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