xref: /linux/mm/vmstat.c (revision c3f2d783a459980eafd24c5af94ccd56a615961f)
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_owner.h>
30 #include <linux/sched/isolation.h>
31 
32 #include "internal.h"
33 
34 #ifdef CONFIG_NUMA
35 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
36 
37 /* zero numa counters within a zone */
zero_zone_numa_counters(struct zone * zone)38 static void zero_zone_numa_counters(struct zone *zone)
39 {
40 	int item, cpu;
41 
42 	for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
43 		atomic_long_set(&zone->vm_numa_event[item], 0);
44 		for_each_online_cpu(cpu) {
45 			per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
46 						= 0;
47 		}
48 	}
49 }
50 
51 /* zero numa counters of all the populated zones */
zero_zones_numa_counters(void)52 static void zero_zones_numa_counters(void)
53 {
54 	struct zone *zone;
55 
56 	for_each_populated_zone(zone)
57 		zero_zone_numa_counters(zone);
58 }
59 
60 /* zero global numa counters */
zero_global_numa_counters(void)61 static void zero_global_numa_counters(void)
62 {
63 	int item;
64 
65 	for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
66 		atomic_long_set(&vm_numa_event[item], 0);
67 }
68 
invalid_numa_statistics(void)69 static void invalid_numa_statistics(void)
70 {
71 	zero_zones_numa_counters();
72 	zero_global_numa_counters();
73 }
74 
75 static DEFINE_MUTEX(vm_numa_stat_lock);
76 
sysctl_vm_numa_stat_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)77 int sysctl_vm_numa_stat_handler(const struct ctl_table *table, int write,
78 		void *buffer, size_t *length, loff_t *ppos)
79 {
80 	int ret, oldval;
81 
82 	mutex_lock(&vm_numa_stat_lock);
83 	if (write)
84 		oldval = sysctl_vm_numa_stat;
85 	ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
86 	if (ret || !write)
87 		goto out;
88 
89 	if (oldval == sysctl_vm_numa_stat)
90 		goto out;
91 	else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
92 		static_branch_enable(&vm_numa_stat_key);
93 		pr_info("enable numa statistics\n");
94 	} else {
95 		static_branch_disable(&vm_numa_stat_key);
96 		invalid_numa_statistics();
97 		pr_info("disable numa statistics, and clear numa counters\n");
98 	}
99 
100 out:
101 	mutex_unlock(&vm_numa_stat_lock);
102 	return ret;
103 }
104 #endif
105 
106 #ifdef CONFIG_VM_EVENT_COUNTERS
107 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
108 EXPORT_PER_CPU_SYMBOL(vm_event_states);
109 
sum_vm_events(unsigned long * ret)110 static void sum_vm_events(unsigned long *ret)
111 {
112 	int cpu;
113 	int i;
114 
115 	memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
116 
117 	for_each_online_cpu(cpu) {
118 		struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
119 
120 		for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
121 			ret[i] += this->event[i];
122 	}
123 }
124 
125 /*
126  * Accumulate the vm event counters across all CPUs.
127  * The result is unavoidably approximate - it can change
128  * during and after execution of this function.
129 */
all_vm_events(unsigned long * ret)130 void all_vm_events(unsigned long *ret)
131 {
132 	cpus_read_lock();
133 	sum_vm_events(ret);
134 	cpus_read_unlock();
135 }
136 EXPORT_SYMBOL_GPL(all_vm_events);
137 
138 /*
139  * Fold the foreign cpu events into our own.
140  *
141  * This is adding to the events on one processor
142  * but keeps the global counts constant.
143  */
vm_events_fold_cpu(int cpu)144 void vm_events_fold_cpu(int cpu)
145 {
146 	struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
147 	int i;
148 
149 	for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
150 		count_vm_events(i, fold_state->event[i]);
151 		fold_state->event[i] = 0;
152 	}
153 }
154 
155 #endif /* CONFIG_VM_EVENT_COUNTERS */
156 
157 /*
158  * Manage combined zone based / global counters
159  *
160  * vm_stat contains the global counters
161  */
162 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
163 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
164 atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
165 EXPORT_SYMBOL(vm_zone_stat);
166 EXPORT_SYMBOL(vm_node_stat);
167 
168 #ifdef CONFIG_NUMA
fold_vm_zone_numa_events(struct zone * zone)169 static void fold_vm_zone_numa_events(struct zone *zone)
170 {
171 	unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
172 	int cpu;
173 	enum numa_stat_item item;
174 
175 	for_each_online_cpu(cpu) {
176 		struct per_cpu_zonestat *pzstats;
177 
178 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
179 		for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
180 			zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
181 	}
182 
183 	for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
184 		zone_numa_event_add(zone_numa_events[item], zone, item);
185 }
186 
fold_vm_numa_events(void)187 void fold_vm_numa_events(void)
188 {
189 	struct zone *zone;
190 
191 	for_each_populated_zone(zone)
192 		fold_vm_zone_numa_events(zone);
193 }
194 #endif
195 
196 #ifdef CONFIG_SMP
197 
calculate_pressure_threshold(struct zone * zone)198 int calculate_pressure_threshold(struct zone *zone)
199 {
200 	int threshold;
201 	int watermark_distance;
202 
203 	/*
204 	 * As vmstats are not up to date, there is drift between the estimated
205 	 * and real values. For high thresholds and a high number of CPUs, it
206 	 * is possible for the min watermark to be breached while the estimated
207 	 * value looks fine. The pressure threshold is a reduced value such
208 	 * that even the maximum amount of drift will not accidentally breach
209 	 * the min watermark
210 	 */
211 	watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
212 	threshold = max(1, (int)(watermark_distance / num_online_cpus()));
213 
214 	/*
215 	 * Maximum threshold is 125
216 	 */
217 	threshold = min(125, threshold);
218 
219 	return threshold;
220 }
221 
calculate_normal_threshold(struct zone * zone)222 int calculate_normal_threshold(struct zone *zone)
223 {
224 	int threshold;
225 	int mem;	/* memory in 128 MB units */
226 
227 	/*
228 	 * The threshold scales with the number of processors and the amount
229 	 * of memory per zone. More memory means that we can defer updates for
230 	 * longer, more processors could lead to more contention.
231  	 * fls() is used to have a cheap way of logarithmic scaling.
232 	 *
233 	 * Some sample thresholds:
234 	 *
235 	 * Threshold	Processors	(fls)	Zonesize	fls(mem)+1
236 	 * ------------------------------------------------------------------
237 	 * 8		1		1	0.9-1 GB	4
238 	 * 16		2		2	0.9-1 GB	4
239 	 * 20 		2		2	1-2 GB		5
240 	 * 24		2		2	2-4 GB		6
241 	 * 28		2		2	4-8 GB		7
242 	 * 32		2		2	8-16 GB		8
243 	 * 4		2		2	<128M		1
244 	 * 30		4		3	2-4 GB		5
245 	 * 48		4		3	8-16 GB		8
246 	 * 32		8		4	1-2 GB		4
247 	 * 32		8		4	0.9-1GB		4
248 	 * 10		16		5	<128M		1
249 	 * 40		16		5	900M		4
250 	 * 70		64		7	2-4 GB		5
251 	 * 84		64		7	4-8 GB		6
252 	 * 108		512		9	4-8 GB		6
253 	 * 125		1024		10	8-16 GB		8
254 	 * 125		1024		10	16-32 GB	9
255 	 */
256 
257 	mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
258 
259 	threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
260 
261 	/*
262 	 * Maximum threshold is 125
263 	 */
264 	threshold = min(125, threshold);
265 
266 	return threshold;
267 }
268 
269 /*
270  * Refresh the thresholds for each zone.
271  */
refresh_zone_stat_thresholds(void)272 void refresh_zone_stat_thresholds(void)
273 {
274 	struct pglist_data *pgdat;
275 	struct zone *zone;
276 	int cpu;
277 	int threshold;
278 
279 	/* Zero current pgdat thresholds */
280 	for_each_online_pgdat(pgdat) {
281 		for_each_online_cpu(cpu) {
282 			per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
283 		}
284 	}
285 
286 	for_each_populated_zone(zone) {
287 		struct pglist_data *pgdat = zone->zone_pgdat;
288 		unsigned long max_drift, tolerate_drift;
289 
290 		threshold = calculate_normal_threshold(zone);
291 
292 		for_each_online_cpu(cpu) {
293 			int pgdat_threshold;
294 
295 			per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
296 							= threshold;
297 
298 			/* Base nodestat threshold on the largest populated zone. */
299 			pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
300 			per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
301 				= max(threshold, pgdat_threshold);
302 		}
303 
304 		/*
305 		 * Only set percpu_drift_mark if there is a danger that
306 		 * NR_FREE_PAGES reports the low watermark is ok when in fact
307 		 * the min watermark could be breached by an allocation
308 		 */
309 		tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
310 		max_drift = num_online_cpus() * threshold;
311 		if (max_drift > tolerate_drift)
312 			zone->percpu_drift_mark = high_wmark_pages(zone) +
313 					max_drift;
314 	}
315 }
316 
set_pgdat_percpu_threshold(pg_data_t * pgdat,int (* calculate_pressure)(struct zone *))317 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
318 				int (*calculate_pressure)(struct zone *))
319 {
320 	struct zone *zone;
321 	int cpu;
322 	int threshold;
323 	int i;
324 
325 	for (i = 0; i < pgdat->nr_zones; i++) {
326 		zone = &pgdat->node_zones[i];
327 		if (!zone->percpu_drift_mark)
328 			continue;
329 
330 		threshold = (*calculate_pressure)(zone);
331 		for_each_online_cpu(cpu)
332 			per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
333 							= threshold;
334 	}
335 }
336 
337 /*
338  * For use when we know that interrupts are disabled,
339  * or when we know that preemption is disabled and that
340  * particular counter cannot be updated from interrupt context.
341  */
__mod_zone_page_state(struct zone * zone,enum zone_stat_item item,long delta)342 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
343 			   long delta)
344 {
345 	struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
346 	s8 __percpu *p = pcp->vm_stat_diff + item;
347 	long x;
348 	long t;
349 
350 	/*
351 	 * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
352 	 * atomicity is provided by IRQs being disabled -- either explicitly
353 	 * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
354 	 * CPU migrations and preemption potentially corrupts a counter so
355 	 * disable preemption.
356 	 */
357 	preempt_disable_nested();
358 
359 	x = delta + __this_cpu_read(*p);
360 
361 	t = __this_cpu_read(pcp->stat_threshold);
362 
363 	if (unlikely(abs(x) > t)) {
364 		zone_page_state_add(x, zone, item);
365 		x = 0;
366 	}
367 	__this_cpu_write(*p, x);
368 
369 	preempt_enable_nested();
370 }
371 EXPORT_SYMBOL(__mod_zone_page_state);
372 
__mod_node_page_state(struct pglist_data * pgdat,enum node_stat_item item,long delta)373 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
374 				long delta)
375 {
376 	struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
377 	s8 __percpu *p = pcp->vm_node_stat_diff + item;
378 	long x;
379 	long t;
380 
381 	if (vmstat_item_in_bytes(item)) {
382 		/*
383 		 * Only cgroups use subpage accounting right now; at
384 		 * the global level, these items still change in
385 		 * multiples of whole pages. Store them as pages
386 		 * internally to keep the per-cpu counters compact.
387 		 */
388 		VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
389 		delta >>= PAGE_SHIFT;
390 	}
391 
392 	/* See __mod_node_page_state */
393 	preempt_disable_nested();
394 
395 	x = delta + __this_cpu_read(*p);
396 
397 	t = __this_cpu_read(pcp->stat_threshold);
398 
399 	if (unlikely(abs(x) > t)) {
400 		node_page_state_add(x, pgdat, item);
401 		x = 0;
402 	}
403 	__this_cpu_write(*p, x);
404 
405 	preempt_enable_nested();
406 }
407 EXPORT_SYMBOL(__mod_node_page_state);
408 
409 /*
410  * Optimized increment and decrement functions.
411  *
412  * These are only for a single page and therefore can take a struct page *
413  * argument instead of struct zone *. This allows the inclusion of the code
414  * generated for page_zone(page) into the optimized functions.
415  *
416  * No overflow check is necessary and therefore the differential can be
417  * incremented or decremented in place which may allow the compilers to
418  * generate better code.
419  * The increment or decrement is known and therefore one boundary check can
420  * be omitted.
421  *
422  * NOTE: These functions are very performance sensitive. Change only
423  * with care.
424  *
425  * Some processors have inc/dec instructions that are atomic vs an interrupt.
426  * However, the code must first determine the differential location in a zone
427  * based on the processor number and then inc/dec the counter. There is no
428  * guarantee without disabling preemption that the processor will not change
429  * in between and therefore the atomicity vs. interrupt cannot be exploited
430  * in a useful way here.
431  */
__inc_zone_state(struct zone * zone,enum zone_stat_item item)432 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
433 {
434 	struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
435 	s8 __percpu *p = pcp->vm_stat_diff + item;
436 	s8 v, t;
437 
438 	/* See __mod_node_page_state */
439 	preempt_disable_nested();
440 
441 	v = __this_cpu_inc_return(*p);
442 	t = __this_cpu_read(pcp->stat_threshold);
443 	if (unlikely(v > t)) {
444 		s8 overstep = t >> 1;
445 
446 		zone_page_state_add(v + overstep, zone, item);
447 		__this_cpu_write(*p, -overstep);
448 	}
449 
450 	preempt_enable_nested();
451 }
452 
__inc_node_state(struct pglist_data * pgdat,enum node_stat_item item)453 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
454 {
455 	struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
456 	s8 __percpu *p = pcp->vm_node_stat_diff + item;
457 	s8 v, t;
458 
459 	VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
460 
461 	/* See __mod_node_page_state */
462 	preempt_disable_nested();
463 
464 	v = __this_cpu_inc_return(*p);
465 	t = __this_cpu_read(pcp->stat_threshold);
466 	if (unlikely(v > t)) {
467 		s8 overstep = t >> 1;
468 
469 		node_page_state_add(v + overstep, pgdat, item);
470 		__this_cpu_write(*p, -overstep);
471 	}
472 
473 	preempt_enable_nested();
474 }
475 
__inc_zone_page_state(struct page * page,enum zone_stat_item item)476 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
477 {
478 	__inc_zone_state(page_zone(page), item);
479 }
480 EXPORT_SYMBOL(__inc_zone_page_state);
481 
__inc_node_page_state(struct page * page,enum node_stat_item item)482 void __inc_node_page_state(struct page *page, enum node_stat_item item)
483 {
484 	__inc_node_state(page_pgdat(page), item);
485 }
486 EXPORT_SYMBOL(__inc_node_page_state);
487 
__dec_zone_state(struct zone * zone,enum zone_stat_item item)488 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
489 {
490 	struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
491 	s8 __percpu *p = pcp->vm_stat_diff + item;
492 	s8 v, t;
493 
494 	/* See __mod_node_page_state */
495 	preempt_disable_nested();
496 
497 	v = __this_cpu_dec_return(*p);
498 	t = __this_cpu_read(pcp->stat_threshold);
499 	if (unlikely(v < - t)) {
500 		s8 overstep = t >> 1;
501 
502 		zone_page_state_add(v - overstep, zone, item);
503 		__this_cpu_write(*p, overstep);
504 	}
505 
506 	preempt_enable_nested();
507 }
508 
__dec_node_state(struct pglist_data * pgdat,enum node_stat_item item)509 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
510 {
511 	struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
512 	s8 __percpu *p = pcp->vm_node_stat_diff + item;
513 	s8 v, t;
514 
515 	VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
516 
517 	/* See __mod_node_page_state */
518 	preempt_disable_nested();
519 
520 	v = __this_cpu_dec_return(*p);
521 	t = __this_cpu_read(pcp->stat_threshold);
522 	if (unlikely(v < - t)) {
523 		s8 overstep = t >> 1;
524 
525 		node_page_state_add(v - overstep, pgdat, item);
526 		__this_cpu_write(*p, overstep);
527 	}
528 
529 	preempt_enable_nested();
530 }
531 
__dec_zone_page_state(struct page * page,enum zone_stat_item item)532 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
533 {
534 	__dec_zone_state(page_zone(page), item);
535 }
536 EXPORT_SYMBOL(__dec_zone_page_state);
537 
__dec_node_page_state(struct page * page,enum node_stat_item item)538 void __dec_node_page_state(struct page *page, enum node_stat_item item)
539 {
540 	__dec_node_state(page_pgdat(page), item);
541 }
542 EXPORT_SYMBOL(__dec_node_page_state);
543 
544 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
545 /*
546  * If we have cmpxchg_local support then we do not need to incur the overhead
547  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
548  *
549  * mod_state() modifies the zone counter state through atomic per cpu
550  * operations.
551  *
552  * Overstep mode specifies how overstep should handled:
553  *     0       No overstepping
554  *     1       Overstepping half of threshold
555  *     -1      Overstepping minus half of threshold
556 */
mod_zone_state(struct zone * zone,enum zone_stat_item item,long delta,int overstep_mode)557 static inline void mod_zone_state(struct zone *zone,
558        enum zone_stat_item item, long delta, int overstep_mode)
559 {
560 	struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
561 	s8 __percpu *p = pcp->vm_stat_diff + item;
562 	long n, t, z;
563 	s8 o;
564 
565 	o = this_cpu_read(*p);
566 	do {
567 		z = 0;  /* overflow to zone counters */
568 
569 		/*
570 		 * The fetching of the stat_threshold is racy. We may apply
571 		 * a counter threshold to the wrong the cpu if we get
572 		 * rescheduled while executing here. However, the next
573 		 * counter update will apply the threshold again and
574 		 * therefore bring the counter under the threshold again.
575 		 *
576 		 * Most of the time the thresholds are the same anyways
577 		 * for all cpus in a zone.
578 		 */
579 		t = this_cpu_read(pcp->stat_threshold);
580 
581 		n = delta + (long)o;
582 
583 		if (abs(n) > t) {
584 			int os = overstep_mode * (t >> 1) ;
585 
586 			/* Overflow must be added to zone counters */
587 			z = n + os;
588 			n = -os;
589 		}
590 	} while (!this_cpu_try_cmpxchg(*p, &o, n));
591 
592 	if (z)
593 		zone_page_state_add(z, zone, item);
594 }
595 
mod_zone_page_state(struct zone * zone,enum zone_stat_item item,long delta)596 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
597 			 long delta)
598 {
599 	mod_zone_state(zone, item, delta, 0);
600 }
601 EXPORT_SYMBOL(mod_zone_page_state);
602 
inc_zone_page_state(struct page * page,enum zone_stat_item item)603 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
604 {
605 	mod_zone_state(page_zone(page), item, 1, 1);
606 }
607 EXPORT_SYMBOL(inc_zone_page_state);
608 
dec_zone_page_state(struct page * page,enum zone_stat_item item)609 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
610 {
611 	mod_zone_state(page_zone(page), item, -1, -1);
612 }
613 EXPORT_SYMBOL(dec_zone_page_state);
614 
mod_node_state(struct pglist_data * pgdat,enum node_stat_item item,int delta,int overstep_mode)615 static inline void mod_node_state(struct pglist_data *pgdat,
616        enum node_stat_item item, int delta, int overstep_mode)
617 {
618 	struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
619 	s8 __percpu *p = pcp->vm_node_stat_diff + item;
620 	long n, t, z;
621 	s8 o;
622 
623 	if (vmstat_item_in_bytes(item)) {
624 		/*
625 		 * Only cgroups use subpage accounting right now; at
626 		 * the global level, these items still change in
627 		 * multiples of whole pages. Store them as pages
628 		 * internally to keep the per-cpu counters compact.
629 		 */
630 		VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
631 		delta >>= PAGE_SHIFT;
632 	}
633 
634 	o = this_cpu_read(*p);
635 	do {
636 		z = 0;  /* overflow to node counters */
637 
638 		/*
639 		 * The fetching of the stat_threshold is racy. We may apply
640 		 * a counter threshold to the wrong the cpu if we get
641 		 * rescheduled while executing here. However, the next
642 		 * counter update will apply the threshold again and
643 		 * therefore bring the counter under the threshold again.
644 		 *
645 		 * Most of the time the thresholds are the same anyways
646 		 * for all cpus in a node.
647 		 */
648 		t = this_cpu_read(pcp->stat_threshold);
649 
650 		n = delta + (long)o;
651 
652 		if (abs(n) > t) {
653 			int os = overstep_mode * (t >> 1) ;
654 
655 			/* Overflow must be added to node counters */
656 			z = n + os;
657 			n = -os;
658 		}
659 	} while (!this_cpu_try_cmpxchg(*p, &o, n));
660 
661 	if (z)
662 		node_page_state_add(z, pgdat, item);
663 }
664 
mod_node_page_state(struct pglist_data * pgdat,enum node_stat_item item,long delta)665 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
666 					long delta)
667 {
668 	mod_node_state(pgdat, item, delta, 0);
669 }
670 EXPORT_SYMBOL(mod_node_page_state);
671 
inc_node_state(struct pglist_data * pgdat,enum node_stat_item item)672 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
673 {
674 	mod_node_state(pgdat, item, 1, 1);
675 }
676 
inc_node_page_state(struct page * page,enum node_stat_item item)677 void inc_node_page_state(struct page *page, enum node_stat_item item)
678 {
679 	mod_node_state(page_pgdat(page), item, 1, 1);
680 }
681 EXPORT_SYMBOL(inc_node_page_state);
682 
dec_node_page_state(struct page * page,enum node_stat_item item)683 void dec_node_page_state(struct page *page, enum node_stat_item item)
684 {
685 	mod_node_state(page_pgdat(page), item, -1, -1);
686 }
687 EXPORT_SYMBOL(dec_node_page_state);
688 #else
689 /*
690  * Use interrupt disable to serialize counter updates
691  */
mod_zone_page_state(struct zone * zone,enum zone_stat_item item,long delta)692 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
693 			 long delta)
694 {
695 	unsigned long flags;
696 
697 	local_irq_save(flags);
698 	__mod_zone_page_state(zone, item, delta);
699 	local_irq_restore(flags);
700 }
701 EXPORT_SYMBOL(mod_zone_page_state);
702 
inc_zone_page_state(struct page * page,enum zone_stat_item item)703 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
704 {
705 	unsigned long flags;
706 	struct zone *zone;
707 
708 	zone = page_zone(page);
709 	local_irq_save(flags);
710 	__inc_zone_state(zone, item);
711 	local_irq_restore(flags);
712 }
713 EXPORT_SYMBOL(inc_zone_page_state);
714 
dec_zone_page_state(struct page * page,enum zone_stat_item item)715 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
716 {
717 	unsigned long flags;
718 
719 	local_irq_save(flags);
720 	__dec_zone_page_state(page, item);
721 	local_irq_restore(flags);
722 }
723 EXPORT_SYMBOL(dec_zone_page_state);
724 
inc_node_state(struct pglist_data * pgdat,enum node_stat_item item)725 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
726 {
727 	unsigned long flags;
728 
729 	local_irq_save(flags);
730 	__inc_node_state(pgdat, item);
731 	local_irq_restore(flags);
732 }
733 EXPORT_SYMBOL(inc_node_state);
734 
mod_node_page_state(struct pglist_data * pgdat,enum node_stat_item item,long delta)735 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
736 					long delta)
737 {
738 	unsigned long flags;
739 
740 	local_irq_save(flags);
741 	__mod_node_page_state(pgdat, item, delta);
742 	local_irq_restore(flags);
743 }
744 EXPORT_SYMBOL(mod_node_page_state);
745 
inc_node_page_state(struct page * page,enum node_stat_item item)746 void inc_node_page_state(struct page *page, enum node_stat_item item)
747 {
748 	unsigned long flags;
749 	struct pglist_data *pgdat;
750 
751 	pgdat = page_pgdat(page);
752 	local_irq_save(flags);
753 	__inc_node_state(pgdat, item);
754 	local_irq_restore(flags);
755 }
756 EXPORT_SYMBOL(inc_node_page_state);
757 
dec_node_page_state(struct page * page,enum node_stat_item item)758 void dec_node_page_state(struct page *page, enum node_stat_item item)
759 {
760 	unsigned long flags;
761 
762 	local_irq_save(flags);
763 	__dec_node_page_state(page, item);
764 	local_irq_restore(flags);
765 }
766 EXPORT_SYMBOL(dec_node_page_state);
767 #endif
768 
769 /*
770  * Fold a differential into the global counters.
771  * Returns the number of counters updated.
772  */
fold_diff(int * zone_diff,int * node_diff)773 static int fold_diff(int *zone_diff, int *node_diff)
774 {
775 	int i;
776 	int changes = 0;
777 
778 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
779 		if (zone_diff[i]) {
780 			atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
781 			changes++;
782 	}
783 
784 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
785 		if (node_diff[i]) {
786 			atomic_long_add(node_diff[i], &vm_node_stat[i]);
787 			changes++;
788 	}
789 	return changes;
790 }
791 
792 /*
793  * Update the zone counters for the current cpu.
794  *
795  * Note that refresh_cpu_vm_stats strives to only access
796  * node local memory. The per cpu pagesets on remote zones are placed
797  * in the memory local to the processor using that pageset. So the
798  * loop over all zones will access a series of cachelines local to
799  * the processor.
800  *
801  * The call to zone_page_state_add updates the cachelines with the
802  * statistics in the remote zone struct as well as the global cachelines
803  * with the global counters. These could cause remote node cache line
804  * bouncing and will have to be only done when necessary.
805  *
806  * The function returns the number of global counters updated.
807  */
refresh_cpu_vm_stats(bool do_pagesets)808 static int refresh_cpu_vm_stats(bool do_pagesets)
809 {
810 	struct pglist_data *pgdat;
811 	struct zone *zone;
812 	int i;
813 	int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
814 	int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
815 	int changes = 0;
816 
817 	for_each_populated_zone(zone) {
818 		struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
819 		struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
820 
821 		for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
822 			int v;
823 
824 			v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
825 			if (v) {
826 
827 				atomic_long_add(v, &zone->vm_stat[i]);
828 				global_zone_diff[i] += v;
829 #ifdef CONFIG_NUMA
830 				/* 3 seconds idle till flush */
831 				__this_cpu_write(pcp->expire, 3);
832 #endif
833 			}
834 		}
835 
836 		if (do_pagesets) {
837 			cond_resched();
838 
839 			changes += decay_pcp_high(zone, this_cpu_ptr(pcp));
840 #ifdef CONFIG_NUMA
841 			/*
842 			 * Deal with draining the remote pageset of this
843 			 * processor
844 			 *
845 			 * Check if there are pages remaining in this pageset
846 			 * if not then there is nothing to expire.
847 			 */
848 			if (!__this_cpu_read(pcp->expire) ||
849 			       !__this_cpu_read(pcp->count))
850 				continue;
851 
852 			/*
853 			 * We never drain zones local to this processor.
854 			 */
855 			if (zone_to_nid(zone) == numa_node_id()) {
856 				__this_cpu_write(pcp->expire, 0);
857 				continue;
858 			}
859 
860 			if (__this_cpu_dec_return(pcp->expire)) {
861 				changes++;
862 				continue;
863 			}
864 
865 			if (__this_cpu_read(pcp->count)) {
866 				drain_zone_pages(zone, this_cpu_ptr(pcp));
867 				changes++;
868 			}
869 #endif
870 		}
871 	}
872 
873 	for_each_online_pgdat(pgdat) {
874 		struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
875 
876 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
877 			int v;
878 
879 			v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
880 			if (v) {
881 				atomic_long_add(v, &pgdat->vm_stat[i]);
882 				global_node_diff[i] += v;
883 			}
884 		}
885 	}
886 
887 	changes += fold_diff(global_zone_diff, global_node_diff);
888 	return changes;
889 }
890 
891 /*
892  * Fold the data for an offline cpu into the global array.
893  * There cannot be any access by the offline cpu and therefore
894  * synchronization is simplified.
895  */
cpu_vm_stats_fold(int cpu)896 void cpu_vm_stats_fold(int cpu)
897 {
898 	struct pglist_data *pgdat;
899 	struct zone *zone;
900 	int i;
901 	int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
902 	int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
903 
904 	for_each_populated_zone(zone) {
905 		struct per_cpu_zonestat *pzstats;
906 
907 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
908 
909 		for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
910 			if (pzstats->vm_stat_diff[i]) {
911 				int v;
912 
913 				v = pzstats->vm_stat_diff[i];
914 				pzstats->vm_stat_diff[i] = 0;
915 				atomic_long_add(v, &zone->vm_stat[i]);
916 				global_zone_diff[i] += v;
917 			}
918 		}
919 #ifdef CONFIG_NUMA
920 		for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
921 			if (pzstats->vm_numa_event[i]) {
922 				unsigned long v;
923 
924 				v = pzstats->vm_numa_event[i];
925 				pzstats->vm_numa_event[i] = 0;
926 				zone_numa_event_add(v, zone, i);
927 			}
928 		}
929 #endif
930 	}
931 
932 	for_each_online_pgdat(pgdat) {
933 		struct per_cpu_nodestat *p;
934 
935 		p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
936 
937 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
938 			if (p->vm_node_stat_diff[i]) {
939 				int v;
940 
941 				v = p->vm_node_stat_diff[i];
942 				p->vm_node_stat_diff[i] = 0;
943 				atomic_long_add(v, &pgdat->vm_stat[i]);
944 				global_node_diff[i] += v;
945 			}
946 	}
947 
948 	fold_diff(global_zone_diff, global_node_diff);
949 }
950 
951 /*
952  * this is only called if !populated_zone(zone), which implies no other users of
953  * pset->vm_stat_diff[] exist.
954  */
drain_zonestat(struct zone * zone,struct per_cpu_zonestat * pzstats)955 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
956 {
957 	unsigned long v;
958 	int i;
959 
960 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
961 		if (pzstats->vm_stat_diff[i]) {
962 			v = pzstats->vm_stat_diff[i];
963 			pzstats->vm_stat_diff[i] = 0;
964 			zone_page_state_add(v, zone, i);
965 		}
966 	}
967 
968 #ifdef CONFIG_NUMA
969 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
970 		if (pzstats->vm_numa_event[i]) {
971 			v = pzstats->vm_numa_event[i];
972 			pzstats->vm_numa_event[i] = 0;
973 			zone_numa_event_add(v, zone, i);
974 		}
975 	}
976 #endif
977 }
978 #endif
979 
980 #ifdef CONFIG_NUMA
981 /*
982  * Determine the per node value of a stat item. This function
983  * is called frequently in a NUMA machine, so try to be as
984  * frugal as possible.
985  */
sum_zone_node_page_state(int node,enum zone_stat_item item)986 unsigned long sum_zone_node_page_state(int node,
987 				 enum zone_stat_item item)
988 {
989 	struct zone *zones = NODE_DATA(node)->node_zones;
990 	int i;
991 	unsigned long count = 0;
992 
993 	for (i = 0; i < MAX_NR_ZONES; i++)
994 		count += zone_page_state(zones + i, item);
995 
996 	return count;
997 }
998 
999 /* Determine the per node value of a numa stat item. */
sum_zone_numa_event_state(int node,enum numa_stat_item item)1000 unsigned long sum_zone_numa_event_state(int node,
1001 				 enum numa_stat_item item)
1002 {
1003 	struct zone *zones = NODE_DATA(node)->node_zones;
1004 	unsigned long count = 0;
1005 	int i;
1006 
1007 	for (i = 0; i < MAX_NR_ZONES; i++)
1008 		count += zone_numa_event_state(zones + i, item);
1009 
1010 	return count;
1011 }
1012 
1013 /*
1014  * Determine the per node value of a stat item.
1015  */
node_page_state_pages(struct pglist_data * pgdat,enum node_stat_item item)1016 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1017 				    enum node_stat_item item)
1018 {
1019 	long x = atomic_long_read(&pgdat->vm_stat[item]);
1020 #ifdef CONFIG_SMP
1021 	if (x < 0)
1022 		x = 0;
1023 #endif
1024 	return x;
1025 }
1026 
node_page_state(struct pglist_data * pgdat,enum node_stat_item item)1027 unsigned long node_page_state(struct pglist_data *pgdat,
1028 			      enum node_stat_item item)
1029 {
1030 	VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1031 
1032 	return node_page_state_pages(pgdat, item);
1033 }
1034 #endif
1035 
1036 /*
1037  * Count number of pages "struct page" and "struct page_ext" consume.
1038  * nr_memmap_boot_pages: # of pages allocated by boot allocator
1039  * nr_memmap_pages: # of pages that were allocated by buddy allocator
1040  */
1041 static atomic_long_t nr_memmap_boot_pages = ATOMIC_LONG_INIT(0);
1042 static atomic_long_t nr_memmap_pages = ATOMIC_LONG_INIT(0);
1043 
memmap_boot_pages_add(long delta)1044 void memmap_boot_pages_add(long delta)
1045 {
1046 	atomic_long_add(delta, &nr_memmap_boot_pages);
1047 }
1048 
memmap_pages_add(long delta)1049 void memmap_pages_add(long delta)
1050 {
1051 	atomic_long_add(delta, &nr_memmap_pages);
1052 }
1053 
1054 #ifdef CONFIG_COMPACTION
1055 
1056 struct contig_page_info {
1057 	unsigned long free_pages;
1058 	unsigned long free_blocks_total;
1059 	unsigned long free_blocks_suitable;
1060 };
1061 
1062 /*
1063  * Calculate the number of free pages in a zone, how many contiguous
1064  * pages are free and how many are large enough to satisfy an allocation of
1065  * the target size. Note that this function makes no attempt to estimate
1066  * how many suitable free blocks there *might* be if MOVABLE pages were
1067  * migrated. Calculating that is possible, but expensive and can be
1068  * figured out from userspace
1069  */
fill_contig_page_info(struct zone * zone,unsigned int suitable_order,struct contig_page_info * info)1070 static void fill_contig_page_info(struct zone *zone,
1071 				unsigned int suitable_order,
1072 				struct contig_page_info *info)
1073 {
1074 	unsigned int order;
1075 
1076 	info->free_pages = 0;
1077 	info->free_blocks_total = 0;
1078 	info->free_blocks_suitable = 0;
1079 
1080 	for (order = 0; order < NR_PAGE_ORDERS; order++) {
1081 		unsigned long blocks;
1082 
1083 		/*
1084 		 * Count number of free blocks.
1085 		 *
1086 		 * Access to nr_free is lockless as nr_free is used only for
1087 		 * diagnostic purposes. Use data_race to avoid KCSAN warning.
1088 		 */
1089 		blocks = data_race(zone->free_area[order].nr_free);
1090 		info->free_blocks_total += blocks;
1091 
1092 		/* Count free base pages */
1093 		info->free_pages += blocks << order;
1094 
1095 		/* Count the suitable free blocks */
1096 		if (order >= suitable_order)
1097 			info->free_blocks_suitable += blocks <<
1098 						(order - suitable_order);
1099 	}
1100 }
1101 
1102 /*
1103  * A fragmentation index only makes sense if an allocation of a requested
1104  * size would fail. If that is true, the fragmentation index indicates
1105  * whether external fragmentation or a lack of memory was the problem.
1106  * The value can be used to determine if page reclaim or compaction
1107  * should be used
1108  */
__fragmentation_index(unsigned int order,struct contig_page_info * info)1109 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1110 {
1111 	unsigned long requested = 1UL << order;
1112 
1113 	if (WARN_ON_ONCE(order > MAX_PAGE_ORDER))
1114 		return 0;
1115 
1116 	if (!info->free_blocks_total)
1117 		return 0;
1118 
1119 	/* Fragmentation index only makes sense when a request would fail */
1120 	if (info->free_blocks_suitable)
1121 		return -1000;
1122 
1123 	/*
1124 	 * Index is between 0 and 1 so return within 3 decimal places
1125 	 *
1126 	 * 0 => allocation would fail due to lack of memory
1127 	 * 1 => allocation would fail due to fragmentation
1128 	 */
1129 	return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1130 }
1131 
1132 /*
1133  * Calculates external fragmentation within a zone wrt the given order.
1134  * It is defined as the percentage of pages found in blocks of size
1135  * less than 1 << order. It returns values in range [0, 100].
1136  */
extfrag_for_order(struct zone * zone,unsigned int order)1137 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1138 {
1139 	struct contig_page_info info;
1140 
1141 	fill_contig_page_info(zone, order, &info);
1142 	if (info.free_pages == 0)
1143 		return 0;
1144 
1145 	return div_u64((info.free_pages -
1146 			(info.free_blocks_suitable << order)) * 100,
1147 			info.free_pages);
1148 }
1149 
1150 /* Same as __fragmentation index but allocs contig_page_info on stack */
fragmentation_index(struct zone * zone,unsigned int order)1151 int fragmentation_index(struct zone *zone, unsigned int order)
1152 {
1153 	struct contig_page_info info;
1154 
1155 	fill_contig_page_info(zone, order, &info);
1156 	return __fragmentation_index(order, &info);
1157 }
1158 #endif
1159 
1160 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1161     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1162 #ifdef CONFIG_ZONE_DMA
1163 #define TEXT_FOR_DMA(xx) xx "_dma",
1164 #else
1165 #define TEXT_FOR_DMA(xx)
1166 #endif
1167 
1168 #ifdef CONFIG_ZONE_DMA32
1169 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1170 #else
1171 #define TEXT_FOR_DMA32(xx)
1172 #endif
1173 
1174 #ifdef CONFIG_HIGHMEM
1175 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1176 #else
1177 #define TEXT_FOR_HIGHMEM(xx)
1178 #endif
1179 
1180 #ifdef CONFIG_ZONE_DEVICE
1181 #define TEXT_FOR_DEVICE(xx) xx "_device",
1182 #else
1183 #define TEXT_FOR_DEVICE(xx)
1184 #endif
1185 
1186 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1187 					TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1188 					TEXT_FOR_DEVICE(xx)
1189 
1190 const char * const vmstat_text[] = {
1191 	/* enum zone_stat_item counters */
1192 	"nr_free_pages",
1193 	"nr_zone_inactive_anon",
1194 	"nr_zone_active_anon",
1195 	"nr_zone_inactive_file",
1196 	"nr_zone_active_file",
1197 	"nr_zone_unevictable",
1198 	"nr_zone_write_pending",
1199 	"nr_mlock",
1200 	"nr_bounce",
1201 #if IS_ENABLED(CONFIG_ZSMALLOC)
1202 	"nr_zspages",
1203 #endif
1204 	"nr_free_cma",
1205 #ifdef CONFIG_UNACCEPTED_MEMORY
1206 	"nr_unaccepted",
1207 #endif
1208 
1209 	/* enum numa_stat_item counters */
1210 #ifdef CONFIG_NUMA
1211 	"numa_hit",
1212 	"numa_miss",
1213 	"numa_foreign",
1214 	"numa_interleave",
1215 	"numa_local",
1216 	"numa_other",
1217 #endif
1218 
1219 	/* enum node_stat_item counters */
1220 	"nr_inactive_anon",
1221 	"nr_active_anon",
1222 	"nr_inactive_file",
1223 	"nr_active_file",
1224 	"nr_unevictable",
1225 	"nr_slab_reclaimable",
1226 	"nr_slab_unreclaimable",
1227 	"nr_isolated_anon",
1228 	"nr_isolated_file",
1229 	"workingset_nodes",
1230 	"workingset_refault_anon",
1231 	"workingset_refault_file",
1232 	"workingset_activate_anon",
1233 	"workingset_activate_file",
1234 	"workingset_restore_anon",
1235 	"workingset_restore_file",
1236 	"workingset_nodereclaim",
1237 	"nr_anon_pages",
1238 	"nr_mapped",
1239 	"nr_file_pages",
1240 	"nr_dirty",
1241 	"nr_writeback",
1242 	"nr_writeback_temp",
1243 	"nr_shmem",
1244 	"nr_shmem_hugepages",
1245 	"nr_shmem_pmdmapped",
1246 	"nr_file_hugepages",
1247 	"nr_file_pmdmapped",
1248 	"nr_anon_transparent_hugepages",
1249 	"nr_vmscan_write",
1250 	"nr_vmscan_immediate_reclaim",
1251 	"nr_dirtied",
1252 	"nr_written",
1253 	"nr_throttled_written",
1254 	"nr_kernel_misc_reclaimable",
1255 	"nr_foll_pin_acquired",
1256 	"nr_foll_pin_released",
1257 	"nr_kernel_stack",
1258 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1259 	"nr_shadow_call_stack",
1260 #endif
1261 	"nr_page_table_pages",
1262 	"nr_sec_page_table_pages",
1263 #ifdef CONFIG_IOMMU_SUPPORT
1264 	"nr_iommu_pages",
1265 #endif
1266 #ifdef CONFIG_SWAP
1267 	"nr_swapcached",
1268 #endif
1269 #ifdef CONFIG_NUMA_BALANCING
1270 	"pgpromote_success",
1271 	"pgpromote_candidate",
1272 #endif
1273 	"pgdemote_kswapd",
1274 	"pgdemote_direct",
1275 	"pgdemote_khugepaged",
1276 	/* system-wide enum vm_stat_item counters */
1277 	"nr_dirty_threshold",
1278 	"nr_dirty_background_threshold",
1279 	"nr_memmap_pages",
1280 	"nr_memmap_boot_pages",
1281 
1282 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1283 	/* enum vm_event_item counters */
1284 	"pgpgin",
1285 	"pgpgout",
1286 	"pswpin",
1287 	"pswpout",
1288 
1289 	TEXTS_FOR_ZONES("pgalloc")
1290 	TEXTS_FOR_ZONES("allocstall")
1291 	TEXTS_FOR_ZONES("pgskip")
1292 
1293 	"pgfree",
1294 	"pgactivate",
1295 	"pgdeactivate",
1296 	"pglazyfree",
1297 
1298 	"pgfault",
1299 	"pgmajfault",
1300 	"pglazyfreed",
1301 
1302 	"pgrefill",
1303 	"pgreuse",
1304 	"pgsteal_kswapd",
1305 	"pgsteal_direct",
1306 	"pgsteal_khugepaged",
1307 	"pgscan_kswapd",
1308 	"pgscan_direct",
1309 	"pgscan_khugepaged",
1310 	"pgscan_direct_throttle",
1311 	"pgscan_anon",
1312 	"pgscan_file",
1313 	"pgsteal_anon",
1314 	"pgsteal_file",
1315 
1316 #ifdef CONFIG_NUMA
1317 	"zone_reclaim_failed",
1318 #endif
1319 	"pginodesteal",
1320 	"slabs_scanned",
1321 	"kswapd_inodesteal",
1322 	"kswapd_low_wmark_hit_quickly",
1323 	"kswapd_high_wmark_hit_quickly",
1324 	"pageoutrun",
1325 
1326 	"pgrotated",
1327 
1328 	"drop_pagecache",
1329 	"drop_slab",
1330 	"oom_kill",
1331 
1332 #ifdef CONFIG_NUMA_BALANCING
1333 	"numa_pte_updates",
1334 	"numa_huge_pte_updates",
1335 	"numa_hint_faults",
1336 	"numa_hint_faults_local",
1337 	"numa_pages_migrated",
1338 #endif
1339 #ifdef CONFIG_MIGRATION
1340 	"pgmigrate_success",
1341 	"pgmigrate_fail",
1342 	"thp_migration_success",
1343 	"thp_migration_fail",
1344 	"thp_migration_split",
1345 #endif
1346 #ifdef CONFIG_COMPACTION
1347 	"compact_migrate_scanned",
1348 	"compact_free_scanned",
1349 	"compact_isolated",
1350 	"compact_stall",
1351 	"compact_fail",
1352 	"compact_success",
1353 	"compact_daemon_wake",
1354 	"compact_daemon_migrate_scanned",
1355 	"compact_daemon_free_scanned",
1356 #endif
1357 
1358 #ifdef CONFIG_HUGETLB_PAGE
1359 	"htlb_buddy_alloc_success",
1360 	"htlb_buddy_alloc_fail",
1361 #endif
1362 #ifdef CONFIG_CMA
1363 	"cma_alloc_success",
1364 	"cma_alloc_fail",
1365 #endif
1366 	"unevictable_pgs_culled",
1367 	"unevictable_pgs_scanned",
1368 	"unevictable_pgs_rescued",
1369 	"unevictable_pgs_mlocked",
1370 	"unevictable_pgs_munlocked",
1371 	"unevictable_pgs_cleared",
1372 	"unevictable_pgs_stranded",
1373 
1374 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1375 	"thp_fault_alloc",
1376 	"thp_fault_fallback",
1377 	"thp_fault_fallback_charge",
1378 	"thp_collapse_alloc",
1379 	"thp_collapse_alloc_failed",
1380 	"thp_file_alloc",
1381 	"thp_file_fallback",
1382 	"thp_file_fallback_charge",
1383 	"thp_file_mapped",
1384 	"thp_split_page",
1385 	"thp_split_page_failed",
1386 	"thp_deferred_split_page",
1387 	"thp_split_pmd",
1388 	"thp_scan_exceed_none_pte",
1389 	"thp_scan_exceed_swap_pte",
1390 	"thp_scan_exceed_share_pte",
1391 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1392 	"thp_split_pud",
1393 #endif
1394 	"thp_zero_page_alloc",
1395 	"thp_zero_page_alloc_failed",
1396 	"thp_swpout",
1397 	"thp_swpout_fallback",
1398 #endif
1399 #ifdef CONFIG_MEMORY_BALLOON
1400 	"balloon_inflate",
1401 	"balloon_deflate",
1402 #ifdef CONFIG_BALLOON_COMPACTION
1403 	"balloon_migrate",
1404 #endif
1405 #endif /* CONFIG_MEMORY_BALLOON */
1406 #ifdef CONFIG_DEBUG_TLBFLUSH
1407 	"nr_tlb_remote_flush",
1408 	"nr_tlb_remote_flush_received",
1409 	"nr_tlb_local_flush_all",
1410 	"nr_tlb_local_flush_one",
1411 #endif /* CONFIG_DEBUG_TLBFLUSH */
1412 
1413 #ifdef CONFIG_SWAP
1414 	"swap_ra",
1415 	"swap_ra_hit",
1416 #ifdef CONFIG_KSM
1417 	"ksm_swpin_copy",
1418 #endif
1419 #endif
1420 #ifdef CONFIG_KSM
1421 	"cow_ksm",
1422 #endif
1423 #ifdef CONFIG_ZSWAP
1424 	"zswpin",
1425 	"zswpout",
1426 	"zswpwb",
1427 #endif
1428 #ifdef CONFIG_X86
1429 	"direct_map_level2_splits",
1430 	"direct_map_level3_splits",
1431 #endif
1432 #ifdef CONFIG_PER_VMA_LOCK_STATS
1433 	"vma_lock_success",
1434 	"vma_lock_abort",
1435 	"vma_lock_retry",
1436 	"vma_lock_miss",
1437 #endif
1438 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1439 };
1440 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1441 
1442 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1443      defined(CONFIG_PROC_FS)
frag_start(struct seq_file * m,loff_t * pos)1444 static void *frag_start(struct seq_file *m, loff_t *pos)
1445 {
1446 	pg_data_t *pgdat;
1447 	loff_t node = *pos;
1448 
1449 	for (pgdat = first_online_pgdat();
1450 	     pgdat && node;
1451 	     pgdat = next_online_pgdat(pgdat))
1452 		--node;
1453 
1454 	return pgdat;
1455 }
1456 
frag_next(struct seq_file * m,void * arg,loff_t * pos)1457 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1458 {
1459 	pg_data_t *pgdat = (pg_data_t *)arg;
1460 
1461 	(*pos)++;
1462 	return next_online_pgdat(pgdat);
1463 }
1464 
frag_stop(struct seq_file * m,void * arg)1465 static void frag_stop(struct seq_file *m, void *arg)
1466 {
1467 }
1468 
1469 /*
1470  * Walk zones in a node and print using a callback.
1471  * If @assert_populated is true, only use callback for zones that are populated.
1472  */
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 *))1473 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1474 		bool assert_populated, bool nolock,
1475 		void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1476 {
1477 	struct zone *zone;
1478 	struct zone *node_zones = pgdat->node_zones;
1479 	unsigned long flags;
1480 
1481 	for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1482 		if (assert_populated && !populated_zone(zone))
1483 			continue;
1484 
1485 		if (!nolock)
1486 			spin_lock_irqsave(&zone->lock, flags);
1487 		print(m, pgdat, zone);
1488 		if (!nolock)
1489 			spin_unlock_irqrestore(&zone->lock, flags);
1490 	}
1491 }
1492 #endif
1493 
1494 #ifdef CONFIG_PROC_FS
frag_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1495 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1496 						struct zone *zone)
1497 {
1498 	int order;
1499 
1500 	seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1501 	for (order = 0; order < NR_PAGE_ORDERS; ++order)
1502 		/*
1503 		 * Access to nr_free is lockless as nr_free is used only for
1504 		 * printing purposes. Use data_race to avoid KCSAN warning.
1505 		 */
1506 		seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1507 	seq_putc(m, '\n');
1508 }
1509 
1510 /*
1511  * This walks the free areas for each zone.
1512  */
frag_show(struct seq_file * m,void * arg)1513 static int frag_show(struct seq_file *m, void *arg)
1514 {
1515 	pg_data_t *pgdat = (pg_data_t *)arg;
1516 	walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1517 	return 0;
1518 }
1519 
pagetypeinfo_showfree_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1520 static void pagetypeinfo_showfree_print(struct seq_file *m,
1521 					pg_data_t *pgdat, struct zone *zone)
1522 {
1523 	int order, mtype;
1524 
1525 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1526 		seq_printf(m, "Node %4d, zone %8s, type %12s ",
1527 					pgdat->node_id,
1528 					zone->name,
1529 					migratetype_names[mtype]);
1530 		for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1531 			unsigned long freecount = 0;
1532 			struct free_area *area;
1533 			struct list_head *curr;
1534 			bool overflow = false;
1535 
1536 			area = &(zone->free_area[order]);
1537 
1538 			list_for_each(curr, &area->free_list[mtype]) {
1539 				/*
1540 				 * Cap the free_list iteration because it might
1541 				 * be really large and we are under a spinlock
1542 				 * so a long time spent here could trigger a
1543 				 * hard lockup detector. Anyway this is a
1544 				 * debugging tool so knowing there is a handful
1545 				 * of pages of this order should be more than
1546 				 * sufficient.
1547 				 */
1548 				if (++freecount >= 100000) {
1549 					overflow = true;
1550 					break;
1551 				}
1552 			}
1553 			seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1554 			spin_unlock_irq(&zone->lock);
1555 			cond_resched();
1556 			spin_lock_irq(&zone->lock);
1557 		}
1558 		seq_putc(m, '\n');
1559 	}
1560 }
1561 
1562 /* Print out the free pages at each order for each migatetype */
pagetypeinfo_showfree(struct seq_file * m,void * arg)1563 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1564 {
1565 	int order;
1566 	pg_data_t *pgdat = (pg_data_t *)arg;
1567 
1568 	/* Print header */
1569 	seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1570 	for (order = 0; order < NR_PAGE_ORDERS; ++order)
1571 		seq_printf(m, "%6d ", order);
1572 	seq_putc(m, '\n');
1573 
1574 	walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1575 }
1576 
pagetypeinfo_showblockcount_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1577 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1578 					pg_data_t *pgdat, struct zone *zone)
1579 {
1580 	int mtype;
1581 	unsigned long pfn;
1582 	unsigned long start_pfn = zone->zone_start_pfn;
1583 	unsigned long end_pfn = zone_end_pfn(zone);
1584 	unsigned long count[MIGRATE_TYPES] = { 0, };
1585 
1586 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1587 		struct page *page;
1588 
1589 		page = pfn_to_online_page(pfn);
1590 		if (!page)
1591 			continue;
1592 
1593 		if (page_zone(page) != zone)
1594 			continue;
1595 
1596 		mtype = get_pageblock_migratetype(page);
1597 
1598 		if (mtype < MIGRATE_TYPES)
1599 			count[mtype]++;
1600 	}
1601 
1602 	/* Print counts */
1603 	seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1604 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1605 		seq_printf(m, "%12lu ", count[mtype]);
1606 	seq_putc(m, '\n');
1607 }
1608 
1609 /* Print out the number of pageblocks for each migratetype */
pagetypeinfo_showblockcount(struct seq_file * m,void * arg)1610 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1611 {
1612 	int mtype;
1613 	pg_data_t *pgdat = (pg_data_t *)arg;
1614 
1615 	seq_printf(m, "\n%-23s", "Number of blocks type ");
1616 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1617 		seq_printf(m, "%12s ", migratetype_names[mtype]);
1618 	seq_putc(m, '\n');
1619 	walk_zones_in_node(m, pgdat, true, false,
1620 		pagetypeinfo_showblockcount_print);
1621 }
1622 
1623 /*
1624  * Print out the number of pageblocks for each migratetype that contain pages
1625  * of other types. This gives an indication of how well fallbacks are being
1626  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1627  * to determine what is going on
1628  */
pagetypeinfo_showmixedcount(struct seq_file * m,pg_data_t * pgdat)1629 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1630 {
1631 #ifdef CONFIG_PAGE_OWNER
1632 	int mtype;
1633 
1634 	if (!static_branch_unlikely(&page_owner_inited))
1635 		return;
1636 
1637 	drain_all_pages(NULL);
1638 
1639 	seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1640 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1641 		seq_printf(m, "%12s ", migratetype_names[mtype]);
1642 	seq_putc(m, '\n');
1643 
1644 	walk_zones_in_node(m, pgdat, true, true,
1645 		pagetypeinfo_showmixedcount_print);
1646 #endif /* CONFIG_PAGE_OWNER */
1647 }
1648 
1649 /*
1650  * This prints out statistics in relation to grouping pages by mobility.
1651  * It is expensive to collect so do not constantly read the file.
1652  */
pagetypeinfo_show(struct seq_file * m,void * arg)1653 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1654 {
1655 	pg_data_t *pgdat = (pg_data_t *)arg;
1656 
1657 	/* check memoryless node */
1658 	if (!node_state(pgdat->node_id, N_MEMORY))
1659 		return 0;
1660 
1661 	seq_printf(m, "Page block order: %d\n", pageblock_order);
1662 	seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1663 	seq_putc(m, '\n');
1664 	pagetypeinfo_showfree(m, pgdat);
1665 	pagetypeinfo_showblockcount(m, pgdat);
1666 	pagetypeinfo_showmixedcount(m, pgdat);
1667 
1668 	return 0;
1669 }
1670 
1671 static const struct seq_operations fragmentation_op = {
1672 	.start	= frag_start,
1673 	.next	= frag_next,
1674 	.stop	= frag_stop,
1675 	.show	= frag_show,
1676 };
1677 
1678 static const struct seq_operations pagetypeinfo_op = {
1679 	.start	= frag_start,
1680 	.next	= frag_next,
1681 	.stop	= frag_stop,
1682 	.show	= pagetypeinfo_show,
1683 };
1684 
is_zone_first_populated(pg_data_t * pgdat,struct zone * zone)1685 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1686 {
1687 	int zid;
1688 
1689 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1690 		struct zone *compare = &pgdat->node_zones[zid];
1691 
1692 		if (populated_zone(compare))
1693 			return zone == compare;
1694 	}
1695 
1696 	return false;
1697 }
1698 
zoneinfo_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1699 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1700 							struct zone *zone)
1701 {
1702 	int i;
1703 	seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1704 	if (is_zone_first_populated(pgdat, zone)) {
1705 		seq_printf(m, "\n  per-node stats");
1706 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1707 			unsigned long pages = node_page_state_pages(pgdat, i);
1708 
1709 			if (vmstat_item_print_in_thp(i))
1710 				pages /= HPAGE_PMD_NR;
1711 			seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1712 				   pages);
1713 		}
1714 	}
1715 	seq_printf(m,
1716 		   "\n  pages free     %lu"
1717 		   "\n        boost    %lu"
1718 		   "\n        min      %lu"
1719 		   "\n        low      %lu"
1720 		   "\n        high     %lu"
1721 		   "\n        spanned  %lu"
1722 		   "\n        present  %lu"
1723 		   "\n        managed  %lu"
1724 		   "\n        cma      %lu",
1725 		   zone_page_state(zone, NR_FREE_PAGES),
1726 		   zone->watermark_boost,
1727 		   min_wmark_pages(zone),
1728 		   low_wmark_pages(zone),
1729 		   high_wmark_pages(zone),
1730 		   zone->spanned_pages,
1731 		   zone->present_pages,
1732 		   zone_managed_pages(zone),
1733 		   zone_cma_pages(zone));
1734 
1735 	seq_printf(m,
1736 		   "\n        protection: (%ld",
1737 		   zone->lowmem_reserve[0]);
1738 	for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1739 		seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1740 	seq_putc(m, ')');
1741 
1742 	/* If unpopulated, no other information is useful */
1743 	if (!populated_zone(zone)) {
1744 		seq_putc(m, '\n');
1745 		return;
1746 	}
1747 
1748 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1749 		seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1750 			   zone_page_state(zone, i));
1751 
1752 #ifdef CONFIG_NUMA
1753 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1754 		seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1755 			   zone_numa_event_state(zone, i));
1756 #endif
1757 
1758 	seq_printf(m, "\n  pagesets");
1759 	for_each_online_cpu(i) {
1760 		struct per_cpu_pages *pcp;
1761 		struct per_cpu_zonestat __maybe_unused *pzstats;
1762 
1763 		pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1764 		seq_printf(m,
1765 			   "\n    cpu: %i"
1766 			   "\n              count: %i"
1767 			   "\n              high:  %i"
1768 			   "\n              batch: %i",
1769 			   i,
1770 			   pcp->count,
1771 			   pcp->high,
1772 			   pcp->batch);
1773 #ifdef CONFIG_SMP
1774 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1775 		seq_printf(m, "\n  vm stats threshold: %d",
1776 				pzstats->stat_threshold);
1777 #endif
1778 	}
1779 	seq_printf(m,
1780 		   "\n  node_unreclaimable:  %u"
1781 		   "\n  start_pfn:           %lu",
1782 		   pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1783 		   zone->zone_start_pfn);
1784 	seq_putc(m, '\n');
1785 }
1786 
1787 /*
1788  * Output information about zones in @pgdat.  All zones are printed regardless
1789  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1790  * set of all zones and userspace would not be aware of such zones if they are
1791  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1792  */
zoneinfo_show(struct seq_file * m,void * arg)1793 static int zoneinfo_show(struct seq_file *m, void *arg)
1794 {
1795 	pg_data_t *pgdat = (pg_data_t *)arg;
1796 	walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1797 	return 0;
1798 }
1799 
1800 static const struct seq_operations zoneinfo_op = {
1801 	.start	= frag_start, /* iterate over all zones. The same as in
1802 			       * fragmentation. */
1803 	.next	= frag_next,
1804 	.stop	= frag_stop,
1805 	.show	= zoneinfo_show,
1806 };
1807 
1808 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1809 			 NR_VM_NUMA_EVENT_ITEMS + \
1810 			 NR_VM_NODE_STAT_ITEMS + \
1811 			 NR_VM_STAT_ITEMS + \
1812 			 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1813 			  NR_VM_EVENT_ITEMS : 0))
1814 
vmstat_start(struct seq_file * m,loff_t * pos)1815 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1816 {
1817 	unsigned long *v;
1818 	int i;
1819 
1820 	if (*pos >= NR_VMSTAT_ITEMS)
1821 		return NULL;
1822 
1823 	BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1824 	fold_vm_numa_events();
1825 	v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1826 	m->private = v;
1827 	if (!v)
1828 		return ERR_PTR(-ENOMEM);
1829 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1830 		v[i] = global_zone_page_state(i);
1831 	v += NR_VM_ZONE_STAT_ITEMS;
1832 
1833 #ifdef CONFIG_NUMA
1834 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1835 		v[i] = global_numa_event_state(i);
1836 	v += NR_VM_NUMA_EVENT_ITEMS;
1837 #endif
1838 
1839 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1840 		v[i] = global_node_page_state_pages(i);
1841 		if (vmstat_item_print_in_thp(i))
1842 			v[i] /= HPAGE_PMD_NR;
1843 	}
1844 	v += NR_VM_NODE_STAT_ITEMS;
1845 
1846 	global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1847 			    v + NR_DIRTY_THRESHOLD);
1848 	v[NR_MEMMAP_PAGES] = atomic_long_read(&nr_memmap_pages);
1849 	v[NR_MEMMAP_BOOT_PAGES] = atomic_long_read(&nr_memmap_boot_pages);
1850 	v += NR_VM_STAT_ITEMS;
1851 
1852 #ifdef CONFIG_VM_EVENT_COUNTERS
1853 	all_vm_events(v);
1854 	v[PGPGIN] /= 2;		/* sectors -> kbytes */
1855 	v[PGPGOUT] /= 2;
1856 #endif
1857 	return (unsigned long *)m->private + *pos;
1858 }
1859 
vmstat_next(struct seq_file * m,void * arg,loff_t * pos)1860 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1861 {
1862 	(*pos)++;
1863 	if (*pos >= NR_VMSTAT_ITEMS)
1864 		return NULL;
1865 	return (unsigned long *)m->private + *pos;
1866 }
1867 
vmstat_show(struct seq_file * m,void * arg)1868 static int vmstat_show(struct seq_file *m, void *arg)
1869 {
1870 	unsigned long *l = arg;
1871 	unsigned long off = l - (unsigned long *)m->private;
1872 
1873 	seq_puts(m, vmstat_text[off]);
1874 	seq_put_decimal_ull(m, " ", *l);
1875 	seq_putc(m, '\n');
1876 
1877 	if (off == NR_VMSTAT_ITEMS - 1) {
1878 		/*
1879 		 * We've come to the end - add any deprecated counters to avoid
1880 		 * breaking userspace which might depend on them being present.
1881 		 */
1882 		seq_puts(m, "nr_unstable 0\n");
1883 	}
1884 	return 0;
1885 }
1886 
vmstat_stop(struct seq_file * m,void * arg)1887 static void vmstat_stop(struct seq_file *m, void *arg)
1888 {
1889 	kfree(m->private);
1890 	m->private = NULL;
1891 }
1892 
1893 static const struct seq_operations vmstat_op = {
1894 	.start	= vmstat_start,
1895 	.next	= vmstat_next,
1896 	.stop	= vmstat_stop,
1897 	.show	= vmstat_show,
1898 };
1899 #endif /* CONFIG_PROC_FS */
1900 
1901 #ifdef CONFIG_SMP
1902 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1903 int sysctl_stat_interval __read_mostly = HZ;
1904 
1905 #ifdef CONFIG_PROC_FS
refresh_vm_stats(struct work_struct * work)1906 static void refresh_vm_stats(struct work_struct *work)
1907 {
1908 	refresh_cpu_vm_stats(true);
1909 }
1910 
vmstat_refresh(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)1911 int vmstat_refresh(const struct ctl_table *table, int write,
1912 		   void *buffer, size_t *lenp, loff_t *ppos)
1913 {
1914 	long val;
1915 	int err;
1916 	int i;
1917 
1918 	/*
1919 	 * The regular update, every sysctl_stat_interval, may come later
1920 	 * than expected: leaving a significant amount in per_cpu buckets.
1921 	 * This is particularly misleading when checking a quantity of HUGE
1922 	 * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1923 	 * which can equally be echo'ed to or cat'ted from (by root),
1924 	 * can be used to update the stats just before reading them.
1925 	 *
1926 	 * Oh, and since global_zone_page_state() etc. are so careful to hide
1927 	 * transiently negative values, report an error here if any of
1928 	 * the stats is negative, so we know to go looking for imbalance.
1929 	 */
1930 	err = schedule_on_each_cpu(refresh_vm_stats);
1931 	if (err)
1932 		return err;
1933 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1934 		/*
1935 		 * Skip checking stats known to go negative occasionally.
1936 		 */
1937 		switch (i) {
1938 		case NR_ZONE_WRITE_PENDING:
1939 		case NR_FREE_CMA_PAGES:
1940 			continue;
1941 		}
1942 		val = atomic_long_read(&vm_zone_stat[i]);
1943 		if (val < 0) {
1944 			pr_warn("%s: %s %ld\n",
1945 				__func__, zone_stat_name(i), val);
1946 		}
1947 	}
1948 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1949 		/*
1950 		 * Skip checking stats known to go negative occasionally.
1951 		 */
1952 		switch (i) {
1953 		case NR_WRITEBACK:
1954 			continue;
1955 		}
1956 		val = atomic_long_read(&vm_node_stat[i]);
1957 		if (val < 0) {
1958 			pr_warn("%s: %s %ld\n",
1959 				__func__, node_stat_name(i), val);
1960 		}
1961 	}
1962 	if (write)
1963 		*ppos += *lenp;
1964 	else
1965 		*lenp = 0;
1966 	return 0;
1967 }
1968 #endif /* CONFIG_PROC_FS */
1969 
vmstat_update(struct work_struct * w)1970 static void vmstat_update(struct work_struct *w)
1971 {
1972 	if (refresh_cpu_vm_stats(true)) {
1973 		/*
1974 		 * Counters were updated so we expect more updates
1975 		 * to occur in the future. Keep on running the
1976 		 * update worker thread.
1977 		 */
1978 		queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1979 				this_cpu_ptr(&vmstat_work),
1980 				round_jiffies_relative(sysctl_stat_interval));
1981 	}
1982 }
1983 
1984 /*
1985  * Check if the diffs for a certain cpu indicate that
1986  * an update is needed.
1987  */
need_update(int cpu)1988 static bool need_update(int cpu)
1989 {
1990 	pg_data_t *last_pgdat = NULL;
1991 	struct zone *zone;
1992 
1993 	for_each_populated_zone(zone) {
1994 		struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1995 		struct per_cpu_nodestat *n;
1996 
1997 		/*
1998 		 * The fast way of checking if there are any vmstat diffs.
1999 		 */
2000 		if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
2001 			return true;
2002 
2003 		if (last_pgdat == zone->zone_pgdat)
2004 			continue;
2005 		last_pgdat = zone->zone_pgdat;
2006 		n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
2007 		if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
2008 			return true;
2009 	}
2010 	return false;
2011 }
2012 
2013 /*
2014  * Switch off vmstat processing and then fold all the remaining differentials
2015  * until the diffs stay at zero. The function is used by NOHZ and can only be
2016  * invoked when tick processing is not active.
2017  */
quiet_vmstat(void)2018 void quiet_vmstat(void)
2019 {
2020 	if (system_state != SYSTEM_RUNNING)
2021 		return;
2022 
2023 	if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
2024 		return;
2025 
2026 	if (!need_update(smp_processor_id()))
2027 		return;
2028 
2029 	/*
2030 	 * Just refresh counters and do not care about the pending delayed
2031 	 * vmstat_update. It doesn't fire that often to matter and canceling
2032 	 * it would be too expensive from this path.
2033 	 * vmstat_shepherd will take care about that for us.
2034 	 */
2035 	refresh_cpu_vm_stats(false);
2036 }
2037 
2038 /*
2039  * Shepherd worker thread that checks the
2040  * differentials of processors that have their worker
2041  * threads for vm statistics updates disabled because of
2042  * inactivity.
2043  */
2044 static void vmstat_shepherd(struct work_struct *w);
2045 
2046 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2047 
vmstat_shepherd(struct work_struct * w)2048 static void vmstat_shepherd(struct work_struct *w)
2049 {
2050 	int cpu;
2051 
2052 	cpus_read_lock();
2053 	/* Check processors whose vmstat worker threads have been disabled */
2054 	for_each_online_cpu(cpu) {
2055 		struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2056 
2057 		/*
2058 		 * In kernel users of vmstat counters either require the precise value and
2059 		 * they are using zone_page_state_snapshot interface or they can live with
2060 		 * an imprecision as the regular flushing can happen at arbitrary time and
2061 		 * cumulative error can grow (see calculate_normal_threshold).
2062 		 *
2063 		 * From that POV the regular flushing can be postponed for CPUs that have
2064 		 * been isolated from the kernel interference without critical
2065 		 * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2066 		 * for all isolated CPUs to avoid interference with the isolated workload.
2067 		 */
2068 		if (cpu_is_isolated(cpu))
2069 			continue;
2070 
2071 		if (!delayed_work_pending(dw) && need_update(cpu))
2072 			queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2073 
2074 		cond_resched();
2075 	}
2076 	cpus_read_unlock();
2077 
2078 	schedule_delayed_work(&shepherd,
2079 		round_jiffies_relative(sysctl_stat_interval));
2080 }
2081 
start_shepherd_timer(void)2082 static void __init start_shepherd_timer(void)
2083 {
2084 	int cpu;
2085 
2086 	for_each_possible_cpu(cpu)
2087 		INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2088 			vmstat_update);
2089 
2090 	schedule_delayed_work(&shepherd,
2091 		round_jiffies_relative(sysctl_stat_interval));
2092 }
2093 
init_cpu_node_state(void)2094 static void __init init_cpu_node_state(void)
2095 {
2096 	int node;
2097 
2098 	for_each_online_node(node) {
2099 		if (!cpumask_empty(cpumask_of_node(node)))
2100 			node_set_state(node, N_CPU);
2101 	}
2102 }
2103 
vmstat_cpu_online(unsigned int cpu)2104 static int vmstat_cpu_online(unsigned int cpu)
2105 {
2106 	refresh_zone_stat_thresholds();
2107 
2108 	if (!node_state(cpu_to_node(cpu), N_CPU)) {
2109 		node_set_state(cpu_to_node(cpu), N_CPU);
2110 	}
2111 
2112 	return 0;
2113 }
2114 
vmstat_cpu_down_prep(unsigned int cpu)2115 static int vmstat_cpu_down_prep(unsigned int cpu)
2116 {
2117 	cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2118 	return 0;
2119 }
2120 
vmstat_cpu_dead(unsigned int cpu)2121 static int vmstat_cpu_dead(unsigned int cpu)
2122 {
2123 	const struct cpumask *node_cpus;
2124 	int node;
2125 
2126 	node = cpu_to_node(cpu);
2127 
2128 	refresh_zone_stat_thresholds();
2129 	node_cpus = cpumask_of_node(node);
2130 	if (!cpumask_empty(node_cpus))
2131 		return 0;
2132 
2133 	node_clear_state(node, N_CPU);
2134 
2135 	return 0;
2136 }
2137 
2138 #endif
2139 
2140 struct workqueue_struct *mm_percpu_wq;
2141 
init_mm_internals(void)2142 void __init init_mm_internals(void)
2143 {
2144 	int ret __maybe_unused;
2145 
2146 	mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2147 
2148 #ifdef CONFIG_SMP
2149 	ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2150 					NULL, vmstat_cpu_dead);
2151 	if (ret < 0)
2152 		pr_err("vmstat: failed to register 'dead' hotplug state\n");
2153 
2154 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2155 					vmstat_cpu_online,
2156 					vmstat_cpu_down_prep);
2157 	if (ret < 0)
2158 		pr_err("vmstat: failed to register 'online' hotplug state\n");
2159 
2160 	cpus_read_lock();
2161 	init_cpu_node_state();
2162 	cpus_read_unlock();
2163 
2164 	start_shepherd_timer();
2165 #endif
2166 #ifdef CONFIG_PROC_FS
2167 	proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2168 	proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2169 	proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2170 	proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2171 #endif
2172 }
2173 
2174 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2175 
2176 /*
2177  * Return an index indicating how much of the available free memory is
2178  * unusable for an allocation of the requested size.
2179  */
unusable_free_index(unsigned int order,struct contig_page_info * info)2180 static int unusable_free_index(unsigned int order,
2181 				struct contig_page_info *info)
2182 {
2183 	/* No free memory is interpreted as all free memory is unusable */
2184 	if (info->free_pages == 0)
2185 		return 1000;
2186 
2187 	/*
2188 	 * Index should be a value between 0 and 1. Return a value to 3
2189 	 * decimal places.
2190 	 *
2191 	 * 0 => no fragmentation
2192 	 * 1 => high fragmentation
2193 	 */
2194 	return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2195 
2196 }
2197 
unusable_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)2198 static void unusable_show_print(struct seq_file *m,
2199 					pg_data_t *pgdat, struct zone *zone)
2200 {
2201 	unsigned int order;
2202 	int index;
2203 	struct contig_page_info info;
2204 
2205 	seq_printf(m, "Node %d, zone %8s ",
2206 				pgdat->node_id,
2207 				zone->name);
2208 	for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2209 		fill_contig_page_info(zone, order, &info);
2210 		index = unusable_free_index(order, &info);
2211 		seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2212 	}
2213 
2214 	seq_putc(m, '\n');
2215 }
2216 
2217 /*
2218  * Display unusable free space index
2219  *
2220  * The unusable free space index measures how much of the available free
2221  * memory cannot be used to satisfy an allocation of a given size and is a
2222  * value between 0 and 1. The higher the value, the more of free memory is
2223  * unusable and by implication, the worse the external fragmentation is. This
2224  * can be expressed as a percentage by multiplying by 100.
2225  */
unusable_show(struct seq_file * m,void * arg)2226 static int unusable_show(struct seq_file *m, void *arg)
2227 {
2228 	pg_data_t *pgdat = (pg_data_t *)arg;
2229 
2230 	/* check memoryless node */
2231 	if (!node_state(pgdat->node_id, N_MEMORY))
2232 		return 0;
2233 
2234 	walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2235 
2236 	return 0;
2237 }
2238 
2239 static const struct seq_operations unusable_sops = {
2240 	.start	= frag_start,
2241 	.next	= frag_next,
2242 	.stop	= frag_stop,
2243 	.show	= unusable_show,
2244 };
2245 
2246 DEFINE_SEQ_ATTRIBUTE(unusable);
2247 
extfrag_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)2248 static void extfrag_show_print(struct seq_file *m,
2249 					pg_data_t *pgdat, struct zone *zone)
2250 {
2251 	unsigned int order;
2252 	int index;
2253 
2254 	/* Alloc on stack as interrupts are disabled for zone walk */
2255 	struct contig_page_info info;
2256 
2257 	seq_printf(m, "Node %d, zone %8s ",
2258 				pgdat->node_id,
2259 				zone->name);
2260 	for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2261 		fill_contig_page_info(zone, order, &info);
2262 		index = __fragmentation_index(order, &info);
2263 		seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2264 	}
2265 
2266 	seq_putc(m, '\n');
2267 }
2268 
2269 /*
2270  * Display fragmentation index for orders that allocations would fail for
2271  */
extfrag_show(struct seq_file * m,void * arg)2272 static int extfrag_show(struct seq_file *m, void *arg)
2273 {
2274 	pg_data_t *pgdat = (pg_data_t *)arg;
2275 
2276 	walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2277 
2278 	return 0;
2279 }
2280 
2281 static const struct seq_operations extfrag_sops = {
2282 	.start	= frag_start,
2283 	.next	= frag_next,
2284 	.stop	= frag_stop,
2285 	.show	= extfrag_show,
2286 };
2287 
2288 DEFINE_SEQ_ATTRIBUTE(extfrag);
2289 
extfrag_debug_init(void)2290 static int __init extfrag_debug_init(void)
2291 {
2292 	struct dentry *extfrag_debug_root;
2293 
2294 	extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2295 
2296 	debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2297 			    &unusable_fops);
2298 
2299 	debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2300 			    &extfrag_fops);
2301 
2302 	return 0;
2303 }
2304 
2305 module_init(extfrag_debug_init);
2306 
2307 #endif
2308