                    zone_page_state(zone, NR_FREE_PAGES),
1013                    min_wmark_pages(zone),
1014                    low_wmark_pages(zone),
1015                    high_wmark_pages(zone),
1016                    zone->pages_scanned,
1017                    zone->spanned_pages,
1018                    zone->present_pages,
1019                    zone->managed_pages);
1020
1021         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1022                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1023                                 zone_page_state(zone, i));
1024
1025         seq_printf(m,
1026                    "\n        protection: (%lu",
1027                    zone->lowmem_reserve[0]);
1028         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1029                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1030         seq_printf(m,
1031                    ")"
1032                    "\n  pagesets");
1033         for_each_online_cpu(i) {
1034                 struct per_cpu_pageset *pageset;
1035
1036                 pageset = per_cpu_ptr(zone->pageset, i);
1037                 seq_printf(m,
1038                            "\n    cpu: %i"
1039                            "\n              count: %i"
1040                            "\n              high:  %i"
1041                            "\n              batch: %i",
1042                            i,
1043                            pageset->pcp.count,
1044                            pageset->pcp.high,
1045                            pageset->pcp.batch);
1046 #ifdef CONFIG_SMP
1047                 seq_printf(m, "\n  vm stats threshold: %d",
1048                                 pageset->stat_threshold);
1049 #endif
1050         }
1051         seq_printf(m,
1052                    "\n  all_unreclaimable: %u"
1053                    "\n  start_pfn:         %lu"
1054                    "\n  inactive_ratio:    %u",
1055                    zone->all_unreclaimable,
1056                    zone->zone_start_pfn,
1057                    zone->inactive_ratio);
1058         seq_putc(m, '\n');
1059 }
1060
1061 /*
1062  * Output information about zones in @pgdat.
1063  */
1064 static int zoneinfo_show(struct seq_file *m, void *arg)
1065 {
1066         pg_data_t *pgdat = (pg_data_t *)arg;
1067         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1068         return 0;
1069 }
1070
1071 static const struct seq_operations zoneinfo_op = {
1072         .start  = frag_start, /* iterate over all zones. The same as in
1073                                * fragmentation. */
1074         .next   = frag_next,
1075         .stop   = frag_stop,
1076         .show   = zoneinfo_show,
1077 };
1078
1079 static int zoneinfo_open(struct inode *inode, struct file *file)
1080 {
1081         return seq_open(file, &zoneinfo_op);
1082 }
1083
1084 static const struct file_operations proc_zoneinfo_file_operations = {
1085         .open           = zoneinfo_open,
1086         .read           = seq_read,
1087         .llseek         = seq_lseek,
1088         .release        = seq_release,
1089 };
1090
1091 enum writeback_stat_item {
1092         NR_DIRTY_THRESHOLD,
1093         NR_DIRTY_BG_THRESHOLD,
1094         NR_VM_WRITEBACK_STAT_ITEMS,
1095 };
1096
1097 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1098 {
1099         unsigned long *v;
1100         int i, stat_items_size;
1101
1102         if (*pos >= ARRAY_SIZE(vmstat_text))
1103                 return NULL;
1104         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1105                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1106
1107 #ifdef CONFIG_VM_EVENT_COUNTERS
1108         stat_items_size += sizeof(struct vm_event_state);
1109 #endif
1110
1111         v = kmalloc(stat_items_size, GFP_KERNEL);
1112         m->private = v;
1113         if (!v)
1114                 return ERR_PTR(-ENOMEM);
1115         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1116                 v[i] = global_page_state(i);
1117         v += NR_VM_ZONE_STAT_ITEMS;
1118
1119         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1120                             v + NR_DIRTY_THRESHOLD);
1121         v += NR_VM_WRITEBACK_STAT_ITEMS;
1122
1123 #ifdef CONFIG_VM_EVENT_COUNTERS
1124         all_vm_events(v);
1125         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1126         v[PGPGOUT] /= 2;
1127 #endif
1128         return (unsigned long *)m->private + *pos;
1129 }
1130
1131 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1132 {
1133         (*pos)++;
1134         if (*pos >= ARRAY_SIZE(vmstat_text))
1135                 return NULL;
1136         return (unsigned long *)m->private + *pos;
1137 }
1138
1139 static int vmstat_show(struct seq_file *m, void *arg)
1140 {
1141         unsigned long *l = arg;
1142         unsigned long off = l - (unsigned long *)m->private;
1143
1144         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1145         return 0;
1146 }
1147
1148 static void vmstat_stop(struct seq_file *m, void *arg)
1149 {
1150         kfree(m->private);
1151         m->private = NULL;
1152 }
1153
1154 static const struct seq_operations vmstat_op = {
1155         .start  = vmstat_start,
1156         .next   = vmstat_next,
1157         .stop   = vmstat_stop,
1158         .show   = vmstat_show,
1159 };
1160
1161 static int vmstat_open(struct inode *inode, struct file *file)
1162 {
1163         return seq_open(file, &vmstat_op);
1164 }
1165
1166 static const struct file_operations proc_vmstat_file_operations = {
1167         .open           = vmstat_open,
1168         .read           = seq_read,
1169         .llseek         = seq_lseek,
1170         .release        = seq_release,
1171 };
1172 #endif /* CONFIG_PROC_FS */
1173
1174 #ifdef CONFIG_SMP
1175 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1176 int sysctl_stat_interval __read_mostly = HZ;
1177
1178 static void vmstat_update(struct work_struct *w)
1179 {
1180         refresh_cpu_vm_stats(smp_processor_id());
1181         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1182                 round_jiffies_relative(sysctl_stat_interval));
1183 }
1184
1185 static void __cpuinit start_cpu_timer(int cpu)
1186 {
1187         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1188
1189         INIT_DEFERRABLE_WORK(work, vmstat_update);
1190         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1191 }
1192
1193 /*
1194  * Use the cpu notifier to insure that the thresholds are recalculated
1195  * when necessary.
1196  */
1197 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1198                 unsigned long action,
1199                 void *hcpu)
1200 {
1201         long cpu = (long)hcpu;
1202
1203         switch (action) {
1204         case CPU_ONLINE:
1205         case CPU_ONLINE_FROZEN:
1206                 refresh_zone_stat_thresholds();
1207                 start_cpu_timer(cpu);
1208                 node_set_state(cpu_to_node(cpu), N_CPU);
1209                 break;
1210         case CPU_DOWN_PREPARE:
1211         case CPU_DOWN_PREPARE_FROZEN:
1212                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1213                 per_cpu(vmstat_work, cpu).work.func = NULL;
1214                 break;
1215         case CPU_DOWN_FAILED:
1216         case CPU_DOWN_FAILED_FROZEN:
1217                 start_cpu_timer(cpu);
1218                 break;
1219         case CPU_DEAD:
1220         case CPU_DEAD_FROZEN:
1221                 refresh_zone_stat_thresholds();
1222                 break;
1223         default:
1224                 break;
1225         }
1226         return NOTIFY_OK;
1227 }
1228
1229 static struct notifier_block __cpuinitdata vmstat_notifier =
1230         { &vmstat_cpuup_callback, NULL, 0 };
1231 #endif
1232
1233 static int __init setup_vmstat(void)
1234 {
1235 #ifdef CONFIG_SMP
1236         int cpu;
1237
1238         register_cpu_notifier(&vmstat_notifier);
1239
1240         for_each_online_cpu(cpu)
1241                 start_cpu_timer(cpu);
1242 #endif
1243 #ifdef CONFIG_PROC_FS
1244         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1245         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1246         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1247         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1248 #endif
1249         return 0;
1250 }
1251 module_init(setup_vmstat)
1252
1253 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1254 #include <linux/debugfs.h>
1255
1256
1257 /*
1258  * Return an index indicating how much of the available free memory is
1259  * unusable for an allocation of the requested size.
1260  */
1261 static int unusable_free_index(unsigned int order,
1262                                 struct contig_page_info *info)
1263 {
1264         /* No free memory is interpreted as all free memory is unusable */
1265         if (info->free_pages == 0)
1266                 return 1000;
1267
1268         /*
1269          * Index should be a value between 0 and 1. Return a value to 3
1270          * decimal places.
1271          *
1272          * 0 => no fragmentation
1273          * 1 => high fragmentation
1274          */
1275         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1276
1277 }
1278
1279 static void unusable_show_print(struct seq_file *m,
1280                                         pg_data_t *pgdat, struct zone *zone)
1281 {
1282         unsigned int order;
1283         int index;
1284         struct contig_page_info info;
1285
1286         seq_printf(m, "Node %d, zone %8s ",
1287                                 pgdat->node_id,
1288                                 zone->name);
1289         for (order = 0; order < MAX_ORDER; ++order) {
1290                 fill_contig_page_info(zone, order, &info);
1291                 index = unusable_free_index(order, &info);
1292                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1293         }
1294
1295         seq_putc(m, '\n');
1296 }
1297
1298 /*
1299  * Display unusable free space index
1300  *
1301  * The unusable free space index measures how much of the available free
1302  * memory cannot be used to satisfy an allocation of a given size and is a
1303  * value between 0 and 1. The higher the value, the more of free memory is
1304  * unusable and by implication, the worse the external fragmentation is. This
1305  * can be expressed as a percentage by multiplying by 100.
1306  */
1307 static int unusable_show(struct seq_file *m, void *arg)
1308 {
1309         pg_data_t *pgdat = (pg_data_t *)arg;
1310
1311         /* check memoryless node */
1312         if (!node_state(pgdat->node_id, N_MEMORY))
1313                 return 0;
1314
1315         walk_zones_in_node(m, pgdat, unusable_show_print);
1316
1317         return 0;
1318 }
1319
1320 static const struct seq_operations unusable_op = {
1321         .start  = frag_start,
1322         .next   = frag_next,
1323         .stop   = frag_stop,
1324         .show   = unusable_show,
1325 };
1326
1327 static int unusable_open(struct inode *inode, struct file *file)
1328 {
1329         return seq_open(file, &unusable_op);
1330 }
1331
1332 static const struct file_operations unusable_file_ops = {
1333         .open           = unusable_open,
1334         .read           = seq_read,
1335         .llseek         = seq_lseek,
1336         .release        = seq_release,
1337 };
1338
1339 static void extfrag_show_print(struct seq_file *m,
1340                                         pg_data_t *pgdat, struct zone *zone)
1341 {
1342         unsigned int order;
1343         int index;
1344
1345         /* Alloc on stack as interrupts are disabled for zone walk */
1346         struct contig_page_info info;
1347
1348         seq_printf(m, "Node %d, zone %8s ",
1349                                 pgdat->node_id,
1350                                 zone->name);
1351         for (order = 0; order < MAX_ORDER; ++order) {
1352                 fill_contig_page_info(zone, order, &info);
1353                 index = __fragmentation_index(order, &info);
1354                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1355         }
1356
1357         seq_putc(m, '\n');
1358 }
1359
1360 /*
1361  * Display fragmentation index for orders that allocations would fail for
1362  */
1363 static int extfrag_show(struct seq_file *m, void *arg)
1364 {
1365         pg_data_t *pgdat = (pg_data_t *)arg;
1366
1367         walk_zones_in_node(m, pgdat, extfrag_show_print);
1368
1369         return 0;
1370 }
1371
1372 static const struct seq_operations extfrag_op = {
1373         .start  = frag_start,
1374         .next   = frag_next,
1375         .stop   = frag_stop,
1376         .show   = extfrag_show,
1377 };
1378
1379 static int extfrag_open(struct inode *inode, struct file *file)
1380 {
1381         return seq_open(file, &extfrag_op);
1382 }
1383
1384 static const struct file_operations extfrag_file_ops = {
1385         .open           = extfrag_open,
1386         .read           = seq_read,
1387         .llseek         = seq_lseek,
1388         .release        = seq_release,
1389 };
1390
1391 static int __init extfrag_debug_init(void)
1392 {
1393         struct dentry *extfrag_debug_root;
1394
1395         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1396         if (!extfrag_debug_root)
1397                 return -ENOMEM;
1398
1399         if (!debugfs_create_file("unusable_index", 0444,
1400                         extfrag_debug_root, NULL, &unusable_file_ops))
1401                 goto fail;
1402
1403         if (!debugfs_create_file("extfrag_index", 0444,
1404                         extfrag_debug_root, NULL, &extfrag_file_ops))
1405                 goto fail;
1406
1407         return 0;
1408 fail:
1409         debugfs_remove_recursive(extfrag_debug_root);
1410         return -ENOMEM;
1411 }
1412
1413 module_init(extfrag_debug_init);
1414 #endif