idle: Remove GENERIC_IDLE_LOOP config switch
[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 #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         "pageoutrun",
774         "allocstall",
775
776         "pgrotated",
777
778 #ifdef CONFIG_NUMA_BALANCING
779         "numa_pte_updates",
780         "numa_hint_faults",
781         "numa_hint_faults_local",
782         "numa_pages_migrated",
783 #endif
784 #ifdef CONFIG_MIGRATION
785         "pgmigrate_success",
786         "pgmigrate_fail",
787 #endif
788 #ifdef CONFIG_COMPACTION
789         "compact_migrate_scanned",
790         "compact_free_scanned",
791         "compact_isolated",
792         "compact_stall",
793         "compact_fail",
794         "compact_success",
795 #endif
796
797 #ifdef CONFIG_HUGETLB_PAGE
798         "htlb_buddy_alloc_success",
799         "htlb_buddy_alloc_fail",
800 #endif
801         "unevictable_pgs_culled",
802         "unevictable_pgs_scanned",
803         "unevictable_pgs_rescued",
804         "unevictable_pgs_mlocked",
805         "unevictable_pgs_munlocked",
806         "unevictable_pgs_cleared",
807         "unevictable_pgs_stranded",
808
809 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
810         "thp_fault_alloc",
811         "thp_fault_fallback",
812         "thp_collapse_alloc",
813         "thp_collapse_alloc_failed",
814         "thp_split",
815         "thp_zero_page_alloc",
816         "thp_zero_page_alloc_failed",
817 #endif
818
819 #endif /* CONFIG_VM_EVENTS_COUNTERS */
820 };
821 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
822
823
824 #ifdef CONFIG_PROC_FS
825 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
826                                                 struct zone *zone)
827 {
828         int order;
829
830         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
831         for (order = 0; order < MAX_ORDER; ++order)
832                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
833         seq_putc(m, '\n');
834 }
835
836 /*
837  * This walks the free areas for each zone.
838  */
839 static int frag_show(struct seq_file *m, void *arg)
840 {
841         pg_data_t *pgdat = (pg_data_t *)arg;
842         walk_zones_in_node(m, pgdat, frag_show_print);
843         return 0;
844 }
845
846 static void pagetypeinfo_showfree_print(struct seq_file *m,
847                                         pg_data_t *pgdat, struct zone *zone)
848 {
849         int order, mtype;
850
851         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
852                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
853                                         pgdat->node_id,
854                                         zone->name,
855                                         migratetype_names[mtype]);
856                 for (order = 0; order < MAX_ORDER; ++order) {
857                         unsigned long freecount = 0;
858                         struct free_area *area;
859                         struct list_head *curr;
860
861                         area = &(zone->free_area[order]);
862
863                         list_for_each(curr, &area->free_list[mtype])
864                                 freecount++;
865                         seq_printf(m, "%6lu ", freecount);
866                 }
867                 seq_putc(m, '\n');
868         }
869 }
870
871 /* Print out the free pages at each order for each migatetype */
872 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
873 {
874         int order;
875         pg_data_t *pgdat = (pg_data_t *)arg;
876
877         /* Print header */
878         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
879         for (order = 0; order < MAX_ORDER; ++order)
880                 seq_printf(m, "%6d ", order);
881         seq_putc(m, '\n');
882
883         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
884
885         return 0;
886 }
887
888 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
889                                         pg_data_t *pgdat, struct zone *zone)
890 {
891         int mtype;
892         unsigned long pfn;
893         unsigned long start_pfn = zone->zone_start_pfn;
894         unsigned long end_pfn = zone_end_pfn(zone);
895         unsigned long count[MIGRATE_TYPES] = { 0, };
896
897         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
898                 struct page *page;
899
900                 if (!pfn_valid(pfn))
901                         continue;
902
903                 page = pfn_to_page(pfn);
904
905                 /* Watch for unexpected holes punched in the memmap */
906                 if (!memmap_valid_within(pfn, page, zone))
907                         continue;
908
909                 mtype = get_pageblock_migratetype(page);
910
911                 if (mtype < MIGRATE_TYPES)
912                         count[mtype]++;
913         }
914
915         /* Print counts */
916         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
917         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
918                 seq_printf(m, "%12lu ", count[mtype]);
919         seq_putc(m, '\n');
920 }
921
922 /* Print out the free pages at each order for each migratetype */
923 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
924 {
925         int mtype;
926         pg_data_t *pgdat = (pg_data_t *)arg;
927
928         seq_printf(m, "\n%-23s", "Number of blocks type ");
929         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
930                 seq_printf(m, "%12s ", migratetype_names[mtype]);
931         seq_putc(m, '\n');
932         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
933
934         return 0;
935 }
936
937 /*
938  * This prints out statistics in relation to grouping pages by mobility.
939  * It is expensive to collect so do not constantly read the file.
940  */
941 static int pagetypeinfo_show(struct seq_file *m, void *arg)
942 {
943         pg_data_t *pgdat = (pg_data_t *)arg;
944
945         /* check memoryless node */
946         if (!node_state(pgdat->node_id, N_MEMORY))
947                 return 0;
948
949         seq_printf(m, "Page block order: %d\n", pageblock_order);
950         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
951         seq_putc(m, '\n');
952         pagetypeinfo_showfree(m, pgdat);
953         pagetypeinfo_showblockcount(m, pgdat);
954
955         return 0;
956 }
957
958 static const struct seq_operations fragmentation_op = {
959         .start  = frag_start,
960         .next   = frag_next,
961         .stop   = frag_stop,
962         .show   = frag_show,
963 };
964
965 static int fragmentation_open(struct inode *inode, struct file *file)
966 {
967         return seq_open(file, &fragmentation_op);
968 }
969
970 static const struct file_operations fragmentation_file_operations = {
971         .open           = fragmentation_open,
972         .read           = seq_read,
973         .llseek         = seq_lseek,
974         .release        = seq_release,
975 };
976
977 static const struct seq_operations pagetypeinfo_op = {
978         .start  = frag_start,
979         .next   = frag_next,
980         .stop   = frag_stop,
981         .show   = pagetypeinfo_show,
982 };
983
984 static int pagetypeinfo_open(struct inode *inode, struct file *file)
985 {
986         return seq_open(file, &pagetypeinfo_op);
987 }
988
989 static const struct file_operations pagetypeinfo_file_ops = {
990         .open           = pagetypeinfo_open,
991         .read           = seq_read,
992         .llseek         = seq_lseek,
993         .release        = seq_release,
994 };
995
996 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
997                                                         struct zone *zone)
998 {
999         int i;
1000         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1001         seq_printf(m,
1002                    "\n  pages free     %lu"
1003                    "\n        min      %lu"
1004                    "\n        low      %lu"
1005                    "\n        high     %lu"
1006                    "\n        scanned  %lu"
1007                    "\n        spanned  %lu"
1008                    "\n        present  %lu"
1009                    "\n        managed  %lu",
1010                    zone_page_state(zone, NR_FREE_PAGES),
1011                    min_wmark_pages(zone),
1012                    low_wmark_pages(zone),
1013                    high_wmark_pages(zone),
1014                    zone->pages_scanned,
1015                    zone->spanned_pages,
1016                    zone->present_pages,
1017                    zone->managed_pages);
1018
1019         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1020                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1021                                 zone_page_state(zone, i));
1022
1023         seq_printf(m,
1024                    "\n        protection: (%lu",
1025                    zone->lowmem_reserve[0]);
1026         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1027                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1028         seq_printf(m,
1029                    ")"
1030                    "\n  pagesets");
1031         for_each_online_cpu(i) {
1032                 struct per_cpu_pageset *pageset;
1033
1034                 pageset = per_cpu_ptr(zone->pageset, i);
1035                 seq_printf(m,
1036                            "\n    cpu: %i"
1037                            "\n              count: %i"
1038                            "\n              high:  %i"
1039                            "\n              batch: %i",
1040                            i,
1041                            pageset->pcp.count,
1042                            pageset->pcp.high,
1043                            pageset->pcp.batch);
1044 #ifdef CONFIG_SMP
1045                 seq_printf(m, "\n  vm stats threshold: %d",
1046                                 pageset->stat_threshold);
1047 #endif
1048         }
1049         seq_printf(m,
1050                    "\n  all_unreclaimable: %u"
1051                    "\n  start_pfn:         %lu"
1052                    "\n  inactive_ratio:    %u",
1053                    zone->all_unreclaimable,
1054                    zone->zone_start_pfn,
1055                    zone->inactive_ratio);
1056         seq_putc(m, '\n');
1057 }
1058
1059 /*
1060  * Output information about zones in @pgdat.
1061  */
1062 static int zoneinfo_show(struct seq_file *m, void *arg)
1063 {
1064         pg_data_t *pgdat = (pg_data_t *)arg;
1065         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1066         return 0;
1067 }
1068
1069 static const struct seq_operations zoneinfo_op = {
1070         .start  = frag_start, /* iterate over all zones. The same as in
1071                                * fragmentation. */
1072         .next   = frag_next,
1073         .stop   = frag_stop,
1074         .show   = zoneinfo_show,
1075 };
1076
1077 static int zoneinfo_open(struct inode *inode, struct file *file)
1078 {
1079         return seq_open(file, &zoneinfo_op);
1080 }
1081
1082 static const struct file_operations proc_zoneinfo_file_operations = {
1083         .open           = zoneinfo_open,
1084         .read           = seq_read,
1085         .llseek         = seq_lseek,
1086         .release        = seq_release,
1087 };
1088
1089 enum writeback_stat_item {
1090         NR_DIRTY_THRESHOLD,
1091         NR_DIRTY_BG_THRESHOLD,
1092         NR_VM_WRITEBACK_STAT_ITEMS,
1093 };
1094
1095 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1096 {
1097         unsigned long *v;
1098         int i, stat_items_size;
1099
1100         if (*pos >= ARRAY_SIZE(vmstat_text))
1101                 return NULL;
1102         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1103                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1104
1105 #ifdef CONFIG_VM_EVENT_COUNTERS
1106         stat_items_size += sizeof(struct vm_event_state);
1107 #endif
1108
1109         v = kmalloc(stat_items_size, GFP_KERNEL);
1110         m->private = v;
1111         if (!v)
1112                 return ERR_PTR(-ENOMEM);
1113         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1114                 v[i] = global_page_state(i);
1115         v += NR_VM_ZONE_STAT_ITEMS;
1116
1117         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1118                             v + NR_DIRTY_THRESHOLD);
1119         v += NR_VM_WRITEBACK_STAT_ITEMS;
1120
1121 #ifdef CONFIG_VM_EVENT_COUNTERS
1122         all_vm_events(v);
1123         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1124         v[PGPGOUT] /= 2;
1125 #endif
1126         return (unsigned long *)m->private + *pos;
1127 }
1128
1129 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1130 {
1131         (*pos)++;
1132         if (*pos >= ARRAY_SIZE(vmstat_text))
1133                 return NULL;
1134         return (unsigned long *)m->private + *pos;
1135 }
1136
1137 static int vmstat_show(struct seq_file *m, void *arg)
1138 {
1139         unsigned long *l = arg;
1140         unsigned long off = l - (unsigned long *)m->private;
1141
1142         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1143         return 0;
1144 }
1145
1146 static void vmstat_stop(struct seq_file *m, void *arg)
1147 {
1148         kfree(m->private);
1149         m->private = NULL;
1150 }
1151
1152 static const struct seq_operations vmstat_op = {
1153         .start  = vmstat_start,
1154         .next   = vmstat_next,
1155         .stop   = vmstat_stop,
1156         .show   = vmstat_show,
1157 };
1158
1159 static int vmstat_open(struct inode *inode, struct file *file)
1160 {
1161         return seq_open(file, &vmstat_op);
1162 }
1163
1164 static const struct file_operations proc_vmstat_file_operations = {
1165         .open           = vmstat_open,
1166         .read           = seq_read,
1167         .llseek         = seq_lseek,
1168         .release        = seq_release,
1169 };
1170 #endif /* CONFIG_PROC_FS */
1171
1172 #ifdef CONFIG_SMP
1173 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1174 int sysctl_stat_interval __read_mostly = HZ;
1175
1176 static void vmstat_update(struct work_struct *w)
1177 {
1178         refresh_cpu_vm_stats(smp_processor_id());
1179         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1180                 round_jiffies_relative(sysctl_stat_interval));
1181 }
1182
1183 static void __cpuinit start_cpu_timer(int cpu)
1184 {
1185         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1186
1187         INIT_DEFERRABLE_WORK(work, vmstat_update);
1188         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1189 }
1190
1191 /*
1192  * Use the cpu notifier to insure that the thresholds are recalculated
1193  * when necessary.
1194  */
1195 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1196                 unsigned long action,
1197                 void *hcpu)
1198 {
1199         long cpu = (long)hcpu;
1200
1201         switch (action) {
1202         case CPU_ONLINE:
1203         case CPU_ONLINE_FROZEN:
1204                 refresh_zone_stat_thresholds();
1205                 start_cpu_timer(cpu);
1206                 node_set_state(cpu_to_node(cpu), N_CPU);
1207                 break;
1208         case CPU_DOWN_PREPARE:
1209         case CPU_DOWN_PREPARE_FROZEN:
1210                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1211                 per_cpu(vmstat_work, cpu).work.func = NULL;
1212                 break;
1213         case CPU_DOWN_FAILED:
1214         case CPU_DOWN_FAILED_FROZEN:
1215                 start_cpu_timer(cpu);
1216                 break;
1217         case CPU_DEAD:
1218         case CPU_DEAD_FROZEN:
1219                 refresh_zone_stat_thresholds();
1220                 break;
1221         default:
1222                 break;
1223         }
1224         return NOTIFY_OK;
1225 }
1226
1227 static struct notifier_block __cpuinitdata vmstat_notifier =
1228         { &vmstat_cpuup_callback, NULL, 0 };
1229 #endif
1230
1231 static int __init setup_vmstat(void)
1232 {
1233 #ifdef CONFIG_SMP
1234         int cpu;
1235
1236         register_cpu_notifier(&vmstat_notifier);
1237
1238         for_each_online_cpu(cpu)
1239                 start_cpu_timer(cpu);
1240 #endif
1241 #ifdef CONFIG_PROC_FS
1242         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1243         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1244         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1245         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1246 #endif
1247         return 0;
1248 }
1249 module_init(setup_vmstat)
1250
1251 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1252 #include <linux/debugfs.h>
1253
1254
1255 /*
1256  * Return an index indicating how much of the available free memory is
1257  * unusable for an allocation of the requested size.
1258  */
1259 static int unusable_free_index(unsigned int order,
1260                                 struct contig_page_info *info)
1261 {
1262         /* No free memory is interpreted as all free memory is unusable */
1263         if (info->free_pages == 0)
1264                 return 1000;
1265
1266         /*
1267          * Index should be a value between 0 and 1. Return a value to 3
1268          * decimal places.
1269          *
1270          * 0 => no fragmentation
1271          * 1 => high fragmentation
1272          */
1273         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1274
1275 }
1276
1277 static void unusable_show_print(struct seq_file *m,
1278                                         pg_data_t *pgdat, struct zone *zone)
1279 {
1280         unsigned int order;
1281         int index;
1282         struct contig_page_info info;
1283
1284         seq_printf(m, "Node %d, zone %8s ",
1285                                 pgdat->node_id,
1286                                 zone->name);
1287         for (order = 0; order < MAX_ORDER; ++order) {
1288                 fill_contig_page_info(zone, order, &info);
1289                 index = unusable_free_index(order, &info);
1290                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1291         }
1292
1293         seq_putc(m, '\n');
1294 }
1295
1296 /*
1297  * Display unusable free space index
1298  *
1299  * The unusable free space index measures how much of the available free
1300  * memory cannot be used to satisfy an allocation of a given size and is a
1301  * value between 0 and 1. The higher the value, the more of free memory is
1302  * unusable and by implication, the worse the external fragmentation is. This
1303  * can be expressed as a percentage by multiplying by 100.
1304  */
1305 static int unusable_show(struct seq_file *m, void *arg)
1306 {
1307         pg_data_t *pgdat = (pg_data_t *)arg;
1308
1309         /* check memoryless node */
1310         if (!node_state(pgdat->node_id, N_MEMORY))
1311                 return 0;
1312
1313         walk_zones_in_node(m, pgdat, unusable_show_print);
1314
1315         return 0;
1316 }
1317
1318 static const struct seq_operations unusable_op = {
1319         .start  = frag_start,
1320         .next   = frag_next,
1321         .stop   = frag_stop,
1322         .show   = unusable_show,
1323 };
1324
1325 static int unusable_open(struct inode *inode, struct file *file)
1326 {
1327         return seq_open(file, &unusable_op);
1328 }
1329
1330 static const struct file_operations unusable_file_ops = {
1331         .open           = unusable_open,
1332         .read           = seq_read,
1333         .llseek         = seq_lseek,
1334         .release        = seq_release,
1335 };
1336
1337 static void extfrag_show_print(struct seq_file *m,
1338                                         pg_data_t *pgdat, struct zone *zone)
1339 {
1340         unsigned int order;
1341         int index;
1342
1343         /* Alloc on stack as interrupts are disabled for zone walk */
1344         struct contig_page_info info;
1345
1346         seq_printf(m, "Node %d, zone %8s ",
1347                                 pgdat->node_id,
1348                                 zone->name);
1349         for (order = 0; order < MAX_ORDER; ++order) {
1350                 fill_contig_page_info(zone, order, &info);
1351                 index = __fragmentation_index(order, &info);
1352                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1353         }
1354
1355         seq_putc(m, '\n');
1356 }
1357
1358 /*
1359  * Display fragmentation index for orders that allocations would fail for
1360  */
1361 static int extfrag_show(struct seq_file *m, void *arg)
1362 {
1363         pg_data_t *pgdat = (pg_data_t *)arg;
1364
1365         walk_zones_in_node(m, pgdat, extfrag_show_print);
1366
1367         return 0;
1368 }
1369
1370 static const struct seq_operations extfrag_op = {
1371         .start  = frag_start,
1372         .next   = frag_next,
1373         .stop   = frag_stop,
1374         .show   = extfrag_show,
1375 };
1376
1377 static int extfrag_open(struct inode *inode, struct file *file)
1378 {
1379         return seq_open(file, &extfrag_op);
1380 }
1381
1382 static const struct file_operations extfrag_file_ops = {
1383         .open           = extfrag_open,
1384         .read           = seq_read,
1385         .llseek         = seq_lseek,
1386         .release        = seq_release,
1387 };
1388
1389 static int __init extfrag_debug_init(void)
1390 {
1391         struct dentry *extfrag_debug_root;
1392
1393         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1394         if (!extfrag_debug_root)
1395                 return -ENOMEM;
1396
1397         if (!debugfs_create_file("unusable_index", 0444,
1398                         extfrag_debug_root, NULL, &unusable_file_ops))
1399                 goto fail;
1400
1401         if (!debugfs_create_file("extfrag_index", 0444,
1402                         extfrag_debug_root, NULL, &extfrag_file_ops))
1403                 goto fail;
1404
1405         return 0;
1406 fail:
1407         debugfs_remove_recursive(extfrag_debug_root);
1408         return -ENOMEM;
1409 }
1410
1411 module_init(extfrag_debug_init);
1412 #endif