mm: compaction: direct compact when a high-order allocation fails
[linux-2.6.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
21 #ifdef CONFIG_VM_EVENT_COUNTERS
22 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
23 EXPORT_PER_CPU_SYMBOL(vm_event_states);
24
25 static void sum_vm_events(unsigned long *ret, const struct cpumask *cpumask)
26 {
27         int cpu;
28         int i;
29
30         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
31
32         for_each_cpu(cpu, cpumask) {
33                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
34
35                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
36                         ret[i] += this->event[i];
37         }
38 }
39
40 /*
41  * Accumulate the vm event counters across all CPUs.
42  * The result is unavoidably approximate - it can change
43  * during and after execution of this function.
44 */
45 void all_vm_events(unsigned long *ret)
46 {
47         get_online_cpus();
48         sum_vm_events(ret, cpu_online_mask);
49         put_online_cpus();
50 }
51 EXPORT_SYMBOL_GPL(all_vm_events);
52
53 #ifdef CONFIG_HOTPLUG
54 /*
55  * Fold the foreign cpu events into our own.
56  *
57  * This is adding to the events on one processor
58  * but keeps the global counts constant.
59  */
60 void vm_events_fold_cpu(int cpu)
61 {
62         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
63         int i;
64
65         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
66                 count_vm_events(i, fold_state->event[i]);
67                 fold_state->event[i] = 0;
68         }
69 }
70 #endif /* CONFIG_HOTPLUG */
71
72 #endif /* CONFIG_VM_EVENT_COUNTERS */
73
74 /*
75  * Manage combined zone based / global counters
76  *
77  * vm_stat contains the global counters
78  */
79 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
80 EXPORT_SYMBOL(vm_stat);
81
82 #ifdef CONFIG_SMP
83
84 static int calculate_threshold(struct zone *zone)
85 {
86         int threshold;
87         int mem;        /* memory in 128 MB units */
88
89         /*
90          * The threshold scales with the number of processors and the amount
91          * of memory per zone. More memory means that we can defer updates for
92          * longer, more processors could lead to more contention.
93          * fls() is used to have a cheap way of logarithmic scaling.
94          *
95          * Some sample thresholds:
96          *
97          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
98          * ------------------------------------------------------------------
99          * 8            1               1       0.9-1 GB        4
100          * 16           2               2       0.9-1 GB        4
101          * 20           2               2       1-2 GB          5
102          * 24           2               2       2-4 GB          6
103          * 28           2               2       4-8 GB          7
104          * 32           2               2       8-16 GB         8
105          * 4            2               2       <128M           1
106          * 30           4               3       2-4 GB          5
107          * 48           4               3       8-16 GB         8
108          * 32           8               4       1-2 GB          4
109          * 32           8               4       0.9-1GB         4
110          * 10           16              5       <128M           1
111          * 40           16              5       900M            4
112          * 70           64              7       2-4 GB          5
113          * 84           64              7       4-8 GB          6
114          * 108          512             9       4-8 GB          6
115          * 125          1024            10      8-16 GB         8
116          * 125          1024            10      16-32 GB        9
117          */
118
119         mem = zone->present_pages >> (27 - PAGE_SHIFT);
120
121         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
122
123         /*
124          * Maximum threshold is 125
125          */
126         threshold = min(125, threshold);
127
128         return threshold;
129 }
130
131 /*
132  * Refresh the thresholds for each zone.
133  */
134 static void refresh_zone_stat_thresholds(void)
135 {
136         struct zone *zone;
137         int cpu;
138         int threshold;
139
140         for_each_populated_zone(zone) {
141                 threshold = calculate_threshold(zone);
142
143                 for_each_online_cpu(cpu)
144                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
145                                                         = threshold;
146         }
147 }
148
149 /*
150  * For use when we know that interrupts are disabled.
151  */
152 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
153                                 int delta)
154 {
155         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
156
157         s8 *p = pcp->vm_stat_diff + item;
158         long x;
159
160         x = delta + *p;
161
162         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
163                 zone_page_state_add(x, zone, item);
164                 x = 0;
165         }
166         *p = x;
167 }
168 EXPORT_SYMBOL(__mod_zone_page_state);
169
170 /*
171  * For an unknown interrupt state
172  */
173 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
174                                         int delta)
175 {
176         unsigned long flags;
177
178         local_irq_save(flags);
179         __mod_zone_page_state(zone, item, delta);
180         local_irq_restore(flags);
181 }
182 EXPORT_SYMBOL(mod_zone_page_state);
183
184 /*
185  * Optimized increment and decrement functions.
186  *
187  * These are only for a single page and therefore can take a struct page *
188  * argument instead of struct zone *. This allows the inclusion of the code
189  * generated for page_zone(page) into the optimized functions.
190  *
191  * No overflow check is necessary and therefore the differential can be
192  * incremented or decremented in place which may allow the compilers to
193  * generate better code.
194  * The increment or decrement is known and therefore one boundary check can
195  * be omitted.
196  *
197  * NOTE: These functions are very performance sensitive. Change only
198  * with care.
199  *
200  * Some processors have inc/dec instructions that are atomic vs an interrupt.
201  * However, the code must first determine the differential location in a zone
202  * based on the processor number and then inc/dec the counter. There is no
203  * guarantee without disabling preemption that the processor will not change
204  * in between and therefore the atomicity vs. interrupt cannot be exploited
205  * in a useful way here.
206  */
207 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
208 {
209         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
210         s8 *p = pcp->vm_stat_diff + item;
211
212         (*p)++;
213
214         if (unlikely(*p > pcp->stat_threshold)) {
215                 int overstep = pcp->stat_threshold / 2;
216
217                 zone_page_state_add(*p + overstep, zone, item);
218                 *p = -overstep;
219         }
220 }
221
222 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
223 {
224         __inc_zone_state(page_zone(page), item);
225 }
226 EXPORT_SYMBOL(__inc_zone_page_state);
227
228 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
229 {
230         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
231         s8 *p = pcp->vm_stat_diff + item;
232
233         (*p)--;
234
235         if (unlikely(*p < - pcp->stat_threshold)) {
236                 int overstep = pcp->stat_threshold / 2;
237
238                 zone_page_state_add(*p - overstep, zone, item);
239                 *p = overstep;
240         }
241 }
242
243 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
244 {
245         __dec_zone_state(page_zone(page), item);
246 }
247 EXPORT_SYMBOL(__dec_zone_page_state);
248
249 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
250 {
251         unsigned long flags;
252
253         local_irq_save(flags);
254         __inc_zone_state(zone, item);
255         local_irq_restore(flags);
256 }
257
258 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
259 {
260         unsigned long flags;
261         struct zone *zone;
262
263         zone = page_zone(page);
264         local_irq_save(flags);
265         __inc_zone_state(zone, item);
266         local_irq_restore(flags);
267 }
268 EXPORT_SYMBOL(inc_zone_page_state);
269
270 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
271 {
272         unsigned long flags;
273
274         local_irq_save(flags);
275         __dec_zone_page_state(page, item);
276         local_irq_restore(flags);
277 }
278 EXPORT_SYMBOL(dec_zone_page_state);
279
280 /*
281  * Update the zone counters for one cpu.
282  *
283  * The cpu specified must be either the current cpu or a processor that
284  * is not online. If it is the current cpu then the execution thread must
285  * be pinned to the current cpu.
286  *
287  * Note that refresh_cpu_vm_stats strives to only access
288  * node local memory. The per cpu pagesets on remote zones are placed
289  * in the memory local to the processor using that pageset. So the
290  * loop over all zones will access a series of cachelines local to
291  * the processor.
292  *
293  * The call to zone_page_state_add updates the cachelines with the
294  * statistics in the remote zone struct as well as the global cachelines
295  * with the global counters. These could cause remote node cache line
296  * bouncing and will have to be only done when necessary.
297  */
298 void refresh_cpu_vm_stats(int cpu)
299 {
300         struct zone *zone;
301         int i;
302         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
303
304         for_each_populated_zone(zone) {
305                 struct per_cpu_pageset *p;
306
307                 p = per_cpu_ptr(zone->pageset, cpu);
308
309                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
310                         if (p->vm_stat_diff[i]) {
311                                 unsigned long flags;
312                                 int v;
313
314                                 local_irq_save(flags);
315                                 v = p->vm_stat_diff[i];
316                                 p->vm_stat_diff[i] = 0;
317                                 local_irq_restore(flags);
318                                 atomic_long_add(v, &zone->vm_stat[i]);
319                                 global_diff[i] += v;
320 #ifdef CONFIG_NUMA
321                                 /* 3 seconds idle till flush */
322                                 p->expire = 3;
323 #endif
324                         }
325                 cond_resched();
326 #ifdef CONFIG_NUMA
327                 /*
328                  * Deal with draining the remote pageset of this
329                  * processor
330                  *
331                  * Check if there are pages remaining in this pageset
332                  * if not then there is nothing to expire.
333                  */
334                 if (!p->expire || !p->pcp.count)
335                         continue;
336
337                 /*
338                  * We never drain zones local to this processor.
339                  */
340                 if (zone_to_nid(zone) == numa_node_id()) {
341                         p->expire = 0;
342                         continue;
343                 }
344
345                 p->expire--;
346                 if (p->expire)
347                         continue;
348
349                 if (p->pcp.count)
350                         drain_zone_pages(zone, &p->pcp);
351 #endif
352         }
353
354         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
355                 if (global_diff[i])
356                         atomic_long_add(global_diff[i], &vm_stat[i]);
357 }
358
359 #endif
360
361 #ifdef CONFIG_NUMA
362 /*
363  * zonelist = the list of zones passed to the allocator
364  * z        = the zone from which the allocation occurred.
365  *
366  * Must be called with interrupts disabled.
367  */
368 void zone_statistics(struct zone *preferred_zone, struct zone *z)
369 {
370         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
371                 __inc_zone_state(z, NUMA_HIT);
372         } else {
373                 __inc_zone_state(z, NUMA_MISS);
374                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
375         }
376         if (z->node == numa_node_id())
377                 __inc_zone_state(z, NUMA_LOCAL);
378         else
379                 __inc_zone_state(z, NUMA_OTHER);
380 }
381 #endif
382
383 #ifdef CONFIG_COMPACTION
384 struct contig_page_info {
385         unsigned long free_pages;
386         unsigned long free_blocks_total;
387         unsigned long free_blocks_suitable;
388 };
389
390 /*
391  * Calculate the number of free pages in a zone, how many contiguous
392  * pages are free and how many are large enough to satisfy an allocation of
393  * the target size. Note that this function makes no attempt to estimate
394  * how many suitable free blocks there *might* be if MOVABLE pages were
395  * migrated. Calculating that is possible, but expensive and can be
396  * figured out from userspace
397  */
398 static void fill_contig_page_info(struct zone *zone,
399                                 unsigned int suitable_order,
400                                 struct contig_page_info *info)
401 {
402         unsigned int order;
403
404         info->free_pages = 0;
405         info->free_blocks_total = 0;
406         info->free_blocks_suitable = 0;
407
408         for (order = 0; order < MAX_ORDER; order++) {
409                 unsigned long blocks;
410
411                 /* Count number of free blocks */
412                 blocks = zone->free_area[order].nr_free;
413                 info->free_blocks_total += blocks;
414
415                 /* Count free base pages */
416                 info->free_pages += blocks << order;
417
418                 /* Count the suitable free blocks */
419                 if (order >= suitable_order)
420                         info->free_blocks_suitable += blocks <<
421                                                 (order - suitable_order);
422         }
423 }
424
425 /*
426  * A fragmentation index only makes sense if an allocation of a requested
427  * size would fail. If that is true, the fragmentation index indicates
428  * whether external fragmentation or a lack of memory was the problem.
429  * The value can be used to determine if page reclaim or compaction
430  * should be used
431  */
432 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
433 {
434         unsigned long requested = 1UL << order;
435
436         if (!info->free_blocks_total)
437                 return 0;
438
439         /* Fragmentation index only makes sense when a request would fail */
440         if (info->free_blocks_suitable)
441                 return -1000;
442
443         /*
444          * Index is between 0 and 1 so return within 3 decimal places
445          *
446          * 0 => allocation would fail due to lack of memory
447          * 1 => allocation would fail due to fragmentation
448          */
449         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
450 }
451
452 /* Same as __fragmentation index but allocs contig_page_info on stack */
453 int fragmentation_index(struct zone *zone, unsigned int order)
454 {
455         struct contig_page_info info;
456
457         fill_contig_page_info(zone, order, &info);
458         return __fragmentation_index(order, &info);
459 }
460 #endif
461
462 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
463 #include <linux/proc_fs.h>
464 #include <linux/seq_file.h>
465
466 static char * const migratetype_names[MIGRATE_TYPES] = {
467         "Unmovable",
468         "Reclaimable",
469         "Movable",
470         "Reserve",
471         "Isolate",
472 };
473
474 static void *frag_start(struct seq_file *m, loff_t *pos)
475 {
476         pg_data_t *pgdat;
477         loff_t node = *pos;
478         for (pgdat = first_online_pgdat();
479              pgdat && node;
480              pgdat = next_online_pgdat(pgdat))
481                 --node;
482
483         return pgdat;
484 }
485
486 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
487 {
488         pg_data_t *pgdat = (pg_data_t *)arg;
489
490         (*pos)++;
491         return next_online_pgdat(pgdat);
492 }
493
494 static void frag_stop(struct seq_file *m, void *arg)
495 {
496 }
497
498 /* Walk all the zones in a node and print using a callback */
499 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
500                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
501 {
502         struct zone *zone;
503         struct zone *node_zones = pgdat->node_zones;
504         unsigned long flags;
505
506         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
507                 if (!populated_zone(zone))
508                         continue;
509
510                 spin_lock_irqsave(&zone->lock, flags);
511                 print(m, pgdat, zone);
512                 spin_unlock_irqrestore(&zone->lock, flags);
513         }
514 }
515 #endif
516
517 #ifdef CONFIG_PROC_FS
518 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
519                                                 struct zone *zone)
520 {
521         int order;
522
523         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
524         for (order = 0; order < MAX_ORDER; ++order)
525                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
526         seq_putc(m, '\n');
527 }
528
529 /*
530  * This walks the free areas for each zone.
531  */
532 static int frag_show(struct seq_file *m, void *arg)
533 {
534         pg_data_t *pgdat = (pg_data_t *)arg;
535         walk_zones_in_node(m, pgdat, frag_show_print);
536         return 0;
537 }
538
539 static void pagetypeinfo_showfree_print(struct seq_file *m,
540                                         pg_data_t *pgdat, struct zone *zone)
541 {
542         int order, mtype;
543
544         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
545                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
546                                         pgdat->node_id,
547                                         zone->name,
548                                         migratetype_names[mtype]);
549                 for (order = 0; order < MAX_ORDER; ++order) {
550                         unsigned long freecount = 0;
551                         struct free_area *area;
552                         struct list_head *curr;
553
554                         area = &(zone->free_area[order]);
555
556                         list_for_each(curr, &area->free_list[mtype])
557                                 freecount++;
558                         seq_printf(m, "%6lu ", freecount);
559                 }
560                 seq_putc(m, '\n');
561         }
562 }
563
564 /* Print out the free pages at each order for each migatetype */
565 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
566 {
567         int order;
568         pg_data_t *pgdat = (pg_data_t *)arg;
569
570         /* Print header */
571         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
572         for (order = 0; order < MAX_ORDER; ++order)
573                 seq_printf(m, "%6d ", order);
574         seq_putc(m, '\n');
575
576         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
577
578         return 0;
579 }
580
581 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
582                                         pg_data_t *pgdat, struct zone *zone)
583 {
584         int mtype;
585         unsigned long pfn;
586         unsigned long start_pfn = zone->zone_start_pfn;
587         unsigned long end_pfn = start_pfn + zone->spanned_pages;
588         unsigned long count[MIGRATE_TYPES] = { 0, };
589
590         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
591                 struct page *page;
592
593                 if (!pfn_valid(pfn))
594                         continue;
595
596                 page = pfn_to_page(pfn);
597
598                 /* Watch for unexpected holes punched in the memmap */
599                 if (!memmap_valid_within(pfn, page, zone))
600                         continue;
601
602                 mtype = get_pageblock_migratetype(page);
603
604                 if (mtype < MIGRATE_TYPES)
605                         count[mtype]++;
606         }
607
608         /* Print counts */
609         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
610         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
611                 seq_printf(m, "%12lu ", count[mtype]);
612         seq_putc(m, '\n');
613 }
614
615 /* Print out the free pages at each order for each migratetype */
616 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
617 {
618         int mtype;
619         pg_data_t *pgdat = (pg_data_t *)arg;
620
621         seq_printf(m, "\n%-23s", "Number of blocks type ");
622         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
623                 seq_printf(m, "%12s ", migratetype_names[mtype]);
624         seq_putc(m, '\n');
625         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
626
627         return 0;
628 }
629
630 /*
631  * This prints out statistics in relation to grouping pages by mobility.
632  * It is expensive to collect so do not constantly read the file.
633  */
634 static int pagetypeinfo_show(struct seq_file *m, void *arg)
635 {
636         pg_data_t *pgdat = (pg_data_t *)arg;
637
638         /* check memoryless node */
639         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
640                 return 0;
641
642         seq_printf(m, "Page block order: %d\n", pageblock_order);
643         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
644         seq_putc(m, '\n');
645         pagetypeinfo_showfree(m, pgdat);
646         pagetypeinfo_showblockcount(m, pgdat);
647
648         return 0;
649 }
650
651 static const struct seq_operations fragmentation_op = {
652         .start  = frag_start,
653         .next   = frag_next,
654         .stop   = frag_stop,
655         .show   = frag_show,
656 };
657
658 static int fragmentation_open(struct inode *inode, struct file *file)
659 {
660         return seq_open(file, &fragmentation_op);
661 }
662
663 static const struct file_operations fragmentation_file_operations = {
664         .open           = fragmentation_open,
665         .read           = seq_read,
666         .llseek         = seq_lseek,
667         .release        = seq_release,
668 };
669
670 static const struct seq_operations pagetypeinfo_op = {
671         .start  = frag_start,
672         .next   = frag_next,
673         .stop   = frag_stop,
674         .show   = pagetypeinfo_show,
675 };
676
677 static int pagetypeinfo_open(struct inode *inode, struct file *file)
678 {
679         return seq_open(file, &pagetypeinfo_op);
680 }
681
682 static const struct file_operations pagetypeinfo_file_ops = {
683         .open           = pagetypeinfo_open,
684         .read           = seq_read,
685         .llseek         = seq_lseek,
686         .release        = seq_release,
687 };
688
689 #ifdef CONFIG_ZONE_DMA
690 #define TEXT_FOR_DMA(xx) xx "_dma",
691 #else
692 #define TEXT_FOR_DMA(xx)
693 #endif
694
695 #ifdef CONFIG_ZONE_DMA32
696 #define TEXT_FOR_DMA32(xx) xx "_dma32",
697 #else
698 #define TEXT_FOR_DMA32(xx)
699 #endif
700
701 #ifdef CONFIG_HIGHMEM
702 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
703 #else
704 #define TEXT_FOR_HIGHMEM(xx)
705 #endif
706
707 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
708                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
709
710 static const char * const vmstat_text[] = {
711         /* Zoned VM counters */
712         "nr_free_pages",
713         "nr_inactive_anon",
714         "nr_active_anon",
715         "nr_inactive_file",
716         "nr_active_file",
717         "nr_unevictable",
718         "nr_mlock",
719         "nr_anon_pages",
720         "nr_mapped",
721         "nr_file_pages",
722         "nr_dirty",
723         "nr_writeback",
724         "nr_slab_reclaimable",
725         "nr_slab_unreclaimable",
726         "nr_page_table_pages",
727         "nr_kernel_stack",
728         "nr_unstable",
729         "nr_bounce",
730         "nr_vmscan_write",
731         "nr_writeback_temp",
732         "nr_isolated_anon",
733         "nr_isolated_file",
734         "nr_shmem",
735 #ifdef CONFIG_NUMA
736         "numa_hit",
737         "numa_miss",
738         "numa_foreign",
739         "numa_interleave",
740         "numa_local",
741         "numa_other",
742 #endif
743
744 #ifdef CONFIG_VM_EVENT_COUNTERS
745         "pgpgin",
746         "pgpgout",
747         "pswpin",
748         "pswpout",
749
750         TEXTS_FOR_ZONES("pgalloc")
751
752         "pgfree",
753         "pgactivate",
754         "pgdeactivate",
755
756         "pgfault",
757         "pgmajfault",
758
759         TEXTS_FOR_ZONES("pgrefill")
760         TEXTS_FOR_ZONES("pgsteal")
761         TEXTS_FOR_ZONES("pgscan_kswapd")
762         TEXTS_FOR_ZONES("pgscan_direct")
763
764 #ifdef CONFIG_NUMA
765         "zone_reclaim_failed",
766 #endif
767         "pginodesteal",
768         "slabs_scanned",
769         "kswapd_steal",
770         "kswapd_inodesteal",
771         "kswapd_low_wmark_hit_quickly",
772         "kswapd_high_wmark_hit_quickly",
773         "kswapd_skip_congestion_wait",
774         "pageoutrun",
775         "allocstall",
776
777         "pgrotated",
778
779 #ifdef CONFIG_COMPACTION
780         "compact_blocks_moved",
781         "compact_pages_moved",
782         "compact_pagemigrate_failed",
783         "compact_stall",
784         "compact_fail",
785         "compact_success",
786 #endif
787
788 #ifdef CONFIG_HUGETLB_PAGE
789         "htlb_buddy_alloc_success",
790         "htlb_buddy_alloc_fail",
791 #endif
792         "unevictable_pgs_culled",
793         "unevictable_pgs_scanned",
794         "unevictable_pgs_rescued",
795         "unevictable_pgs_mlocked",
796         "unevictable_pgs_munlocked",
797         "unevictable_pgs_cleared",
798         "unevictable_pgs_stranded",
799         "unevictable_pgs_mlockfreed",
800 #endif
801 };
802
803 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
804                                                         struct zone *zone)
805 {
806         int i;
807         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
808         seq_printf(m,
809                    "\n  pages free     %lu"
810                    "\n        min      %lu"
811                    "\n        low      %lu"
812                    "\n        high     %lu"
813                    "\n        scanned  %lu"
814                    "\n        spanned  %lu"
815                    "\n        present  %lu",
816                    zone_page_state(zone, NR_FREE_PAGES),
817                    min_wmark_pages(zone),
818                    low_wmark_pages(zone),
819                    high_wmark_pages(zone),
820                    zone->pages_scanned,
821                    zone->spanned_pages,
822                    zone->present_pages);
823
824         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
825                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
826                                 zone_page_state(zone, i));
827
828         seq_printf(m,
829                    "\n        protection: (%lu",
830                    zone->lowmem_reserve[0]);
831         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
832                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
833         seq_printf(m,
834                    ")"
835                    "\n  pagesets");
836         for_each_online_cpu(i) {
837                 struct per_cpu_pageset *pageset;
838
839                 pageset = per_cpu_ptr(zone->pageset, i);
840                 seq_printf(m,
841                            "\n    cpu: %i"
842                            "\n              count: %i"
843                            "\n              high:  %i"
844                            "\n              batch: %i",
845                            i,
846                            pageset->pcp.count,
847                            pageset->pcp.high,
848                            pageset->pcp.batch);
849 #ifdef CONFIG_SMP
850                 seq_printf(m, "\n  vm stats threshold: %d",
851                                 pageset->stat_threshold);
852 #endif
853         }
854         seq_printf(m,
855                    "\n  all_unreclaimable: %u"
856                    "\n  prev_priority:     %i"
857                    "\n  start_pfn:         %lu"
858                    "\n  inactive_ratio:    %u",
859                    zone->all_unreclaimable,
860                    zone->prev_priority,
861                    zone->zone_start_pfn,
862                    zone->inactive_ratio);
863         seq_putc(m, '\n');
864 }
865
866 /*
867  * Output information about zones in @pgdat.
868  */
869 static int zoneinfo_show(struct seq_file *m, void *arg)
870 {
871         pg_data_t *pgdat = (pg_data_t *)arg;
872         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
873         return 0;
874 }
875
876 static const struct seq_operations zoneinfo_op = {
877         .start  = frag_start, /* iterate over all zones. The same as in
878                                * fragmentation. */
879         .next   = frag_next,
880         .stop   = frag_stop,
881         .show   = zoneinfo_show,
882 };
883
884 static int zoneinfo_open(struct inode *inode, struct file *file)
885 {
886         return seq_open(file, &zoneinfo_op);
887 }
888
889 static const struct file_operations proc_zoneinfo_file_operations = {
890         .open           = zoneinfo_open,
891         .read           = seq_read,
892         .llseek         = seq_lseek,
893         .release        = seq_release,
894 };
895
896 static void *vmstat_start(struct seq_file *m, loff_t *pos)
897 {
898         unsigned long *v;
899 #ifdef CONFIG_VM_EVENT_COUNTERS
900         unsigned long *e;
901 #endif
902         int i;
903
904         if (*pos >= ARRAY_SIZE(vmstat_text))
905                 return NULL;
906
907 #ifdef CONFIG_VM_EVENT_COUNTERS
908         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
909                         + sizeof(struct vm_event_state), GFP_KERNEL);
910 #else
911         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
912                         GFP_KERNEL);
913 #endif
914         m->private = v;
915         if (!v)
916                 return ERR_PTR(-ENOMEM);
917         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
918                 v[i] = global_page_state(i);
919 #ifdef CONFIG_VM_EVENT_COUNTERS
920         e = v + NR_VM_ZONE_STAT_ITEMS;
921         all_vm_events(e);
922         e[PGPGIN] /= 2;         /* sectors -> kbytes */
923         e[PGPGOUT] /= 2;
924 #endif
925         return v + *pos;
926 }
927
928 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
929 {
930         (*pos)++;
931         if (*pos >= ARRAY_SIZE(vmstat_text))
932                 return NULL;
933         return (unsigned long *)m->private + *pos;
934 }
935
936 static int vmstat_show(struct seq_file *m, void *arg)
937 {
938         unsigned long *l = arg;
939         unsigned long off = l - (unsigned long *)m->private;
940
941         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
942         return 0;
943 }
944
945 static void vmstat_stop(struct seq_file *m, void *arg)
946 {
947         kfree(m->private);
948         m->private = NULL;
949 }
950
951 static const struct seq_operations vmstat_op = {
952         .start  = vmstat_start,
953         .next   = vmstat_next,
954         .stop   = vmstat_stop,
955         .show   = vmstat_show,
956 };
957
958 static int vmstat_open(struct inode *inode, struct file *file)
959 {
960         return seq_open(file, &vmstat_op);
961 }
962
963 static const struct file_operations proc_vmstat_file_operations = {
964         .open           = vmstat_open,
965         .read           = seq_read,
966         .llseek         = seq_lseek,
967         .release        = seq_release,
968 };
969 #endif /* CONFIG_PROC_FS */
970
971 #ifdef CONFIG_SMP
972 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
973 int sysctl_stat_interval __read_mostly = HZ;
974
975 static void vmstat_update(struct work_struct *w)
976 {
977         refresh_cpu_vm_stats(smp_processor_id());
978         schedule_delayed_work(&__get_cpu_var(vmstat_work),
979                 round_jiffies_relative(sysctl_stat_interval));
980 }
981
982 static void __cpuinit start_cpu_timer(int cpu)
983 {
984         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
985
986         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
987         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
988 }
989
990 /*
991  * Use the cpu notifier to insure that the thresholds are recalculated
992  * when necessary.
993  */
994 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
995                 unsigned long action,
996                 void *hcpu)
997 {
998         long cpu = (long)hcpu;
999
1000         switch (action) {
1001         case CPU_ONLINE:
1002         case CPU_ONLINE_FROZEN:
1003                 start_cpu_timer(cpu);
1004                 node_set_state(cpu_to_node(cpu), N_CPU);
1005                 break;
1006         case CPU_DOWN_PREPARE:
1007         case CPU_DOWN_PREPARE_FROZEN:
1008                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
1009                 per_cpu(vmstat_work, cpu).work.func = NULL;
1010                 break;
1011         case CPU_DOWN_FAILED:
1012         case CPU_DOWN_FAILED_FROZEN:
1013                 start_cpu_timer(cpu);
1014                 break;
1015         case CPU_DEAD:
1016         case CPU_DEAD_FROZEN:
1017                 refresh_zone_stat_thresholds();
1018                 break;
1019         default:
1020                 break;
1021         }
1022         return NOTIFY_OK;
1023 }
1024
1025 static struct notifier_block __cpuinitdata vmstat_notifier =
1026         { &vmstat_cpuup_callback, NULL, 0 };
1027 #endif
1028
1029 static int __init setup_vmstat(void)
1030 {
1031 #ifdef CONFIG_SMP
1032         int cpu;
1033
1034         refresh_zone_stat_thresholds();
1035         register_cpu_notifier(&vmstat_notifier);
1036
1037         for_each_online_cpu(cpu)
1038                 start_cpu_timer(cpu);
1039 #endif
1040 #ifdef CONFIG_PROC_FS
1041         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1042         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1043         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1044         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1045 #endif
1046         return 0;
1047 }
1048 module_init(setup_vmstat)
1049
1050 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1051 #include <linux/debugfs.h>
1052
1053 static struct dentry *extfrag_debug_root;
1054
1055 /*
1056  * Return an index indicating how much of the available free memory is
1057  * unusable for an allocation of the requested size.
1058  */
1059 static int unusable_free_index(unsigned int order,
1060                                 struct contig_page_info *info)
1061 {
1062         /* No free memory is interpreted as all free memory is unusable */
1063         if (info->free_pages == 0)
1064                 return 1000;
1065
1066         /*
1067          * Index should be a value between 0 and 1. Return a value to 3
1068          * decimal places.
1069          *
1070          * 0 => no fragmentation
1071          * 1 => high fragmentation
1072          */
1073         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1074
1075 }
1076
1077 static void unusable_show_print(struct seq_file *m,
1078                                         pg_data_t *pgdat, struct zone *zone)
1079 {
1080         unsigned int order;
1081         int index;
1082         struct contig_page_info info;
1083
1084         seq_printf(m, "Node %d, zone %8s ",
1085                                 pgdat->node_id,
1086                                 zone->name);
1087         for (order = 0; order < MAX_ORDER; ++order) {
1088                 fill_contig_page_info(zone, order, &info);
1089                 index = unusable_free_index(order, &info);
1090                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1091         }
1092
1093         seq_putc(m, '\n');
1094 }
1095
1096 /*
1097  * Display unusable free space index
1098  *
1099  * The unusable free space index measures how much of the available free
1100  * memory cannot be used to satisfy an allocation of a given size and is a
1101  * value between 0 and 1. The higher the value, the more of free memory is
1102  * unusable and by implication, the worse the external fragmentation is. This
1103  * can be expressed as a percentage by multiplying by 100.
1104  */
1105 static int unusable_show(struct seq_file *m, void *arg)
1106 {
1107         pg_data_t *pgdat = (pg_data_t *)arg;
1108
1109         /* check memoryless node */
1110         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1111                 return 0;
1112
1113         walk_zones_in_node(m, pgdat, unusable_show_print);
1114
1115         return 0;
1116 }
1117
1118 static const struct seq_operations unusable_op = {
1119         .start  = frag_start,
1120         .next   = frag_next,
1121         .stop   = frag_stop,
1122         .show   = unusable_show,
1123 };
1124
1125 static int unusable_open(struct inode *inode, struct file *file)
1126 {
1127         return seq_open(file, &unusable_op);
1128 }
1129
1130 static const struct file_operations unusable_file_ops = {
1131         .open           = unusable_open,
1132         .read           = seq_read,
1133         .llseek         = seq_lseek,
1134         .release        = seq_release,
1135 };
1136
1137 static void extfrag_show_print(struct seq_file *m,
1138                                         pg_data_t *pgdat, struct zone *zone)
1139 {
1140         unsigned int order;
1141         int index;
1142
1143         /* Alloc on stack as interrupts are disabled for zone walk */
1144         struct contig_page_info info;
1145
1146         seq_printf(m, "Node %d, zone %8s ",
1147                                 pgdat->node_id,
1148                                 zone->name);
1149         for (order = 0; order < MAX_ORDER; ++order) {
1150                 fill_contig_page_info(zone, order, &info);
1151                 index = __fragmentation_index(order, &info);
1152                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1153         }
1154
1155         seq_putc(m, '\n');
1156 }
1157
1158 /*
1159  * Display fragmentation index for orders that allocations would fail for
1160  */
1161 static int extfrag_show(struct seq_file *m, void *arg)
1162 {
1163         pg_data_t *pgdat = (pg_data_t *)arg;
1164
1165         walk_zones_in_node(m, pgdat, extfrag_show_print);
1166
1167         return 0;
1168 }
1169
1170 static const struct seq_operations extfrag_op = {
1171         .start  = frag_start,
1172         .next   = frag_next,
1173         .stop   = frag_stop,
1174         .show   = extfrag_show,
1175 };
1176
1177 static int extfrag_open(struct inode *inode, struct file *file)
1178 {
1179         return seq_open(file, &extfrag_op);
1180 }
1181
1182 static const struct file_operations extfrag_file_ops = {
1183         .open           = extfrag_open,
1184         .read           = seq_read,
1185         .llseek         = seq_lseek,
1186         .release        = seq_release,
1187 };
1188
1189 static int __init extfrag_debug_init(void)
1190 {
1191         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1192         if (!extfrag_debug_root)
1193                 return -ENOMEM;
1194
1195         if (!debugfs_create_file("unusable_index", 0444,
1196                         extfrag_debug_root, NULL, &unusable_file_ops))
1197                 return -ENOMEM;
1198
1199         if (!debugfs_create_file("extfrag_index", 0444,
1200                         extfrag_debug_root, NULL, &extfrag_file_ops))
1201                 return -ENOMEM;
1202
1203         return 0;
1204 }
1205
1206 module_init(extfrag_debug_init);
1207 #endif