xref: /openbmc/linux/mm/vmstat.c (revision ded1ffea)
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(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)77 int sysctl_vm_numa_stat_handler(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 o, n, t, z;
563 
564 	do {
565 		z = 0;  /* overflow to zone counters */
566 
567 		/*
568 		 * The fetching of the stat_threshold is racy. We may apply
569 		 * a counter threshold to the wrong the cpu if we get
570 		 * rescheduled while executing here. However, the next
571 		 * counter update will apply the threshold again and
572 		 * therefore bring the counter under the threshold again.
573 		 *
574 		 * Most of the time the thresholds are the same anyways
575 		 * for all cpus in a zone.
576 		 */
577 		t = this_cpu_read(pcp->stat_threshold);
578 
579 		o = this_cpu_read(*p);
580 		n = delta + o;
581 
582 		if (abs(n) > t) {
583 			int os = overstep_mode * (t >> 1) ;
584 
585 			/* Overflow must be added to zone counters */
586 			z = n + os;
587 			n = -os;
588 		}
589 	} while (this_cpu_cmpxchg(*p, o, n) != o);
590 
591 	if (z)
592 		zone_page_state_add(z, zone, item);
593 }
594 
mod_zone_page_state(struct zone * zone,enum zone_stat_item item,long delta)595 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
596 			 long delta)
597 {
598 	mod_zone_state(zone, item, delta, 0);
599 }
600 EXPORT_SYMBOL(mod_zone_page_state);
601 
inc_zone_page_state(struct page * page,enum zone_stat_item item)602 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
603 {
604 	mod_zone_state(page_zone(page), item, 1, 1);
605 }
606 EXPORT_SYMBOL(inc_zone_page_state);
607 
dec_zone_page_state(struct page * page,enum zone_stat_item item)608 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
609 {
610 	mod_zone_state(page_zone(page), item, -1, -1);
611 }
612 EXPORT_SYMBOL(dec_zone_page_state);
613 
mod_node_state(struct pglist_data * pgdat,enum node_stat_item item,int delta,int overstep_mode)614 static inline void mod_node_state(struct pglist_data *pgdat,
615        enum node_stat_item item, int delta, int overstep_mode)
616 {
617 	struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
618 	s8 __percpu *p = pcp->vm_node_stat_diff + item;
619 	long o, n, t, z;
620 
621 	if (vmstat_item_in_bytes(item)) {
622 		/*
623 		 * Only cgroups use subpage accounting right now; at
624 		 * the global level, these items still change in
625 		 * multiples of whole pages. Store them as pages
626 		 * internally to keep the per-cpu counters compact.
627 		 */
628 		VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
629 		delta >>= PAGE_SHIFT;
630 	}
631 
632 	do {
633 		z = 0;  /* overflow to node counters */
634 
635 		/*
636 		 * The fetching of the stat_threshold is racy. We may apply
637 		 * a counter threshold to the wrong the cpu if we get
638 		 * rescheduled while executing here. However, the next
639 		 * counter update will apply the threshold again and
640 		 * therefore bring the counter under the threshold again.
641 		 *
642 		 * Most of the time the thresholds are the same anyways
643 		 * for all cpus in a node.
644 		 */
645 		t = this_cpu_read(pcp->stat_threshold);
646 
647 		o = this_cpu_read(*p);
648 		n = delta + o;
649 
650 		if (abs(n) > t) {
651 			int os = overstep_mode * (t >> 1) ;
652 
653 			/* Overflow must be added to node counters */
654 			z = n + os;
655 			n = -os;
656 		}
657 	} while (this_cpu_cmpxchg(*p, o, n) != o);
658 
659 	if (z)
660 		node_page_state_add(z, pgdat, item);
661 }
662 
mod_node_page_state(struct pglist_data * pgdat,enum node_stat_item item,long delta)663 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
664 					long delta)
665 {
666 	mod_node_state(pgdat, item, delta, 0);
667 }
668 EXPORT_SYMBOL(mod_node_page_state);
669 
inc_node_state(struct pglist_data * pgdat,enum node_stat_item item)670 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
671 {
672 	mod_node_state(pgdat, item, 1, 1);
673 }
674 
inc_node_page_state(struct page * page,enum node_stat_item item)675 void inc_node_page_state(struct page *page, enum node_stat_item item)
676 {
677 	mod_node_state(page_pgdat(page), item, 1, 1);
678 }
679 EXPORT_SYMBOL(inc_node_page_state);
680 
dec_node_page_state(struct page * page,enum node_stat_item item)681 void dec_node_page_state(struct page *page, enum node_stat_item item)
682 {
683 	mod_node_state(page_pgdat(page), item, -1, -1);
684 }
685 EXPORT_SYMBOL(dec_node_page_state);
686 #else
687 /*
688  * Use interrupt disable to serialize counter updates
689  */
mod_zone_page_state(struct zone * zone,enum zone_stat_item item,long delta)690 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
691 			 long delta)
692 {
693 	unsigned long flags;
694 
695 	local_irq_save(flags);
696 	__mod_zone_page_state(zone, item, delta);
697 	local_irq_restore(flags);
698 }
699 EXPORT_SYMBOL(mod_zone_page_state);
700 
inc_zone_page_state(struct page * page,enum zone_stat_item item)701 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
702 {
703 	unsigned long flags;
704 	struct zone *zone;
705 
706 	zone = page_zone(page);
707 	local_irq_save(flags);
708 	__inc_zone_state(zone, item);
709 	local_irq_restore(flags);
710 }
711 EXPORT_SYMBOL(inc_zone_page_state);
712 
dec_zone_page_state(struct page * page,enum zone_stat_item item)713 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
714 {
715 	unsigned long flags;
716 
717 	local_irq_save(flags);
718 	__dec_zone_page_state(page, item);
719 	local_irq_restore(flags);
720 }
721 EXPORT_SYMBOL(dec_zone_page_state);
722 
inc_node_state(struct pglist_data * pgdat,enum node_stat_item item)723 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
724 {
725 	unsigned long flags;
726 
727 	local_irq_save(flags);
728 	__inc_node_state(pgdat, item);
729 	local_irq_restore(flags);
730 }
731 EXPORT_SYMBOL(inc_node_state);
732 
mod_node_page_state(struct pglist_data * pgdat,enum node_stat_item item,long delta)733 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
734 					long delta)
735 {
736 	unsigned long flags;
737 
738 	local_irq_save(flags);
739 	__mod_node_page_state(pgdat, item, delta);
740 	local_irq_restore(flags);
741 }
742 EXPORT_SYMBOL(mod_node_page_state);
743 
inc_node_page_state(struct page * page,enum node_stat_item item)744 void inc_node_page_state(struct page *page, enum node_stat_item item)
745 {
746 	unsigned long flags;
747 	struct pglist_data *pgdat;
748 
749 	pgdat = page_pgdat(page);
750 	local_irq_save(flags);
751 	__inc_node_state(pgdat, item);
752 	local_irq_restore(flags);
753 }
754 EXPORT_SYMBOL(inc_node_page_state);
755 
dec_node_page_state(struct page * page,enum node_stat_item item)756 void dec_node_page_state(struct page *page, enum node_stat_item item)
757 {
758 	unsigned long flags;
759 
760 	local_irq_save(flags);
761 	__dec_node_page_state(page, item);
762 	local_irq_restore(flags);
763 }
764 EXPORT_SYMBOL(dec_node_page_state);
765 #endif
766 
767 /*
768  * Fold a differential into the global counters.
769  * Returns the number of counters updated.
770  */
fold_diff(int * zone_diff,int * node_diff)771 static int fold_diff(int *zone_diff, int *node_diff)
772 {
773 	int i;
774 	int changes = 0;
775 
776 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
777 		if (zone_diff[i]) {
778 			atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
779 			changes++;
780 	}
781 
782 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
783 		if (node_diff[i]) {
784 			atomic_long_add(node_diff[i], &vm_node_stat[i]);
785 			changes++;
786 	}
787 	return changes;
788 }
789 
790 /*
791  * Update the zone counters for the current cpu.
792  *
793  * Note that refresh_cpu_vm_stats strives to only access
794  * node local memory. The per cpu pagesets on remote zones are placed
795  * in the memory local to the processor using that pageset. So the
796  * loop over all zones will access a series of cachelines local to
797  * the processor.
798  *
799  * The call to zone_page_state_add updates the cachelines with the
800  * statistics in the remote zone struct as well as the global cachelines
801  * with the global counters. These could cause remote node cache line
802  * bouncing and will have to be only done when necessary.
803  *
804  * The function returns the number of global counters updated.
805  */
refresh_cpu_vm_stats(bool do_pagesets)806 static int refresh_cpu_vm_stats(bool do_pagesets)
807 {
808 	struct pglist_data *pgdat;
809 	struct zone *zone;
810 	int i;
811 	int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
812 	int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
813 	int changes = 0;
814 
815 	for_each_populated_zone(zone) {
816 		struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
817 #ifdef CONFIG_NUMA
818 		struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
819 #endif
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 #ifdef CONFIG_NUMA
836 
837 		if (do_pagesets) {
838 			cond_resched();
839 			/*
840 			 * Deal with draining the remote pageset of this
841 			 * processor
842 			 *
843 			 * Check if there are pages remaining in this pageset
844 			 * if not then there is nothing to expire.
845 			 */
846 			if (!__this_cpu_read(pcp->expire) ||
847 			       !__this_cpu_read(pcp->count))
848 				continue;
849 
850 			/*
851 			 * We never drain zones local to this processor.
852 			 */
853 			if (zone_to_nid(zone) == numa_node_id()) {
854 				__this_cpu_write(pcp->expire, 0);
855 				continue;
856 			}
857 
858 			if (__this_cpu_dec_return(pcp->expire))
859 				continue;
860 
861 			if (__this_cpu_read(pcp->count)) {
862 				drain_zone_pages(zone, this_cpu_ptr(pcp));
863 				changes++;
864 			}
865 		}
866 #endif
867 	}
868 
869 	for_each_online_pgdat(pgdat) {
870 		struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
871 
872 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
873 			int v;
874 
875 			v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
876 			if (v) {
877 				atomic_long_add(v, &pgdat->vm_stat[i]);
878 				global_node_diff[i] += v;
879 			}
880 		}
881 	}
882 
883 	changes += fold_diff(global_zone_diff, global_node_diff);
884 	return changes;
885 }
886 
887 /*
888  * Fold the data for an offline cpu into the global array.
889  * There cannot be any access by the offline cpu and therefore
890  * synchronization is simplified.
891  */
cpu_vm_stats_fold(int cpu)892 void cpu_vm_stats_fold(int cpu)
893 {
894 	struct pglist_data *pgdat;
895 	struct zone *zone;
896 	int i;
897 	int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
898 	int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
899 
900 	for_each_populated_zone(zone) {
901 		struct per_cpu_zonestat *pzstats;
902 
903 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
904 
905 		for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
906 			if (pzstats->vm_stat_diff[i]) {
907 				int v;
908 
909 				v = pzstats->vm_stat_diff[i];
910 				pzstats->vm_stat_diff[i] = 0;
911 				atomic_long_add(v, &zone->vm_stat[i]);
912 				global_zone_diff[i] += v;
913 			}
914 		}
915 #ifdef CONFIG_NUMA
916 		for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
917 			if (pzstats->vm_numa_event[i]) {
918 				unsigned long v;
919 
920 				v = pzstats->vm_numa_event[i];
921 				pzstats->vm_numa_event[i] = 0;
922 				zone_numa_event_add(v, zone, i);
923 			}
924 		}
925 #endif
926 	}
927 
928 	for_each_online_pgdat(pgdat) {
929 		struct per_cpu_nodestat *p;
930 
931 		p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
932 
933 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
934 			if (p->vm_node_stat_diff[i]) {
935 				int v;
936 
937 				v = p->vm_node_stat_diff[i];
938 				p->vm_node_stat_diff[i] = 0;
939 				atomic_long_add(v, &pgdat->vm_stat[i]);
940 				global_node_diff[i] += v;
941 			}
942 	}
943 
944 	fold_diff(global_zone_diff, global_node_diff);
945 }
946 
947 /*
948  * this is only called if !populated_zone(zone), which implies no other users of
949  * pset->vm_stat_diff[] exist.
950  */
drain_zonestat(struct zone * zone,struct per_cpu_zonestat * pzstats)951 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
952 {
953 	unsigned long v;
954 	int i;
955 
956 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
957 		if (pzstats->vm_stat_diff[i]) {
958 			v = pzstats->vm_stat_diff[i];
959 			pzstats->vm_stat_diff[i] = 0;
960 			zone_page_state_add(v, zone, i);
961 		}
962 	}
963 
964 #ifdef CONFIG_NUMA
965 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
966 		if (pzstats->vm_numa_event[i]) {
967 			v = pzstats->vm_numa_event[i];
968 			pzstats->vm_numa_event[i] = 0;
969 			zone_numa_event_add(v, zone, i);
970 		}
971 	}
972 #endif
973 }
974 #endif
975 
976 #ifdef CONFIG_NUMA
977 /*
978  * Determine the per node value of a stat item. This function
979  * is called frequently in a NUMA machine, so try to be as
980  * frugal as possible.
981  */
sum_zone_node_page_state(int node,enum zone_stat_item item)982 unsigned long sum_zone_node_page_state(int node,
983 				 enum zone_stat_item item)
984 {
985 	struct zone *zones = NODE_DATA(node)->node_zones;
986 	int i;
987 	unsigned long count = 0;
988 
989 	for (i = 0; i < MAX_NR_ZONES; i++)
990 		count += zone_page_state(zones + i, item);
991 
992 	return count;
993 }
994 
995 /* Determine the per node value of a numa stat item. */
sum_zone_numa_event_state(int node,enum numa_stat_item item)996 unsigned long sum_zone_numa_event_state(int node,
997 				 enum numa_stat_item item)
998 {
999 	struct zone *zones = NODE_DATA(node)->node_zones;
1000 	unsigned long count = 0;
1001 	int i;
1002 
1003 	for (i = 0; i < MAX_NR_ZONES; i++)
1004 		count += zone_numa_event_state(zones + i, item);
1005 
1006 	return count;
1007 }
1008 
1009 /*
1010  * Determine the per node value of a stat item.
1011  */
node_page_state_pages(struct pglist_data * pgdat,enum node_stat_item item)1012 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1013 				    enum node_stat_item item)
1014 {
1015 	long x = atomic_long_read(&pgdat->vm_stat[item]);
1016 #ifdef CONFIG_SMP
1017 	if (x < 0)
1018 		x = 0;
1019 #endif
1020 	return x;
1021 }
1022 
node_page_state(struct pglist_data * pgdat,enum node_stat_item item)1023 unsigned long node_page_state(struct pglist_data *pgdat,
1024 			      enum node_stat_item item)
1025 {
1026 	VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1027 
1028 	return node_page_state_pages(pgdat, item);
1029 }
1030 #endif
1031 
1032 #ifdef CONFIG_COMPACTION
1033 
1034 struct contig_page_info {
1035 	unsigned long free_pages;
1036 	unsigned long free_blocks_total;
1037 	unsigned long free_blocks_suitable;
1038 };
1039 
1040 /*
1041  * Calculate the number of free pages in a zone, how many contiguous
1042  * pages are free and how many are large enough to satisfy an allocation of
1043  * the target size. Note that this function makes no attempt to estimate
1044  * how many suitable free blocks there *might* be if MOVABLE pages were
1045  * migrated. Calculating that is possible, but expensive and can be
1046  * figured out from userspace
1047  */
fill_contig_page_info(struct zone * zone,unsigned int suitable_order,struct contig_page_info * info)1048 static void fill_contig_page_info(struct zone *zone,
1049 				unsigned int suitable_order,
1050 				struct contig_page_info *info)
1051 {
1052 	unsigned int order;
1053 
1054 	info->free_pages = 0;
1055 	info->free_blocks_total = 0;
1056 	info->free_blocks_suitable = 0;
1057 
1058 	for (order = 0; order < NR_PAGE_ORDERS; order++) {
1059 		unsigned long blocks;
1060 
1061 		/*
1062 		 * Count number of free blocks.
1063 		 *
1064 		 * Access to nr_free is lockless as nr_free is used only for
1065 		 * diagnostic purposes. Use data_race to avoid KCSAN warning.
1066 		 */
1067 		blocks = data_race(zone->free_area[order].nr_free);
1068 		info->free_blocks_total += blocks;
1069 
1070 		/* Count free base pages */
1071 		info->free_pages += blocks << order;
1072 
1073 		/* Count the suitable free blocks */
1074 		if (order >= suitable_order)
1075 			info->free_blocks_suitable += blocks <<
1076 						(order - suitable_order);
1077 	}
1078 }
1079 
1080 /*
1081  * A fragmentation index only makes sense if an allocation of a requested
1082  * size would fail. If that is true, the fragmentation index indicates
1083  * whether external fragmentation or a lack of memory was the problem.
1084  * The value can be used to determine if page reclaim or compaction
1085  * should be used
1086  */
__fragmentation_index(unsigned int order,struct contig_page_info * info)1087 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1088 {
1089 	unsigned long requested = 1UL << order;
1090 
1091 	if (WARN_ON_ONCE(order > MAX_ORDER))
1092 		return 0;
1093 
1094 	if (!info->free_blocks_total)
1095 		return 0;
1096 
1097 	/* Fragmentation index only makes sense when a request would fail */
1098 	if (info->free_blocks_suitable)
1099 		return -1000;
1100 
1101 	/*
1102 	 * Index is between 0 and 1 so return within 3 decimal places
1103 	 *
1104 	 * 0 => allocation would fail due to lack of memory
1105 	 * 1 => allocation would fail due to fragmentation
1106 	 */
1107 	return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1108 }
1109 
1110 /*
1111  * Calculates external fragmentation within a zone wrt the given order.
1112  * It is defined as the percentage of pages found in blocks of size
1113  * less than 1 << order. It returns values in range [0, 100].
1114  */
extfrag_for_order(struct zone * zone,unsigned int order)1115 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1116 {
1117 	struct contig_page_info info;
1118 
1119 	fill_contig_page_info(zone, order, &info);
1120 	if (info.free_pages == 0)
1121 		return 0;
1122 
1123 	return div_u64((info.free_pages -
1124 			(info.free_blocks_suitable << order)) * 100,
1125 			info.free_pages);
1126 }
1127 
1128 /* Same as __fragmentation index but allocs contig_page_info on stack */
fragmentation_index(struct zone * zone,unsigned int order)1129 int fragmentation_index(struct zone *zone, unsigned int order)
1130 {
1131 	struct contig_page_info info;
1132 
1133 	fill_contig_page_info(zone, order, &info);
1134 	return __fragmentation_index(order, &info);
1135 }
1136 #endif
1137 
1138 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1139     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1140 #ifdef CONFIG_ZONE_DMA
1141 #define TEXT_FOR_DMA(xx) xx "_dma",
1142 #else
1143 #define TEXT_FOR_DMA(xx)
1144 #endif
1145 
1146 #ifdef CONFIG_ZONE_DMA32
1147 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1148 #else
1149 #define TEXT_FOR_DMA32(xx)
1150 #endif
1151 
1152 #ifdef CONFIG_HIGHMEM
1153 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1154 #else
1155 #define TEXT_FOR_HIGHMEM(xx)
1156 #endif
1157 
1158 #ifdef CONFIG_ZONE_DEVICE
1159 #define TEXT_FOR_DEVICE(xx) xx "_device",
1160 #else
1161 #define TEXT_FOR_DEVICE(xx)
1162 #endif
1163 
1164 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1165 					TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1166 					TEXT_FOR_DEVICE(xx)
1167 
1168 const char * const vmstat_text[] = {
1169 	/* enum zone_stat_item counters */
1170 	"nr_free_pages",
1171 	"nr_zone_inactive_anon",
1172 	"nr_zone_active_anon",
1173 	"nr_zone_inactive_file",
1174 	"nr_zone_active_file",
1175 	"nr_zone_unevictable",
1176 	"nr_zone_write_pending",
1177 	"nr_mlock",
1178 	"nr_bounce",
1179 #if IS_ENABLED(CONFIG_ZSMALLOC)
1180 	"nr_zspages",
1181 #endif
1182 	"nr_free_cma",
1183 #ifdef CONFIG_UNACCEPTED_MEMORY
1184 	"nr_unaccepted",
1185 #endif
1186 
1187 	/* enum numa_stat_item counters */
1188 #ifdef CONFIG_NUMA
1189 	"numa_hit",
1190 	"numa_miss",
1191 	"numa_foreign",
1192 	"numa_interleave",
1193 	"numa_local",
1194 	"numa_other",
1195 #endif
1196 
1197 	/* enum node_stat_item counters */
1198 	"nr_inactive_anon",
1199 	"nr_active_anon",
1200 	"nr_inactive_file",
1201 	"nr_active_file",
1202 	"nr_unevictable",
1203 	"nr_slab_reclaimable",
1204 	"nr_slab_unreclaimable",
1205 	"nr_isolated_anon",
1206 	"nr_isolated_file",
1207 	"workingset_nodes",
1208 	"workingset_refault_anon",
1209 	"workingset_refault_file",
1210 	"workingset_activate_anon",
1211 	"workingset_activate_file",
1212 	"workingset_restore_anon",
1213 	"workingset_restore_file",
1214 	"workingset_nodereclaim",
1215 	"nr_anon_pages",
1216 	"nr_mapped",
1217 	"nr_file_pages",
1218 	"nr_dirty",
1219 	"nr_writeback",
1220 	"nr_writeback_temp",
1221 	"nr_shmem",
1222 	"nr_shmem_hugepages",
1223 	"nr_shmem_pmdmapped",
1224 	"nr_file_hugepages",
1225 	"nr_file_pmdmapped",
1226 	"nr_anon_transparent_hugepages",
1227 	"nr_vmscan_write",
1228 	"nr_vmscan_immediate_reclaim",
1229 	"nr_dirtied",
1230 	"nr_written",
1231 	"nr_throttled_written",
1232 	"nr_kernel_misc_reclaimable",
1233 	"nr_foll_pin_acquired",
1234 	"nr_foll_pin_released",
1235 	"nr_kernel_stack",
1236 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1237 	"nr_shadow_call_stack",
1238 #endif
1239 	"nr_page_table_pages",
1240 	"nr_sec_page_table_pages",
1241 #ifdef CONFIG_SWAP
1242 	"nr_swapcached",
1243 #endif
1244 #ifdef CONFIG_NUMA_BALANCING
1245 	"pgpromote_success",
1246 	"pgpromote_candidate",
1247 #endif
1248 
1249 	/* enum writeback_stat_item counters */
1250 	"nr_dirty_threshold",
1251 	"nr_dirty_background_threshold",
1252 
1253 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1254 	/* enum vm_event_item counters */
1255 	"pgpgin",
1256 	"pgpgout",
1257 	"pswpin",
1258 	"pswpout",
1259 
1260 	TEXTS_FOR_ZONES("pgalloc")
1261 	TEXTS_FOR_ZONES("allocstall")
1262 	TEXTS_FOR_ZONES("pgskip")
1263 
1264 	"pgfree",
1265 	"pgactivate",
1266 	"pgdeactivate",
1267 	"pglazyfree",
1268 
1269 	"pgfault",
1270 	"pgmajfault",
1271 	"pglazyfreed",
1272 
1273 	"pgrefill",
1274 	"pgreuse",
1275 	"pgsteal_kswapd",
1276 	"pgsteal_direct",
1277 	"pgsteal_khugepaged",
1278 	"pgdemote_kswapd",
1279 	"pgdemote_direct",
1280 	"pgdemote_khugepaged",
1281 	"pgscan_kswapd",
1282 	"pgscan_direct",
1283 	"pgscan_khugepaged",
1284 	"pgscan_direct_throttle",
1285 	"pgscan_anon",
1286 	"pgscan_file",
1287 	"pgsteal_anon",
1288 	"pgsteal_file",
1289 
1290 #ifdef CONFIG_NUMA
1291 	"zone_reclaim_failed",
1292 #endif
1293 	"pginodesteal",
1294 	"slabs_scanned",
1295 	"kswapd_inodesteal",
1296 	"kswapd_low_wmark_hit_quickly",
1297 	"kswapd_high_wmark_hit_quickly",
1298 	"pageoutrun",
1299 
1300 	"pgrotated",
1301 
1302 	"drop_pagecache",
1303 	"drop_slab",
1304 	"oom_kill",
1305 
1306 #ifdef CONFIG_NUMA_BALANCING
1307 	"numa_pte_updates",
1308 	"numa_huge_pte_updates",
1309 	"numa_hint_faults",
1310 	"numa_hint_faults_local",
1311 	"numa_pages_migrated",
1312 #endif
1313 #ifdef CONFIG_MIGRATION
1314 	"pgmigrate_success",
1315 	"pgmigrate_fail",
1316 	"thp_migration_success",
1317 	"thp_migration_fail",
1318 	"thp_migration_split",
1319 #endif
1320 #ifdef CONFIG_COMPACTION
1321 	"compact_migrate_scanned",
1322 	"compact_free_scanned",
1323 	"compact_isolated",
1324 	"compact_stall",
1325 	"compact_fail",
1326 	"compact_success",
1327 	"compact_daemon_wake",
1328 	"compact_daemon_migrate_scanned",
1329 	"compact_daemon_free_scanned",
1330 #endif
1331 
1332 #ifdef CONFIG_HUGETLB_PAGE
1333 	"htlb_buddy_alloc_success",
1334 	"htlb_buddy_alloc_fail",
1335 #endif
1336 #ifdef CONFIG_CMA
1337 	"cma_alloc_success",
1338 	"cma_alloc_fail",
1339 #endif
1340 	"unevictable_pgs_culled",
1341 	"unevictable_pgs_scanned",
1342 	"unevictable_pgs_rescued",
1343 	"unevictable_pgs_mlocked",
1344 	"unevictable_pgs_munlocked",
1345 	"unevictable_pgs_cleared",
1346 	"unevictable_pgs_stranded",
1347 
1348 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1349 	"thp_fault_alloc",
1350 	"thp_fault_fallback",
1351 	"thp_fault_fallback_charge",
1352 	"thp_collapse_alloc",
1353 	"thp_collapse_alloc_failed",
1354 	"thp_file_alloc",
1355 	"thp_file_fallback",
1356 	"thp_file_fallback_charge",
1357 	"thp_file_mapped",
1358 	"thp_split_page",
1359 	"thp_split_page_failed",
1360 	"thp_deferred_split_page",
1361 	"thp_split_pmd",
1362 	"thp_scan_exceed_none_pte",
1363 	"thp_scan_exceed_swap_pte",
1364 	"thp_scan_exceed_share_pte",
1365 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1366 	"thp_split_pud",
1367 #endif
1368 	"thp_zero_page_alloc",
1369 	"thp_zero_page_alloc_failed",
1370 	"thp_swpout",
1371 	"thp_swpout_fallback",
1372 #endif
1373 #ifdef CONFIG_MEMORY_BALLOON
1374 	"balloon_inflate",
1375 	"balloon_deflate",
1376 #ifdef CONFIG_BALLOON_COMPACTION
1377 	"balloon_migrate",
1378 #endif
1379 #endif /* CONFIG_MEMORY_BALLOON */
1380 #ifdef CONFIG_DEBUG_TLBFLUSH
1381 	"nr_tlb_remote_flush",
1382 	"nr_tlb_remote_flush_received",
1383 	"nr_tlb_local_flush_all",
1384 	"nr_tlb_local_flush_one",
1385 #endif /* CONFIG_DEBUG_TLBFLUSH */
1386 
1387 #ifdef CONFIG_SWAP
1388 	"swap_ra",
1389 	"swap_ra_hit",
1390 #ifdef CONFIG_KSM
1391 	"ksm_swpin_copy",
1392 #endif
1393 #endif
1394 #ifdef CONFIG_KSM
1395 	"cow_ksm",
1396 #endif
1397 #ifdef CONFIG_ZSWAP
1398 	"zswpin",
1399 	"zswpout",
1400 #endif
1401 #ifdef CONFIG_X86
1402 	"direct_map_level2_splits",
1403 	"direct_map_level3_splits",
1404 #endif
1405 #ifdef CONFIG_PER_VMA_LOCK_STATS
1406 	"vma_lock_success",
1407 	"vma_lock_abort",
1408 	"vma_lock_retry",
1409 	"vma_lock_miss",
1410 #endif
1411 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1412 };
1413 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1414 
1415 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1416      defined(CONFIG_PROC_FS)
frag_start(struct seq_file * m,loff_t * pos)1417 static void *frag_start(struct seq_file *m, loff_t *pos)
1418 {
1419 	pg_data_t *pgdat;
1420 	loff_t node = *pos;
1421 
1422 	for (pgdat = first_online_pgdat();
1423 	     pgdat && node;
1424 	     pgdat = next_online_pgdat(pgdat))
1425 		--node;
1426 
1427 	return pgdat;
1428 }
1429 
frag_next(struct seq_file * m,void * arg,loff_t * pos)1430 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1431 {
1432 	pg_data_t *pgdat = (pg_data_t *)arg;
1433 
1434 	(*pos)++;
1435 	return next_online_pgdat(pgdat);
1436 }
1437 
frag_stop(struct seq_file * m,void * arg)1438 static void frag_stop(struct seq_file *m, void *arg)
1439 {
1440 }
1441 
1442 /*
1443  * Walk zones in a node and print using a callback.
1444  * If @assert_populated is true, only use callback for zones that are populated.
1445  */
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 *))1446 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1447 		bool assert_populated, bool nolock,
1448 		void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1449 {
1450 	struct zone *zone;
1451 	struct zone *node_zones = pgdat->node_zones;
1452 	unsigned long flags;
1453 
1454 	for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1455 		if (assert_populated && !populated_zone(zone))
1456 			continue;
1457 
1458 		if (!nolock)
1459 			spin_lock_irqsave(&zone->lock, flags);
1460 		print(m, pgdat, zone);
1461 		if (!nolock)
1462 			spin_unlock_irqrestore(&zone->lock, flags);
1463 	}
1464 }
1465 #endif
1466 
1467 #ifdef CONFIG_PROC_FS
frag_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1468 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1469 						struct zone *zone)
1470 {
1471 	int order;
1472 
1473 	seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1474 	for (order = 0; order < NR_PAGE_ORDERS; ++order)
1475 		/*
1476 		 * Access to nr_free is lockless as nr_free is used only for
1477 		 * printing purposes. Use data_race to avoid KCSAN warning.
1478 		 */
1479 		seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1480 	seq_putc(m, '\n');
1481 }
1482 
1483 /*
1484  * This walks the free areas for each zone.
1485  */
frag_show(struct seq_file * m,void * arg)1486 static int frag_show(struct seq_file *m, void *arg)
1487 {
1488 	pg_data_t *pgdat = (pg_data_t *)arg;
1489 	walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1490 	return 0;
1491 }
1492 
pagetypeinfo_showfree_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1493 static void pagetypeinfo_showfree_print(struct seq_file *m,
1494 					pg_data_t *pgdat, struct zone *zone)
1495 {
1496 	int order, mtype;
1497 
1498 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1499 		seq_printf(m, "Node %4d, zone %8s, type %12s ",
1500 					pgdat->node_id,
1501 					zone->name,
1502 					migratetype_names[mtype]);
1503 		for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1504 			unsigned long freecount = 0;
1505 			struct free_area *area;
1506 			struct list_head *curr;
1507 			bool overflow = false;
1508 
1509 			area = &(zone->free_area[order]);
1510 
1511 			list_for_each(curr, &area->free_list[mtype]) {
1512 				/*
1513 				 * Cap the free_list iteration because it might
1514 				 * be really large and we are under a spinlock
1515 				 * so a long time spent here could trigger a
1516 				 * hard lockup detector. Anyway this is a
1517 				 * debugging tool so knowing there is a handful
1518 				 * of pages of this order should be more than
1519 				 * sufficient.
1520 				 */
1521 				if (++freecount >= 100000) {
1522 					overflow = true;
1523 					break;
1524 				}
1525 			}
1526 			seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1527 			spin_unlock_irq(&zone->lock);
1528 			cond_resched();
1529 			spin_lock_irq(&zone->lock);
1530 		}
1531 		seq_putc(m, '\n');
1532 	}
1533 }
1534 
1535 /* Print out the free pages at each order for each migatetype */
pagetypeinfo_showfree(struct seq_file * m,void * arg)1536 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1537 {
1538 	int order;
1539 	pg_data_t *pgdat = (pg_data_t *)arg;
1540 
1541 	/* Print header */
1542 	seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1543 	for (order = 0; order < NR_PAGE_ORDERS; ++order)
1544 		seq_printf(m, "%6d ", order);
1545 	seq_putc(m, '\n');
1546 
1547 	walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1548 }
1549 
pagetypeinfo_showblockcount_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1550 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1551 					pg_data_t *pgdat, struct zone *zone)
1552 {
1553 	int mtype;
1554 	unsigned long pfn;
1555 	unsigned long start_pfn = zone->zone_start_pfn;
1556 	unsigned long end_pfn = zone_end_pfn(zone);
1557 	unsigned long count[MIGRATE_TYPES] = { 0, };
1558 
1559 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1560 		struct page *page;
1561 
1562 		page = pfn_to_online_page(pfn);
1563 		if (!page)
1564 			continue;
1565 
1566 		if (page_zone(page) != zone)
1567 			continue;
1568 
1569 		mtype = get_pageblock_migratetype(page);
1570 
1571 		if (mtype < MIGRATE_TYPES)
1572 			count[mtype]++;
1573 	}
1574 
1575 	/* Print counts */
1576 	seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1577 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1578 		seq_printf(m, "%12lu ", count[mtype]);
1579 	seq_putc(m, '\n');
1580 }
1581 
1582 /* Print out the number of pageblocks for each migratetype */
pagetypeinfo_showblockcount(struct seq_file * m,void * arg)1583 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1584 {
1585 	int mtype;
1586 	pg_data_t *pgdat = (pg_data_t *)arg;
1587 
1588 	seq_printf(m, "\n%-23s", "Number of blocks type ");
1589 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1590 		seq_printf(m, "%12s ", migratetype_names[mtype]);
1591 	seq_putc(m, '\n');
1592 	walk_zones_in_node(m, pgdat, true, false,
1593 		pagetypeinfo_showblockcount_print);
1594 }
1595 
1596 /*
1597  * Print out the number of pageblocks for each migratetype that contain pages
1598  * of other types. This gives an indication of how well fallbacks are being
1599  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1600  * to determine what is going on
1601  */
pagetypeinfo_showmixedcount(struct seq_file * m,pg_data_t * pgdat)1602 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1603 {
1604 #ifdef CONFIG_PAGE_OWNER
1605 	int mtype;
1606 
1607 	if (!static_branch_unlikely(&page_owner_inited))
1608 		return;
1609 
1610 	drain_all_pages(NULL);
1611 
1612 	seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1613 	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1614 		seq_printf(m, "%12s ", migratetype_names[mtype]);
1615 	seq_putc(m, '\n');
1616 
1617 	walk_zones_in_node(m, pgdat, true, true,
1618 		pagetypeinfo_showmixedcount_print);
1619 #endif /* CONFIG_PAGE_OWNER */
1620 }
1621 
1622 /*
1623  * This prints out statistics in relation to grouping pages by mobility.
1624  * It is expensive to collect so do not constantly read the file.
1625  */
pagetypeinfo_show(struct seq_file * m,void * arg)1626 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1627 {
1628 	pg_data_t *pgdat = (pg_data_t *)arg;
1629 
1630 	/* check memoryless node */
1631 	if (!node_state(pgdat->node_id, N_MEMORY))
1632 		return 0;
1633 
1634 	seq_printf(m, "Page block order: %d\n", pageblock_order);
1635 	seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1636 	seq_putc(m, '\n');
1637 	pagetypeinfo_showfree(m, pgdat);
1638 	pagetypeinfo_showblockcount(m, pgdat);
1639 	pagetypeinfo_showmixedcount(m, pgdat);
1640 
1641 	return 0;
1642 }
1643 
1644 static const struct seq_operations fragmentation_op = {
1645 	.start	= frag_start,
1646 	.next	= frag_next,
1647 	.stop	= frag_stop,
1648 	.show	= frag_show,
1649 };
1650 
1651 static const struct seq_operations pagetypeinfo_op = {
1652 	.start	= frag_start,
1653 	.next	= frag_next,
1654 	.stop	= frag_stop,
1655 	.show	= pagetypeinfo_show,
1656 };
1657 
is_zone_first_populated(pg_data_t * pgdat,struct zone * zone)1658 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1659 {
1660 	int zid;
1661 
1662 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1663 		struct zone *compare = &pgdat->node_zones[zid];
1664 
1665 		if (populated_zone(compare))
1666 			return zone == compare;
1667 	}
1668 
1669 	return false;
1670 }
1671 
zoneinfo_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)1672 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1673 							struct zone *zone)
1674 {
1675 	int i;
1676 	seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1677 	if (is_zone_first_populated(pgdat, zone)) {
1678 		seq_printf(m, "\n  per-node stats");
1679 		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1680 			unsigned long pages = node_page_state_pages(pgdat, i);
1681 
1682 			if (vmstat_item_print_in_thp(i))
1683 				pages /= HPAGE_PMD_NR;
1684 			seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1685 				   pages);
1686 		}
1687 	}
1688 	seq_printf(m,
1689 		   "\n  pages free     %lu"
1690 		   "\n        boost    %lu"
1691 		   "\n        min      %lu"
1692 		   "\n        low      %lu"
1693 		   "\n        high     %lu"
1694 		   "\n        spanned  %lu"
1695 		   "\n        present  %lu"
1696 		   "\n        managed  %lu"
1697 		   "\n        cma      %lu",
1698 		   zone_page_state(zone, NR_FREE_PAGES),
1699 		   zone->watermark_boost,
1700 		   min_wmark_pages(zone),
1701 		   low_wmark_pages(zone),
1702 		   high_wmark_pages(zone),
1703 		   zone->spanned_pages,
1704 		   zone->present_pages,
1705 		   zone_managed_pages(zone),
1706 		   zone_cma_pages(zone));
1707 
1708 	seq_printf(m,
1709 		   "\n        protection: (%ld",
1710 		   zone->lowmem_reserve[0]);
1711 	for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1712 		seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1713 	seq_putc(m, ')');
1714 
1715 	/* If unpopulated, no other information is useful */
1716 	if (!populated_zone(zone)) {
1717 		seq_putc(m, '\n');
1718 		return;
1719 	}
1720 
1721 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1722 		seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1723 			   zone_page_state(zone, i));
1724 
1725 #ifdef CONFIG_NUMA
1726 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1727 		seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1728 			   zone_numa_event_state(zone, i));
1729 #endif
1730 
1731 	seq_printf(m, "\n  pagesets");
1732 	for_each_online_cpu(i) {
1733 		struct per_cpu_pages *pcp;
1734 		struct per_cpu_zonestat __maybe_unused *pzstats;
1735 
1736 		pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1737 		seq_printf(m,
1738 			   "\n    cpu: %i"
1739 			   "\n              count: %i"
1740 			   "\n              high:  %i"
1741 			   "\n              batch: %i",
1742 			   i,
1743 			   pcp->count,
1744 			   pcp->high,
1745 			   pcp->batch);
1746 #ifdef CONFIG_SMP
1747 		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1748 		seq_printf(m, "\n  vm stats threshold: %d",
1749 				pzstats->stat_threshold);
1750 #endif
1751 	}
1752 	seq_printf(m,
1753 		   "\n  node_unreclaimable:  %u"
1754 		   "\n  start_pfn:           %lu",
1755 		   pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1756 		   zone->zone_start_pfn);
1757 	seq_putc(m, '\n');
1758 }
1759 
1760 /*
1761  * Output information about zones in @pgdat.  All zones are printed regardless
1762  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1763  * set of all zones and userspace would not be aware of such zones if they are
1764  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1765  */
zoneinfo_show(struct seq_file * m,void * arg)1766 static int zoneinfo_show(struct seq_file *m, void *arg)
1767 {
1768 	pg_data_t *pgdat = (pg_data_t *)arg;
1769 	walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1770 	return 0;
1771 }
1772 
1773 static const struct seq_operations zoneinfo_op = {
1774 	.start	= frag_start, /* iterate over all zones. The same as in
1775 			       * fragmentation. */
1776 	.next	= frag_next,
1777 	.stop	= frag_stop,
1778 	.show	= zoneinfo_show,
1779 };
1780 
1781 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1782 			 NR_VM_NUMA_EVENT_ITEMS + \
1783 			 NR_VM_NODE_STAT_ITEMS + \
1784 			 NR_VM_WRITEBACK_STAT_ITEMS + \
1785 			 (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1786 			  NR_VM_EVENT_ITEMS : 0))
1787 
vmstat_start(struct seq_file * m,loff_t * pos)1788 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1789 {
1790 	unsigned long *v;
1791 	int i;
1792 
1793 	if (*pos >= NR_VMSTAT_ITEMS)
1794 		return NULL;
1795 
1796 	BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1797 	fold_vm_numa_events();
1798 	v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1799 	m->private = v;
1800 	if (!v)
1801 		return ERR_PTR(-ENOMEM);
1802 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1803 		v[i] = global_zone_page_state(i);
1804 	v += NR_VM_ZONE_STAT_ITEMS;
1805 
1806 #ifdef CONFIG_NUMA
1807 	for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1808 		v[i] = global_numa_event_state(i);
1809 	v += NR_VM_NUMA_EVENT_ITEMS;
1810 #endif
1811 
1812 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1813 		v[i] = global_node_page_state_pages(i);
1814 		if (vmstat_item_print_in_thp(i))
1815 			v[i] /= HPAGE_PMD_NR;
1816 	}
1817 	v += NR_VM_NODE_STAT_ITEMS;
1818 
1819 	global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1820 			    v + NR_DIRTY_THRESHOLD);
1821 	v += NR_VM_WRITEBACK_STAT_ITEMS;
1822 
1823 #ifdef CONFIG_VM_EVENT_COUNTERS
1824 	all_vm_events(v);
1825 	v[PGPGIN] /= 2;		/* sectors -> kbytes */
1826 	v[PGPGOUT] /= 2;
1827 #endif
1828 	return (unsigned long *)m->private + *pos;
1829 }
1830 
vmstat_next(struct seq_file * m,void * arg,loff_t * pos)1831 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1832 {
1833 	(*pos)++;
1834 	if (*pos >= NR_VMSTAT_ITEMS)
1835 		return NULL;
1836 	return (unsigned long *)m->private + *pos;
1837 }
1838 
vmstat_show(struct seq_file * m,void * arg)1839 static int vmstat_show(struct seq_file *m, void *arg)
1840 {
1841 	unsigned long *l = arg;
1842 	unsigned long off = l - (unsigned long *)m->private;
1843 
1844 	seq_puts(m, vmstat_text[off]);
1845 	seq_put_decimal_ull(m, " ", *l);
1846 	seq_putc(m, '\n');
1847 
1848 	if (off == NR_VMSTAT_ITEMS - 1) {
1849 		/*
1850 		 * We've come to the end - add any deprecated counters to avoid
1851 		 * breaking userspace which might depend on them being present.
1852 		 */
1853 		seq_puts(m, "nr_unstable 0\n");
1854 	}
1855 	return 0;
1856 }
1857 
vmstat_stop(struct seq_file * m,void * arg)1858 static void vmstat_stop(struct seq_file *m, void *arg)
1859 {
1860 	kfree(m->private);
1861 	m->private = NULL;
1862 }
1863 
1864 static const struct seq_operations vmstat_op = {
1865 	.start	= vmstat_start,
1866 	.next	= vmstat_next,
1867 	.stop	= vmstat_stop,
1868 	.show	= vmstat_show,
1869 };
1870 #endif /* CONFIG_PROC_FS */
1871 
1872 #ifdef CONFIG_SMP
1873 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1874 int sysctl_stat_interval __read_mostly = HZ;
1875 
1876 #ifdef CONFIG_PROC_FS
refresh_vm_stats(struct work_struct * work)1877 static void refresh_vm_stats(struct work_struct *work)
1878 {
1879 	refresh_cpu_vm_stats(true);
1880 }
1881 
vmstat_refresh(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)1882 int vmstat_refresh(struct ctl_table *table, int write,
1883 		   void *buffer, size_t *lenp, loff_t *ppos)
1884 {
1885 	long val;
1886 	int err;
1887 	int i;
1888 
1889 	/*
1890 	 * The regular update, every sysctl_stat_interval, may come later
1891 	 * than expected: leaving a significant amount in per_cpu buckets.
1892 	 * This is particularly misleading when checking a quantity of HUGE
1893 	 * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1894 	 * which can equally be echo'ed to or cat'ted from (by root),
1895 	 * can be used to update the stats just before reading them.
1896 	 *
1897 	 * Oh, and since global_zone_page_state() etc. are so careful to hide
1898 	 * transiently negative values, report an error here if any of
1899 	 * the stats is negative, so we know to go looking for imbalance.
1900 	 */
1901 	err = schedule_on_each_cpu(refresh_vm_stats);
1902 	if (err)
1903 		return err;
1904 	for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1905 		/*
1906 		 * Skip checking stats known to go negative occasionally.
1907 		 */
1908 		switch (i) {
1909 		case NR_ZONE_WRITE_PENDING:
1910 		case NR_FREE_CMA_PAGES:
1911 			continue;
1912 		}
1913 		val = atomic_long_read(&vm_zone_stat[i]);
1914 		if (val < 0) {
1915 			pr_warn("%s: %s %ld\n",
1916 				__func__, zone_stat_name(i), val);
1917 		}
1918 	}
1919 	for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1920 		/*
1921 		 * Skip checking stats known to go negative occasionally.
1922 		 */
1923 		switch (i) {
1924 		case NR_WRITEBACK:
1925 			continue;
1926 		}
1927 		val = atomic_long_read(&vm_node_stat[i]);
1928 		if (val < 0) {
1929 			pr_warn("%s: %s %ld\n",
1930 				__func__, node_stat_name(i), val);
1931 		}
1932 	}
1933 	if (write)
1934 		*ppos += *lenp;
1935 	else
1936 		*lenp = 0;
1937 	return 0;
1938 }
1939 #endif /* CONFIG_PROC_FS */
1940 
vmstat_update(struct work_struct * w)1941 static void vmstat_update(struct work_struct *w)
1942 {
1943 	if (refresh_cpu_vm_stats(true)) {
1944 		/*
1945 		 * Counters were updated so we expect more updates
1946 		 * to occur in the future. Keep on running the
1947 		 * update worker thread.
1948 		 */
1949 		queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1950 				this_cpu_ptr(&vmstat_work),
1951 				round_jiffies_relative(sysctl_stat_interval));
1952 	}
1953 }
1954 
1955 /*
1956  * Check if the diffs for a certain cpu indicate that
1957  * an update is needed.
1958  */
need_update(int cpu)1959 static bool need_update(int cpu)
1960 {
1961 	pg_data_t *last_pgdat = NULL;
1962 	struct zone *zone;
1963 
1964 	for_each_populated_zone(zone) {
1965 		struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1966 		struct per_cpu_nodestat *n;
1967 
1968 		/*
1969 		 * The fast way of checking if there are any vmstat diffs.
1970 		 */
1971 		if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
1972 			return true;
1973 
1974 		if (last_pgdat == zone->zone_pgdat)
1975 			continue;
1976 		last_pgdat = zone->zone_pgdat;
1977 		n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1978 		if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
1979 			return true;
1980 	}
1981 	return false;
1982 }
1983 
1984 /*
1985  * Switch off vmstat processing and then fold all the remaining differentials
1986  * until the diffs stay at zero. The function is used by NOHZ and can only be
1987  * invoked when tick processing is not active.
1988  */
quiet_vmstat(void)1989 void quiet_vmstat(void)
1990 {
1991 	if (system_state != SYSTEM_RUNNING)
1992 		return;
1993 
1994 	if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1995 		return;
1996 
1997 	if (!need_update(smp_processor_id()))
1998 		return;
1999 
2000 	/*
2001 	 * Just refresh counters and do not care about the pending delayed
2002 	 * vmstat_update. It doesn't fire that often to matter and canceling
2003 	 * it would be too expensive from this path.
2004 	 * vmstat_shepherd will take care about that for us.
2005 	 */
2006 	refresh_cpu_vm_stats(false);
2007 }
2008 
2009 /*
2010  * Shepherd worker thread that checks the
2011  * differentials of processors that have their worker
2012  * threads for vm statistics updates disabled because of
2013  * inactivity.
2014  */
2015 static void vmstat_shepherd(struct work_struct *w);
2016 
2017 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2018 
vmstat_shepherd(struct work_struct * w)2019 static void vmstat_shepherd(struct work_struct *w)
2020 {
2021 	int cpu;
2022 
2023 	cpus_read_lock();
2024 	/* Check processors whose vmstat worker threads have been disabled */
2025 	for_each_online_cpu(cpu) {
2026 		struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2027 
2028 		/*
2029 		 * In kernel users of vmstat counters either require the precise value and
2030 		 * they are using zone_page_state_snapshot interface or they can live with
2031 		 * an imprecision as the regular flushing can happen at arbitrary time and
2032 		 * cumulative error can grow (see calculate_normal_threshold).
2033 		 *
2034 		 * From that POV the regular flushing can be postponed for CPUs that have
2035 		 * been isolated from the kernel interference without critical
2036 		 * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2037 		 * for all isolated CPUs to avoid interference with the isolated workload.
2038 		 */
2039 		if (cpu_is_isolated(cpu))
2040 			continue;
2041 
2042 		if (!delayed_work_pending(dw) && need_update(cpu))
2043 			queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2044 
2045 		cond_resched();
2046 	}
2047 	cpus_read_unlock();
2048 
2049 	schedule_delayed_work(&shepherd,
2050 		round_jiffies_relative(sysctl_stat_interval));
2051 }
2052 
start_shepherd_timer(void)2053 static void __init start_shepherd_timer(void)
2054 {
2055 	int cpu;
2056 
2057 	for_each_possible_cpu(cpu)
2058 		INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2059 			vmstat_update);
2060 
2061 	schedule_delayed_work(&shepherd,
2062 		round_jiffies_relative(sysctl_stat_interval));
2063 }
2064 
init_cpu_node_state(void)2065 static void __init init_cpu_node_state(void)
2066 {
2067 	int node;
2068 
2069 	for_each_online_node(node) {
2070 		if (!cpumask_empty(cpumask_of_node(node)))
2071 			node_set_state(node, N_CPU);
2072 	}
2073 }
2074 
vmstat_cpu_online(unsigned int cpu)2075 static int vmstat_cpu_online(unsigned int cpu)
2076 {
2077 	refresh_zone_stat_thresholds();
2078 
2079 	if (!node_state(cpu_to_node(cpu), N_CPU)) {
2080 		node_set_state(cpu_to_node(cpu), N_CPU);
2081 	}
2082 
2083 	return 0;
2084 }
2085 
vmstat_cpu_down_prep(unsigned int cpu)2086 static int vmstat_cpu_down_prep(unsigned int cpu)
2087 {
2088 	cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2089 	return 0;
2090 }
2091 
vmstat_cpu_dead(unsigned int cpu)2092 static int vmstat_cpu_dead(unsigned int cpu)
2093 {
2094 	const struct cpumask *node_cpus;
2095 	int node;
2096 
2097 	node = cpu_to_node(cpu);
2098 
2099 	refresh_zone_stat_thresholds();
2100 	node_cpus = cpumask_of_node(node);
2101 	if (!cpumask_empty(node_cpus))
2102 		return 0;
2103 
2104 	node_clear_state(node, N_CPU);
2105 
2106 	return 0;
2107 }
2108 
2109 #endif
2110 
2111 struct workqueue_struct *mm_percpu_wq;
2112 
init_mm_internals(void)2113 void __init init_mm_internals(void)
2114 {
2115 	int ret __maybe_unused;
2116 
2117 	mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2118 
2119 #ifdef CONFIG_SMP
2120 	ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2121 					NULL, vmstat_cpu_dead);
2122 	if (ret < 0)
2123 		pr_err("vmstat: failed to register 'dead' hotplug state\n");
2124 
2125 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2126 					vmstat_cpu_online,
2127 					vmstat_cpu_down_prep);
2128 	if (ret < 0)
2129 		pr_err("vmstat: failed to register 'online' hotplug state\n");
2130 
2131 	cpus_read_lock();
2132 	init_cpu_node_state();
2133 	cpus_read_unlock();
2134 
2135 	start_shepherd_timer();
2136 #endif
2137 #ifdef CONFIG_PROC_FS
2138 	proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2139 	proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2140 	proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2141 	proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2142 #endif
2143 }
2144 
2145 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2146 
2147 /*
2148  * Return an index indicating how much of the available free memory is
2149  * unusable for an allocation of the requested size.
2150  */
unusable_free_index(unsigned int order,struct contig_page_info * info)2151 static int unusable_free_index(unsigned int order,
2152 				struct contig_page_info *info)
2153 {
2154 	/* No free memory is interpreted as all free memory is unusable */
2155 	if (info->free_pages == 0)
2156 		return 1000;
2157 
2158 	/*
2159 	 * Index should be a value between 0 and 1. Return a value to 3
2160 	 * decimal places.
2161 	 *
2162 	 * 0 => no fragmentation
2163 	 * 1 => high fragmentation
2164 	 */
2165 	return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2166 
2167 }
2168 
unusable_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)2169 static void unusable_show_print(struct seq_file *m,
2170 					pg_data_t *pgdat, struct zone *zone)
2171 {
2172 	unsigned int order;
2173 	int index;
2174 	struct contig_page_info info;
2175 
2176 	seq_printf(m, "Node %d, zone %8s ",
2177 				pgdat->node_id,
2178 				zone->name);
2179 	for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2180 		fill_contig_page_info(zone, order, &info);
2181 		index = unusable_free_index(order, &info);
2182 		seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2183 	}
2184 
2185 	seq_putc(m, '\n');
2186 }
2187 
2188 /*
2189  * Display unusable free space index
2190  *
2191  * The unusable free space index measures how much of the available free
2192  * memory cannot be used to satisfy an allocation of a given size and is a
2193  * value between 0 and 1. The higher the value, the more of free memory is
2194  * unusable and by implication, the worse the external fragmentation is. This
2195  * can be expressed as a percentage by multiplying by 100.
2196  */
unusable_show(struct seq_file * m,void * arg)2197 static int unusable_show(struct seq_file *m, void *arg)
2198 {
2199 	pg_data_t *pgdat = (pg_data_t *)arg;
2200 
2201 	/* check memoryless node */
2202 	if (!node_state(pgdat->node_id, N_MEMORY))
2203 		return 0;
2204 
2205 	walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2206 
2207 	return 0;
2208 }
2209 
2210 static const struct seq_operations unusable_sops = {
2211 	.start	= frag_start,
2212 	.next	= frag_next,
2213 	.stop	= frag_stop,
2214 	.show	= unusable_show,
2215 };
2216 
2217 DEFINE_SEQ_ATTRIBUTE(unusable);
2218 
extfrag_show_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)2219 static void extfrag_show_print(struct seq_file *m,
2220 					pg_data_t *pgdat, struct zone *zone)
2221 {
2222 	unsigned int order;
2223 	int index;
2224 
2225 	/* Alloc on stack as interrupts are disabled for zone walk */
2226 	struct contig_page_info info;
2227 
2228 	seq_printf(m, "Node %d, zone %8s ",
2229 				pgdat->node_id,
2230 				zone->name);
2231 	for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2232 		fill_contig_page_info(zone, order, &info);
2233 		index = __fragmentation_index(order, &info);
2234 		seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2235 	}
2236 
2237 	seq_putc(m, '\n');
2238 }
2239 
2240 /*
2241  * Display fragmentation index for orders that allocations would fail for
2242  */
extfrag_show(struct seq_file * m,void * arg)2243 static int extfrag_show(struct seq_file *m, void *arg)
2244 {
2245 	pg_data_t *pgdat = (pg_data_t *)arg;
2246 
2247 	walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2248 
2249 	return 0;
2250 }
2251 
2252 static const struct seq_operations extfrag_sops = {
2253 	.start	= frag_start,
2254 	.next	= frag_next,
2255 	.stop	= frag_stop,
2256 	.show	= extfrag_show,
2257 };
2258 
2259 DEFINE_SEQ_ATTRIBUTE(extfrag);
2260 
extfrag_debug_init(void)2261 static int __init extfrag_debug_init(void)
2262 {
2263 	struct dentry *extfrag_debug_root;
2264 
2265 	extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2266 
2267 	debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2268 			    &unusable_fops);
2269 
2270 	debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2271 			    &extfrag_fops);
2272 
2273 	return 0;
2274 }
2275 
2276 module_init(extfrag_debug_init);
2277 #endif
2278