dm ioctl: use kmalloc if possible
[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->present_pages >> (27 - PAGE_SHIFT);
146
147         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
148
149         /*
150          * Maximum threshold is 125
151          */
152         threshold = min(125, threshold);
153
154         return threshold;
155 }
156
157 /*
158  * Refresh the thresholds for each zone.
159  */
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         "Isolate",
632 };
633
634 static void *frag_start(struct seq_file *m, loff_t *pos)
635 {
636         pg_data_t *pgdat;
637         loff_t node = *pos;
638         for (pgdat = first_online_pgdat();
639              pgdat && node;
640              pgdat = next_online_pgdat(pgdat))
641                 --node;
642
643         return pgdat;
644 }
645
646 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
647 {
648         pg_data_t *pgdat = (pg_data_t *)arg;
649
650         (*pos)++;
651         return next_online_pgdat(pgdat);
652 }
653
654 static void frag_stop(struct seq_file *m, void *arg)
655 {
656 }
657
658 /* Walk all the zones in a node and print using a callback */
659 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
660                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
661 {
662         struct zone *zone;
663         struct zone *node_zones = pgdat->node_zones;
664         unsigned long flags;
665
666         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
667                 if (!populated_zone(zone))
668                         continue;
669
670                 spin_lock_irqsave(&zone->lock, flags);
671                 print(m, pgdat, zone);
672                 spin_unlock_irqrestore(&zone->lock, flags);
673         }
674 }
675 #endif
676
677 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
678 #ifdef CONFIG_ZONE_DMA
679 #define TEXT_FOR_DMA(xx) xx "_dma",
680 #else
681 #define TEXT_FOR_DMA(xx)
682 #endif
683
684 #ifdef CONFIG_ZONE_DMA32
685 #define TEXT_FOR_DMA32(xx) xx "_dma32",
686 #else
687 #define TEXT_FOR_DMA32(xx)
688 #endif
689
690 #ifdef CONFIG_HIGHMEM
691 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
692 #else
693 #define TEXT_FOR_HIGHMEM(xx)
694 #endif
695
696 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
697                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
698
699 const char * const vmstat_text[] = {
700         /* Zoned VM counters */
701         "nr_free_pages",
702         "nr_inactive_anon",
703         "nr_active_anon",
704         "nr_inactive_file",
705         "nr_active_file",
706         "nr_unevictable",
707         "nr_mlock",
708         "nr_anon_pages",
709         "nr_mapped",
710         "nr_file_pages",
711         "nr_dirty",
712         "nr_writeback",
713         "nr_slab_reclaimable",
714         "nr_slab_unreclaimable",
715         "nr_page_table_pages",
716         "nr_kernel_stack",
717         "nr_unstable",
718         "nr_bounce",
719         "nr_vmscan_write",
720         "nr_vmscan_immediate_reclaim",
721         "nr_writeback_temp",
722         "nr_isolated_anon",
723         "nr_isolated_file",
724         "nr_shmem",
725         "nr_dirtied",
726         "nr_written",
727
728 #ifdef CONFIG_NUMA
729         "numa_hit",
730         "numa_miss",
731         "numa_foreign",
732         "numa_interleave",
733         "numa_local",
734         "numa_other",
735 #endif
736         "nr_anon_transparent_hugepages",
737         "nr_free_cma",
738         "nr_dirty_threshold",
739         "nr_dirty_background_threshold",
740
741 #ifdef CONFIG_VM_EVENT_COUNTERS
742         "pgpgin",
743         "pgpgout",
744         "pswpin",
745         "pswpout",
746
747         TEXTS_FOR_ZONES("pgalloc")
748
749         "pgfree",
750         "pgactivate",
751         "pgdeactivate",
752
753         "pgfault",
754         "pgmajfault",
755
756         TEXTS_FOR_ZONES("pgrefill")
757         TEXTS_FOR_ZONES("pgsteal_kswapd")
758         TEXTS_FOR_ZONES("pgsteal_direct")
759         TEXTS_FOR_ZONES("pgscan_kswapd")
760         TEXTS_FOR_ZONES("pgscan_direct")
761         "pgscan_direct_throttle",
762
763 #ifdef CONFIG_NUMA
764         "zone_reclaim_failed",
765 #endif
766         "pginodesteal",
767         "slabs_scanned",
768         "kswapd_inodesteal",
769         "kswapd_low_wmark_hit_quickly",
770         "kswapd_high_wmark_hit_quickly",
771         "kswapd_skip_congestion_wait",
772         "pageoutrun",
773         "allocstall",
774
775         "pgrotated",
776
777 #ifdef CONFIG_COMPACTION
778         "compact_blocks_moved",
779         "compact_pages_moved",
780         "compact_pagemigrate_failed",
781         "compact_stall",
782         "compact_fail",
783         "compact_success",
784 #endif
785
786 #ifdef CONFIG_HUGETLB_PAGE
787         "htlb_buddy_alloc_success",
788         "htlb_buddy_alloc_fail",
789 #endif
790         "unevictable_pgs_culled",
791         "unevictable_pgs_scanned",
792         "unevictable_pgs_rescued",
793         "unevictable_pgs_mlocked",
794         "unevictable_pgs_munlocked",
795         "unevictable_pgs_cleared",
796         "unevictable_pgs_stranded",
797
798 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
799         "thp_fault_alloc",
800         "thp_fault_fallback",
801         "thp_collapse_alloc",
802         "thp_collapse_alloc_failed",
803         "thp_split",
804 #endif
805
806 #endif /* CONFIG_VM_EVENTS_COUNTERS */
807 };
808 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
809
810
811 #ifdef CONFIG_PROC_FS
812 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
813                                                 struct zone *zone)
814 {
815         int order;
816
817         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
818         for (order = 0; order < MAX_ORDER; ++order)
819                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
820         seq_putc(m, '\n');
821 }
822
823 /*
824  * This walks the free areas for each zone.
825  */
826 static int frag_show(struct seq_file *m, void *arg)
827 {
828         pg_data_t *pgdat = (pg_data_t *)arg;
829         walk_zones_in_node(m, pgdat, frag_show_print);
830         return 0;
831 }
832
833 static void pagetypeinfo_showfree_print(struct seq_file *m,
834                                         pg_data_t *pgdat, struct zone *zone)
835 {
836         int order, mtype;
837
838         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
839                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
840                                         pgdat->node_id,
841                                         zone->name,
842                                         migratetype_names[mtype]);
843                 for (order = 0; order < MAX_ORDER; ++order) {
844                         unsigned long freecount = 0;
845                         struct free_area *area;
846                         struct list_head *curr;
847
848                         area = &(zone->free_area[order]);
849
850                         list_for_each(curr, &area->free_list[mtype])
851                                 freecount++;
852                         seq_printf(m, "%6lu ", freecount);
853                 }
854                 seq_putc(m, '\n');
855         }
856 }
857
858 /* Print out the free pages at each order for each migatetype */
859 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
860 {
861         int order;
862         pg_data_t *pgdat = (pg_data_t *)arg;
863
864         /* Print header */
865         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
866         for (order = 0; order < MAX_ORDER; ++order)
867                 seq_printf(m, "%6d ", order);
868         seq_putc(m, '\n');
869
870         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
871
872         return 0;
873 }
874
875 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
876                                         pg_data_t *pgdat, struct zone *zone)
877 {
878         int mtype;
879         unsigned long pfn;
880         unsigned long start_pfn = zone->zone_start_pfn;
881         unsigned long end_pfn = start_pfn + zone->spanned_pages;
882         unsigned long count[MIGRATE_TYPES] = { 0, };
883
884         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
885                 struct page *page;
886
887                 if (!pfn_valid(pfn))
888                         continue;
889
890                 page = pfn_to_page(pfn);
891
892                 /* Watch for unexpected holes punched in the memmap */
893                 if (!memmap_valid_within(pfn, page, zone))
894                         continue;
895
896                 mtype = get_pageblock_migratetype(page);
897
898                 if (mtype < MIGRATE_TYPES)
899                         count[mtype]++;
900         }
901
902         /* Print counts */
903         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
904         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
905                 seq_printf(m, "%12lu ", count[mtype]);
906         seq_putc(m, '\n');
907 }
908
909 /* Print out the free pages at each order for each migratetype */
910 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
911 {
912         int mtype;
913         pg_data_t *pgdat = (pg_data_t *)arg;
914
915         seq_printf(m, "\n%-23s", "Number of blocks type ");
916         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
917                 seq_printf(m, "%12s ", migratetype_names[mtype]);
918         seq_putc(m, '\n');
919         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
920
921         return 0;
922 }
923
924 /*
925  * This prints out statistics in relation to grouping pages by mobility.
926  * It is expensive to collect so do not constantly read the file.
927  */
928 static int pagetypeinfo_show(struct seq_file *m, void *arg)
929 {
930         pg_data_t *pgdat = (pg_data_t *)arg;
931
932         /* check memoryless node */
933         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
934                 return 0;
935
936         seq_printf(m, "Page block order: %d\n", pageblock_order);
937         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
938         seq_putc(m, '\n');
939         pagetypeinfo_showfree(m, pgdat);
940         pagetypeinfo_showblockcount(m, pgdat);
941
942         return 0;
943 }
944
945 static const struct seq_operations fragmentation_op = {
946         .start  = frag_start,
947         .next   = frag_next,
948         .stop   = frag_stop,
949         .show   = frag_show,
950 };
951
952 static int fragmentation_open(struct inode *inode, struct file *file)
953 {
954         return seq_open(file, &fragmentation_op);
955 }
956
957 static const struct file_operations fragmentation_file_operations = {
958         .open           = fragmentation_open,
959         .read           = seq_read,
960         .llseek         = seq_lseek,
961         .release        = seq_release,
962 };
963
964 static const struct seq_operations pagetypeinfo_op = {
965         .start  = frag_start,
966         .next   = frag_next,
967         .stop   = frag_stop,
968         .show   = pagetypeinfo_show,
969 };
970
971 static int pagetypeinfo_open(struct inode *inode, struct file *file)
972 {
973         return seq_open(file, &pagetypeinfo_op);
974 }
975
976 static const struct file_operations pagetypeinfo_file_ops = {
977         .open           = pagetypeinfo_open,
978         .read           = seq_read,
979         .llseek         = seq_lseek,
980         .release        = seq_release,
981 };
982
983 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
984                                                         struct zone *zone)
985 {
986         int i;
987         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
988         seq_printf(m,
989                    "\n  pages free     %lu"
990                    "\n        min      %lu"
991                    "\n        low      %lu"
992                    "\n        high     %lu"
993                    "\n        scanned  %lu"
994                    "\n        spanned  %lu"
995                    "\n        present  %lu",
996                    zone_page_state(zone, NR_FREE_PAGES),
997                    min_wmark_pages(zone),
998                    low_wmark_pages(zone),
999                    high_wmark_pages(zone),
1000                    zone->pages_scanned,
1001                    zone->spanned_pages,
1002                    zone->present_pages);
1003
1004         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1005                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1006                                 zone_page_state(zone, i));
1007
1008         seq_printf(m,
1009                    "\n        protection: (%lu",
1010                    zone->lowmem_reserve[0]);
1011         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1012                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1013         seq_printf(m,
1014                    ")"
1015                    "\n  pagesets");
1016         for_each_online_cpu(i) {
1017                 struct per_cpu_pageset *pageset;
1018
1019                 pageset = per_cpu_ptr(zone->pageset, i);
1020                 seq_printf(m,
1021                            "\n    cpu: %i"
1022                            "\n              count: %i"
1023                            "\n              high:  %i"
1024                            "\n              batch: %i",
1025                            i,
1026                            pageset->pcp.count,
1027                            pageset->pcp.high,
1028                            pageset->pcp.batch);
1029 #ifdef CONFIG_SMP
1030                 seq_printf(m, "\n  vm stats threshold: %d",
1031                                 pageset->stat_threshold);
1032 #endif
1033         }
1034         seq_printf(m,
1035                    "\n  all_unreclaimable: %u"
1036                    "\n  start_pfn:         %lu"
1037                    "\n  inactive_ratio:    %u",
1038                    zone->all_unreclaimable,
1039                    zone->zone_start_pfn,
1040                    zone->inactive_ratio);
1041         seq_putc(m, '\n');
1042 }
1043
1044 /*
1045  * Output information about zones in @pgdat.
1046  */
1047 static int zoneinfo_show(struct seq_file *m, void *arg)
1048 {
1049         pg_data_t *pgdat = (pg_data_t *)arg;
1050         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1051         return 0;
1052 }
1053
1054 static const struct seq_operations zoneinfo_op = {
1055         .start  = frag_start, /* iterate over all zones. The same as in
1056                                * fragmentation. */
1057         .next   = frag_next,
1058         .stop   = frag_stop,
1059         .show   = zoneinfo_show,
1060 };
1061
1062 static int zoneinfo_open(struct inode *inode, struct file *file)
1063 {
1064         return seq_open(file, &zoneinfo_op);
1065 }
1066
1067 static const struct file_operations proc_zoneinfo_file_operations = {
1068         .open           = zoneinfo_open,
1069         .read           = seq_read,
1070         .llseek         = seq_lseek,
1071         .release        = seq_release,
1072 };
1073
1074 enum writeback_stat_item {
1075         NR_DIRTY_THRESHOLD,
1076         NR_DIRTY_BG_THRESHOLD,
1077         NR_VM_WRITEBACK_STAT_ITEMS,
1078 };
1079
1080 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1081 {
1082         unsigned long *v;
1083         int i, stat_items_size;
1084
1085         if (*pos >= ARRAY_SIZE(vmstat_text))
1086                 return NULL;
1087         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1088                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1089
1090 #ifdef CONFIG_VM_EVENT_COUNTERS
1091         stat_items_size += sizeof(struct vm_event_state);
1092 #endif
1093
1094         v = kmalloc(stat_items_size, GFP_KERNEL);
1095         m->private = v;
1096         if (!v)
1097                 return ERR_PTR(-ENOMEM);
1098         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1099                 v[i] = global_page_state(i);
1100         v += NR_VM_ZONE_STAT_ITEMS;
1101
1102         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1103                             v + NR_DIRTY_THRESHOLD);
1104         v += NR_VM_WRITEBACK_STAT_ITEMS;
1105
1106 #ifdef CONFIG_VM_EVENT_COUNTERS
1107         all_vm_events(v);
1108         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1109         v[PGPGOUT] /= 2;
1110 #endif
1111         return (unsigned long *)m->private + *pos;
1112 }
1113
1114 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1115 {
1116         (*pos)++;
1117         if (*pos >= ARRAY_SIZE(vmstat_text))
1118                 return NULL;
1119         return (unsigned long *)m->private + *pos;
1120 }
1121
1122 static int vmstat_show(struct seq_file *m, void *arg)
1123 {
1124         unsigned long *l = arg;
1125         unsigned long off = l - (unsigned long *)m->private;
1126
1127         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1128         return 0;
1129 }
1130
1131 static void vmstat_stop(struct seq_file *m, void *arg)
1132 {
1133         kfree(m->private);
1134         m->private = NULL;
1135 }
1136
1137 static const struct seq_operations vmstat_op = {
1138         .start  = vmstat_start,
1139         .next   = vmstat_next,
1140         .stop   = vmstat_stop,
1141         .show   = vmstat_show,
1142 };
1143
1144 static int vmstat_open(struct inode *inode, struct file *file)
1145 {
1146         return seq_open(file, &vmstat_op);
1147 }
1148
1149 static const struct file_operations proc_vmstat_file_operations = {
1150         .open           = vmstat_open,
1151         .read           = seq_read,
1152         .llseek         = seq_lseek,
1153         .release        = seq_release,
1154 };
1155 #endif /* CONFIG_PROC_FS */
1156
1157 #ifdef CONFIG_SMP
1158 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1159 int sysctl_stat_interval __read_mostly = HZ;
1160
1161 static void vmstat_update(struct work_struct *w)
1162 {
1163         refresh_cpu_vm_stats(smp_processor_id());
1164         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1165                 round_jiffies_relative(sysctl_stat_interval));
1166 }
1167
1168 static void __cpuinit start_cpu_timer(int cpu)
1169 {
1170         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1171
1172         INIT_DEFERRABLE_WORK(work, vmstat_update);
1173         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1174 }
1175
1176 /*
1177  * Use the cpu notifier to insure that the thresholds are recalculated
1178  * when necessary.
1179  */
1180 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1181                 unsigned long action,
1182                 void *hcpu)
1183 {
1184         long cpu = (long)hcpu;
1185
1186         switch (action) {
1187         case CPU_ONLINE:
1188         case CPU_ONLINE_FROZEN:
1189                 refresh_zone_stat_thresholds();
1190                 start_cpu_timer(cpu);
1191                 node_set_state(cpu_to_node(cpu), N_CPU);
1192                 break;
1193         case CPU_DOWN_PREPARE:
1194         case CPU_DOWN_PREPARE_FROZEN:
1195                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1196                 per_cpu(vmstat_work, cpu).work.func = NULL;
1197                 break;
1198         case CPU_DOWN_FAILED:
1199         case CPU_DOWN_FAILED_FROZEN:
1200                 start_cpu_timer(cpu);
1201                 break;
1202         case CPU_DEAD:
1203         case CPU_DEAD_FROZEN:
1204                 refresh_zone_stat_thresholds();
1205                 break;
1206         default:
1207                 break;
1208         }
1209         return NOTIFY_OK;
1210 }
1211
1212 static struct notifier_block __cpuinitdata vmstat_notifier =
1213         { &vmstat_cpuup_callback, NULL, 0 };
1214 #endif
1215
1216 static int __init setup_vmstat(void)
1217 {
1218 #ifdef CONFIG_SMP
1219         int cpu;
1220
1221         register_cpu_notifier(&vmstat_notifier);
1222
1223         for_each_online_cpu(cpu)
1224                 start_cpu_timer(cpu);
1225 #endif
1226 #ifdef CONFIG_PROC_FS
1227         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1228         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1229         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1230         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1231 #endif
1232         return 0;
1233 }
1234 module_init(setup_vmstat)
1235
1236 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1237 #include <linux/debugfs.h>
1238
1239
1240 /*
1241  * Return an index indicating how much of the available free memory is
1242  * unusable for an allocation of the requested size.
1243  */
1244 static int unusable_free_index(unsigned int order,
1245                                 struct contig_page_info *info)
1246 {
1247         /* No free memory is interpreted as all free memory is unusable */
1248         if (info->free_pages == 0)
1249                 return 1000;
1250
1251         /*
1252          * Index should be a value between 0 and 1. Return a value to 3
1253          * decimal places.
1254          *
1255          * 0 => no fragmentation
1256          * 1 => high fragmentation
1257          */
1258         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1259
1260 }
1261
1262 static void unusable_show_print(struct seq_file *m,
1263                                         pg_data_t *pgdat, struct zone *zone)
1264 {
1265         unsigned int order;
1266         int index;
1267         struct contig_page_info info;
1268
1269         seq_printf(m, "Node %d, zone %8s ",
1270                                 pgdat->node_id,
1271                                 zone->name);
1272         for (order = 0; order < MAX_ORDER; ++order) {
1273                 fill_contig_page_info(zone, order, &info);
1274                 index = unusable_free_index(order, &info);
1275                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1276         }
1277
1278         seq_putc(m, '\n');
1279 }
1280
1281 /*
1282  * Display unusable free space index
1283  *
1284  * The unusable free space index measures how much of the available free
1285  * memory cannot be used to satisfy an allocation of a given size and is a
1286  * value between 0 and 1. The higher the value, the more of free memory is
1287  * unusable and by implication, the worse the external fragmentation is. This
1288  * can be expressed as a percentage by multiplying by 100.
1289  */
1290 static int unusable_show(struct seq_file *m, void *arg)
1291 {
1292         pg_data_t *pgdat = (pg_data_t *)arg;
1293
1294         /* check memoryless node */
1295         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1296                 return 0;
1297
1298         walk_zones_in_node(m, pgdat, unusable_show_print);
1299
1300         return 0;
1301 }
1302
1303 static const struct seq_operations unusable_op = {
1304         .start  = frag_start,
1305         .next   = frag_next,
1306         .stop   = frag_stop,
1307         .show   = unusable_show,
1308 };
1309
1310 static int unusable_open(struct inode *inode, struct file *file)
1311 {
1312         return seq_open(file, &unusable_op);
1313 }
1314
1315 static const struct file_operations unusable_file_ops = {
1316         .open           = unusable_open,
1317         .read           = seq_read,
1318         .llseek         = seq_lseek,
1319         .release        = seq_release,
1320 };
1321
1322 static void extfrag_show_print(struct seq_file *m,
1323                                         pg_data_t *pgdat, struct zone *zone)
1324 {
1325         unsigned int order;
1326         int index;
1327
1328         /* Alloc on stack as interrupts are disabled for zone walk */
1329         struct contig_page_info info;
1330
1331         seq_printf(m, "Node %d, zone %8s ",
1332                                 pgdat->node_id,
1333                                 zone->name);
1334         for (order = 0; order < MAX_ORDER; ++order) {
1335                 fill_contig_page_info(zone, order, &info);
1336                 index = __fragmentation_index(order, &info);
1337                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1338         }
1339
1340         seq_putc(m, '\n');
1341 }
1342
1343 /*
1344  * Display fragmentation index for orders that allocations would fail for
1345  */
1346 static int extfrag_show(struct seq_file *m, void *arg)
1347 {
1348         pg_data_t *pgdat = (pg_data_t *)arg;
1349
1350         walk_zones_in_node(m, pgdat, extfrag_show_print);
1351
1352         return 0;
1353 }
1354
1355 static const struct seq_operations extfrag_op = {
1356         .start  = frag_start,
1357         .next   = frag_next,
1358         .stop   = frag_stop,
1359         .show   = extfrag_show,
1360 };
1361
1362 static int extfrag_open(struct inode *inode, struct file *file)
1363 {
1364         return seq_open(file, &extfrag_op);
1365 }
1366
1367 static const struct file_operations extfrag_file_ops = {
1368         .open           = extfrag_open,
1369         .read           = seq_read,
1370         .llseek         = seq_lseek,
1371         .release        = seq_release,
1372 };
1373
1374 static int __init extfrag_debug_init(void)
1375 {
1376         struct dentry *extfrag_debug_root;
1377
1378         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1379         if (!extfrag_debug_root)
1380                 return -ENOMEM;
1381
1382         if (!debugfs_create_file("unusable_index", 0444,
1383                         extfrag_debug_root, NULL, &unusable_file_ops))
1384                 goto fail;
1385
1386         if (!debugfs_create_file("extfrag_index", 0444,
1387                         extfrag_debug_root, NULL, &extfrag_file_ops))
1388                 goto fail;
1389
1390         return 0;
1391 fail:
1392         debugfs_remove_recursive(extfrag_debug_root);
1393         return -ENOMEM;
1394 }
1395
1396 module_init(extfrag_debug_init);
1397 #endif