Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0-only
2 : /*
3 : * linux/mm/vmstat.c
4 : *
5 : * Manages VM statistics
6 : * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 : *
8 : * zoned VM statistics
9 : * Copyright (C) 2006 Silicon Graphics, Inc.,
10 : * Christoph Lameter <christoph@lameter.com>
11 : * Copyright (C) 2008-2014 Christoph Lameter
12 : */
13 : #include <linux/fs.h>
14 : #include <linux/mm.h>
15 : #include <linux/err.h>
16 : #include <linux/module.h>
17 : #include <linux/slab.h>
18 : #include <linux/cpu.h>
19 : #include <linux/cpumask.h>
20 : #include <linux/vmstat.h>
21 : #include <linux/proc_fs.h>
22 : #include <linux/seq_file.h>
23 : #include <linux/debugfs.h>
24 : #include <linux/sched.h>
25 : #include <linux/math64.h>
26 : #include <linux/writeback.h>
27 : #include <linux/compaction.h>
28 : #include <linux/mm_inline.h>
29 : #include <linux/page_ext.h>
30 : #include <linux/page_owner.h>
31 :
32 : #include "internal.h"
33 :
34 : #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
35 :
36 : #ifdef CONFIG_NUMA
37 : int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
38 :
39 : /* zero numa counters within a zone */
40 0 : static void zero_zone_numa_counters(struct zone *zone)
41 : {
42 0 : int item, cpu;
43 :
44 0 : for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
45 0 : atomic_long_set(&zone->vm_numa_stat[item], 0);
46 0 : for_each_online_cpu(cpu)
47 0 : per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
48 0 : = 0;
49 : }
50 0 : }
51 :
52 : /* zero numa counters of all the populated zones */
53 0 : static void zero_zones_numa_counters(void)
54 : {
55 0 : struct zone *zone;
56 :
57 0 : for_each_populated_zone(zone)
58 0 : zero_zone_numa_counters(zone);
59 0 : }
60 :
61 : /* zero global numa counters */
62 0 : static void zero_global_numa_counters(void)
63 : {
64 0 : int item;
65 :
66 0 : for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
67 0 : atomic_long_set(&vm_numa_stat[item], 0);
68 0 : }
69 :
70 0 : static void invalid_numa_statistics(void)
71 : {
72 0 : zero_zones_numa_counters();
73 0 : zero_global_numa_counters();
74 : }
75 :
76 : static DEFINE_MUTEX(vm_numa_stat_lock);
77 :
78 0 : int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
79 : void *buffer, size_t *length, loff_t *ppos)
80 : {
81 0 : int ret, oldval;
82 :
83 0 : mutex_lock(&vm_numa_stat_lock);
84 0 : if (write)
85 0 : oldval = sysctl_vm_numa_stat;
86 0 : ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
87 0 : if (ret || !write)
88 0 : goto out;
89 :
90 0 : if (oldval == sysctl_vm_numa_stat)
91 0 : goto out;
92 0 : else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
93 0 : static_branch_enable(&vm_numa_stat_key);
94 0 : pr_info("enable numa statistics\n");
95 : } else {
96 0 : static_branch_disable(&vm_numa_stat_key);
97 0 : invalid_numa_statistics();
98 0 : pr_info("disable numa statistics, and clear numa counters\n");
99 : }
100 :
101 0 : out:
102 0 : mutex_unlock(&vm_numa_stat_lock);
103 0 : return ret;
104 : }
105 : #endif
106 :
107 : #ifdef CONFIG_VM_EVENT_COUNTERS
108 : DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
109 : EXPORT_PER_CPU_SYMBOL(vm_event_states);
110 :
111 0 : static void sum_vm_events(unsigned long *ret)
112 : {
113 0 : int cpu;
114 0 : int i;
115 :
116 0 : memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
117 :
118 0 : for_each_online_cpu(cpu) {
119 0 : struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
120 :
121 0 : for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
122 0 : ret[i] += this->event[i];
123 : }
124 0 : }
125 :
126 : /*
127 : * Accumulate the vm event counters across all CPUs.
128 : * The result is unavoidably approximate - it can change
129 : * during and after execution of this function.
130 : */
131 0 : void all_vm_events(unsigned long *ret)
132 : {
133 0 : get_online_cpus();
134 0 : sum_vm_events(ret);
135 0 : put_online_cpus();
136 0 : }
137 : EXPORT_SYMBOL_GPL(all_vm_events);
138 :
139 : /*
140 : * Fold the foreign cpu events into our own.
141 : *
142 : * This is adding to the events on one processor
143 : * but keeps the global counts constant.
144 : */
145 0 : void vm_events_fold_cpu(int cpu)
146 : {
147 0 : struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
148 0 : int i;
149 :
150 0 : for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
151 0 : count_vm_events(i, fold_state->event[i]);
152 0 : fold_state->event[i] = 0;
153 : }
154 0 : }
155 :
156 : #endif /* CONFIG_VM_EVENT_COUNTERS */
157 :
158 : /*
159 : * Manage combined zone based / global counters
160 : *
161 : * vm_stat contains the global counters
162 : */
163 : atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
164 : atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
165 : atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
166 : EXPORT_SYMBOL(vm_zone_stat);
167 : EXPORT_SYMBOL(vm_numa_stat);
168 : EXPORT_SYMBOL(vm_node_stat);
169 :
170 : #ifdef CONFIG_SMP
171 :
172 0 : int calculate_pressure_threshold(struct zone *zone)
173 : {
174 0 : int threshold;
175 0 : int watermark_distance;
176 :
177 : /*
178 : * As vmstats are not up to date, there is drift between the estimated
179 : * and real values. For high thresholds and a high number of CPUs, it
180 : * is possible for the min watermark to be breached while the estimated
181 : * value looks fine. The pressure threshold is a reduced value such
182 : * that even the maximum amount of drift will not accidentally breach
183 : * the min watermark
184 : */
185 0 : watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
186 0 : threshold = max(1, (int)(watermark_distance / num_online_cpus()));
187 :
188 : /*
189 : * Maximum threshold is 125
190 : */
191 0 : threshold = min(125, threshold);
192 :
193 0 : return threshold;
194 : }
195 :
196 5 : int calculate_normal_threshold(struct zone *zone)
197 : {
198 5 : int threshold;
199 5 : int mem; /* memory in 128 MB units */
200 :
201 : /*
202 : * The threshold scales with the number of processors and the amount
203 : * of memory per zone. More memory means that we can defer updates for
204 : * longer, more processors could lead to more contention.
205 : * fls() is used to have a cheap way of logarithmic scaling.
206 : *
207 : * Some sample thresholds:
208 : *
209 : * Threshold Processors (fls) Zonesize fls(mem+1)
210 : * ------------------------------------------------------------------
211 : * 8 1 1 0.9-1 GB 4
212 : * 16 2 2 0.9-1 GB 4
213 : * 20 2 2 1-2 GB 5
214 : * 24 2 2 2-4 GB 6
215 : * 28 2 2 4-8 GB 7
216 : * 32 2 2 8-16 GB 8
217 : * 4 2 2 <128M 1
218 : * 30 4 3 2-4 GB 5
219 : * 48 4 3 8-16 GB 8
220 : * 32 8 4 1-2 GB 4
221 : * 32 8 4 0.9-1GB 4
222 : * 10 16 5 <128M 1
223 : * 40 16 5 900M 4
224 : * 70 64 7 2-4 GB 5
225 : * 84 64 7 4-8 GB 6
226 : * 108 512 9 4-8 GB 6
227 : * 125 1024 10 8-16 GB 8
228 : * 125 1024 10 16-32 GB 9
229 : */
230 :
231 5 : mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
232 :
233 5 : threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
234 :
235 : /*
236 : * Maximum threshold is 125
237 : */
238 5 : threshold = min(125, threshold);
239 :
240 5 : return threshold;
241 : }
242 :
243 : /*
244 : * Refresh the thresholds for each zone.
245 : */
246 4 : void refresh_zone_stat_thresholds(void)
247 : {
248 4 : struct pglist_data *pgdat;
249 4 : struct zone *zone;
250 4 : int cpu;
251 4 : int threshold;
252 :
253 : /* Zero current pgdat thresholds */
254 8 : for_each_online_pgdat(pgdat) {
255 17 : for_each_online_cpu(cpu) {
256 13 : per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
257 : }
258 : }
259 :
260 16 : for_each_populated_zone(zone) {
261 4 : struct pglist_data *pgdat = zone->zone_pgdat;
262 4 : unsigned long max_drift, tolerate_drift;
263 :
264 4 : threshold = calculate_normal_threshold(zone);
265 :
266 21 : for_each_online_cpu(cpu) {
267 13 : int pgdat_threshold;
268 :
269 13 : per_cpu_ptr(zone->pageset, cpu)->stat_threshold
270 13 : = threshold;
271 :
272 : /* Base nodestat threshold on the largest populated zone. */
273 13 : pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
274 13 : per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
275 13 : = max(threshold, pgdat_threshold);
276 : }
277 :
278 : /*
279 : * Only set percpu_drift_mark if there is a danger that
280 : * NR_FREE_PAGES reports the low watermark is ok when in fact
281 : * the min watermark could be breached by an allocation
282 : */
283 4 : tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
284 4 : max_drift = num_online_cpus() * threshold;
285 4 : if (max_drift > tolerate_drift)
286 3 : zone->percpu_drift_mark = high_wmark_pages(zone) +
287 : max_drift;
288 : }
289 4 : }
290 :
291 1 : void set_pgdat_percpu_threshold(pg_data_t *pgdat,
292 : int (*calculate_pressure)(struct zone *))
293 : {
294 1 : struct zone *zone;
295 1 : int cpu;
296 1 : int threshold;
297 1 : int i;
298 :
299 2 : for (i = 0; i < pgdat->nr_zones; i++) {
300 1 : zone = &pgdat->node_zones[i];
301 1 : if (!zone->percpu_drift_mark)
302 0 : continue;
303 :
304 1 : threshold = (*calculate_pressure)(zone);
305 6 : for_each_online_cpu(cpu)
306 4 : per_cpu_ptr(zone->pageset, cpu)->stat_threshold
307 4 : = threshold;
308 : }
309 1 : }
310 :
311 : /*
312 : * For use when we know that interrupts are disabled,
313 : * or when we know that preemption is disabled and that
314 : * particular counter cannot be updated from interrupt context.
315 : */
316 257179 : void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
317 : long delta)
318 : {
319 257179 : struct per_cpu_pageset __percpu *pcp = zone->pageset;
320 257179 : s8 __percpu *p = pcp->vm_stat_diff + item;
321 257179 : long x;
322 257179 : long t;
323 :
324 257179 : x = delta + __this_cpu_read(*p);
325 :
326 257179 : t = __this_cpu_read(pcp->stat_threshold);
327 :
328 257179 : if (unlikely(abs(x) > t)) {
329 12587 : zone_page_state_add(x, zone, item);
330 12587 : x = 0;
331 : }
332 257179 : __this_cpu_write(*p, x);
333 257179 : }
334 : EXPORT_SYMBOL(__mod_zone_page_state);
335 :
336 561713 : void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
337 : long delta)
338 : {
339 561713 : struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
340 561713 : s8 __percpu *p = pcp->vm_node_stat_diff + item;
341 561713 : long x;
342 561713 : long t;
343 :
344 561713 : if (vmstat_item_in_bytes(item)) {
345 : /*
346 : * Only cgroups use subpage accounting right now; at
347 : * the global level, these items still change in
348 : * multiples of whole pages. Store them as pages
349 : * internally to keep the per-cpu counters compact.
350 : */
351 0 : VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
352 0 : delta >>= PAGE_SHIFT;
353 : }
354 :
355 561713 : x = delta + __this_cpu_read(*p);
356 :
357 561713 : t = __this_cpu_read(pcp->stat_threshold);
358 :
359 561713 : if (unlikely(abs(x) > t)) {
360 15270 : node_page_state_add(x, pgdat, item);
361 15270 : x = 0;
362 : }
363 561713 : __this_cpu_write(*p, x);
364 561713 : }
365 : EXPORT_SYMBOL(__mod_node_page_state);
366 :
367 : /*
368 : * Optimized increment and decrement functions.
369 : *
370 : * These are only for a single page and therefore can take a struct page *
371 : * argument instead of struct zone *. This allows the inclusion of the code
372 : * generated for page_zone(page) into the optimized functions.
373 : *
374 : * No overflow check is necessary and therefore the differential can be
375 : * incremented or decremented in place which may allow the compilers to
376 : * generate better code.
377 : * The increment or decrement is known and therefore one boundary check can
378 : * be omitted.
379 : *
380 : * NOTE: These functions are very performance sensitive. Change only
381 : * with care.
382 : *
383 : * Some processors have inc/dec instructions that are atomic vs an interrupt.
384 : * However, the code must first determine the differential location in a zone
385 : * based on the processor number and then inc/dec the counter. There is no
386 : * guarantee without disabling preemption that the processor will not change
387 : * in between and therefore the atomicity vs. interrupt cannot be exploited
388 : * in a useful way here.
389 : */
390 2515 : void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
391 : {
392 2515 : struct per_cpu_pageset __percpu *pcp = zone->pageset;
393 2515 : s8 __percpu *p = pcp->vm_stat_diff + item;
394 2515 : s8 v, t;
395 :
396 2515 : v = __this_cpu_inc_return(*p);
397 2515 : t = __this_cpu_read(pcp->stat_threshold);
398 2515 : if (unlikely(v > t)) {
399 43 : s8 overstep = t >> 1;
400 :
401 43 : zone_page_state_add(v + overstep, zone, item);
402 2515 : __this_cpu_write(*p, -overstep);
403 : }
404 2515 : }
405 :
406 2515 : void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
407 : {
408 2515 : struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
409 2515 : s8 __percpu *p = pcp->vm_node_stat_diff + item;
410 2515 : s8 v, t;
411 :
412 2515 : VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
413 :
414 2515 : v = __this_cpu_inc_return(*p);
415 2515 : t = __this_cpu_read(pcp->stat_threshold);
416 2515 : if (unlikely(v > t)) {
417 54 : s8 overstep = t >> 1;
418 :
419 54 : node_page_state_add(v + overstep, pgdat, item);
420 2515 : __this_cpu_write(*p, -overstep);
421 : }
422 2515 : }
423 :
424 2515 : void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
425 : {
426 2515 : __inc_zone_state(page_zone(page), item);
427 2515 : }
428 : EXPORT_SYMBOL(__inc_zone_page_state);
429 :
430 2515 : void __inc_node_page_state(struct page *page, enum node_stat_item item)
431 : {
432 2515 : __inc_node_state(page_pgdat(page), item);
433 2515 : }
434 : EXPORT_SYMBOL(__inc_node_page_state);
435 :
436 0 : void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
437 : {
438 0 : struct per_cpu_pageset __percpu *pcp = zone->pageset;
439 0 : s8 __percpu *p = pcp->vm_stat_diff + item;
440 0 : s8 v, t;
441 :
442 0 : v = __this_cpu_dec_return(*p);
443 0 : t = __this_cpu_read(pcp->stat_threshold);
444 0 : if (unlikely(v < - t)) {
445 0 : s8 overstep = t >> 1;
446 :
447 0 : zone_page_state_add(v - overstep, zone, item);
448 0 : __this_cpu_write(*p, overstep);
449 : }
450 0 : }
451 :
452 0 : void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
453 : {
454 0 : struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
455 0 : s8 __percpu *p = pcp->vm_node_stat_diff + item;
456 0 : s8 v, t;
457 :
458 0 : VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
459 :
460 0 : v = __this_cpu_dec_return(*p);
461 0 : t = __this_cpu_read(pcp->stat_threshold);
462 0 : if (unlikely(v < - t)) {
463 0 : s8 overstep = t >> 1;
464 :
465 0 : node_page_state_add(v - overstep, pgdat, item);
466 0 : __this_cpu_write(*p, overstep);
467 : }
468 0 : }
469 :
470 0 : void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
471 : {
472 0 : __dec_zone_state(page_zone(page), item);
473 0 : }
474 : EXPORT_SYMBOL(__dec_zone_page_state);
475 :
476 0 : void __dec_node_page_state(struct page *page, enum node_stat_item item)
477 : {
478 0 : __dec_node_state(page_pgdat(page), item);
479 0 : }
480 : EXPORT_SYMBOL(__dec_node_page_state);
481 :
482 : #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
483 : /*
484 : * If we have cmpxchg_local support then we do not need to incur the overhead
485 : * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
486 : *
487 : * mod_state() modifies the zone counter state through atomic per cpu
488 : * operations.
489 : *
490 : * Overstep mode specifies how overstep should handled:
491 : * 0 No overstepping
492 : * 1 Overstepping half of threshold
493 : * -1 Overstepping minus half of threshold
494 : */
495 4351 : static inline void mod_zone_state(struct zone *zone,
496 : enum zone_stat_item item, long delta, int overstep_mode)
497 : {
498 4351 : struct per_cpu_pageset __percpu *pcp = zone->pageset;
499 4351 : s8 __percpu *p = pcp->vm_stat_diff + item;
500 4351 : long o, n, t, z;
501 :
502 4351 : do {
503 4351 : z = 0; /* overflow to zone counters */
504 :
505 : /*
506 : * The fetching of the stat_threshold is racy. We may apply
507 : * a counter threshold to the wrong the cpu if we get
508 : * rescheduled while executing here. However, the next
509 : * counter update will apply the threshold again and
510 : * therefore bring the counter under the threshold again.
511 : *
512 : * Most of the time the thresholds are the same anyways
513 : * for all cpus in a zone.
514 : */
515 4351 : t = this_cpu_read(pcp->stat_threshold);
516 :
517 4351 : o = this_cpu_read(*p);
518 4351 : n = delta + o;
519 :
520 4351 : if (abs(n) > t) {
521 34 : int os = overstep_mode * (t >> 1) ;
522 :
523 : /* Overflow must be added to zone counters */
524 34 : z = n + os;
525 34 : n = -os;
526 : }
527 4351 : } while (this_cpu_cmpxchg(*p, o, n) != o);
528 :
529 4351 : if (z)
530 34 : zone_page_state_add(z, zone, item);
531 4351 : }
532 :
533 1 : void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
534 : long delta)
535 : {
536 1 : mod_zone_state(zone, item, delta, 0);
537 1 : }
538 : EXPORT_SYMBOL(mod_zone_page_state);
539 :
540 1284 : void inc_zone_page_state(struct page *page, enum zone_stat_item item)
541 : {
542 1284 : mod_zone_state(page_zone(page), item, 1, 1);
543 1284 : }
544 : EXPORT_SYMBOL(inc_zone_page_state);
545 :
546 3066 : void dec_zone_page_state(struct page *page, enum zone_stat_item item)
547 : {
548 3066 : mod_zone_state(page_zone(page), item, -1, -1);
549 3066 : }
550 : EXPORT_SYMBOL(dec_zone_page_state);
551 :
552 107430 : static inline void mod_node_state(struct pglist_data *pgdat,
553 : enum node_stat_item item, int delta, int overstep_mode)
554 : {
555 107430 : struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
556 107430 : s8 __percpu *p = pcp->vm_node_stat_diff + item;
557 107430 : long o, n, t, z;
558 :
559 107430 : if (vmstat_item_in_bytes(item)) {
560 : /*
561 : * Only cgroups use subpage accounting right now; at
562 : * the global level, these items still change in
563 : * multiples of whole pages. Store them as pages
564 : * internally to keep the per-cpu counters compact.
565 : */
566 44785 : VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
567 44785 : delta >>= PAGE_SHIFT;
568 : }
569 :
570 107430 : do {
571 107430 : z = 0; /* overflow to node counters */
572 :
573 : /*
574 : * The fetching of the stat_threshold is racy. We may apply
575 : * a counter threshold to the wrong the cpu if we get
576 : * rescheduled while executing here. However, the next
577 : * counter update will apply the threshold again and
578 : * therefore bring the counter under the threshold again.
579 : *
580 : * Most of the time the thresholds are the same anyways
581 : * for all cpus in a node.
582 : */
583 107430 : t = this_cpu_read(pcp->stat_threshold);
584 :
585 107430 : o = this_cpu_read(*p);
586 107431 : n = delta + o;
587 :
588 107431 : if (abs(n) > t) {
589 8370 : int os = overstep_mode * (t >> 1) ;
590 :
591 : /* Overflow must be added to node counters */
592 8370 : z = n + os;
593 8370 : n = -os;
594 : }
595 107431 : } while (this_cpu_cmpxchg(*p, o, n) != o);
596 :
597 107431 : if (z)
598 8370 : node_page_state_add(z, pgdat, item);
599 107431 : }
600 :
601 106142 : void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
602 : long delta)
603 : {
604 106142 : mod_node_state(pgdat, item, delta, 0);
605 106146 : }
606 : EXPORT_SYMBOL(mod_node_page_state);
607 :
608 0 : void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
609 : {
610 0 : mod_node_state(pgdat, item, 1, 1);
611 0 : }
612 :
613 1284 : void inc_node_page_state(struct page *page, enum node_stat_item item)
614 : {
615 1284 : mod_node_state(page_pgdat(page), item, 1, 1);
616 1284 : }
617 : EXPORT_SYMBOL(inc_node_page_state);
618 :
619 2 : void dec_node_page_state(struct page *page, enum node_stat_item item)
620 : {
621 2 : mod_node_state(page_pgdat(page), item, -1, -1);
622 2 : }
623 : EXPORT_SYMBOL(dec_node_page_state);
624 : #else
625 : /*
626 : * Use interrupt disable to serialize counter updates
627 : */
628 : void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
629 : long delta)
630 : {
631 : unsigned long flags;
632 :
633 : local_irq_save(flags);
634 : __mod_zone_page_state(zone, item, delta);
635 : local_irq_restore(flags);
636 : }
637 : EXPORT_SYMBOL(mod_zone_page_state);
638 :
639 : void inc_zone_page_state(struct page *page, enum zone_stat_item item)
640 : {
641 : unsigned long flags;
642 : struct zone *zone;
643 :
644 : zone = page_zone(page);
645 : local_irq_save(flags);
646 : __inc_zone_state(zone, item);
647 : local_irq_restore(flags);
648 : }
649 : EXPORT_SYMBOL(inc_zone_page_state);
650 :
651 : void dec_zone_page_state(struct page *page, enum zone_stat_item item)
652 : {
653 : unsigned long flags;
654 :
655 : local_irq_save(flags);
656 : __dec_zone_page_state(page, item);
657 : local_irq_restore(flags);
658 : }
659 : EXPORT_SYMBOL(dec_zone_page_state);
660 :
661 : void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
662 : {
663 : unsigned long flags;
664 :
665 : local_irq_save(flags);
666 : __inc_node_state(pgdat, item);
667 : local_irq_restore(flags);
668 : }
669 : EXPORT_SYMBOL(inc_node_state);
670 :
671 : void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
672 : long delta)
673 : {
674 : unsigned long flags;
675 :
676 : local_irq_save(flags);
677 : __mod_node_page_state(pgdat, item, delta);
678 : local_irq_restore(flags);
679 : }
680 : EXPORT_SYMBOL(mod_node_page_state);
681 :
682 : void inc_node_page_state(struct page *page, enum node_stat_item item)
683 : {
684 : unsigned long flags;
685 : struct pglist_data *pgdat;
686 :
687 : pgdat = page_pgdat(page);
688 : local_irq_save(flags);
689 : __inc_node_state(pgdat, item);
690 : local_irq_restore(flags);
691 : }
692 : EXPORT_SYMBOL(inc_node_page_state);
693 :
694 : void dec_node_page_state(struct page *page, enum node_stat_item item)
695 : {
696 : unsigned long flags;
697 :
698 : local_irq_save(flags);
699 : __dec_node_page_state(page, item);
700 : local_irq_restore(flags);
701 : }
702 : EXPORT_SYMBOL(dec_node_page_state);
703 : #endif
704 :
705 : /*
706 : * Fold a differential into the global counters.
707 : * Returns the number of counters updated.
708 : */
709 : #ifdef CONFIG_NUMA
710 412 : static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
711 : {
712 412 : int i;
713 412 : int changes = 0;
714 :
715 4532 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
716 4120 : if (zone_diff[i]) {
717 1231 : atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
718 1231 : changes++;
719 : }
720 :
721 2884 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
722 2472 : if (numa_diff[i]) {
723 797 : atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
724 797 : changes++;
725 : }
726 :
727 16061 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
728 15649 : if (node_diff[i]) {
729 2618 : atomic_long_add(node_diff[i], &vm_node_stat[i]);
730 2618 : changes++;
731 : }
732 412 : return changes;
733 : }
734 : #else
735 : static int fold_diff(int *zone_diff, int *node_diff)
736 : {
737 : int i;
738 : int changes = 0;
739 :
740 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
741 : if (zone_diff[i]) {
742 : atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
743 : changes++;
744 : }
745 :
746 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
747 : if (node_diff[i]) {
748 : atomic_long_add(node_diff[i], &vm_node_stat[i]);
749 : changes++;
750 : }
751 : return changes;
752 : }
753 : #endif /* CONFIG_NUMA */
754 :
755 : /*
756 : * Update the zone counters for the current cpu.
757 : *
758 : * Note that refresh_cpu_vm_stats strives to only access
759 : * node local memory. The per cpu pagesets on remote zones are placed
760 : * in the memory local to the processor using that pageset. So the
761 : * loop over all zones will access a series of cachelines local to
762 : * the processor.
763 : *
764 : * The call to zone_page_state_add updates the cachelines with the
765 : * statistics in the remote zone struct as well as the global cachelines
766 : * with the global counters. These could cause remote node cache line
767 : * bouncing and will have to be only done when necessary.
768 : *
769 : * The function returns the number of global counters updated.
770 : */
771 412 : static int refresh_cpu_vm_stats(bool do_pagesets)
772 : {
773 412 : struct pglist_data *pgdat;
774 412 : struct zone *zone;
775 412 : int i;
776 412 : int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
777 : #ifdef CONFIG_NUMA
778 412 : int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
779 : #endif
780 412 : int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
781 412 : int changes = 0;
782 :
783 1648 : for_each_populated_zone(zone) {
784 412 : struct per_cpu_pageset __percpu *p = zone->pageset;
785 :
786 4532 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
787 4120 : int v;
788 :
789 4120 : v = this_cpu_xchg(p->vm_stat_diff[i], 0);
790 4120 : if (v) {
791 :
792 1231 : atomic_long_add(v, &zone->vm_stat[i]);
793 1231 : global_zone_diff[i] += v;
794 : #ifdef CONFIG_NUMA
795 : /* 3 seconds idle till flush */
796 4120 : __this_cpu_write(p->expire, 3);
797 : #endif
798 : }
799 : }
800 : #ifdef CONFIG_NUMA
801 2884 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
802 2472 : int v;
803 :
804 2472 : v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
805 2472 : if (v) {
806 :
807 797 : atomic_long_add(v, &zone->vm_numa_stat[i]);
808 797 : global_numa_diff[i] += v;
809 2472 : __this_cpu_write(p->expire, 3);
810 : }
811 : }
812 :
813 412 : if (do_pagesets) {
814 157 : cond_resched();
815 : /*
816 : * Deal with draining the remote pageset of this
817 : * processor
818 : *
819 : * Check if there are pages remaining in this pageset
820 : * if not then there is nothing to expire.
821 : */
822 157 : if (!__this_cpu_read(p->expire) ||
823 155 : !__this_cpu_read(p->pcp.count))
824 2 : continue;
825 :
826 : /*
827 : * We never drain zones local to this processor.
828 : */
829 155 : if (zone_to_nid(zone) == numa_node_id()) {
830 155 : __this_cpu_write(p->expire, 0);
831 155 : continue;
832 : }
833 :
834 0 : if (__this_cpu_dec_return(p->expire))
835 0 : continue;
836 :
837 0 : if (__this_cpu_read(p->pcp.count)) {
838 0 : drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
839 0 : changes++;
840 : }
841 : }
842 : #endif
843 : }
844 :
845 824 : for_each_online_pgdat(pgdat) {
846 412 : struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
847 :
848 16065 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
849 15653 : int v;
850 :
851 15653 : v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
852 15653 : if (v) {
853 2619 : atomic_long_add(v, &pgdat->vm_stat[i]);
854 2619 : global_node_diff[i] += v;
855 : }
856 : }
857 : }
858 :
859 : #ifdef CONFIG_NUMA
860 412 : changes += fold_diff(global_zone_diff, global_numa_diff,
861 : global_node_diff);
862 : #else
863 : changes += fold_diff(global_zone_diff, global_node_diff);
864 : #endif
865 412 : return changes;
866 : }
867 :
868 : /*
869 : * Fold the data for an offline cpu into the global array.
870 : * There cannot be any access by the offline cpu and therefore
871 : * synchronization is simplified.
872 : */
873 0 : void cpu_vm_stats_fold(int cpu)
874 : {
875 0 : struct pglist_data *pgdat;
876 0 : struct zone *zone;
877 0 : int i;
878 0 : int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
879 : #ifdef CONFIG_NUMA
880 0 : int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
881 : #endif
882 0 : int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
883 :
884 0 : for_each_populated_zone(zone) {
885 0 : struct per_cpu_pageset *p;
886 :
887 0 : p = per_cpu_ptr(zone->pageset, cpu);
888 :
889 0 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
890 0 : if (p->vm_stat_diff[i]) {
891 0 : int v;
892 :
893 0 : v = p->vm_stat_diff[i];
894 0 : p->vm_stat_diff[i] = 0;
895 0 : atomic_long_add(v, &zone->vm_stat[i]);
896 0 : global_zone_diff[i] += v;
897 : }
898 :
899 : #ifdef CONFIG_NUMA
900 0 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
901 0 : if (p->vm_numa_stat_diff[i]) {
902 0 : int v;
903 :
904 0 : v = p->vm_numa_stat_diff[i];
905 0 : p->vm_numa_stat_diff[i] = 0;
906 0 : atomic_long_add(v, &zone->vm_numa_stat[i]);
907 0 : global_numa_diff[i] += v;
908 : }
909 : #endif
910 : }
911 :
912 0 : for_each_online_pgdat(pgdat) {
913 0 : struct per_cpu_nodestat *p;
914 :
915 0 : p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
916 :
917 0 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
918 0 : if (p->vm_node_stat_diff[i]) {
919 0 : int v;
920 :
921 0 : v = p->vm_node_stat_diff[i];
922 0 : p->vm_node_stat_diff[i] = 0;
923 0 : atomic_long_add(v, &pgdat->vm_stat[i]);
924 0 : global_node_diff[i] += v;
925 : }
926 : }
927 :
928 : #ifdef CONFIG_NUMA
929 0 : fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
930 : #else
931 : fold_diff(global_zone_diff, global_node_diff);
932 : #endif
933 0 : }
934 :
935 : /*
936 : * this is only called if !populated_zone(zone), which implies no other users of
937 : * pset->vm_stat_diff[] exsist.
938 : */
939 0 : void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
940 : {
941 0 : int i;
942 :
943 0 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
944 0 : if (pset->vm_stat_diff[i]) {
945 0 : int v = pset->vm_stat_diff[i];
946 0 : pset->vm_stat_diff[i] = 0;
947 0 : atomic_long_add(v, &zone->vm_stat[i]);
948 0 : atomic_long_add(v, &vm_zone_stat[i]);
949 : }
950 :
951 : #ifdef CONFIG_NUMA
952 0 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
953 0 : if (pset->vm_numa_stat_diff[i]) {
954 0 : int v = pset->vm_numa_stat_diff[i];
955 :
956 0 : pset->vm_numa_stat_diff[i] = 0;
957 0 : atomic_long_add(v, &zone->vm_numa_stat[i]);
958 0 : atomic_long_add(v, &vm_numa_stat[i]);
959 : }
960 : #endif
961 0 : }
962 : #endif
963 :
964 : #ifdef CONFIG_NUMA
965 375138 : void __inc_numa_state(struct zone *zone,
966 : enum numa_stat_item item)
967 : {
968 375138 : struct per_cpu_pageset __percpu *pcp = zone->pageset;
969 375138 : u16 __percpu *p = pcp->vm_numa_stat_diff + item;
970 375138 : u16 v;
971 :
972 375138 : v = __this_cpu_inc_return(*p);
973 :
974 375138 : if (unlikely(v > NUMA_STATS_THRESHOLD)) {
975 0 : zone_numa_state_add(v, zone, item);
976 375138 : __this_cpu_write(*p, 0);
977 : }
978 375138 : }
979 :
980 : /*
981 : * Determine the per node value of a stat item. This function
982 : * is called frequently in a NUMA machine, so try to be as
983 : * frugal as possible.
984 : */
985 0 : unsigned long sum_zone_node_page_state(int node,
986 : enum zone_stat_item item)
987 : {
988 0 : struct zone *zones = NODE_DATA(node)->node_zones;
989 0 : int i;
990 0 : unsigned long count = 0;
991 :
992 0 : for (i = 0; i < MAX_NR_ZONES; i++)
993 0 : count += zone_page_state(zones + i, item);
994 :
995 0 : return count;
996 : }
997 :
998 : /*
999 : * Determine the per node value of a numa stat item. To avoid deviation,
1000 : * the per cpu stat number in vm_numa_stat_diff[] is also included.
1001 : */
1002 0 : unsigned long sum_zone_numa_state(int node,
1003 : enum numa_stat_item item)
1004 : {
1005 0 : struct zone *zones = NODE_DATA(node)->node_zones;
1006 0 : int i;
1007 0 : unsigned long count = 0;
1008 :
1009 0 : for (i = 0; i < MAX_NR_ZONES; i++)
1010 0 : count += zone_numa_state_snapshot(zones + i, item);
1011 :
1012 0 : return count;
1013 : }
1014 :
1015 : /*
1016 : * Determine the per node value of a stat item.
1017 : */
1018 6124 : unsigned long node_page_state_pages(struct pglist_data *pgdat,
1019 : enum node_stat_item item)
1020 : {
1021 0 : long x = atomic_long_read(&pgdat->vm_stat[item]);
1022 : #ifdef CONFIG_SMP
1023 6124 : if (x < 0)
1024 : x = 0;
1025 : #endif
1026 6124 : return x;
1027 : }
1028 :
1029 6124 : unsigned long node_page_state(struct pglist_data *pgdat,
1030 : enum node_stat_item item)
1031 : {
1032 6124 : VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1033 :
1034 6124 : return node_page_state_pages(pgdat, item);
1035 : }
1036 : #endif
1037 :
1038 : #ifdef CONFIG_COMPACTION
1039 :
1040 : struct contig_page_info {
1041 : unsigned long free_pages;
1042 : unsigned long free_blocks_total;
1043 : unsigned long free_blocks_suitable;
1044 : };
1045 :
1046 : /*
1047 : * Calculate the number of free pages in a zone, how many contiguous
1048 : * pages are free and how many are large enough to satisfy an allocation of
1049 : * the target size. Note that this function makes no attempt to estimate
1050 : * how many suitable free blocks there *might* be if MOVABLE pages were
1051 : * migrated. Calculating that is possible, but expensive and can be
1052 : * figured out from userspace
1053 : */
1054 234 : static void fill_contig_page_info(struct zone *zone,
1055 : unsigned int suitable_order,
1056 : struct contig_page_info *info)
1057 : {
1058 234 : unsigned int order;
1059 :
1060 234 : info->free_pages = 0;
1061 234 : info->free_blocks_total = 0;
1062 234 : info->free_blocks_suitable = 0;
1063 :
1064 2808 : for (order = 0; order < MAX_ORDER; order++) {
1065 2574 : unsigned long blocks;
1066 :
1067 : /* Count number of free blocks */
1068 2574 : blocks = zone->free_area[order].nr_free;
1069 2574 : info->free_blocks_total += blocks;
1070 :
1071 : /* Count free base pages */
1072 2574 : info->free_pages += blocks << order;
1073 :
1074 : /* Count the suitable free blocks */
1075 2574 : if (order >= suitable_order)
1076 468 : info->free_blocks_suitable += blocks <<
1077 468 : (order - suitable_order);
1078 : }
1079 234 : }
1080 :
1081 : /*
1082 : * A fragmentation index only makes sense if an allocation of a requested
1083 : * size would fail. If that is true, the fragmentation index indicates
1084 : * whether external fragmentation or a lack of memory was the problem.
1085 : * The value can be used to determine if page reclaim or compaction
1086 : * should be used
1087 : */
1088 0 : static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1089 : {
1090 0 : unsigned long requested = 1UL << order;
1091 :
1092 0 : if (WARN_ON_ONCE(order >= MAX_ORDER))
1093 : return 0;
1094 :
1095 0 : if (!info->free_blocks_total)
1096 : return 0;
1097 :
1098 : /* Fragmentation index only makes sense when a request would fail */
1099 0 : if (info->free_blocks_suitable)
1100 : return -1000;
1101 :
1102 : /*
1103 : * Index is between 0 and 1 so return within 3 decimal places
1104 : *
1105 : * 0 => allocation would fail due to lack of memory
1106 : * 1 => allocation would fail due to fragmentation
1107 : */
1108 0 : return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1109 : }
1110 :
1111 : /*
1112 : * Calculates external fragmentation within a zone wrt the given order.
1113 : * It is defined as the percentage of pages found in blocks of size
1114 : * less than 1 << order. It returns values in range [0, 100].
1115 : */
1116 234 : unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1117 : {
1118 234 : struct contig_page_info info;
1119 :
1120 234 : fill_contig_page_info(zone, order, &info);
1121 234 : if (info.free_pages == 0)
1122 : return 0;
1123 :
1124 78 : return div_u64((info.free_pages -
1125 78 : (info.free_blocks_suitable << order)) * 100,
1126 : info.free_pages);
1127 : }
1128 :
1129 : /* Same as __fragmentation index but allocs contig_page_info on stack */
1130 0 : int fragmentation_index(struct zone *zone, unsigned int order)
1131 : {
1132 0 : struct contig_page_info info;
1133 :
1134 0 : fill_contig_page_info(zone, order, &info);
1135 0 : return __fragmentation_index(order, &info);
1136 : }
1137 : #endif
1138 :
1139 : #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1140 : defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1141 : #ifdef CONFIG_ZONE_DMA
1142 : #define TEXT_FOR_DMA(xx) xx "_dma",
1143 : #else
1144 : #define TEXT_FOR_DMA(xx)
1145 : #endif
1146 :
1147 : #ifdef CONFIG_ZONE_DMA32
1148 : #define TEXT_FOR_DMA32(xx) xx "_dma32",
1149 : #else
1150 : #define TEXT_FOR_DMA32(xx)
1151 : #endif
1152 :
1153 : #ifdef CONFIG_HIGHMEM
1154 : #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1155 : #else
1156 : #define TEXT_FOR_HIGHMEM(xx)
1157 : #endif
1158 :
1159 : #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1160 : TEXT_FOR_HIGHMEM(xx) xx "_movable",
1161 :
1162 : const char * const vmstat_text[] = {
1163 : /* enum zone_stat_item counters */
1164 : "nr_free_pages",
1165 : "nr_zone_inactive_anon",
1166 : "nr_zone_active_anon",
1167 : "nr_zone_inactive_file",
1168 : "nr_zone_active_file",
1169 : "nr_zone_unevictable",
1170 : "nr_zone_write_pending",
1171 : "nr_mlock",
1172 : "nr_bounce",
1173 : #if IS_ENABLED(CONFIG_ZSMALLOC)
1174 : "nr_zspages",
1175 : #endif
1176 : "nr_free_cma",
1177 :
1178 : /* enum numa_stat_item counters */
1179 : #ifdef CONFIG_NUMA
1180 : "numa_hit",
1181 : "numa_miss",
1182 : "numa_foreign",
1183 : "numa_interleave",
1184 : "numa_local",
1185 : "numa_other",
1186 : #endif
1187 :
1188 : /* enum node_stat_item counters */
1189 : "nr_inactive_anon",
1190 : "nr_active_anon",
1191 : "nr_inactive_file",
1192 : "nr_active_file",
1193 : "nr_unevictable",
1194 : "nr_slab_reclaimable",
1195 : "nr_slab_unreclaimable",
1196 : "nr_isolated_anon",
1197 : "nr_isolated_file",
1198 : "workingset_nodes",
1199 : "workingset_refault_anon",
1200 : "workingset_refault_file",
1201 : "workingset_activate_anon",
1202 : "workingset_activate_file",
1203 : "workingset_restore_anon",
1204 : "workingset_restore_file",
1205 : "workingset_nodereclaim",
1206 : "nr_anon_pages",
1207 : "nr_mapped",
1208 : "nr_file_pages",
1209 : "nr_dirty",
1210 : "nr_writeback",
1211 : "nr_writeback_temp",
1212 : "nr_shmem",
1213 : "nr_shmem_hugepages",
1214 : "nr_shmem_pmdmapped",
1215 : "nr_file_hugepages",
1216 : "nr_file_pmdmapped",
1217 : "nr_anon_transparent_hugepages",
1218 : "nr_vmscan_write",
1219 : "nr_vmscan_immediate_reclaim",
1220 : "nr_dirtied",
1221 : "nr_written",
1222 : "nr_kernel_misc_reclaimable",
1223 : "nr_foll_pin_acquired",
1224 : "nr_foll_pin_released",
1225 : "nr_kernel_stack",
1226 : #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1227 : "nr_shadow_call_stack",
1228 : #endif
1229 : "nr_page_table_pages",
1230 : #ifdef CONFIG_SWAP
1231 : "nr_swapcached",
1232 : #endif
1233 :
1234 : /* enum writeback_stat_item counters */
1235 : "nr_dirty_threshold",
1236 : "nr_dirty_background_threshold",
1237 :
1238 : #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1239 : /* enum vm_event_item counters */
1240 : "pgpgin",
1241 : "pgpgout",
1242 : "pswpin",
1243 : "pswpout",
1244 :
1245 : TEXTS_FOR_ZONES("pgalloc")
1246 : TEXTS_FOR_ZONES("allocstall")
1247 : TEXTS_FOR_ZONES("pgskip")
1248 :
1249 : "pgfree",
1250 : "pgactivate",
1251 : "pgdeactivate",
1252 : "pglazyfree",
1253 :
1254 : "pgfault",
1255 : "pgmajfault",
1256 : "pglazyfreed",
1257 :
1258 : "pgrefill",
1259 : "pgreuse",
1260 : "pgsteal_kswapd",
1261 : "pgsteal_direct",
1262 : "pgscan_kswapd",
1263 : "pgscan_direct",
1264 : "pgscan_direct_throttle",
1265 : "pgscan_anon",
1266 : "pgscan_file",
1267 : "pgsteal_anon",
1268 : "pgsteal_file",
1269 :
1270 : #ifdef CONFIG_NUMA
1271 : "zone_reclaim_failed",
1272 : #endif
1273 : "pginodesteal",
1274 : "slabs_scanned",
1275 : "kswapd_inodesteal",
1276 : "kswapd_low_wmark_hit_quickly",
1277 : "kswapd_high_wmark_hit_quickly",
1278 : "pageoutrun",
1279 :
1280 : "pgrotated",
1281 :
1282 : "drop_pagecache",
1283 : "drop_slab",
1284 : "oom_kill",
1285 :
1286 : #ifdef CONFIG_NUMA_BALANCING
1287 : "numa_pte_updates",
1288 : "numa_huge_pte_updates",
1289 : "numa_hint_faults",
1290 : "numa_hint_faults_local",
1291 : "numa_pages_migrated",
1292 : #endif
1293 : #ifdef CONFIG_MIGRATION
1294 : "pgmigrate_success",
1295 : "pgmigrate_fail",
1296 : "thp_migration_success",
1297 : "thp_migration_fail",
1298 : "thp_migration_split",
1299 : #endif
1300 : #ifdef CONFIG_COMPACTION
1301 : "compact_migrate_scanned",
1302 : "compact_free_scanned",
1303 : "compact_isolated",
1304 : "compact_stall",
1305 : "compact_fail",
1306 : "compact_success",
1307 : "compact_daemon_wake",
1308 : "compact_daemon_migrate_scanned",
1309 : "compact_daemon_free_scanned",
1310 : #endif
1311 :
1312 : #ifdef CONFIG_HUGETLB_PAGE
1313 : "htlb_buddy_alloc_success",
1314 : "htlb_buddy_alloc_fail",
1315 : #endif
1316 : "unevictable_pgs_culled",
1317 : "unevictable_pgs_scanned",
1318 : "unevictable_pgs_rescued",
1319 : "unevictable_pgs_mlocked",
1320 : "unevictable_pgs_munlocked",
1321 : "unevictable_pgs_cleared",
1322 : "unevictable_pgs_stranded",
1323 :
1324 : #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1325 : "thp_fault_alloc",
1326 : "thp_fault_fallback",
1327 : "thp_fault_fallback_charge",
1328 : "thp_collapse_alloc",
1329 : "thp_collapse_alloc_failed",
1330 : "thp_file_alloc",
1331 : "thp_file_fallback",
1332 : "thp_file_fallback_charge",
1333 : "thp_file_mapped",
1334 : "thp_split_page",
1335 : "thp_split_page_failed",
1336 : "thp_deferred_split_page",
1337 : "thp_split_pmd",
1338 : #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1339 : "thp_split_pud",
1340 : #endif
1341 : "thp_zero_page_alloc",
1342 : "thp_zero_page_alloc_failed",
1343 : "thp_swpout",
1344 : "thp_swpout_fallback",
1345 : #endif
1346 : #ifdef CONFIG_MEMORY_BALLOON
1347 : "balloon_inflate",
1348 : "balloon_deflate",
1349 : #ifdef CONFIG_BALLOON_COMPACTION
1350 : "balloon_migrate",
1351 : #endif
1352 : #endif /* CONFIG_MEMORY_BALLOON */
1353 : #ifdef CONFIG_DEBUG_TLBFLUSH
1354 : "nr_tlb_remote_flush",
1355 : "nr_tlb_remote_flush_received",
1356 : "nr_tlb_local_flush_all",
1357 : "nr_tlb_local_flush_one",
1358 : #endif /* CONFIG_DEBUG_TLBFLUSH */
1359 :
1360 : #ifdef CONFIG_DEBUG_VM_VMACACHE
1361 : "vmacache_find_calls",
1362 : "vmacache_find_hits",
1363 : #endif
1364 : #ifdef CONFIG_SWAP
1365 : "swap_ra",
1366 : "swap_ra_hit",
1367 : #endif
1368 : #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1369 : };
1370 : #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1371 :
1372 : #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1373 : defined(CONFIG_PROC_FS)
1374 0 : static void *frag_start(struct seq_file *m, loff_t *pos)
1375 : {
1376 0 : pg_data_t *pgdat;
1377 0 : loff_t node = *pos;
1378 :
1379 0 : for (pgdat = first_online_pgdat();
1380 0 : pgdat && node;
1381 0 : pgdat = next_online_pgdat(pgdat))
1382 0 : --node;
1383 :
1384 0 : return pgdat;
1385 : }
1386 :
1387 0 : static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1388 : {
1389 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1390 :
1391 0 : (*pos)++;
1392 0 : return next_online_pgdat(pgdat);
1393 : }
1394 :
1395 0 : static void frag_stop(struct seq_file *m, void *arg)
1396 : {
1397 0 : }
1398 :
1399 : /*
1400 : * Walk zones in a node and print using a callback.
1401 : * If @assert_populated is true, only use callback for zones that are populated.
1402 : */
1403 0 : static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1404 : bool assert_populated, bool nolock,
1405 : void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1406 : {
1407 0 : struct zone *zone;
1408 0 : struct zone *node_zones = pgdat->node_zones;
1409 0 : unsigned long flags;
1410 :
1411 0 : for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1412 0 : if (assert_populated && !populated_zone(zone))
1413 0 : continue;
1414 :
1415 0 : if (!nolock)
1416 0 : spin_lock_irqsave(&zone->lock, flags);
1417 0 : print(m, pgdat, zone);
1418 0 : if (!nolock)
1419 0 : spin_unlock_irqrestore(&zone->lock, flags);
1420 : }
1421 0 : }
1422 : #endif
1423 :
1424 : #ifdef CONFIG_PROC_FS
1425 0 : static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1426 : struct zone *zone)
1427 : {
1428 0 : int order;
1429 :
1430 0 : seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1431 0 : for (order = 0; order < MAX_ORDER; ++order)
1432 0 : seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1433 0 : seq_putc(m, '\n');
1434 0 : }
1435 :
1436 : /*
1437 : * This walks the free areas for each zone.
1438 : */
1439 0 : static int frag_show(struct seq_file *m, void *arg)
1440 : {
1441 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1442 0 : walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1443 0 : return 0;
1444 : }
1445 :
1446 0 : static void pagetypeinfo_showfree_print(struct seq_file *m,
1447 : pg_data_t *pgdat, struct zone *zone)
1448 : {
1449 0 : int order, mtype;
1450 :
1451 0 : for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1452 0 : seq_printf(m, "Node %4d, zone %8s, type %12s ",
1453 : pgdat->node_id,
1454 : zone->name,
1455 : migratetype_names[mtype]);
1456 0 : for (order = 0; order < MAX_ORDER; ++order) {
1457 0 : unsigned long freecount = 0;
1458 0 : struct free_area *area;
1459 0 : struct list_head *curr;
1460 0 : bool overflow = false;
1461 :
1462 0 : area = &(zone->free_area[order]);
1463 :
1464 0 : list_for_each(curr, &area->free_list[mtype]) {
1465 : /*
1466 : * Cap the free_list iteration because it might
1467 : * be really large and we are under a spinlock
1468 : * so a long time spent here could trigger a
1469 : * hard lockup detector. Anyway this is a
1470 : * debugging tool so knowing there is a handful
1471 : * of pages of this order should be more than
1472 : * sufficient.
1473 : */
1474 0 : if (++freecount >= 100000) {
1475 : overflow = true;
1476 : break;
1477 : }
1478 : }
1479 0 : seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1480 0 : spin_unlock_irq(&zone->lock);
1481 0 : cond_resched();
1482 0 : spin_lock_irq(&zone->lock);
1483 : }
1484 0 : seq_putc(m, '\n');
1485 : }
1486 0 : }
1487 :
1488 : /* Print out the free pages at each order for each migatetype */
1489 0 : static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1490 : {
1491 0 : int order;
1492 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1493 :
1494 : /* Print header */
1495 0 : seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1496 0 : for (order = 0; order < MAX_ORDER; ++order)
1497 0 : seq_printf(m, "%6d ", order);
1498 0 : seq_putc(m, '\n');
1499 :
1500 0 : walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1501 :
1502 0 : return 0;
1503 : }
1504 :
1505 0 : static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1506 : pg_data_t *pgdat, struct zone *zone)
1507 : {
1508 0 : int mtype;
1509 0 : unsigned long pfn;
1510 0 : unsigned long start_pfn = zone->zone_start_pfn;
1511 0 : unsigned long end_pfn = zone_end_pfn(zone);
1512 0 : unsigned long count[MIGRATE_TYPES] = { 0, };
1513 :
1514 0 : for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1515 0 : struct page *page;
1516 :
1517 0 : page = pfn_to_online_page(pfn);
1518 0 : if (!page)
1519 0 : continue;
1520 :
1521 0 : if (page_zone(page) != zone)
1522 0 : continue;
1523 :
1524 0 : mtype = get_pageblock_migratetype(page);
1525 :
1526 0 : if (mtype < MIGRATE_TYPES)
1527 0 : count[mtype]++;
1528 : }
1529 :
1530 : /* Print counts */
1531 0 : seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1532 0 : for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1533 0 : seq_printf(m, "%12lu ", count[mtype]);
1534 0 : seq_putc(m, '\n');
1535 0 : }
1536 :
1537 : /* Print out the number of pageblocks for each migratetype */
1538 0 : static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1539 : {
1540 0 : int mtype;
1541 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1542 :
1543 0 : seq_printf(m, "\n%-23s", "Number of blocks type ");
1544 0 : for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1545 0 : seq_printf(m, "%12s ", migratetype_names[mtype]);
1546 0 : seq_putc(m, '\n');
1547 0 : walk_zones_in_node(m, pgdat, true, false,
1548 : pagetypeinfo_showblockcount_print);
1549 :
1550 0 : return 0;
1551 : }
1552 :
1553 : /*
1554 : * Print out the number of pageblocks for each migratetype that contain pages
1555 : * of other types. This gives an indication of how well fallbacks are being
1556 : * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1557 : * to determine what is going on
1558 : */
1559 0 : static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1560 : {
1561 : #ifdef CONFIG_PAGE_OWNER
1562 : int mtype;
1563 :
1564 : if (!static_branch_unlikely(&page_owner_inited))
1565 : return;
1566 :
1567 : drain_all_pages(NULL);
1568 :
1569 : seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1570 : for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1571 : seq_printf(m, "%12s ", migratetype_names[mtype]);
1572 : seq_putc(m, '\n');
1573 :
1574 : walk_zones_in_node(m, pgdat, true, true,
1575 : pagetypeinfo_showmixedcount_print);
1576 : #endif /* CONFIG_PAGE_OWNER */
1577 0 : }
1578 :
1579 : /*
1580 : * This prints out statistics in relation to grouping pages by mobility.
1581 : * It is expensive to collect so do not constantly read the file.
1582 : */
1583 0 : static int pagetypeinfo_show(struct seq_file *m, void *arg)
1584 : {
1585 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1586 :
1587 : /* check memoryless node */
1588 0 : if (!node_state(pgdat->node_id, N_MEMORY))
1589 : return 0;
1590 :
1591 0 : seq_printf(m, "Page block order: %d\n", pageblock_order);
1592 0 : seq_printf(m, "Pages per block: %lu\n", pageblock_nr_pages);
1593 0 : seq_putc(m, '\n');
1594 0 : pagetypeinfo_showfree(m, pgdat);
1595 0 : pagetypeinfo_showblockcount(m, pgdat);
1596 0 : pagetypeinfo_showmixedcount(m, pgdat);
1597 :
1598 0 : return 0;
1599 : }
1600 :
1601 : static const struct seq_operations fragmentation_op = {
1602 : .start = frag_start,
1603 : .next = frag_next,
1604 : .stop = frag_stop,
1605 : .show = frag_show,
1606 : };
1607 :
1608 : static const struct seq_operations pagetypeinfo_op = {
1609 : .start = frag_start,
1610 : .next = frag_next,
1611 : .stop = frag_stop,
1612 : .show = pagetypeinfo_show,
1613 : };
1614 :
1615 0 : static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1616 : {
1617 0 : int zid;
1618 :
1619 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1620 0 : struct zone *compare = &pgdat->node_zones[zid];
1621 :
1622 0 : if (populated_zone(compare))
1623 0 : return zone == compare;
1624 : }
1625 :
1626 : return false;
1627 : }
1628 :
1629 0 : static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1630 : struct zone *zone)
1631 : {
1632 0 : int i;
1633 0 : seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1634 0 : if (is_zone_first_populated(pgdat, zone)) {
1635 0 : seq_printf(m, "\n per-node stats");
1636 0 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1637 0 : unsigned long pages = node_page_state_pages(pgdat, i);
1638 :
1639 0 : if (vmstat_item_print_in_thp(i))
1640 0 : pages /= HPAGE_PMD_NR;
1641 0 : seq_printf(m, "\n %-12s %lu", node_stat_name(i),
1642 : pages);
1643 : }
1644 : }
1645 0 : seq_printf(m,
1646 : "\n pages free %lu"
1647 : "\n min %lu"
1648 : "\n low %lu"
1649 : "\n high %lu"
1650 : "\n spanned %lu"
1651 : "\n present %lu"
1652 : "\n managed %lu"
1653 : "\n cma %lu",
1654 : zone_page_state(zone, NR_FREE_PAGES),
1655 0 : min_wmark_pages(zone),
1656 0 : low_wmark_pages(zone),
1657 0 : high_wmark_pages(zone),
1658 : zone->spanned_pages,
1659 : zone->present_pages,
1660 : zone_managed_pages(zone),
1661 : zone_cma_pages(zone));
1662 :
1663 0 : seq_printf(m,
1664 : "\n protection: (%ld",
1665 : zone->lowmem_reserve[0]);
1666 0 : for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1667 0 : seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1668 0 : seq_putc(m, ')');
1669 :
1670 : /* If unpopulated, no other information is useful */
1671 0 : if (!populated_zone(zone)) {
1672 0 : seq_putc(m, '\n');
1673 0 : return;
1674 : }
1675 :
1676 0 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1677 0 : seq_printf(m, "\n %-12s %lu", zone_stat_name(i),
1678 : zone_page_state(zone, i));
1679 :
1680 : #ifdef CONFIG_NUMA
1681 0 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1682 0 : seq_printf(m, "\n %-12s %lu", numa_stat_name(i),
1683 : zone_numa_state_snapshot(zone, i));
1684 : #endif
1685 :
1686 0 : seq_printf(m, "\n pagesets");
1687 0 : for_each_online_cpu(i) {
1688 0 : struct per_cpu_pageset *pageset;
1689 :
1690 0 : pageset = per_cpu_ptr(zone->pageset, i);
1691 0 : seq_printf(m,
1692 : "\n cpu: %i"
1693 : "\n count: %i"
1694 : "\n high: %i"
1695 : "\n batch: %i",
1696 : i,
1697 : pageset->pcp.count,
1698 : pageset->pcp.high,
1699 : pageset->pcp.batch);
1700 : #ifdef CONFIG_SMP
1701 0 : seq_printf(m, "\n vm stats threshold: %d",
1702 0 : pageset->stat_threshold);
1703 : #endif
1704 : }
1705 0 : seq_printf(m,
1706 : "\n node_unreclaimable: %u"
1707 : "\n start_pfn: %lu",
1708 0 : pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1709 : zone->zone_start_pfn);
1710 0 : seq_putc(m, '\n');
1711 : }
1712 :
1713 : /*
1714 : * Output information about zones in @pgdat. All zones are printed regardless
1715 : * of whether they are populated or not: lowmem_reserve_ratio operates on the
1716 : * set of all zones and userspace would not be aware of such zones if they are
1717 : * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1718 : */
1719 0 : static int zoneinfo_show(struct seq_file *m, void *arg)
1720 : {
1721 0 : pg_data_t *pgdat = (pg_data_t *)arg;
1722 0 : walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1723 0 : return 0;
1724 : }
1725 :
1726 : static const struct seq_operations zoneinfo_op = {
1727 : .start = frag_start, /* iterate over all zones. The same as in
1728 : * fragmentation. */
1729 : .next = frag_next,
1730 : .stop = frag_stop,
1731 : .show = zoneinfo_show,
1732 : };
1733 :
1734 : #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1735 : NR_VM_NUMA_STAT_ITEMS + \
1736 : NR_VM_NODE_STAT_ITEMS + \
1737 : NR_VM_WRITEBACK_STAT_ITEMS + \
1738 : (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1739 : NR_VM_EVENT_ITEMS : 0))
1740 :
1741 0 : static void *vmstat_start(struct seq_file *m, loff_t *pos)
1742 : {
1743 0 : unsigned long *v;
1744 0 : int i;
1745 :
1746 0 : if (*pos >= NR_VMSTAT_ITEMS)
1747 : return NULL;
1748 :
1749 0 : BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1750 0 : v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1751 0 : m->private = v;
1752 0 : if (!v)
1753 0 : return ERR_PTR(-ENOMEM);
1754 0 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1755 0 : v[i] = global_zone_page_state(i);
1756 0 : v += NR_VM_ZONE_STAT_ITEMS;
1757 :
1758 : #ifdef CONFIG_NUMA
1759 0 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1760 0 : v[i] = global_numa_state(i);
1761 0 : v += NR_VM_NUMA_STAT_ITEMS;
1762 : #endif
1763 :
1764 0 : for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1765 0 : v[i] = global_node_page_state_pages(i);
1766 0 : if (vmstat_item_print_in_thp(i))
1767 0 : v[i] /= HPAGE_PMD_NR;
1768 : }
1769 0 : v += NR_VM_NODE_STAT_ITEMS;
1770 :
1771 0 : global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1772 : v + NR_DIRTY_THRESHOLD);
1773 0 : v += NR_VM_WRITEBACK_STAT_ITEMS;
1774 :
1775 : #ifdef CONFIG_VM_EVENT_COUNTERS
1776 0 : all_vm_events(v);
1777 0 : v[PGPGIN] /= 2; /* sectors -> kbytes */
1778 0 : v[PGPGOUT] /= 2;
1779 : #endif
1780 0 : return (unsigned long *)m->private + *pos;
1781 : }
1782 :
1783 0 : static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1784 : {
1785 0 : (*pos)++;
1786 0 : if (*pos >= NR_VMSTAT_ITEMS)
1787 : return NULL;
1788 0 : return (unsigned long *)m->private + *pos;
1789 : }
1790 :
1791 0 : static int vmstat_show(struct seq_file *m, void *arg)
1792 : {
1793 0 : unsigned long *l = arg;
1794 0 : unsigned long off = l - (unsigned long *)m->private;
1795 :
1796 0 : seq_puts(m, vmstat_text[off]);
1797 0 : seq_put_decimal_ull(m, " ", *l);
1798 0 : seq_putc(m, '\n');
1799 :
1800 0 : if (off == NR_VMSTAT_ITEMS - 1) {
1801 : /*
1802 : * We've come to the end - add any deprecated counters to avoid
1803 : * breaking userspace which might depend on them being present.
1804 : */
1805 0 : seq_puts(m, "nr_unstable 0\n");
1806 : }
1807 0 : return 0;
1808 : }
1809 :
1810 0 : static void vmstat_stop(struct seq_file *m, void *arg)
1811 : {
1812 0 : kfree(m->private);
1813 0 : m->private = NULL;
1814 0 : }
1815 :
1816 : static const struct seq_operations vmstat_op = {
1817 : .start = vmstat_start,
1818 : .next = vmstat_next,
1819 : .stop = vmstat_stop,
1820 : .show = vmstat_show,
1821 : };
1822 : #endif /* CONFIG_PROC_FS */
1823 :
1824 : #ifdef CONFIG_SMP
1825 : static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1826 : int sysctl_stat_interval __read_mostly = HZ;
1827 :
1828 : #ifdef CONFIG_PROC_FS
1829 0 : static void refresh_vm_stats(struct work_struct *work)
1830 : {
1831 0 : refresh_cpu_vm_stats(true);
1832 0 : }
1833 :
1834 0 : int vmstat_refresh(struct ctl_table *table, int write,
1835 : void *buffer, size_t *lenp, loff_t *ppos)
1836 : {
1837 0 : long val;
1838 0 : int err;
1839 0 : int i;
1840 :
1841 : /*
1842 : * The regular update, every sysctl_stat_interval, may come later
1843 : * than expected: leaving a significant amount in per_cpu buckets.
1844 : * This is particularly misleading when checking a quantity of HUGE
1845 : * pages, immediately after running a test. /proc/sys/vm/stat_refresh,
1846 : * which can equally be echo'ed to or cat'ted from (by root),
1847 : * can be used to update the stats just before reading them.
1848 : *
1849 : * Oh, and since global_zone_page_state() etc. are so careful to hide
1850 : * transiently negative values, report an error here if any of
1851 : * the stats is negative, so we know to go looking for imbalance.
1852 : */
1853 0 : err = schedule_on_each_cpu(refresh_vm_stats);
1854 0 : if (err)
1855 : return err;
1856 0 : for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1857 0 : val = atomic_long_read(&vm_zone_stat[i]);
1858 0 : if (val < 0) {
1859 0 : pr_warn("%s: %s %ld\n",
1860 : __func__, zone_stat_name(i), val);
1861 0 : err = -EINVAL;
1862 : }
1863 : }
1864 : #ifdef CONFIG_NUMA
1865 0 : for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1866 0 : val = atomic_long_read(&vm_numa_stat[i]);
1867 0 : if (val < 0) {
1868 0 : pr_warn("%s: %s %ld\n",
1869 : __func__, numa_stat_name(i), val);
1870 0 : err = -EINVAL;
1871 : }
1872 : }
1873 : #endif
1874 0 : if (err)
1875 : return err;
1876 0 : if (write)
1877 0 : *ppos += *lenp;
1878 : else
1879 0 : *lenp = 0;
1880 : return 0;
1881 : }
1882 : #endif /* CONFIG_PROC_FS */
1883 :
1884 157 : static void vmstat_update(struct work_struct *w)
1885 : {
1886 157 : if (refresh_cpu_vm_stats(true)) {
1887 : /*
1888 : * Counters were updated so we expect more updates
1889 : * to occur in the future. Keep on running the
1890 : * update worker thread.
1891 : */
1892 145 : queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1893 145 : this_cpu_ptr(&vmstat_work),
1894 : round_jiffies_relative(sysctl_stat_interval));
1895 : }
1896 157 : }
1897 :
1898 : /*
1899 : * Switch off vmstat processing and then fold all the remaining differentials
1900 : * until the diffs stay at zero. The function is used by NOHZ and can only be
1901 : * invoked when tick processing is not active.
1902 : */
1903 : /*
1904 : * Check if the diffs for a certain cpu indicate that
1905 : * an update is needed.
1906 : */
1907 986 : static bool need_update(int cpu)
1908 : {
1909 986 : pg_data_t *last_pgdat = NULL;
1910 986 : struct zone *zone;
1911 :
1912 3131 : for_each_populated_zone(zone) {
1913 986 : struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1914 986 : struct per_cpu_nodestat *n;
1915 : /*
1916 : * The fast way of checking if there are any vmstat diffs.
1917 : */
1918 986 : if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1919 : sizeof(p->vm_stat_diff[0])))
1920 : return true;
1921 : #ifdef CONFIG_NUMA
1922 725 : if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1923 : sizeof(p->vm_numa_stat_diff[0])))
1924 : return true;
1925 : #endif
1926 715 : if (last_pgdat == zone->zone_pgdat)
1927 0 : continue;
1928 715 : last_pgdat = zone->zone_pgdat;
1929 715 : n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1930 715 : if (memchr_inv(n->vm_node_stat_diff, 0, NR_VM_NODE_STAT_ITEMS *
1931 : sizeof(n->vm_node_stat_diff[0])))
1932 : return true;
1933 : }
1934 : return false;
1935 : }
1936 :
1937 : /*
1938 : * Switch off vmstat processing and then fold all the remaining differentials
1939 : * until the diffs stay at zero. The function is used by NOHZ and can only be
1940 : * invoked when tick processing is not active.
1941 : */
1942 1160 : void quiet_vmstat(void)
1943 : {
1944 1160 : if (system_state != SYSTEM_RUNNING)
1945 : return;
1946 :
1947 1124 : if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1948 : return;
1949 :
1950 963 : if (!need_update(smp_processor_id()))
1951 : return;
1952 :
1953 : /*
1954 : * Just refresh counters and do not care about the pending delayed
1955 : * vmstat_update. It doesn't fire that often to matter and canceling
1956 : * it would be too expensive from this path.
1957 : * vmstat_shepherd will take care about that for us.
1958 : */
1959 255 : refresh_cpu_vm_stats(false);
1960 : }
1961 :
1962 : /*
1963 : * Shepherd worker thread that checks the
1964 : * differentials of processors that have their worker
1965 : * threads for vm statistics updates disabled because of
1966 : * inactivity.
1967 : */
1968 : static void vmstat_shepherd(struct work_struct *w);
1969 :
1970 : static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1971 :
1972 39 : static void vmstat_shepherd(struct work_struct *w)
1973 : {
1974 39 : int cpu;
1975 :
1976 39 : get_online_cpus();
1977 : /* Check processors whose vmstat worker threads have been disabled */
1978 234 : for_each_online_cpu(cpu) {
1979 156 : struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1980 :
1981 156 : if (!delayed_work_pending(dw) && need_update(cpu))
1982 16 : queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1983 :
1984 156 : cond_resched();
1985 : }
1986 39 : put_online_cpus();
1987 :
1988 39 : schedule_delayed_work(&shepherd,
1989 : round_jiffies_relative(sysctl_stat_interval));
1990 39 : }
1991 :
1992 1 : static void __init start_shepherd_timer(void)
1993 : {
1994 1 : int cpu;
1995 :
1996 6 : for_each_possible_cpu(cpu)
1997 5 : INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1998 : vmstat_update);
1999 :
2000 1 : schedule_delayed_work(&shepherd,
2001 : round_jiffies_relative(sysctl_stat_interval));
2002 1 : }
2003 :
2004 1 : static void __init init_cpu_node_state(void)
2005 : {
2006 1 : int node;
2007 :
2008 2 : for_each_online_node(node) {
2009 1 : if (cpumask_weight(cpumask_of_node(node)) > 0)
2010 2 : node_set_state(node, N_CPU);
2011 : }
2012 1 : }
2013 :
2014 3 : static int vmstat_cpu_online(unsigned int cpu)
2015 : {
2016 3 : refresh_zone_stat_thresholds();
2017 3 : node_set_state(cpu_to_node(cpu), N_CPU);
2018 3 : return 0;
2019 : }
2020 :
2021 0 : static int vmstat_cpu_down_prep(unsigned int cpu)
2022 : {
2023 0 : cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2024 0 : return 0;
2025 : }
2026 :
2027 0 : static int vmstat_cpu_dead(unsigned int cpu)
2028 : {
2029 0 : const struct cpumask *node_cpus;
2030 0 : int node;
2031 :
2032 0 : node = cpu_to_node(cpu);
2033 :
2034 0 : refresh_zone_stat_thresholds();
2035 0 : node_cpus = cpumask_of_node(node);
2036 0 : if (cpumask_weight(node_cpus) > 0)
2037 : return 0;
2038 :
2039 0 : node_clear_state(node, N_CPU);
2040 0 : return 0;
2041 : }
2042 :
2043 : #endif
2044 :
2045 : struct workqueue_struct *mm_percpu_wq;
2046 :
2047 1 : void __init init_mm_internals(void)
2048 : {
2049 1 : int ret __maybe_unused;
2050 :
2051 1 : mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2052 :
2053 : #ifdef CONFIG_SMP
2054 1 : ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2055 : NULL, vmstat_cpu_dead);
2056 1 : if (ret < 0)
2057 0 : pr_err("vmstat: failed to register 'dead' hotplug state\n");
2058 :
2059 1 : ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2060 : vmstat_cpu_online,
2061 : vmstat_cpu_down_prep);
2062 1 : if (ret < 0)
2063 0 : pr_err("vmstat: failed to register 'online' hotplug state\n");
2064 :
2065 1 : get_online_cpus();
2066 1 : init_cpu_node_state();
2067 1 : put_online_cpus();
2068 :
2069 1 : start_shepherd_timer();
2070 : #endif
2071 : #ifdef CONFIG_PROC_FS
2072 1 : proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2073 1 : proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2074 1 : proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2075 1 : proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2076 : #endif
2077 1 : }
2078 :
2079 : #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2080 :
2081 : /*
2082 : * Return an index indicating how much of the available free memory is
2083 : * unusable for an allocation of the requested size.
2084 : */
2085 0 : static int unusable_free_index(unsigned int order,
2086 : struct contig_page_info *info)
2087 : {
2088 : /* No free memory is interpreted as all free memory is unusable */
2089 0 : if (info->free_pages == 0)
2090 : return 1000;
2091 :
2092 : /*
2093 : * Index should be a value between 0 and 1. Return a value to 3
2094 : * decimal places.
2095 : *
2096 : * 0 => no fragmentation
2097 : * 1 => high fragmentation
2098 : */
2099 0 : return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2100 :
2101 : }
2102 :
2103 0 : static void unusable_show_print(struct seq_file *m,
2104 : pg_data_t *pgdat, struct zone *zone)
2105 : {
2106 0 : unsigned int order;
2107 0 : int index;
2108 0 : struct contig_page_info info;
2109 :
2110 0 : seq_printf(m, "Node %d, zone %8s ",
2111 : pgdat->node_id,
2112 : zone->name);
2113 0 : for (order = 0; order < MAX_ORDER; ++order) {
2114 0 : fill_contig_page_info(zone, order, &info);
2115 0 : index = unusable_free_index(order, &info);
2116 0 : seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2117 : }
2118 :
2119 0 : seq_putc(m, '\n');
2120 0 : }
2121 :
2122 : /*
2123 : * Display unusable free space index
2124 : *
2125 : * The unusable free space index measures how much of the available free
2126 : * memory cannot be used to satisfy an allocation of a given size and is a
2127 : * value between 0 and 1. The higher the value, the more of free memory is
2128 : * unusable and by implication, the worse the external fragmentation is. This
2129 : * can be expressed as a percentage by multiplying by 100.
2130 : */
2131 0 : static int unusable_show(struct seq_file *m, void *arg)
2132 : {
2133 0 : pg_data_t *pgdat = (pg_data_t *)arg;
2134 :
2135 : /* check memoryless node */
2136 0 : if (!node_state(pgdat->node_id, N_MEMORY))
2137 : return 0;
2138 :
2139 0 : walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2140 :
2141 0 : return 0;
2142 : }
2143 :
2144 : static const struct seq_operations unusable_sops = {
2145 : .start = frag_start,
2146 : .next = frag_next,
2147 : .stop = frag_stop,
2148 : .show = unusable_show,
2149 : };
2150 :
2151 0 : DEFINE_SEQ_ATTRIBUTE(unusable);
2152 :
2153 0 : static void extfrag_show_print(struct seq_file *m,
2154 : pg_data_t *pgdat, struct zone *zone)
2155 : {
2156 0 : unsigned int order;
2157 0 : int index;
2158 :
2159 : /* Alloc on stack as interrupts are disabled for zone walk */
2160 0 : struct contig_page_info info;
2161 :
2162 0 : seq_printf(m, "Node %d, zone %8s ",
2163 : pgdat->node_id,
2164 : zone->name);
2165 0 : for (order = 0; order < MAX_ORDER; ++order) {
2166 0 : fill_contig_page_info(zone, order, &info);
2167 0 : index = __fragmentation_index(order, &info);
2168 0 : seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2169 : }
2170 :
2171 0 : seq_putc(m, '\n');
2172 0 : }
2173 :
2174 : /*
2175 : * Display fragmentation index for orders that allocations would fail for
2176 : */
2177 0 : static int extfrag_show(struct seq_file *m, void *arg)
2178 : {
2179 0 : pg_data_t *pgdat = (pg_data_t *)arg;
2180 :
2181 0 : walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2182 :
2183 0 : return 0;
2184 : }
2185 :
2186 : static const struct seq_operations extfrag_sops = {
2187 : .start = frag_start,
2188 : .next = frag_next,
2189 : .stop = frag_stop,
2190 : .show = extfrag_show,
2191 : };
2192 :
2193 0 : DEFINE_SEQ_ATTRIBUTE(extfrag);
2194 :
2195 1 : static int __init extfrag_debug_init(void)
2196 : {
2197 1 : struct dentry *extfrag_debug_root;
2198 :
2199 1 : extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2200 :
2201 1 : debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2202 : &unusable_fops);
2203 :
2204 1 : debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2205 : &extfrag_fops);
2206 :
2207 1 : return 0;
2208 : }
2209 :
2210 : module_init(extfrag_debug_init);
2211 : #endif
|
