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