Loading...
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 * Copyright (C) 2008-2014 Christoph Lameter
11 */
12#include <linux/fs.h>
13#include <linux/mm.h>
14#include <linux/err.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/cpu.h>
18#include <linux/cpumask.h>
19#include <linux/vmstat.h>
20#include <linux/proc_fs.h>
21#include <linux/seq_file.h>
22#include <linux/debugfs.h>
23#include <linux/sched.h>
24#include <linux/math64.h>
25#include <linux/writeback.h>
26#include <linux/compaction.h>
27#include <linux/mm_inline.h>
28#include <linux/page_ext.h>
29#include <linux/page_owner.h>
30
31#include "internal.h"
32
33#define NUMA_STATS_THRESHOLD (U16_MAX - 2)
34
35#ifdef CONFIG_NUMA
36int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
37
38/* zero numa counters within a zone */
39static void zero_zone_numa_counters(struct zone *zone)
40{
41 int item, cpu;
42
43 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
44 atomic_long_set(&zone->vm_numa_stat[item], 0);
45 for_each_online_cpu(cpu)
46 per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
47 = 0;
48 }
49}
50
51/* zero numa counters of all the populated zones */
52static void zero_zones_numa_counters(void)
53{
54 struct zone *zone;
55
56 for_each_populated_zone(zone)
57 zero_zone_numa_counters(zone);
58}
59
60/* zero global numa counters */
61static void zero_global_numa_counters(void)
62{
63 int item;
64
65 for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
66 atomic_long_set(&vm_numa_stat[item], 0);
67}
68
69static void invalid_numa_statistics(void)
70{
71 zero_zones_numa_counters();
72 zero_global_numa_counters();
73}
74
75static DEFINE_MUTEX(vm_numa_stat_lock);
76
77int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
78 void __user *buffer, size_t *length, loff_t *ppos)
79{
80 int ret, oldval;
81
82 mutex_lock(&vm_numa_stat_lock);
83 if (write)
84 oldval = sysctl_vm_numa_stat;
85 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
86 if (ret || !write)
87 goto out;
88
89 if (oldval == sysctl_vm_numa_stat)
90 goto out;
91 else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
92 static_branch_enable(&vm_numa_stat_key);
93 pr_info("enable numa statistics\n");
94 } else {
95 static_branch_disable(&vm_numa_stat_key);
96 invalid_numa_statistics();
97 pr_info("disable numa statistics, and clear numa counters\n");
98 }
99
100out:
101 mutex_unlock(&vm_numa_stat_lock);
102 return ret;
103}
104#endif
105
106#ifdef CONFIG_VM_EVENT_COUNTERS
107DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
108EXPORT_PER_CPU_SYMBOL(vm_event_states);
109
110static void sum_vm_events(unsigned long *ret)
111{
112 int cpu;
113 int i;
114
115 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
116
117 for_each_online_cpu(cpu) {
118 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
119
120 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
121 ret[i] += this->event[i];
122 }
123}
124
125/*
126 * Accumulate the vm event counters across all CPUs.
127 * The result is unavoidably approximate - it can change
128 * during and after execution of this function.
129*/
130void all_vm_events(unsigned long *ret)
131{
132 get_online_cpus();
133 sum_vm_events(ret);
134 put_online_cpus();
135}
136EXPORT_SYMBOL_GPL(all_vm_events);
137
138/*
139 * Fold the foreign cpu events into our own.
140 *
141 * This is adding to the events on one processor
142 * but keeps the global counts constant.
143 */
144void vm_events_fold_cpu(int cpu)
145{
146 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
147 int i;
148
149 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
150 count_vm_events(i, fold_state->event[i]);
151 fold_state->event[i] = 0;
152 }
153}
154
155#endif /* CONFIG_VM_EVENT_COUNTERS */
156
157/*
158 * Manage combined zone based / global counters
159 *
160 * vm_stat contains the global counters
161 */
162atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
163atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
164atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
165EXPORT_SYMBOL(vm_zone_stat);
166EXPORT_SYMBOL(vm_numa_stat);
167EXPORT_SYMBOL(vm_node_stat);
168
169#ifdef CONFIG_SMP
170
171int calculate_pressure_threshold(struct zone *zone)
172{
173 int threshold;
174 int watermark_distance;
175
176 /*
177 * As vmstats are not up to date, there is drift between the estimated
178 * and real values. For high thresholds and a high number of CPUs, it
179 * is possible for the min watermark to be breached while the estimated
180 * value looks fine. The pressure threshold is a reduced value such
181 * that even the maximum amount of drift will not accidentally breach
182 * the min watermark
183 */
184 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
185 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
186
187 /*
188 * Maximum threshold is 125
189 */
190 threshold = min(125, threshold);
191
192 return threshold;
193}
194
195int calculate_normal_threshold(struct zone *zone)
196{
197 int threshold;
198 int mem; /* memory in 128 MB units */
199
200 /*
201 * The threshold scales with the number of processors and the amount
202 * of memory per zone. More memory means that we can defer updates for
203 * longer, more processors could lead to more contention.
204 * fls() is used to have a cheap way of logarithmic scaling.
205 *
206 * Some sample thresholds:
207 *
208 * Threshold Processors (fls) Zonesize fls(mem+1)
209 * ------------------------------------------------------------------
210 * 8 1 1 0.9-1 GB 4
211 * 16 2 2 0.9-1 GB 4
212 * 20 2 2 1-2 GB 5
213 * 24 2 2 2-4 GB 6
214 * 28 2 2 4-8 GB 7
215 * 32 2 2 8-16 GB 8
216 * 4 2 2 <128M 1
217 * 30 4 3 2-4 GB 5
218 * 48 4 3 8-16 GB 8
219 * 32 8 4 1-2 GB 4
220 * 32 8 4 0.9-1GB 4
221 * 10 16 5 <128M 1
222 * 40 16 5 900M 4
223 * 70 64 7 2-4 GB 5
224 * 84 64 7 4-8 GB 6
225 * 108 512 9 4-8 GB 6
226 * 125 1024 10 8-16 GB 8
227 * 125 1024 10 16-32 GB 9
228 */
229
230 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
231
232 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
233
234 /*
235 * Maximum threshold is 125
236 */
237 threshold = min(125, threshold);
238
239 return threshold;
240}
241
242/*
243 * Refresh the thresholds for each zone.
244 */
245void refresh_zone_stat_thresholds(void)
246{
247 struct pglist_data *pgdat;
248 struct zone *zone;
249 int cpu;
250 int threshold;
251
252 /* Zero current pgdat thresholds */
253 for_each_online_pgdat(pgdat) {
254 for_each_online_cpu(cpu) {
255 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
256 }
257 }
258
259 for_each_populated_zone(zone) {
260 struct pglist_data *pgdat = zone->zone_pgdat;
261 unsigned long max_drift, tolerate_drift;
262
263 threshold = calculate_normal_threshold(zone);
264
265 for_each_online_cpu(cpu) {
266 int pgdat_threshold;
267
268 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
269 = threshold;
270
271 /* Base nodestat threshold on the largest populated zone. */
272 pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
273 per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
274 = max(threshold, pgdat_threshold);
275 }
276
277 /*
278 * Only set percpu_drift_mark if there is a danger that
279 * NR_FREE_PAGES reports the low watermark is ok when in fact
280 * the min watermark could be breached by an allocation
281 */
282 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
283 max_drift = num_online_cpus() * threshold;
284 if (max_drift > tolerate_drift)
285 zone->percpu_drift_mark = high_wmark_pages(zone) +
286 max_drift;
287 }
288}
289
290void set_pgdat_percpu_threshold(pg_data_t *pgdat,
291 int (*calculate_pressure)(struct zone *))
292{
293 struct zone *zone;
294 int cpu;
295 int threshold;
296 int i;
297
298 for (i = 0; i < pgdat->nr_zones; i++) {
299 zone = &pgdat->node_zones[i];
300 if (!zone->percpu_drift_mark)
301 continue;
302
303 threshold = (*calculate_pressure)(zone);
304 for_each_online_cpu(cpu)
305 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
306 = threshold;
307 }
308}
309
310/*
311 * For use when we know that interrupts are disabled,
312 * or when we know that preemption is disabled and that
313 * particular counter cannot be updated from interrupt context.
314 */
315void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
316 long delta)
317{
318 struct per_cpu_pageset __percpu *pcp = zone->pageset;
319 s8 __percpu *p = pcp->vm_stat_diff + item;
320 long x;
321 long t;
322
323 x = delta + __this_cpu_read(*p);
324
325 t = __this_cpu_read(pcp->stat_threshold);
326
327 if (unlikely(x > t || x < -t)) {
328 zone_page_state_add(x, zone, item);
329 x = 0;
330 }
331 __this_cpu_write(*p, x);
332}
333EXPORT_SYMBOL(__mod_zone_page_state);
334
335void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
336 long delta)
337{
338 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
339 s8 __percpu *p = pcp->vm_node_stat_diff + item;
340 long x;
341 long t;
342
343 x = delta + __this_cpu_read(*p);
344
345 t = __this_cpu_read(pcp->stat_threshold);
346
347 if (unlikely(x > t || x < -t)) {
348 node_page_state_add(x, pgdat, item);
349 x = 0;
350 }
351 __this_cpu_write(*p, x);
352}
353EXPORT_SYMBOL(__mod_node_page_state);
354
355/*
356 * Optimized increment and decrement functions.
357 *
358 * These are only for a single page and therefore can take a struct page *
359 * argument instead of struct zone *. This allows the inclusion of the code
360 * generated for page_zone(page) into the optimized functions.
361 *
362 * No overflow check is necessary and therefore the differential can be
363 * incremented or decremented in place which may allow the compilers to
364 * generate better code.
365 * The increment or decrement is known and therefore one boundary check can
366 * be omitted.
367 *
368 * NOTE: These functions are very performance sensitive. Change only
369 * with care.
370 *
371 * Some processors have inc/dec instructions that are atomic vs an interrupt.
372 * However, the code must first determine the differential location in a zone
373 * based on the processor number and then inc/dec the counter. There is no
374 * guarantee without disabling preemption that the processor will not change
375 * in between and therefore the atomicity vs. interrupt cannot be exploited
376 * in a useful way here.
377 */
378void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
379{
380 struct per_cpu_pageset __percpu *pcp = zone->pageset;
381 s8 __percpu *p = pcp->vm_stat_diff + item;
382 s8 v, t;
383
384 v = __this_cpu_inc_return(*p);
385 t = __this_cpu_read(pcp->stat_threshold);
386 if (unlikely(v > t)) {
387 s8 overstep = t >> 1;
388
389 zone_page_state_add(v + overstep, zone, item);
390 __this_cpu_write(*p, -overstep);
391 }
392}
393
394void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
395{
396 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
397 s8 __percpu *p = pcp->vm_node_stat_diff + item;
398 s8 v, t;
399
400 v = __this_cpu_inc_return(*p);
401 t = __this_cpu_read(pcp->stat_threshold);
402 if (unlikely(v > t)) {
403 s8 overstep = t >> 1;
404
405 node_page_state_add(v + overstep, pgdat, item);
406 __this_cpu_write(*p, -overstep);
407 }
408}
409
410void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
411{
412 __inc_zone_state(page_zone(page), item);
413}
414EXPORT_SYMBOL(__inc_zone_page_state);
415
416void __inc_node_page_state(struct page *page, enum node_stat_item item)
417{
418 __inc_node_state(page_pgdat(page), item);
419}
420EXPORT_SYMBOL(__inc_node_page_state);
421
422void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
423{
424 struct per_cpu_pageset __percpu *pcp = zone->pageset;
425 s8 __percpu *p = pcp->vm_stat_diff + item;
426 s8 v, t;
427
428 v = __this_cpu_dec_return(*p);
429 t = __this_cpu_read(pcp->stat_threshold);
430 if (unlikely(v < - t)) {
431 s8 overstep = t >> 1;
432
433 zone_page_state_add(v - overstep, zone, item);
434 __this_cpu_write(*p, overstep);
435 }
436}
437
438void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
439{
440 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
441 s8 __percpu *p = pcp->vm_node_stat_diff + item;
442 s8 v, t;
443
444 v = __this_cpu_dec_return(*p);
445 t = __this_cpu_read(pcp->stat_threshold);
446 if (unlikely(v < - t)) {
447 s8 overstep = t >> 1;
448
449 node_page_state_add(v - overstep, pgdat, item);
450 __this_cpu_write(*p, overstep);
451 }
452}
453
454void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
455{
456 __dec_zone_state(page_zone(page), item);
457}
458EXPORT_SYMBOL(__dec_zone_page_state);
459
460void __dec_node_page_state(struct page *page, enum node_stat_item item)
461{
462 __dec_node_state(page_pgdat(page), item);
463}
464EXPORT_SYMBOL(__dec_node_page_state);
465
466#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
467/*
468 * If we have cmpxchg_local support then we do not need to incur the overhead
469 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
470 *
471 * mod_state() modifies the zone counter state through atomic per cpu
472 * operations.
473 *
474 * Overstep mode specifies how overstep should handled:
475 * 0 No overstepping
476 * 1 Overstepping half of threshold
477 * -1 Overstepping minus half of threshold
478*/
479static inline void mod_zone_state(struct zone *zone,
480 enum zone_stat_item item, long delta, int overstep_mode)
481{
482 struct per_cpu_pageset __percpu *pcp = zone->pageset;
483 s8 __percpu *p = pcp->vm_stat_diff + item;
484 long o, n, t, z;
485
486 do {
487 z = 0; /* overflow to zone counters */
488
489 /*
490 * The fetching of the stat_threshold is racy. We may apply
491 * a counter threshold to the wrong the cpu if we get
492 * rescheduled while executing here. However, the next
493 * counter update will apply the threshold again and
494 * therefore bring the counter under the threshold again.
495 *
496 * Most of the time the thresholds are the same anyways
497 * for all cpus in a zone.
498 */
499 t = this_cpu_read(pcp->stat_threshold);
500
501 o = this_cpu_read(*p);
502 n = delta + o;
503
504 if (n > t || n < -t) {
505 int os = overstep_mode * (t >> 1) ;
506
507 /* Overflow must be added to zone counters */
508 z = n + os;
509 n = -os;
510 }
511 } while (this_cpu_cmpxchg(*p, o, n) != o);
512
513 if (z)
514 zone_page_state_add(z, zone, item);
515}
516
517void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
518 long delta)
519{
520 mod_zone_state(zone, item, delta, 0);
521}
522EXPORT_SYMBOL(mod_zone_page_state);
523
524void inc_zone_page_state(struct page *page, enum zone_stat_item item)
525{
526 mod_zone_state(page_zone(page), item, 1, 1);
527}
528EXPORT_SYMBOL(inc_zone_page_state);
529
530void dec_zone_page_state(struct page *page, enum zone_stat_item item)
531{
532 mod_zone_state(page_zone(page), item, -1, -1);
533}
534EXPORT_SYMBOL(dec_zone_page_state);
535
536static inline void mod_node_state(struct pglist_data *pgdat,
537 enum node_stat_item item, int delta, int overstep_mode)
538{
539 struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
540 s8 __percpu *p = pcp->vm_node_stat_diff + item;
541 long o, n, t, z;
542
543 do {
544 z = 0; /* overflow to node counters */
545
546 /*
547 * The fetching of the stat_threshold is racy. We may apply
548 * a counter threshold to the wrong the cpu if we get
549 * rescheduled while executing here. However, the next
550 * counter update will apply the threshold again and
551 * therefore bring the counter under the threshold again.
552 *
553 * Most of the time the thresholds are the same anyways
554 * for all cpus in a node.
555 */
556 t = this_cpu_read(pcp->stat_threshold);
557
558 o = this_cpu_read(*p);
559 n = delta + o;
560
561 if (n > t || n < -t) {
562 int os = overstep_mode * (t >> 1) ;
563
564 /* Overflow must be added to node counters */
565 z = n + os;
566 n = -os;
567 }
568 } while (this_cpu_cmpxchg(*p, o, n) != o);
569
570 if (z)
571 node_page_state_add(z, pgdat, item);
572}
573
574void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
575 long delta)
576{
577 mod_node_state(pgdat, item, delta, 0);
578}
579EXPORT_SYMBOL(mod_node_page_state);
580
581void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
582{
583 mod_node_state(pgdat, item, 1, 1);
584}
585
586void inc_node_page_state(struct page *page, enum node_stat_item item)
587{
588 mod_node_state(page_pgdat(page), item, 1, 1);
589}
590EXPORT_SYMBOL(inc_node_page_state);
591
592void dec_node_page_state(struct page *page, enum node_stat_item item)
593{
594 mod_node_state(page_pgdat(page), item, -1, -1);
595}
596EXPORT_SYMBOL(dec_node_page_state);
597#else
598/*
599 * Use interrupt disable to serialize counter updates
600 */
601void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
602 long delta)
603{
604 unsigned long flags;
605
606 local_irq_save(flags);
607 __mod_zone_page_state(zone, item, delta);
608 local_irq_restore(flags);
609}
610EXPORT_SYMBOL(mod_zone_page_state);
611
612void inc_zone_page_state(struct page *page, enum zone_stat_item item)
613{
614 unsigned long flags;
615 struct zone *zone;
616
617 zone = page_zone(page);
618 local_irq_save(flags);
619 __inc_zone_state(zone, item);
620 local_irq_restore(flags);
621}
622EXPORT_SYMBOL(inc_zone_page_state);
623
624void dec_zone_page_state(struct page *page, enum zone_stat_item item)
625{
626 unsigned long flags;
627
628 local_irq_save(flags);
629 __dec_zone_page_state(page, item);
630 local_irq_restore(flags);
631}
632EXPORT_SYMBOL(dec_zone_page_state);
633
634void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
635{
636 unsigned long flags;
637
638 local_irq_save(flags);
639 __inc_node_state(pgdat, item);
640 local_irq_restore(flags);
641}
642EXPORT_SYMBOL(inc_node_state);
643
644void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
645 long delta)
646{
647 unsigned long flags;
648
649 local_irq_save(flags);
650 __mod_node_page_state(pgdat, item, delta);
651 local_irq_restore(flags);
652}
653EXPORT_SYMBOL(mod_node_page_state);
654
655void inc_node_page_state(struct page *page, enum node_stat_item item)
656{
657 unsigned long flags;
658 struct pglist_data *pgdat;
659
660 pgdat = page_pgdat(page);
661 local_irq_save(flags);
662 __inc_node_state(pgdat, item);
663 local_irq_restore(flags);
664}
665EXPORT_SYMBOL(inc_node_page_state);
666
667void dec_node_page_state(struct page *page, enum node_stat_item item)
668{
669 unsigned long flags;
670
671 local_irq_save(flags);
672 __dec_node_page_state(page, item);
673 local_irq_restore(flags);
674}
675EXPORT_SYMBOL(dec_node_page_state);
676#endif
677
678/*
679 * Fold a differential into the global counters.
680 * Returns the number of counters updated.
681 */
682#ifdef CONFIG_NUMA
683static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
684{
685 int i;
686 int changes = 0;
687
688 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
689 if (zone_diff[i]) {
690 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
691 changes++;
692 }
693
694 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
695 if (numa_diff[i]) {
696 atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
697 changes++;
698 }
699
700 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
701 if (node_diff[i]) {
702 atomic_long_add(node_diff[i], &vm_node_stat[i]);
703 changes++;
704 }
705 return changes;
706}
707#else
708static int fold_diff(int *zone_diff, int *node_diff)
709{
710 int i;
711 int changes = 0;
712
713 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
714 if (zone_diff[i]) {
715 atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
716 changes++;
717 }
718
719 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
720 if (node_diff[i]) {
721 atomic_long_add(node_diff[i], &vm_node_stat[i]);
722 changes++;
723 }
724 return changes;
725}
726#endif /* CONFIG_NUMA */
727
728/*
729 * Update the zone counters for the current cpu.
730 *
731 * Note that refresh_cpu_vm_stats strives to only access
732 * node local memory. The per cpu pagesets on remote zones are placed
733 * in the memory local to the processor using that pageset. So the
734 * loop over all zones will access a series of cachelines local to
735 * the processor.
736 *
737 * The call to zone_page_state_add updates the cachelines with the
738 * statistics in the remote zone struct as well as the global cachelines
739 * with the global counters. These could cause remote node cache line
740 * bouncing and will have to be only done when necessary.
741 *
742 * The function returns the number of global counters updated.
743 */
744static int refresh_cpu_vm_stats(bool do_pagesets)
745{
746 struct pglist_data *pgdat;
747 struct zone *zone;
748 int i;
749 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
750#ifdef CONFIG_NUMA
751 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
752#endif
753 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
754 int changes = 0;
755
756 for_each_populated_zone(zone) {
757 struct per_cpu_pageset __percpu *p = zone->pageset;
758
759 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
760 int v;
761
762 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
763 if (v) {
764
765 atomic_long_add(v, &zone->vm_stat[i]);
766 global_zone_diff[i] += v;
767#ifdef CONFIG_NUMA
768 /* 3 seconds idle till flush */
769 __this_cpu_write(p->expire, 3);
770#endif
771 }
772 }
773#ifdef CONFIG_NUMA
774 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
775 int v;
776
777 v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
778 if (v) {
779
780 atomic_long_add(v, &zone->vm_numa_stat[i]);
781 global_numa_diff[i] += v;
782 __this_cpu_write(p->expire, 3);
783 }
784 }
785
786 if (do_pagesets) {
787 cond_resched();
788 /*
789 * Deal with draining the remote pageset of this
790 * processor
791 *
792 * Check if there are pages remaining in this pageset
793 * if not then there is nothing to expire.
794 */
795 if (!__this_cpu_read(p->expire) ||
796 !__this_cpu_read(p->pcp.count))
797 continue;
798
799 /*
800 * We never drain zones local to this processor.
801 */
802 if (zone_to_nid(zone) == numa_node_id()) {
803 __this_cpu_write(p->expire, 0);
804 continue;
805 }
806
807 if (__this_cpu_dec_return(p->expire))
808 continue;
809
810 if (__this_cpu_read(p->pcp.count)) {
811 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
812 changes++;
813 }
814 }
815#endif
816 }
817
818 for_each_online_pgdat(pgdat) {
819 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
820
821 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
822 int v;
823
824 v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
825 if (v) {
826 atomic_long_add(v, &pgdat->vm_stat[i]);
827 global_node_diff[i] += v;
828 }
829 }
830 }
831
832#ifdef CONFIG_NUMA
833 changes += fold_diff(global_zone_diff, global_numa_diff,
834 global_node_diff);
835#else
836 changes += fold_diff(global_zone_diff, global_node_diff);
837#endif
838 return changes;
839}
840
841/*
842 * Fold the data for an offline cpu into the global array.
843 * There cannot be any access by the offline cpu and therefore
844 * synchronization is simplified.
845 */
846void cpu_vm_stats_fold(int cpu)
847{
848 struct pglist_data *pgdat;
849 struct zone *zone;
850 int i;
851 int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
852#ifdef CONFIG_NUMA
853 int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
854#endif
855 int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
856
857 for_each_populated_zone(zone) {
858 struct per_cpu_pageset *p;
859
860 p = per_cpu_ptr(zone->pageset, cpu);
861
862 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
863 if (p->vm_stat_diff[i]) {
864 int v;
865
866 v = p->vm_stat_diff[i];
867 p->vm_stat_diff[i] = 0;
868 atomic_long_add(v, &zone->vm_stat[i]);
869 global_zone_diff[i] += v;
870 }
871
872#ifdef CONFIG_NUMA
873 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
874 if (p->vm_numa_stat_diff[i]) {
875 int v;
876
877 v = p->vm_numa_stat_diff[i];
878 p->vm_numa_stat_diff[i] = 0;
879 atomic_long_add(v, &zone->vm_numa_stat[i]);
880 global_numa_diff[i] += v;
881 }
882#endif
883 }
884
885 for_each_online_pgdat(pgdat) {
886 struct per_cpu_nodestat *p;
887
888 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
889
890 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
891 if (p->vm_node_stat_diff[i]) {
892 int v;
893
894 v = p->vm_node_stat_diff[i];
895 p->vm_node_stat_diff[i] = 0;
896 atomic_long_add(v, &pgdat->vm_stat[i]);
897 global_node_diff[i] += v;
898 }
899 }
900
901#ifdef CONFIG_NUMA
902 fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
903#else
904 fold_diff(global_zone_diff, global_node_diff);
905#endif
906}
907
908/*
909 * this is only called if !populated_zone(zone), which implies no other users of
910 * pset->vm_stat_diff[] exsist.
911 */
912void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
913{
914 int i;
915
916 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
917 if (pset->vm_stat_diff[i]) {
918 int v = pset->vm_stat_diff[i];
919 pset->vm_stat_diff[i] = 0;
920 atomic_long_add(v, &zone->vm_stat[i]);
921 atomic_long_add(v, &vm_zone_stat[i]);
922 }
923
924#ifdef CONFIG_NUMA
925 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
926 if (pset->vm_numa_stat_diff[i]) {
927 int v = pset->vm_numa_stat_diff[i];
928
929 pset->vm_numa_stat_diff[i] = 0;
930 atomic_long_add(v, &zone->vm_numa_stat[i]);
931 atomic_long_add(v, &vm_numa_stat[i]);
932 }
933#endif
934}
935#endif
936
937#ifdef CONFIG_NUMA
938void __inc_numa_state(struct zone *zone,
939 enum numa_stat_item item)
940{
941 struct per_cpu_pageset __percpu *pcp = zone->pageset;
942 u16 __percpu *p = pcp->vm_numa_stat_diff + item;
943 u16 v;
944
945 v = __this_cpu_inc_return(*p);
946
947 if (unlikely(v > NUMA_STATS_THRESHOLD)) {
948 zone_numa_state_add(v, zone, item);
949 __this_cpu_write(*p, 0);
950 }
951}
952
953/*
954 * Determine the per node value of a stat item. This function
955 * is called frequently in a NUMA machine, so try to be as
956 * frugal as possible.
957 */
958unsigned long sum_zone_node_page_state(int node,
959 enum zone_stat_item item)
960{
961 struct zone *zones = NODE_DATA(node)->node_zones;
962 int i;
963 unsigned long count = 0;
964
965 for (i = 0; i < MAX_NR_ZONES; i++)
966 count += zone_page_state(zones + i, item);
967
968 return count;
969}
970
971/*
972 * Determine the per node value of a numa stat item. To avoid deviation,
973 * the per cpu stat number in vm_numa_stat_diff[] is also included.
974 */
975unsigned long sum_zone_numa_state(int node,
976 enum numa_stat_item item)
977{
978 struct zone *zones = NODE_DATA(node)->node_zones;
979 int i;
980 unsigned long count = 0;
981
982 for (i = 0; i < MAX_NR_ZONES; i++)
983 count += zone_numa_state_snapshot(zones + i, item);
984
985 return count;
986}
987
988/*
989 * Determine the per node value of a stat item.
990 */
991unsigned long node_page_state(struct pglist_data *pgdat,
992 enum node_stat_item item)
993{
994 long x = atomic_long_read(&pgdat->vm_stat[item]);
995#ifdef CONFIG_SMP
996 if (x < 0)
997 x = 0;
998#endif
999 return x;
1000}
1001#endif
1002
1003#ifdef CONFIG_COMPACTION
1004
1005struct contig_page_info {
1006 unsigned long free_pages;
1007 unsigned long free_blocks_total;
1008 unsigned long free_blocks_suitable;
1009};
1010
1011/*
1012 * Calculate the number of free pages in a zone, how many contiguous
1013 * pages are free and how many are large enough to satisfy an allocation of
1014 * the target size. Note that this function makes no attempt to estimate
1015 * how many suitable free blocks there *might* be if MOVABLE pages were
1016 * migrated. Calculating that is possible, but expensive and can be
1017 * figured out from userspace
1018 */
1019static void fill_contig_page_info(struct zone *zone,
1020 unsigned int suitable_order,
1021 struct contig_page_info *info)
1022{
1023 unsigned int order;
1024
1025 info->free_pages = 0;
1026 info->free_blocks_total = 0;
1027 info->free_blocks_suitable = 0;
1028
1029 for (order = 0; order < MAX_ORDER; order++) {
1030 unsigned long blocks;
1031
1032 /* Count number of free blocks */
1033 blocks = zone->free_area[order].nr_free;
1034 info->free_blocks_total += blocks;
1035
1036 /* Count free base pages */
1037 info->free_pages += blocks << order;
1038
1039 /* Count the suitable free blocks */
1040 if (order >= suitable_order)
1041 info->free_blocks_suitable += blocks <<
1042 (order - suitable_order);
1043 }
1044}
1045
1046/*
1047 * A fragmentation index only makes sense if an allocation of a requested
1048 * size would fail. If that is true, the fragmentation index indicates
1049 * whether external fragmentation or a lack of memory was the problem.
1050 * The value can be used to determine if page reclaim or compaction
1051 * should be used
1052 */
1053static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1054{
1055 unsigned long requested = 1UL << order;
1056
1057 if (WARN_ON_ONCE(order >= MAX_ORDER))
1058 return 0;
1059
1060 if (!info->free_blocks_total)
1061 return 0;
1062
1063 /* Fragmentation index only makes sense when a request would fail */
1064 if (info->free_blocks_suitable)
1065 return -1000;
1066
1067 /*
1068 * Index is between 0 and 1 so return within 3 decimal places
1069 *
1070 * 0 => allocation would fail due to lack of memory
1071 * 1 => allocation would fail due to fragmentation
1072 */
1073 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1074}
1075
1076/* Same as __fragmentation index but allocs contig_page_info on stack */
1077int fragmentation_index(struct zone *zone, unsigned int order)
1078{
1079 struct contig_page_info info;
1080
1081 fill_contig_page_info(zone, order, &info);
1082 return __fragmentation_index(order, &info);
1083}
1084#endif
1085
1086#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
1087#ifdef CONFIG_ZONE_DMA
1088#define TEXT_FOR_DMA(xx) xx "_dma",
1089#else
1090#define TEXT_FOR_DMA(xx)
1091#endif
1092
1093#ifdef CONFIG_ZONE_DMA32
1094#define TEXT_FOR_DMA32(xx) xx "_dma32",
1095#else
1096#define TEXT_FOR_DMA32(xx)
1097#endif
1098
1099#ifdef CONFIG_HIGHMEM
1100#define TEXT_FOR_HIGHMEM(xx) xx "_high",
1101#else
1102#define TEXT_FOR_HIGHMEM(xx)
1103#endif
1104
1105#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1106 TEXT_FOR_HIGHMEM(xx) xx "_movable",
1107
1108const char * const vmstat_text[] = {
1109 /* enum zone_stat_item countes */
1110 "nr_free_pages",
1111 "nr_zone_inactive_anon",
1112 "nr_zone_active_anon",
1113 "nr_zone_inactive_file",
1114 "nr_zone_active_file",
1115 "nr_zone_unevictable",
1116 "nr_zone_write_pending",
1117 "nr_mlock",
1118 "nr_page_table_pages",
1119 "nr_kernel_stack",
1120 "nr_bounce",
1121#if IS_ENABLED(CONFIG_ZSMALLOC)
1122 "nr_zspages",
1123#endif
1124 "nr_free_cma",
1125
1126 /* enum numa_stat_item counters */
1127#ifdef CONFIG_NUMA
1128 "numa_hit",
1129 "numa_miss",
1130 "numa_foreign",
1131 "numa_interleave",
1132 "numa_local",
1133 "numa_other",
1134#endif
1135
1136 /* Node-based counters */
1137 "nr_inactive_anon",
1138 "nr_active_anon",
1139 "nr_inactive_file",
1140 "nr_active_file",
1141 "nr_unevictable",
1142 "nr_slab_reclaimable",
1143 "nr_slab_unreclaimable",
1144 "nr_isolated_anon",
1145 "nr_isolated_file",
1146 "workingset_refault",
1147 "workingset_activate",
1148 "workingset_nodereclaim",
1149 "nr_anon_pages",
1150 "nr_mapped",
1151 "nr_file_pages",
1152 "nr_dirty",
1153 "nr_writeback",
1154 "nr_writeback_temp",
1155 "nr_shmem",
1156 "nr_shmem_hugepages",
1157 "nr_shmem_pmdmapped",
1158 "nr_anon_transparent_hugepages",
1159 "nr_unstable",
1160 "nr_vmscan_write",
1161 "nr_vmscan_immediate_reclaim",
1162 "nr_dirtied",
1163 "nr_written",
1164 "", /* nr_indirectly_reclaimable */
1165
1166 /* enum writeback_stat_item counters */
1167 "nr_dirty_threshold",
1168 "nr_dirty_background_threshold",
1169
1170#ifdef CONFIG_VM_EVENT_COUNTERS
1171 /* enum vm_event_item counters */
1172 "pgpgin",
1173 "pgpgout",
1174 "pswpin",
1175 "pswpout",
1176
1177 TEXTS_FOR_ZONES("pgalloc")
1178 TEXTS_FOR_ZONES("allocstall")
1179 TEXTS_FOR_ZONES("pgskip")
1180
1181 "pgfree",
1182 "pgactivate",
1183 "pgdeactivate",
1184 "pglazyfree",
1185
1186 "pgfault",
1187 "pgmajfault",
1188 "pglazyfreed",
1189
1190 "pgrefill",
1191 "pgsteal_kswapd",
1192 "pgsteal_direct",
1193 "pgscan_kswapd",
1194 "pgscan_direct",
1195 "pgscan_direct_throttle",
1196
1197#ifdef CONFIG_NUMA
1198 "zone_reclaim_failed",
1199#endif
1200 "pginodesteal",
1201 "slabs_scanned",
1202 "kswapd_inodesteal",
1203 "kswapd_low_wmark_hit_quickly",
1204 "kswapd_high_wmark_hit_quickly",
1205 "pageoutrun",
1206
1207 "pgrotated",
1208
1209 "drop_pagecache",
1210 "drop_slab",
1211 "oom_kill",
1212
1213#ifdef CONFIG_NUMA_BALANCING
1214 "numa_pte_updates",
1215 "numa_huge_pte_updates",
1216 "numa_hint_faults",
1217 "numa_hint_faults_local",
1218 "numa_pages_migrated",
1219#endif
1220#ifdef CONFIG_MIGRATION
1221 "pgmigrate_success",
1222 "pgmigrate_fail",
1223#endif
1224#ifdef CONFIG_COMPACTION
1225 "compact_migrate_scanned",
1226 "compact_free_scanned",
1227 "compact_isolated",
1228 "compact_stall",
1229 "compact_fail",
1230 "compact_success",
1231 "compact_daemon_wake",
1232 "compact_daemon_migrate_scanned",
1233 "compact_daemon_free_scanned",
1234#endif
1235
1236#ifdef CONFIG_HUGETLB_PAGE
1237 "htlb_buddy_alloc_success",
1238 "htlb_buddy_alloc_fail",
1239#endif
1240 "unevictable_pgs_culled",
1241 "unevictable_pgs_scanned",
1242 "unevictable_pgs_rescued",
1243 "unevictable_pgs_mlocked",
1244 "unevictable_pgs_munlocked",
1245 "unevictable_pgs_cleared",
1246 "unevictable_pgs_stranded",
1247
1248#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1249 "thp_fault_alloc",
1250 "thp_fault_fallback",
1251 "thp_collapse_alloc",
1252 "thp_collapse_alloc_failed",
1253 "thp_file_alloc",
1254 "thp_file_mapped",
1255 "thp_split_page",
1256 "thp_split_page_failed",
1257 "thp_deferred_split_page",
1258 "thp_split_pmd",
1259#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1260 "thp_split_pud",
1261#endif
1262 "thp_zero_page_alloc",
1263 "thp_zero_page_alloc_failed",
1264 "thp_swpout",
1265 "thp_swpout_fallback",
1266#endif
1267#ifdef CONFIG_MEMORY_BALLOON
1268 "balloon_inflate",
1269 "balloon_deflate",
1270#ifdef CONFIG_BALLOON_COMPACTION
1271 "balloon_migrate",
1272#endif
1273#endif /* CONFIG_MEMORY_BALLOON */
1274#ifdef CONFIG_DEBUG_TLBFLUSH
1275#ifdef CONFIG_SMP
1276 "nr_tlb_remote_flush",
1277 "nr_tlb_remote_flush_received",
1278#endif /* CONFIG_SMP */
1279 "nr_tlb_local_flush_all",
1280 "nr_tlb_local_flush_one",
1281#endif /* CONFIG_DEBUG_TLBFLUSH */
1282
1283#ifdef CONFIG_DEBUG_VM_VMACACHE
1284 "vmacache_find_calls",
1285 "vmacache_find_hits",
1286 "vmacache_full_flushes",
1287#endif
1288#ifdef CONFIG_SWAP
1289 "swap_ra",
1290 "swap_ra_hit",
1291#endif
1292#endif /* CONFIG_VM_EVENTS_COUNTERS */
1293};
1294#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
1295
1296#if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1297 defined(CONFIG_PROC_FS)
1298static void *frag_start(struct seq_file *m, loff_t *pos)
1299{
1300 pg_data_t *pgdat;
1301 loff_t node = *pos;
1302
1303 for (pgdat = first_online_pgdat();
1304 pgdat && node;
1305 pgdat = next_online_pgdat(pgdat))
1306 --node;
1307
1308 return pgdat;
1309}
1310
1311static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1312{
1313 pg_data_t *pgdat = (pg_data_t *)arg;
1314
1315 (*pos)++;
1316 return next_online_pgdat(pgdat);
1317}
1318
1319static void frag_stop(struct seq_file *m, void *arg)
1320{
1321}
1322
1323/*
1324 * Walk zones in a node and print using a callback.
1325 * If @assert_populated is true, only use callback for zones that are populated.
1326 */
1327static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1328 bool assert_populated, bool nolock,
1329 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1330{
1331 struct zone *zone;
1332 struct zone *node_zones = pgdat->node_zones;
1333 unsigned long flags;
1334
1335 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1336 if (assert_populated && !populated_zone(zone))
1337 continue;
1338
1339 if (!nolock)
1340 spin_lock_irqsave(&zone->lock, flags);
1341 print(m, pgdat, zone);
1342 if (!nolock)
1343 spin_unlock_irqrestore(&zone->lock, flags);
1344 }
1345}
1346#endif
1347
1348#ifdef CONFIG_PROC_FS
1349static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1350 struct zone *zone)
1351{
1352 int order;
1353
1354 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1355 for (order = 0; order < MAX_ORDER; ++order)
1356 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1357 seq_putc(m, '\n');
1358}
1359
1360/*
1361 * This walks the free areas for each zone.
1362 */
1363static int frag_show(struct seq_file *m, void *arg)
1364{
1365 pg_data_t *pgdat = (pg_data_t *)arg;
1366 walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1367 return 0;
1368}
1369
1370static void pagetypeinfo_showfree_print(struct seq_file *m,
1371 pg_data_t *pgdat, struct zone *zone)
1372{
1373 int order, mtype;
1374
1375 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1376 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1377 pgdat->node_id,
1378 zone->name,
1379 migratetype_names[mtype]);
1380 for (order = 0; order < MAX_ORDER; ++order) {
1381 unsigned long freecount = 0;
1382 struct free_area *area;
1383 struct list_head *curr;
1384
1385 area = &(zone->free_area[order]);
1386
1387 list_for_each(curr, &area->free_list[mtype])
1388 freecount++;
1389 seq_printf(m, "%6lu ", freecount);
1390 }
1391 seq_putc(m, '\n');
1392 }
1393}
1394
1395/* Print out the free pages at each order for each migatetype */
1396static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1397{
1398 int order;
1399 pg_data_t *pgdat = (pg_data_t *)arg;
1400
1401 /* Print header */
1402 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1403 for (order = 0; order < MAX_ORDER; ++order)
1404 seq_printf(m, "%6d ", order);
1405 seq_putc(m, '\n');
1406
1407 walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1408
1409 return 0;
1410}
1411
1412static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1413 pg_data_t *pgdat, struct zone *zone)
1414{
1415 int mtype;
1416 unsigned long pfn;
1417 unsigned long start_pfn = zone->zone_start_pfn;
1418 unsigned long end_pfn = zone_end_pfn(zone);
1419 unsigned long count[MIGRATE_TYPES] = { 0, };
1420
1421 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1422 struct page *page;
1423
1424 page = pfn_to_online_page(pfn);
1425 if (!page)
1426 continue;
1427
1428 /* Watch for unexpected holes punched in the memmap */
1429 if (!memmap_valid_within(pfn, page, zone))
1430 continue;
1431
1432 if (page_zone(page) != zone)
1433 continue;
1434
1435 mtype = get_pageblock_migratetype(page);
1436
1437 if (mtype < MIGRATE_TYPES)
1438 count[mtype]++;
1439 }
1440
1441 /* Print counts */
1442 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1443 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1444 seq_printf(m, "%12lu ", count[mtype]);
1445 seq_putc(m, '\n');
1446}
1447
1448/* Print out the number of pageblocks for each migratetype */
1449static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1450{
1451 int mtype;
1452 pg_data_t *pgdat = (pg_data_t *)arg;
1453
1454 seq_printf(m, "\n%-23s", "Number of blocks type ");
1455 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1456 seq_printf(m, "%12s ", migratetype_names[mtype]);
1457 seq_putc(m, '\n');
1458 walk_zones_in_node(m, pgdat, true, false,
1459 pagetypeinfo_showblockcount_print);
1460
1461 return 0;
1462}
1463
1464/*
1465 * Print out the number of pageblocks for each migratetype that contain pages
1466 * of other types. This gives an indication of how well fallbacks are being
1467 * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1468 * to determine what is going on
1469 */
1470static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1471{
1472#ifdef CONFIG_PAGE_OWNER
1473 int mtype;
1474
1475 if (!static_branch_unlikely(&page_owner_inited))
1476 return;
1477
1478 drain_all_pages(NULL);
1479
1480 seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1481 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1482 seq_printf(m, "%12s ", migratetype_names[mtype]);
1483 seq_putc(m, '\n');
1484
1485 walk_zones_in_node(m, pgdat, true, true,
1486 pagetypeinfo_showmixedcount_print);
1487#endif /* CONFIG_PAGE_OWNER */
1488}
1489
1490/*
1491 * This prints out statistics in relation to grouping pages by mobility.
1492 * It is expensive to collect so do not constantly read the file.
1493 */
1494static int pagetypeinfo_show(struct seq_file *m, void *arg)
1495{
1496 pg_data_t *pgdat = (pg_data_t *)arg;
1497
1498 /* check memoryless node */
1499 if (!node_state(pgdat->node_id, N_MEMORY))
1500 return 0;
1501
1502 seq_printf(m, "Page block order: %d\n", pageblock_order);
1503 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1504 seq_putc(m, '\n');
1505 pagetypeinfo_showfree(m, pgdat);
1506 pagetypeinfo_showblockcount(m, pgdat);
1507 pagetypeinfo_showmixedcount(m, pgdat);
1508
1509 return 0;
1510}
1511
1512static const struct seq_operations fragmentation_op = {
1513 .start = frag_start,
1514 .next = frag_next,
1515 .stop = frag_stop,
1516 .show = frag_show,
1517};
1518
1519static int fragmentation_open(struct inode *inode, struct file *file)
1520{
1521 return seq_open(file, &fragmentation_op);
1522}
1523
1524static const struct file_operations buddyinfo_file_operations = {
1525 .open = fragmentation_open,
1526 .read = seq_read,
1527 .llseek = seq_lseek,
1528 .release = seq_release,
1529};
1530
1531static const struct seq_operations pagetypeinfo_op = {
1532 .start = frag_start,
1533 .next = frag_next,
1534 .stop = frag_stop,
1535 .show = pagetypeinfo_show,
1536};
1537
1538static int pagetypeinfo_open(struct inode *inode, struct file *file)
1539{
1540 return seq_open(file, &pagetypeinfo_op);
1541}
1542
1543static const struct file_operations pagetypeinfo_file_operations = {
1544 .open = pagetypeinfo_open,
1545 .read = seq_read,
1546 .llseek = seq_lseek,
1547 .release = seq_release,
1548};
1549
1550static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1551{
1552 int zid;
1553
1554 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1555 struct zone *compare = &pgdat->node_zones[zid];
1556
1557 if (populated_zone(compare))
1558 return zone == compare;
1559 }
1560
1561 return false;
1562}
1563
1564static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1565 struct zone *zone)
1566{
1567 int i;
1568 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1569 if (is_zone_first_populated(pgdat, zone)) {
1570 seq_printf(m, "\n per-node stats");
1571 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1572 seq_printf(m, "\n %-12s %lu",
1573 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS +
1574 NR_VM_NUMA_STAT_ITEMS],
1575 node_page_state(pgdat, i));
1576 }
1577 }
1578 seq_printf(m,
1579 "\n pages free %lu"
1580 "\n min %lu"
1581 "\n low %lu"
1582 "\n high %lu"
1583 "\n spanned %lu"
1584 "\n present %lu"
1585 "\n managed %lu",
1586 zone_page_state(zone, NR_FREE_PAGES),
1587 min_wmark_pages(zone),
1588 low_wmark_pages(zone),
1589 high_wmark_pages(zone),
1590 zone->spanned_pages,
1591 zone->present_pages,
1592 zone->managed_pages);
1593
1594 seq_printf(m,
1595 "\n protection: (%ld",
1596 zone->lowmem_reserve[0]);
1597 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1598 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1599 seq_putc(m, ')');
1600
1601 /* If unpopulated, no other information is useful */
1602 if (!populated_zone(zone)) {
1603 seq_putc(m, '\n');
1604 return;
1605 }
1606
1607 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1608 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1609 zone_page_state(zone, i));
1610
1611#ifdef CONFIG_NUMA
1612 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1613 seq_printf(m, "\n %-12s %lu",
1614 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
1615 zone_numa_state_snapshot(zone, i));
1616#endif
1617
1618 seq_printf(m, "\n pagesets");
1619 for_each_online_cpu(i) {
1620 struct per_cpu_pageset *pageset;
1621
1622 pageset = per_cpu_ptr(zone->pageset, i);
1623 seq_printf(m,
1624 "\n cpu: %i"
1625 "\n count: %i"
1626 "\n high: %i"
1627 "\n batch: %i",
1628 i,
1629 pageset->pcp.count,
1630 pageset->pcp.high,
1631 pageset->pcp.batch);
1632#ifdef CONFIG_SMP
1633 seq_printf(m, "\n vm stats threshold: %d",
1634 pageset->stat_threshold);
1635#endif
1636 }
1637 seq_printf(m,
1638 "\n node_unreclaimable: %u"
1639 "\n start_pfn: %lu",
1640 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1641 zone->zone_start_pfn);
1642 seq_putc(m, '\n');
1643}
1644
1645/*
1646 * Output information about zones in @pgdat. All zones are printed regardless
1647 * of whether they are populated or not: lowmem_reserve_ratio operates on the
1648 * set of all zones and userspace would not be aware of such zones if they are
1649 * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1650 */
1651static int zoneinfo_show(struct seq_file *m, void *arg)
1652{
1653 pg_data_t *pgdat = (pg_data_t *)arg;
1654 walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1655 return 0;
1656}
1657
1658static const struct seq_operations zoneinfo_op = {
1659 .start = frag_start, /* iterate over all zones. The same as in
1660 * fragmentation. */
1661 .next = frag_next,
1662 .stop = frag_stop,
1663 .show = zoneinfo_show,
1664};
1665
1666static int zoneinfo_open(struct inode *inode, struct file *file)
1667{
1668 return seq_open(file, &zoneinfo_op);
1669}
1670
1671static const struct file_operations zoneinfo_file_operations = {
1672 .open = zoneinfo_open,
1673 .read = seq_read,
1674 .llseek = seq_lseek,
1675 .release = seq_release,
1676};
1677
1678enum writeback_stat_item {
1679 NR_DIRTY_THRESHOLD,
1680 NR_DIRTY_BG_THRESHOLD,
1681 NR_VM_WRITEBACK_STAT_ITEMS,
1682};
1683
1684static void *vmstat_start(struct seq_file *m, loff_t *pos)
1685{
1686 unsigned long *v;
1687 int i, stat_items_size;
1688
1689 if (*pos >= ARRAY_SIZE(vmstat_text))
1690 return NULL;
1691 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1692 NR_VM_NUMA_STAT_ITEMS * sizeof(unsigned long) +
1693 NR_VM_NODE_STAT_ITEMS * sizeof(unsigned long) +
1694 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1695
1696#ifdef CONFIG_VM_EVENT_COUNTERS
1697 stat_items_size += sizeof(struct vm_event_state);
1698#endif
1699
1700 v = kmalloc(stat_items_size, GFP_KERNEL);
1701 m->private = v;
1702 if (!v)
1703 return ERR_PTR(-ENOMEM);
1704 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1705 v[i] = global_zone_page_state(i);
1706 v += NR_VM_ZONE_STAT_ITEMS;
1707
1708#ifdef CONFIG_NUMA
1709 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1710 v[i] = global_numa_state(i);
1711 v += NR_VM_NUMA_STAT_ITEMS;
1712#endif
1713
1714 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1715 v[i] = global_node_page_state(i);
1716 v += NR_VM_NODE_STAT_ITEMS;
1717
1718 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1719 v + NR_DIRTY_THRESHOLD);
1720 v += NR_VM_WRITEBACK_STAT_ITEMS;
1721
1722#ifdef CONFIG_VM_EVENT_COUNTERS
1723 all_vm_events(v);
1724 v[PGPGIN] /= 2; /* sectors -> kbytes */
1725 v[PGPGOUT] /= 2;
1726#endif
1727 return (unsigned long *)m->private + *pos;
1728}
1729
1730static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1731{
1732 (*pos)++;
1733 if (*pos >= ARRAY_SIZE(vmstat_text))
1734 return NULL;
1735 return (unsigned long *)m->private + *pos;
1736}
1737
1738static int vmstat_show(struct seq_file *m, void *arg)
1739{
1740 unsigned long *l = arg;
1741 unsigned long off = l - (unsigned long *)m->private;
1742
1743 /* Skip hidden vmstat items. */
1744 if (*vmstat_text[off] == '\0')
1745 return 0;
1746
1747 seq_puts(m, vmstat_text[off]);
1748 seq_put_decimal_ull(m, " ", *l);
1749 seq_putc(m, '\n');
1750 return 0;
1751}
1752
1753static void vmstat_stop(struct seq_file *m, void *arg)
1754{
1755 kfree(m->private);
1756 m->private = NULL;
1757}
1758
1759static const struct seq_operations vmstat_op = {
1760 .start = vmstat_start,
1761 .next = vmstat_next,
1762 .stop = vmstat_stop,
1763 .show = vmstat_show,
1764};
1765
1766static int vmstat_open(struct inode *inode, struct file *file)
1767{
1768 return seq_open(file, &vmstat_op);
1769}
1770
1771static const struct file_operations vmstat_file_operations = {
1772 .open = vmstat_open,
1773 .read = seq_read,
1774 .llseek = seq_lseek,
1775 .release = seq_release,
1776};
1777#endif /* CONFIG_PROC_FS */
1778
1779#ifdef CONFIG_SMP
1780static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1781int sysctl_stat_interval __read_mostly = HZ;
1782
1783#ifdef CONFIG_PROC_FS
1784static void refresh_vm_stats(struct work_struct *work)
1785{
1786 refresh_cpu_vm_stats(true);
1787}
1788
1789int vmstat_refresh(struct ctl_table *table, int write,
1790 void __user *buffer, size_t *lenp, loff_t *ppos)
1791{
1792 long val;
1793 int err;
1794 int i;
1795
1796 /*
1797 * The regular update, every sysctl_stat_interval, may come later
1798 * than expected: leaving a significant amount in per_cpu buckets.
1799 * This is particularly misleading when checking a quantity of HUGE
1800 * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1801 * which can equally be echo'ed to or cat'ted from (by root),
1802 * can be used to update the stats just before reading them.
1803 *
1804 * Oh, and since global_zone_page_state() etc. are so careful to hide
1805 * transiently negative values, report an error here if any of
1806 * the stats is negative, so we know to go looking for imbalance.
1807 */
1808 err = schedule_on_each_cpu(refresh_vm_stats);
1809 if (err)
1810 return err;
1811 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1812 val = atomic_long_read(&vm_zone_stat[i]);
1813 if (val < 0) {
1814 pr_warn("%s: %s %ld\n",
1815 __func__, vmstat_text[i], val);
1816 err = -EINVAL;
1817 }
1818 }
1819#ifdef CONFIG_NUMA
1820 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1821 val = atomic_long_read(&vm_numa_stat[i]);
1822 if (val < 0) {
1823 pr_warn("%s: %s %ld\n",
1824 __func__, vmstat_text[i + NR_VM_ZONE_STAT_ITEMS], val);
1825 err = -EINVAL;
1826 }
1827 }
1828#endif
1829 if (err)
1830 return err;
1831 if (write)
1832 *ppos += *lenp;
1833 else
1834 *lenp = 0;
1835 return 0;
1836}
1837#endif /* CONFIG_PROC_FS */
1838
1839static void vmstat_update(struct work_struct *w)
1840{
1841 if (refresh_cpu_vm_stats(true)) {
1842 /*
1843 * Counters were updated so we expect more updates
1844 * to occur in the future. Keep on running the
1845 * update worker thread.
1846 */
1847 preempt_disable();
1848 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1849 this_cpu_ptr(&vmstat_work),
1850 round_jiffies_relative(sysctl_stat_interval));
1851 preempt_enable();
1852 }
1853}
1854
1855/*
1856 * Switch off vmstat processing and then fold all the remaining differentials
1857 * until the diffs stay at zero. The function is used by NOHZ and can only be
1858 * invoked when tick processing is not active.
1859 */
1860/*
1861 * Check if the diffs for a certain cpu indicate that
1862 * an update is needed.
1863 */
1864static bool need_update(int cpu)
1865{
1866 struct zone *zone;
1867
1868 for_each_populated_zone(zone) {
1869 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1870
1871 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1872#ifdef CONFIG_NUMA
1873 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1874#endif
1875
1876 /*
1877 * The fast way of checking if there are any vmstat diffs.
1878 * This works because the diffs are byte sized items.
1879 */
1880 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1881 return true;
1882#ifdef CONFIG_NUMA
1883 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS))
1884 return true;
1885#endif
1886 }
1887 return false;
1888}
1889
1890/*
1891 * Switch off vmstat processing and then fold all the remaining differentials
1892 * until the diffs stay at zero. The function is used by NOHZ and can only be
1893 * invoked when tick processing is not active.
1894 */
1895void quiet_vmstat(void)
1896{
1897 if (system_state != SYSTEM_RUNNING)
1898 return;
1899
1900 if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1901 return;
1902
1903 if (!need_update(smp_processor_id()))
1904 return;
1905
1906 /*
1907 * Just refresh counters and do not care about the pending delayed
1908 * vmstat_update. It doesn't fire that often to matter and canceling
1909 * it would be too expensive from this path.
1910 * vmstat_shepherd will take care about that for us.
1911 */
1912 refresh_cpu_vm_stats(false);
1913}
1914
1915/*
1916 * Shepherd worker thread that checks the
1917 * differentials of processors that have their worker
1918 * threads for vm statistics updates disabled because of
1919 * inactivity.
1920 */
1921static void vmstat_shepherd(struct work_struct *w);
1922
1923static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1924
1925static void vmstat_shepherd(struct work_struct *w)
1926{
1927 int cpu;
1928
1929 get_online_cpus();
1930 /* Check processors whose vmstat worker threads have been disabled */
1931 for_each_online_cpu(cpu) {
1932 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1933
1934 if (!delayed_work_pending(dw) && need_update(cpu))
1935 queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1936 }
1937 put_online_cpus();
1938
1939 schedule_delayed_work(&shepherd,
1940 round_jiffies_relative(sysctl_stat_interval));
1941}
1942
1943static void __init start_shepherd_timer(void)
1944{
1945 int cpu;
1946
1947 for_each_possible_cpu(cpu)
1948 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1949 vmstat_update);
1950
1951 schedule_delayed_work(&shepherd,
1952 round_jiffies_relative(sysctl_stat_interval));
1953}
1954
1955static void __init init_cpu_node_state(void)
1956{
1957 int node;
1958
1959 for_each_online_node(node) {
1960 if (cpumask_weight(cpumask_of_node(node)) > 0)
1961 node_set_state(node, N_CPU);
1962 }
1963}
1964
1965static int vmstat_cpu_online(unsigned int cpu)
1966{
1967 refresh_zone_stat_thresholds();
1968 node_set_state(cpu_to_node(cpu), N_CPU);
1969 return 0;
1970}
1971
1972static int vmstat_cpu_down_prep(unsigned int cpu)
1973{
1974 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1975 return 0;
1976}
1977
1978static int vmstat_cpu_dead(unsigned int cpu)
1979{
1980 const struct cpumask *node_cpus;
1981 int node;
1982
1983 node = cpu_to_node(cpu);
1984
1985 refresh_zone_stat_thresholds();
1986 node_cpus = cpumask_of_node(node);
1987 if (cpumask_weight(node_cpus) > 0)
1988 return 0;
1989
1990 node_clear_state(node, N_CPU);
1991 return 0;
1992}
1993
1994#endif
1995
1996struct workqueue_struct *mm_percpu_wq;
1997
1998void __init init_mm_internals(void)
1999{
2000 int ret __maybe_unused;
2001
2002 mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2003
2004#ifdef CONFIG_SMP
2005 ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2006 NULL, vmstat_cpu_dead);
2007 if (ret < 0)
2008 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2009
2010 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2011 vmstat_cpu_online,
2012 vmstat_cpu_down_prep);
2013 if (ret < 0)
2014 pr_err("vmstat: failed to register 'online' hotplug state\n");
2015
2016 get_online_cpus();
2017 init_cpu_node_state();
2018 put_online_cpus();
2019
2020 start_shepherd_timer();
2021#endif
2022#ifdef CONFIG_PROC_FS
2023 proc_create("buddyinfo", 0444, NULL, &buddyinfo_file_operations);
2024 proc_create("pagetypeinfo", 0444, NULL, &pagetypeinfo_file_operations);
2025 proc_create("vmstat", 0444, NULL, &vmstat_file_operations);
2026 proc_create("zoneinfo", 0444, NULL, &zoneinfo_file_operations);
2027#endif
2028}
2029
2030#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2031
2032/*
2033 * Return an index indicating how much of the available free memory is
2034 * unusable for an allocation of the requested size.
2035 */
2036static int unusable_free_index(unsigned int order,
2037 struct contig_page_info *info)
2038{
2039 /* No free memory is interpreted as all free memory is unusable */
2040 if (info->free_pages == 0)
2041 return 1000;
2042
2043 /*
2044 * Index should be a value between 0 and 1. Return a value to 3
2045 * decimal places.
2046 *
2047 * 0 => no fragmentation
2048 * 1 => high fragmentation
2049 */
2050 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2051
2052}
2053
2054static void unusable_show_print(struct seq_file *m,
2055 pg_data_t *pgdat, struct zone *zone)
2056{
2057 unsigned int order;
2058 int index;
2059 struct contig_page_info info;
2060
2061 seq_printf(m, "Node %d, zone %8s ",
2062 pgdat->node_id,
2063 zone->name);
2064 for (order = 0; order < MAX_ORDER; ++order) {
2065 fill_contig_page_info(zone, order, &info);
2066 index = unusable_free_index(order, &info);
2067 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2068 }
2069
2070 seq_putc(m, '\n');
2071}
2072
2073/*
2074 * Display unusable free space index
2075 *
2076 * The unusable free space index measures how much of the available free
2077 * memory cannot be used to satisfy an allocation of a given size and is a
2078 * value between 0 and 1. The higher the value, the more of free memory is
2079 * unusable and by implication, the worse the external fragmentation is. This
2080 * can be expressed as a percentage by multiplying by 100.
2081 */
2082static int unusable_show(struct seq_file *m, void *arg)
2083{
2084 pg_data_t *pgdat = (pg_data_t *)arg;
2085
2086 /* check memoryless node */
2087 if (!node_state(pgdat->node_id, N_MEMORY))
2088 return 0;
2089
2090 walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2091
2092 return 0;
2093}
2094
2095static const struct seq_operations unusable_op = {
2096 .start = frag_start,
2097 .next = frag_next,
2098 .stop = frag_stop,
2099 .show = unusable_show,
2100};
2101
2102static int unusable_open(struct inode *inode, struct file *file)
2103{
2104 return seq_open(file, &unusable_op);
2105}
2106
2107static const struct file_operations unusable_file_ops = {
2108 .open = unusable_open,
2109 .read = seq_read,
2110 .llseek = seq_lseek,
2111 .release = seq_release,
2112};
2113
2114static void extfrag_show_print(struct seq_file *m,
2115 pg_data_t *pgdat, struct zone *zone)
2116{
2117 unsigned int order;
2118 int index;
2119
2120 /* Alloc on stack as interrupts are disabled for zone walk */
2121 struct contig_page_info info;
2122
2123 seq_printf(m, "Node %d, zone %8s ",
2124 pgdat->node_id,
2125 zone->name);
2126 for (order = 0; order < MAX_ORDER; ++order) {
2127 fill_contig_page_info(zone, order, &info);
2128 index = __fragmentation_index(order, &info);
2129 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2130 }
2131
2132 seq_putc(m, '\n');
2133}
2134
2135/*
2136 * Display fragmentation index for orders that allocations would fail for
2137 */
2138static int extfrag_show(struct seq_file *m, void *arg)
2139{
2140 pg_data_t *pgdat = (pg_data_t *)arg;
2141
2142 walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2143
2144 return 0;
2145}
2146
2147static const struct seq_operations extfrag_op = {
2148 .start = frag_start,
2149 .next = frag_next,
2150 .stop = frag_stop,
2151 .show = extfrag_show,
2152};
2153
2154static int extfrag_open(struct inode *inode, struct file *file)
2155{
2156 return seq_open(file, &extfrag_op);
2157}
2158
2159static const struct file_operations extfrag_file_ops = {
2160 .open = extfrag_open,
2161 .read = seq_read,
2162 .llseek = seq_lseek,
2163 .release = seq_release,
2164};
2165
2166static int __init extfrag_debug_init(void)
2167{
2168 struct dentry *extfrag_debug_root;
2169
2170 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2171 if (!extfrag_debug_root)
2172 return -ENOMEM;
2173
2174 if (!debugfs_create_file("unusable_index", 0444,
2175 extfrag_debug_root, NULL, &unusable_file_ops))
2176 goto fail;
2177
2178 if (!debugfs_create_file("extfrag_index", 0444,
2179 extfrag_debug_root, NULL, &extfrag_file_ops))
2180 goto fail;
2181
2182 return 0;
2183fail:
2184 debugfs_remove_recursive(extfrag_debug_root);
2185 return -ENOMEM;
2186}
2187
2188module_init(extfrag_debug_init);
2189#endif
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#include <linux/mm_inline.h>
23
24#include "internal.h"
25
26#ifdef CONFIG_VM_EVENT_COUNTERS
27DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
28EXPORT_PER_CPU_SYMBOL(vm_event_states);
29
30static void sum_vm_events(unsigned long *ret)
31{
32 int cpu;
33 int i;
34
35 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
36
37 for_each_online_cpu(cpu) {
38 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
39
40 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
41 ret[i] += this->event[i];
42 }
43}
44
45/*
46 * Accumulate the vm event counters across all CPUs.
47 * The result is unavoidably approximate - it can change
48 * during and after execution of this function.
49*/
50void all_vm_events(unsigned long *ret)
51{
52 get_online_cpus();
53 sum_vm_events(ret);
54 put_online_cpus();
55}
56EXPORT_SYMBOL_GPL(all_vm_events);
57
58/*
59 * Fold the foreign cpu events into our own.
60 *
61 * This is adding to the events on one processor
62 * but keeps the global counts constant.
63 */
64void vm_events_fold_cpu(int cpu)
65{
66 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
67 int i;
68
69 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
70 count_vm_events(i, fold_state->event[i]);
71 fold_state->event[i] = 0;
72 }
73}
74
75#endif /* CONFIG_VM_EVENT_COUNTERS */
76
77/*
78 * Manage combined zone based / global counters
79 *
80 * vm_stat contains the global counters
81 */
82atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
83EXPORT_SYMBOL(vm_stat);
84
85#ifdef CONFIG_SMP
86
87int calculate_pressure_threshold(struct zone *zone)
88{
89 int threshold;
90 int watermark_distance;
91
92 /*
93 * As vmstats are not up to date, there is drift between the estimated
94 * and real values. For high thresholds and a high number of CPUs, it
95 * is possible for the min watermark to be breached while the estimated
96 * value looks fine. The pressure threshold is a reduced value such
97 * that even the maximum amount of drift will not accidentally breach
98 * the min watermark
99 */
100 watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
101 threshold = max(1, (int)(watermark_distance / num_online_cpus()));
102
103 /*
104 * Maximum threshold is 125
105 */
106 threshold = min(125, threshold);
107
108 return threshold;
109}
110
111int calculate_normal_threshold(struct zone *zone)
112{
113 int threshold;
114 int mem; /* memory in 128 MB units */
115
116 /*
117 * The threshold scales with the number of processors and the amount
118 * of memory per zone. More memory means that we can defer updates for
119 * longer, more processors could lead to more contention.
120 * fls() is used to have a cheap way of logarithmic scaling.
121 *
122 * Some sample thresholds:
123 *
124 * Threshold Processors (fls) Zonesize fls(mem+1)
125 * ------------------------------------------------------------------
126 * 8 1 1 0.9-1 GB 4
127 * 16 2 2 0.9-1 GB 4
128 * 20 2 2 1-2 GB 5
129 * 24 2 2 2-4 GB 6
130 * 28 2 2 4-8 GB 7
131 * 32 2 2 8-16 GB 8
132 * 4 2 2 <128M 1
133 * 30 4 3 2-4 GB 5
134 * 48 4 3 8-16 GB 8
135 * 32 8 4 1-2 GB 4
136 * 32 8 4 0.9-1GB 4
137 * 10 16 5 <128M 1
138 * 40 16 5 900M 4
139 * 70 64 7 2-4 GB 5
140 * 84 64 7 4-8 GB 6
141 * 108 512 9 4-8 GB 6
142 * 125 1024 10 8-16 GB 8
143 * 125 1024 10 16-32 GB 9
144 */
145
146 mem = zone->managed_pages >> (27 - PAGE_SHIFT);
147
148 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
149
150 /*
151 * Maximum threshold is 125
152 */
153 threshold = min(125, threshold);
154
155 return threshold;
156}
157
158/*
159 * Refresh the thresholds for each zone.
160 */
161void refresh_zone_stat_thresholds(void)
162{
163 struct zone *zone;
164 int cpu;
165 int threshold;
166
167 for_each_populated_zone(zone) {
168 unsigned long max_drift, tolerate_drift;
169
170 threshold = calculate_normal_threshold(zone);
171
172 for_each_online_cpu(cpu)
173 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
174 = threshold;
175
176 /*
177 * Only set percpu_drift_mark if there is a danger that
178 * NR_FREE_PAGES reports the low watermark is ok when in fact
179 * the min watermark could be breached by an allocation
180 */
181 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
182 max_drift = num_online_cpus() * threshold;
183 if (max_drift > tolerate_drift)
184 zone->percpu_drift_mark = high_wmark_pages(zone) +
185 max_drift;
186 }
187}
188
189void set_pgdat_percpu_threshold(pg_data_t *pgdat,
190 int (*calculate_pressure)(struct zone *))
191{
192 struct zone *zone;
193 int cpu;
194 int threshold;
195 int i;
196
197 for (i = 0; i < pgdat->nr_zones; i++) {
198 zone = &pgdat->node_zones[i];
199 if (!zone->percpu_drift_mark)
200 continue;
201
202 threshold = (*calculate_pressure)(zone);
203 for_each_possible_cpu(cpu)
204 per_cpu_ptr(zone->pageset, cpu)->stat_threshold
205 = threshold;
206 }
207}
208
209/*
210 * For use when we know that interrupts are disabled.
211 */
212void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
213 int delta)
214{
215 struct per_cpu_pageset __percpu *pcp = zone->pageset;
216 s8 __percpu *p = pcp->vm_stat_diff + item;
217 long x;
218 long t;
219
220 x = delta + __this_cpu_read(*p);
221
222 t = __this_cpu_read(pcp->stat_threshold);
223
224 if (unlikely(x > t || x < -t)) {
225 zone_page_state_add(x, zone, item);
226 x = 0;
227 }
228 __this_cpu_write(*p, x);
229}
230EXPORT_SYMBOL(__mod_zone_page_state);
231
232/*
233 * Optimized increment and decrement functions.
234 *
235 * These are only for a single page and therefore can take a struct page *
236 * argument instead of struct zone *. This allows the inclusion of the code
237 * generated for page_zone(page) into the optimized functions.
238 *
239 * No overflow check is necessary and therefore the differential can be
240 * incremented or decremented in place which may allow the compilers to
241 * generate better code.
242 * The increment or decrement is known and therefore one boundary check can
243 * be omitted.
244 *
245 * NOTE: These functions are very performance sensitive. Change only
246 * with care.
247 *
248 * Some processors have inc/dec instructions that are atomic vs an interrupt.
249 * However, the code must first determine the differential location in a zone
250 * based on the processor number and then inc/dec the counter. There is no
251 * guarantee without disabling preemption that the processor will not change
252 * in between and therefore the atomicity vs. interrupt cannot be exploited
253 * in a useful way here.
254 */
255void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
256{
257 struct per_cpu_pageset __percpu *pcp = zone->pageset;
258 s8 __percpu *p = pcp->vm_stat_diff + item;
259 s8 v, t;
260
261 v = __this_cpu_inc_return(*p);
262 t = __this_cpu_read(pcp->stat_threshold);
263 if (unlikely(v > t)) {
264 s8 overstep = t >> 1;
265
266 zone_page_state_add(v + overstep, zone, item);
267 __this_cpu_write(*p, -overstep);
268 }
269}
270
271void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
272{
273 __inc_zone_state(page_zone(page), item);
274}
275EXPORT_SYMBOL(__inc_zone_page_state);
276
277void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
278{
279 struct per_cpu_pageset __percpu *pcp = zone->pageset;
280 s8 __percpu *p = pcp->vm_stat_diff + item;
281 s8 v, t;
282
283 v = __this_cpu_dec_return(*p);
284 t = __this_cpu_read(pcp->stat_threshold);
285 if (unlikely(v < - t)) {
286 s8 overstep = t >> 1;
287
288 zone_page_state_add(v - overstep, zone, item);
289 __this_cpu_write(*p, overstep);
290 }
291}
292
293void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
294{
295 __dec_zone_state(page_zone(page), item);
296}
297EXPORT_SYMBOL(__dec_zone_page_state);
298
299#ifdef CONFIG_HAVE_CMPXCHG_LOCAL
300/*
301 * If we have cmpxchg_local support then we do not need to incur the overhead
302 * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
303 *
304 * mod_state() modifies the zone counter state through atomic per cpu
305 * operations.
306 *
307 * Overstep mode specifies how overstep should handled:
308 * 0 No overstepping
309 * 1 Overstepping half of threshold
310 * -1 Overstepping minus half of threshold
311*/
312static inline void mod_state(struct zone *zone,
313 enum zone_stat_item item, int delta, int overstep_mode)
314{
315 struct per_cpu_pageset __percpu *pcp = zone->pageset;
316 s8 __percpu *p = pcp->vm_stat_diff + item;
317 long o, n, t, z;
318
319 do {
320 z = 0; /* overflow to zone counters */
321
322 /*
323 * The fetching of the stat_threshold is racy. We may apply
324 * a counter threshold to the wrong the cpu if we get
325 * rescheduled while executing here. However, the next
326 * counter update will apply the threshold again and
327 * therefore bring the counter under the threshold again.
328 *
329 * Most of the time the thresholds are the same anyways
330 * for all cpus in a zone.
331 */
332 t = this_cpu_read(pcp->stat_threshold);
333
334 o = this_cpu_read(*p);
335 n = delta + o;
336
337 if (n > t || n < -t) {
338 int os = overstep_mode * (t >> 1) ;
339
340 /* Overflow must be added to zone counters */
341 z = n + os;
342 n = -os;
343 }
344 } while (this_cpu_cmpxchg(*p, o, n) != o);
345
346 if (z)
347 zone_page_state_add(z, zone, item);
348}
349
350void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
351 int delta)
352{
353 mod_state(zone, item, delta, 0);
354}
355EXPORT_SYMBOL(mod_zone_page_state);
356
357void inc_zone_state(struct zone *zone, enum zone_stat_item item)
358{
359 mod_state(zone, item, 1, 1);
360}
361
362void inc_zone_page_state(struct page *page, enum zone_stat_item item)
363{
364 mod_state(page_zone(page), item, 1, 1);
365}
366EXPORT_SYMBOL(inc_zone_page_state);
367
368void dec_zone_page_state(struct page *page, enum zone_stat_item item)
369{
370 mod_state(page_zone(page), item, -1, -1);
371}
372EXPORT_SYMBOL(dec_zone_page_state);
373#else
374/*
375 * Use interrupt disable to serialize counter updates
376 */
377void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
378 int delta)
379{
380 unsigned long flags;
381
382 local_irq_save(flags);
383 __mod_zone_page_state(zone, item, delta);
384 local_irq_restore(flags);
385}
386EXPORT_SYMBOL(mod_zone_page_state);
387
388void inc_zone_state(struct zone *zone, enum zone_stat_item item)
389{
390 unsigned long flags;
391
392 local_irq_save(flags);
393 __inc_zone_state(zone, item);
394 local_irq_restore(flags);
395}
396
397void inc_zone_page_state(struct page *page, enum zone_stat_item item)
398{
399 unsigned long flags;
400 struct zone *zone;
401
402 zone = page_zone(page);
403 local_irq_save(flags);
404 __inc_zone_state(zone, item);
405 local_irq_restore(flags);
406}
407EXPORT_SYMBOL(inc_zone_page_state);
408
409void dec_zone_page_state(struct page *page, enum zone_stat_item item)
410{
411 unsigned long flags;
412
413 local_irq_save(flags);
414 __dec_zone_page_state(page, item);
415 local_irq_restore(flags);
416}
417EXPORT_SYMBOL(dec_zone_page_state);
418#endif
419
420static inline void fold_diff(int *diff)
421{
422 int i;
423
424 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
425 if (diff[i])
426 atomic_long_add(diff[i], &vm_stat[i]);
427}
428
429/*
430 * Update the zone counters for the current cpu.
431 *
432 * Note that refresh_cpu_vm_stats strives to only access
433 * node local memory. The per cpu pagesets on remote zones are placed
434 * in the memory local to the processor using that pageset. So the
435 * loop over all zones will access a series of cachelines local to
436 * the processor.
437 *
438 * The call to zone_page_state_add updates the cachelines with the
439 * statistics in the remote zone struct as well as the global cachelines
440 * with the global counters. These could cause remote node cache line
441 * bouncing and will have to be only done when necessary.
442 */
443static void refresh_cpu_vm_stats(void)
444{
445 struct zone *zone;
446 int i;
447 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
448
449 for_each_populated_zone(zone) {
450 struct per_cpu_pageset __percpu *p = zone->pageset;
451
452 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
453 int v;
454
455 v = this_cpu_xchg(p->vm_stat_diff[i], 0);
456 if (v) {
457
458 atomic_long_add(v, &zone->vm_stat[i]);
459 global_diff[i] += v;
460#ifdef CONFIG_NUMA
461 /* 3 seconds idle till flush */
462 __this_cpu_write(p->expire, 3);
463#endif
464 }
465 }
466 cond_resched();
467#ifdef CONFIG_NUMA
468 /*
469 * Deal with draining the remote pageset of this
470 * processor
471 *
472 * Check if there are pages remaining in this pageset
473 * if not then there is nothing to expire.
474 */
475 if (!__this_cpu_read(p->expire) ||
476 !__this_cpu_read(p->pcp.count))
477 continue;
478
479 /*
480 * We never drain zones local to this processor.
481 */
482 if (zone_to_nid(zone) == numa_node_id()) {
483 __this_cpu_write(p->expire, 0);
484 continue;
485 }
486
487
488 if (__this_cpu_dec_return(p->expire))
489 continue;
490
491 if (__this_cpu_read(p->pcp.count))
492 drain_zone_pages(zone, __this_cpu_ptr(&p->pcp));
493#endif
494 }
495 fold_diff(global_diff);
496}
497
498/*
499 * Fold the data for an offline cpu into the global array.
500 * There cannot be any access by the offline cpu and therefore
501 * synchronization is simplified.
502 */
503void cpu_vm_stats_fold(int cpu)
504{
505 struct zone *zone;
506 int i;
507 int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
508
509 for_each_populated_zone(zone) {
510 struct per_cpu_pageset *p;
511
512 p = per_cpu_ptr(zone->pageset, cpu);
513
514 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
515 if (p->vm_stat_diff[i]) {
516 int v;
517
518 v = p->vm_stat_diff[i];
519 p->vm_stat_diff[i] = 0;
520 atomic_long_add(v, &zone->vm_stat[i]);
521 global_diff[i] += v;
522 }
523 }
524
525 fold_diff(global_diff);
526}
527
528/*
529 * this is only called if !populated_zone(zone), which implies no other users of
530 * pset->vm_stat_diff[] exsist.
531 */
532void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
533{
534 int i;
535
536 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
537 if (pset->vm_stat_diff[i]) {
538 int v = pset->vm_stat_diff[i];
539 pset->vm_stat_diff[i] = 0;
540 atomic_long_add(v, &zone->vm_stat[i]);
541 atomic_long_add(v, &vm_stat[i]);
542 }
543}
544#endif
545
546#ifdef CONFIG_NUMA
547/*
548 * zonelist = the list of zones passed to the allocator
549 * z = the zone from which the allocation occurred.
550 *
551 * Must be called with interrupts disabled.
552 *
553 * When __GFP_OTHER_NODE is set assume the node of the preferred
554 * zone is the local node. This is useful for daemons who allocate
555 * memory on behalf of other processes.
556 */
557void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
558{
559 if (z->zone_pgdat == preferred_zone->zone_pgdat) {
560 __inc_zone_state(z, NUMA_HIT);
561 } else {
562 __inc_zone_state(z, NUMA_MISS);
563 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
564 }
565 if (z->node == ((flags & __GFP_OTHER_NODE) ?
566 preferred_zone->node : numa_node_id()))
567 __inc_zone_state(z, NUMA_LOCAL);
568 else
569 __inc_zone_state(z, NUMA_OTHER);
570}
571#endif
572
573#ifdef CONFIG_COMPACTION
574
575struct contig_page_info {
576 unsigned long free_pages;
577 unsigned long free_blocks_total;
578 unsigned long free_blocks_suitable;
579};
580
581/*
582 * Calculate the number of free pages in a zone, how many contiguous
583 * pages are free and how many are large enough to satisfy an allocation of
584 * the target size. Note that this function makes no attempt to estimate
585 * how many suitable free blocks there *might* be if MOVABLE pages were
586 * migrated. Calculating that is possible, but expensive and can be
587 * figured out from userspace
588 */
589static void fill_contig_page_info(struct zone *zone,
590 unsigned int suitable_order,
591 struct contig_page_info *info)
592{
593 unsigned int order;
594
595 info->free_pages = 0;
596 info->free_blocks_total = 0;
597 info->free_blocks_suitable = 0;
598
599 for (order = 0; order < MAX_ORDER; order++) {
600 unsigned long blocks;
601
602 /* Count number of free blocks */
603 blocks = zone->free_area[order].nr_free;
604 info->free_blocks_total += blocks;
605
606 /* Count free base pages */
607 info->free_pages += blocks << order;
608
609 /* Count the suitable free blocks */
610 if (order >= suitable_order)
611 info->free_blocks_suitable += blocks <<
612 (order - suitable_order);
613 }
614}
615
616/*
617 * A fragmentation index only makes sense if an allocation of a requested
618 * size would fail. If that is true, the fragmentation index indicates
619 * whether external fragmentation or a lack of memory was the problem.
620 * The value can be used to determine if page reclaim or compaction
621 * should be used
622 */
623static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
624{
625 unsigned long requested = 1UL << order;
626
627 if (!info->free_blocks_total)
628 return 0;
629
630 /* Fragmentation index only makes sense when a request would fail */
631 if (info->free_blocks_suitable)
632 return -1000;
633
634 /*
635 * Index is between 0 and 1 so return within 3 decimal places
636 *
637 * 0 => allocation would fail due to lack of memory
638 * 1 => allocation would fail due to fragmentation
639 */
640 return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
641}
642
643/* Same as __fragmentation index but allocs contig_page_info on stack */
644int fragmentation_index(struct zone *zone, unsigned int order)
645{
646 struct contig_page_info info;
647
648 fill_contig_page_info(zone, order, &info);
649 return __fragmentation_index(order, &info);
650}
651#endif
652
653#if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
654#include <linux/proc_fs.h>
655#include <linux/seq_file.h>
656
657static char * const migratetype_names[MIGRATE_TYPES] = {
658 "Unmovable",
659 "Reclaimable",
660 "Movable",
661 "Reserve",
662#ifdef CONFIG_CMA
663 "CMA",
664#endif
665#ifdef CONFIG_MEMORY_ISOLATION
666 "Isolate",
667#endif
668};
669
670static void *frag_start(struct seq_file *m, loff_t *pos)
671{
672 pg_data_t *pgdat;
673 loff_t node = *pos;
674 for (pgdat = first_online_pgdat();
675 pgdat && node;
676 pgdat = next_online_pgdat(pgdat))
677 --node;
678
679 return pgdat;
680}
681
682static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
683{
684 pg_data_t *pgdat = (pg_data_t *)arg;
685
686 (*pos)++;
687 return next_online_pgdat(pgdat);
688}
689
690static void frag_stop(struct seq_file *m, void *arg)
691{
692}
693
694/* Walk all the zones in a node and print using a callback */
695static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
696 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
697{
698 struct zone *zone;
699 struct zone *node_zones = pgdat->node_zones;
700 unsigned long flags;
701
702 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
703 if (!populated_zone(zone))
704 continue;
705
706 spin_lock_irqsave(&zone->lock, flags);
707 print(m, pgdat, zone);
708 spin_unlock_irqrestore(&zone->lock, flags);
709 }
710}
711#endif
712
713#if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
714#ifdef CONFIG_ZONE_DMA
715#define TEXT_FOR_DMA(xx) xx "_dma",
716#else
717#define TEXT_FOR_DMA(xx)
718#endif
719
720#ifdef CONFIG_ZONE_DMA32
721#define TEXT_FOR_DMA32(xx) xx "_dma32",
722#else
723#define TEXT_FOR_DMA32(xx)
724#endif
725
726#ifdef CONFIG_HIGHMEM
727#define TEXT_FOR_HIGHMEM(xx) xx "_high",
728#else
729#define TEXT_FOR_HIGHMEM(xx)
730#endif
731
732#define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
733 TEXT_FOR_HIGHMEM(xx) xx "_movable",
734
735const char * const vmstat_text[] = {
736 /* Zoned VM counters */
737 "nr_free_pages",
738 "nr_alloc_batch",
739 "nr_inactive_anon",
740 "nr_active_anon",
741 "nr_inactive_file",
742 "nr_active_file",
743 "nr_unevictable",
744 "nr_mlock",
745 "nr_anon_pages",
746 "nr_mapped",
747 "nr_file_pages",
748 "nr_dirty",
749 "nr_writeback",
750 "nr_slab_reclaimable",
751 "nr_slab_unreclaimable",
752 "nr_page_table_pages",
753 "nr_kernel_stack",
754 "nr_unstable",
755 "nr_bounce",
756 "nr_vmscan_write",
757 "nr_vmscan_immediate_reclaim",
758 "nr_writeback_temp",
759 "nr_isolated_anon",
760 "nr_isolated_file",
761 "nr_shmem",
762 "nr_dirtied",
763 "nr_written",
764
765#ifdef CONFIG_NUMA
766 "numa_hit",
767 "numa_miss",
768 "numa_foreign",
769 "numa_interleave",
770 "numa_local",
771 "numa_other",
772#endif
773 "workingset_refault",
774 "workingset_activate",
775 "workingset_nodereclaim",
776 "nr_anon_transparent_hugepages",
777 "nr_free_cma",
778 "nr_dirty_threshold",
779 "nr_dirty_background_threshold",
780
781#ifdef CONFIG_VM_EVENT_COUNTERS
782 "pgpgin",
783 "pgpgout",
784 "pswpin",
785 "pswpout",
786
787 TEXTS_FOR_ZONES("pgalloc")
788
789 "pgfree",
790 "pgactivate",
791 "pgdeactivate",
792
793 "pgfault",
794 "pgmajfault",
795
796 TEXTS_FOR_ZONES("pgrefill")
797 TEXTS_FOR_ZONES("pgsteal_kswapd")
798 TEXTS_FOR_ZONES("pgsteal_direct")
799 TEXTS_FOR_ZONES("pgscan_kswapd")
800 TEXTS_FOR_ZONES("pgscan_direct")
801 "pgscan_direct_throttle",
802
803#ifdef CONFIG_NUMA
804 "zone_reclaim_failed",
805#endif
806 "pginodesteal",
807 "slabs_scanned",
808 "kswapd_inodesteal",
809 "kswapd_low_wmark_hit_quickly",
810 "kswapd_high_wmark_hit_quickly",
811 "pageoutrun",
812 "allocstall",
813
814 "pgrotated",
815
816 "drop_pagecache",
817 "drop_slab",
818
819#ifdef CONFIG_NUMA_BALANCING
820 "numa_pte_updates",
821 "numa_huge_pte_updates",
822 "numa_hint_faults",
823 "numa_hint_faults_local",
824 "numa_pages_migrated",
825#endif
826#ifdef CONFIG_MIGRATION
827 "pgmigrate_success",
828 "pgmigrate_fail",
829#endif
830#ifdef CONFIG_COMPACTION
831 "compact_migrate_scanned",
832 "compact_free_scanned",
833 "compact_isolated",
834 "compact_stall",
835 "compact_fail",
836 "compact_success",
837#endif
838
839#ifdef CONFIG_HUGETLB_PAGE
840 "htlb_buddy_alloc_success",
841 "htlb_buddy_alloc_fail",
842#endif
843 "unevictable_pgs_culled",
844 "unevictable_pgs_scanned",
845 "unevictable_pgs_rescued",
846 "unevictable_pgs_mlocked",
847 "unevictable_pgs_munlocked",
848 "unevictable_pgs_cleared",
849 "unevictable_pgs_stranded",
850
851#ifdef CONFIG_TRANSPARENT_HUGEPAGE
852 "thp_fault_alloc",
853 "thp_fault_fallback",
854 "thp_collapse_alloc",
855 "thp_collapse_alloc_failed",
856 "thp_split",
857 "thp_zero_page_alloc",
858 "thp_zero_page_alloc_failed",
859#endif
860#ifdef CONFIG_DEBUG_TLBFLUSH
861#ifdef CONFIG_SMP
862 "nr_tlb_remote_flush",
863 "nr_tlb_remote_flush_received",
864#endif /* CONFIG_SMP */
865 "nr_tlb_local_flush_all",
866 "nr_tlb_local_flush_one",
867#endif /* CONFIG_DEBUG_TLBFLUSH */
868
869#endif /* CONFIG_VM_EVENTS_COUNTERS */
870};
871#endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
872
873
874#ifdef CONFIG_PROC_FS
875static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
876 struct zone *zone)
877{
878 int order;
879
880 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
881 for (order = 0; order < MAX_ORDER; ++order)
882 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
883 seq_putc(m, '\n');
884}
885
886/*
887 * This walks the free areas for each zone.
888 */
889static int frag_show(struct seq_file *m, void *arg)
890{
891 pg_data_t *pgdat = (pg_data_t *)arg;
892 walk_zones_in_node(m, pgdat, frag_show_print);
893 return 0;
894}
895
896static void pagetypeinfo_showfree_print(struct seq_file *m,
897 pg_data_t *pgdat, struct zone *zone)
898{
899 int order, mtype;
900
901 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
902 seq_printf(m, "Node %4d, zone %8s, type %12s ",
903 pgdat->node_id,
904 zone->name,
905 migratetype_names[mtype]);
906 for (order = 0; order < MAX_ORDER; ++order) {
907 unsigned long freecount = 0;
908 struct free_area *area;
909 struct list_head *curr;
910
911 area = &(zone->free_area[order]);
912
913 list_for_each(curr, &area->free_list[mtype])
914 freecount++;
915 seq_printf(m, "%6lu ", freecount);
916 }
917 seq_putc(m, '\n');
918 }
919}
920
921/* Print out the free pages at each order for each migatetype */
922static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
923{
924 int order;
925 pg_data_t *pgdat = (pg_data_t *)arg;
926
927 /* Print header */
928 seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
929 for (order = 0; order < MAX_ORDER; ++order)
930 seq_printf(m, "%6d ", order);
931 seq_putc(m, '\n');
932
933 walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
934
935 return 0;
936}
937
938static void pagetypeinfo_showblockcount_print(struct seq_file *m,
939 pg_data_t *pgdat, struct zone *zone)
940{
941 int mtype;
942 unsigned long pfn;
943 unsigned long start_pfn = zone->zone_start_pfn;
944 unsigned long end_pfn = zone_end_pfn(zone);
945 unsigned long count[MIGRATE_TYPES] = { 0, };
946
947 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
948 struct page *page;
949
950 if (!pfn_valid(pfn))
951 continue;
952
953 page = pfn_to_page(pfn);
954
955 /* Watch for unexpected holes punched in the memmap */
956 if (!memmap_valid_within(pfn, page, zone))
957 continue;
958
959 mtype = get_pageblock_migratetype(page);
960
961 if (mtype < MIGRATE_TYPES)
962 count[mtype]++;
963 }
964
965 /* Print counts */
966 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
967 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
968 seq_printf(m, "%12lu ", count[mtype]);
969 seq_putc(m, '\n');
970}
971
972/* Print out the free pages at each order for each migratetype */
973static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
974{
975 int mtype;
976 pg_data_t *pgdat = (pg_data_t *)arg;
977
978 seq_printf(m, "\n%-23s", "Number of blocks type ");
979 for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
980 seq_printf(m, "%12s ", migratetype_names[mtype]);
981 seq_putc(m, '\n');
982 walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
983
984 return 0;
985}
986
987/*
988 * This prints out statistics in relation to grouping pages by mobility.
989 * It is expensive to collect so do not constantly read the file.
990 */
991static int pagetypeinfo_show(struct seq_file *m, void *arg)
992{
993 pg_data_t *pgdat = (pg_data_t *)arg;
994
995 /* check memoryless node */
996 if (!node_state(pgdat->node_id, N_MEMORY))
997 return 0;
998
999 seq_printf(m, "Page block order: %d\n", pageblock_order);
1000 seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1001 seq_putc(m, '\n');
1002 pagetypeinfo_showfree(m, pgdat);
1003 pagetypeinfo_showblockcount(m, pgdat);
1004
1005 return 0;
1006}
1007
1008static const struct seq_operations fragmentation_op = {
1009 .start = frag_start,
1010 .next = frag_next,
1011 .stop = frag_stop,
1012 .show = frag_show,
1013};
1014
1015static int fragmentation_open(struct inode *inode, struct file *file)
1016{
1017 return seq_open(file, &fragmentation_op);
1018}
1019
1020static const struct file_operations fragmentation_file_operations = {
1021 .open = fragmentation_open,
1022 .read = seq_read,
1023 .llseek = seq_lseek,
1024 .release = seq_release,
1025};
1026
1027static const struct seq_operations pagetypeinfo_op = {
1028 .start = frag_start,
1029 .next = frag_next,
1030 .stop = frag_stop,
1031 .show = pagetypeinfo_show,
1032};
1033
1034static int pagetypeinfo_open(struct inode *inode, struct file *file)
1035{
1036 return seq_open(file, &pagetypeinfo_op);
1037}
1038
1039static const struct file_operations pagetypeinfo_file_ops = {
1040 .open = pagetypeinfo_open,
1041 .read = seq_read,
1042 .llseek = seq_lseek,
1043 .release = seq_release,
1044};
1045
1046static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1047 struct zone *zone)
1048{
1049 int i;
1050 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1051 seq_printf(m,
1052 "\n pages free %lu"
1053 "\n min %lu"
1054 "\n low %lu"
1055 "\n high %lu"
1056 "\n scanned %lu"
1057 "\n spanned %lu"
1058 "\n present %lu"
1059 "\n managed %lu",
1060 zone_page_state(zone, NR_FREE_PAGES),
1061 min_wmark_pages(zone),
1062 low_wmark_pages(zone),
1063 high_wmark_pages(zone),
1064 zone->pages_scanned,
1065 zone->spanned_pages,
1066 zone->present_pages,
1067 zone->managed_pages);
1068
1069 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1070 seq_printf(m, "\n %-12s %lu", vmstat_text[i],
1071 zone_page_state(zone, i));
1072
1073 seq_printf(m,
1074 "\n protection: (%lu",
1075 zone->lowmem_reserve[0]);
1076 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1077 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1078 seq_printf(m,
1079 ")"
1080 "\n pagesets");
1081 for_each_online_cpu(i) {
1082 struct per_cpu_pageset *pageset;
1083
1084 pageset = per_cpu_ptr(zone->pageset, i);
1085 seq_printf(m,
1086 "\n cpu: %i"
1087 "\n count: %i"
1088 "\n high: %i"
1089 "\n batch: %i",
1090 i,
1091 pageset->pcp.count,
1092 pageset->pcp.high,
1093 pageset->pcp.batch);
1094#ifdef CONFIG_SMP
1095 seq_printf(m, "\n vm stats threshold: %d",
1096 pageset->stat_threshold);
1097#endif
1098 }
1099 seq_printf(m,
1100 "\n all_unreclaimable: %u"
1101 "\n start_pfn: %lu"
1102 "\n inactive_ratio: %u",
1103 !zone_reclaimable(zone),
1104 zone->zone_start_pfn,
1105 zone->inactive_ratio);
1106 seq_putc(m, '\n');
1107}
1108
1109/*
1110 * Output information about zones in @pgdat.
1111 */
1112static int zoneinfo_show(struct seq_file *m, void *arg)
1113{
1114 pg_data_t *pgdat = (pg_data_t *)arg;
1115 walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1116 return 0;
1117}
1118
1119static const struct seq_operations zoneinfo_op = {
1120 .start = frag_start, /* iterate over all zones. The same as in
1121 * fragmentation. */
1122 .next = frag_next,
1123 .stop = frag_stop,
1124 .show = zoneinfo_show,
1125};
1126
1127static int zoneinfo_open(struct inode *inode, struct file *file)
1128{
1129 return seq_open(file, &zoneinfo_op);
1130}
1131
1132static const struct file_operations proc_zoneinfo_file_operations = {
1133 .open = zoneinfo_open,
1134 .read = seq_read,
1135 .llseek = seq_lseek,
1136 .release = seq_release,
1137};
1138
1139enum writeback_stat_item {
1140 NR_DIRTY_THRESHOLD,
1141 NR_DIRTY_BG_THRESHOLD,
1142 NR_VM_WRITEBACK_STAT_ITEMS,
1143};
1144
1145static void *vmstat_start(struct seq_file *m, loff_t *pos)
1146{
1147 unsigned long *v;
1148 int i, stat_items_size;
1149
1150 if (*pos >= ARRAY_SIZE(vmstat_text))
1151 return NULL;
1152 stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1153 NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1154
1155#ifdef CONFIG_VM_EVENT_COUNTERS
1156 stat_items_size += sizeof(struct vm_event_state);
1157#endif
1158
1159 v = kmalloc(stat_items_size, GFP_KERNEL);
1160 m->private = v;
1161 if (!v)
1162 return ERR_PTR(-ENOMEM);
1163 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1164 v[i] = global_page_state(i);
1165 v += NR_VM_ZONE_STAT_ITEMS;
1166
1167 global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1168 v + NR_DIRTY_THRESHOLD);
1169 v += NR_VM_WRITEBACK_STAT_ITEMS;
1170
1171#ifdef CONFIG_VM_EVENT_COUNTERS
1172 all_vm_events(v);
1173 v[PGPGIN] /= 2; /* sectors -> kbytes */
1174 v[PGPGOUT] /= 2;
1175#endif
1176 return (unsigned long *)m->private + *pos;
1177}
1178
1179static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1180{
1181 (*pos)++;
1182 if (*pos >= ARRAY_SIZE(vmstat_text))
1183 return NULL;
1184 return (unsigned long *)m->private + *pos;
1185}
1186
1187static int vmstat_show(struct seq_file *m, void *arg)
1188{
1189 unsigned long *l = arg;
1190 unsigned long off = l - (unsigned long *)m->private;
1191
1192 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1193 return 0;
1194}
1195
1196static void vmstat_stop(struct seq_file *m, void *arg)
1197{
1198 kfree(m->private);
1199 m->private = NULL;
1200}
1201
1202static const struct seq_operations vmstat_op = {
1203 .start = vmstat_start,
1204 .next = vmstat_next,
1205 .stop = vmstat_stop,
1206 .show = vmstat_show,
1207};
1208
1209static int vmstat_open(struct inode *inode, struct file *file)
1210{
1211 return seq_open(file, &vmstat_op);
1212}
1213
1214static const struct file_operations proc_vmstat_file_operations = {
1215 .open = vmstat_open,
1216 .read = seq_read,
1217 .llseek = seq_lseek,
1218 .release = seq_release,
1219};
1220#endif /* CONFIG_PROC_FS */
1221
1222#ifdef CONFIG_SMP
1223static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1224int sysctl_stat_interval __read_mostly = HZ;
1225
1226static void vmstat_update(struct work_struct *w)
1227{
1228 refresh_cpu_vm_stats();
1229 schedule_delayed_work(&__get_cpu_var(vmstat_work),
1230 round_jiffies_relative(sysctl_stat_interval));
1231}
1232
1233static void start_cpu_timer(int cpu)
1234{
1235 struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1236
1237 INIT_DEFERRABLE_WORK(work, vmstat_update);
1238 schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1239}
1240
1241static void vmstat_cpu_dead(int node)
1242{
1243 int cpu;
1244
1245 get_online_cpus();
1246 for_each_online_cpu(cpu)
1247 if (cpu_to_node(cpu) == node)
1248 goto end;
1249
1250 node_clear_state(node, N_CPU);
1251end:
1252 put_online_cpus();
1253}
1254
1255/*
1256 * Use the cpu notifier to insure that the thresholds are recalculated
1257 * when necessary.
1258 */
1259static int vmstat_cpuup_callback(struct notifier_block *nfb,
1260 unsigned long action,
1261 void *hcpu)
1262{
1263 long cpu = (long)hcpu;
1264
1265 switch (action) {
1266 case CPU_ONLINE:
1267 case CPU_ONLINE_FROZEN:
1268 refresh_zone_stat_thresholds();
1269 start_cpu_timer(cpu);
1270 node_set_state(cpu_to_node(cpu), N_CPU);
1271 break;
1272 case CPU_DOWN_PREPARE:
1273 case CPU_DOWN_PREPARE_FROZEN:
1274 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1275 per_cpu(vmstat_work, cpu).work.func = NULL;
1276 break;
1277 case CPU_DOWN_FAILED:
1278 case CPU_DOWN_FAILED_FROZEN:
1279 start_cpu_timer(cpu);
1280 break;
1281 case CPU_DEAD:
1282 case CPU_DEAD_FROZEN:
1283 refresh_zone_stat_thresholds();
1284 vmstat_cpu_dead(cpu_to_node(cpu));
1285 break;
1286 default:
1287 break;
1288 }
1289 return NOTIFY_OK;
1290}
1291
1292static struct notifier_block vmstat_notifier =
1293 { &vmstat_cpuup_callback, NULL, 0 };
1294#endif
1295
1296static int __init setup_vmstat(void)
1297{
1298#ifdef CONFIG_SMP
1299 int cpu;
1300
1301 cpu_notifier_register_begin();
1302 __register_cpu_notifier(&vmstat_notifier);
1303
1304 for_each_online_cpu(cpu) {
1305 start_cpu_timer(cpu);
1306 node_set_state(cpu_to_node(cpu), N_CPU);
1307 }
1308 cpu_notifier_register_done();
1309#endif
1310#ifdef CONFIG_PROC_FS
1311 proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1312 proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1313 proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1314 proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1315#endif
1316 return 0;
1317}
1318module_init(setup_vmstat)
1319
1320#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1321#include <linux/debugfs.h>
1322
1323
1324/*
1325 * Return an index indicating how much of the available free memory is
1326 * unusable for an allocation of the requested size.
1327 */
1328static int unusable_free_index(unsigned int order,
1329 struct contig_page_info *info)
1330{
1331 /* No free memory is interpreted as all free memory is unusable */
1332 if (info->free_pages == 0)
1333 return 1000;
1334
1335 /*
1336 * Index should be a value between 0 and 1. Return a value to 3
1337 * decimal places.
1338 *
1339 * 0 => no fragmentation
1340 * 1 => high fragmentation
1341 */
1342 return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1343
1344}
1345
1346static void unusable_show_print(struct seq_file *m,
1347 pg_data_t *pgdat, struct zone *zone)
1348{
1349 unsigned int order;
1350 int index;
1351 struct contig_page_info info;
1352
1353 seq_printf(m, "Node %d, zone %8s ",
1354 pgdat->node_id,
1355 zone->name);
1356 for (order = 0; order < MAX_ORDER; ++order) {
1357 fill_contig_page_info(zone, order, &info);
1358 index = unusable_free_index(order, &info);
1359 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1360 }
1361
1362 seq_putc(m, '\n');
1363}
1364
1365/*
1366 * Display unusable free space index
1367 *
1368 * The unusable free space index measures how much of the available free
1369 * memory cannot be used to satisfy an allocation of a given size and is a
1370 * value between 0 and 1. The higher the value, the more of free memory is
1371 * unusable and by implication, the worse the external fragmentation is. This
1372 * can be expressed as a percentage by multiplying by 100.
1373 */
1374static int unusable_show(struct seq_file *m, void *arg)
1375{
1376 pg_data_t *pgdat = (pg_data_t *)arg;
1377
1378 /* check memoryless node */
1379 if (!node_state(pgdat->node_id, N_MEMORY))
1380 return 0;
1381
1382 walk_zones_in_node(m, pgdat, unusable_show_print);
1383
1384 return 0;
1385}
1386
1387static const struct seq_operations unusable_op = {
1388 .start = frag_start,
1389 .next = frag_next,
1390 .stop = frag_stop,
1391 .show = unusable_show,
1392};
1393
1394static int unusable_open(struct inode *inode, struct file *file)
1395{
1396 return seq_open(file, &unusable_op);
1397}
1398
1399static const struct file_operations unusable_file_ops = {
1400 .open = unusable_open,
1401 .read = seq_read,
1402 .llseek = seq_lseek,
1403 .release = seq_release,
1404};
1405
1406static void extfrag_show_print(struct seq_file *m,
1407 pg_data_t *pgdat, struct zone *zone)
1408{
1409 unsigned int order;
1410 int index;
1411
1412 /* Alloc on stack as interrupts are disabled for zone walk */
1413 struct contig_page_info info;
1414
1415 seq_printf(m, "Node %d, zone %8s ",
1416 pgdat->node_id,
1417 zone->name);
1418 for (order = 0; order < MAX_ORDER; ++order) {
1419 fill_contig_page_info(zone, order, &info);
1420 index = __fragmentation_index(order, &info);
1421 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1422 }
1423
1424 seq_putc(m, '\n');
1425}
1426
1427/*
1428 * Display fragmentation index for orders that allocations would fail for
1429 */
1430static int extfrag_show(struct seq_file *m, void *arg)
1431{
1432 pg_data_t *pgdat = (pg_data_t *)arg;
1433
1434 walk_zones_in_node(m, pgdat, extfrag_show_print);
1435
1436 return 0;
1437}
1438
1439static const struct seq_operations extfrag_op = {
1440 .start = frag_start,
1441 .next = frag_next,
1442 .stop = frag_stop,
1443 .show = extfrag_show,
1444};
1445
1446static int extfrag_open(struct inode *inode, struct file *file)
1447{
1448 return seq_open(file, &extfrag_op);
1449}
1450
1451static const struct file_operations extfrag_file_ops = {
1452 .open = extfrag_open,
1453 .read = seq_read,
1454 .llseek = seq_lseek,
1455 .release = seq_release,
1456};
1457
1458static int __init extfrag_debug_init(void)
1459{
1460 struct dentry *extfrag_debug_root;
1461
1462 extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1463 if (!extfrag_debug_root)
1464 return -ENOMEM;
1465
1466 if (!debugfs_create_file("unusable_index", 0444,
1467 extfrag_debug_root, NULL, &unusable_file_ops))
1468 goto fail;
1469
1470 if (!debugfs_create_file("extfrag_index", 0444,
1471 extfrag_debug_root, NULL, &extfrag_file_ops))
1472 goto fail;
1473
1474 return 0;
1475fail:
1476 debugfs_remove_recursive(extfrag_debug_root);
1477 return -ENOMEM;
1478}
1479
1480module_init(extfrag_debug_init);
1481#endif