xref: /openbmc/linux/mm/damon/core.c (revision 31eeb6b0)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Data Access Monitor
4  *
5  * Author: SeongJae Park <sjpark@amazon.de>
6  */
7 
8 #define pr_fmt(fmt) "damon: " fmt
9 
10 #include <linux/damon.h>
11 #include <linux/delay.h>
12 #include <linux/kthread.h>
13 #include <linux/mm.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 
17 #define CREATE_TRACE_POINTS
18 #include <trace/events/damon.h>
19 
20 #ifdef CONFIG_DAMON_KUNIT_TEST
21 #undef DAMON_MIN_REGION
22 #define DAMON_MIN_REGION 1
23 #endif
24 
25 static DEFINE_MUTEX(damon_lock);
26 static int nr_running_ctxs;
27 
28 /*
29  * Construct a damon_region struct
30  *
31  * Returns the pointer to the new struct if success, or NULL otherwise
32  */
33 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
34 {
35 	struct damon_region *region;
36 
37 	region = kmalloc(sizeof(*region), GFP_KERNEL);
38 	if (!region)
39 		return NULL;
40 
41 	region->ar.start = start;
42 	region->ar.end = end;
43 	region->nr_accesses = 0;
44 	INIT_LIST_HEAD(&region->list);
45 
46 	region->age = 0;
47 	region->last_nr_accesses = 0;
48 
49 	return region;
50 }
51 
52 void damon_add_region(struct damon_region *r, struct damon_target *t)
53 {
54 	list_add_tail(&r->list, &t->regions_list);
55 	t->nr_regions++;
56 }
57 
58 static void damon_del_region(struct damon_region *r, struct damon_target *t)
59 {
60 	list_del(&r->list);
61 	t->nr_regions--;
62 }
63 
64 static void damon_free_region(struct damon_region *r)
65 {
66 	kfree(r);
67 }
68 
69 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
70 {
71 	damon_del_region(r, t);
72 	damon_free_region(r);
73 }
74 
75 struct damos *damon_new_scheme(
76 		unsigned long min_sz_region, unsigned long max_sz_region,
77 		unsigned int min_nr_accesses, unsigned int max_nr_accesses,
78 		unsigned int min_age_region, unsigned int max_age_region,
79 		enum damos_action action, struct damos_quota *quota,
80 		struct damos_watermarks *wmarks)
81 {
82 	struct damos *scheme;
83 
84 	scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
85 	if (!scheme)
86 		return NULL;
87 	scheme->min_sz_region = min_sz_region;
88 	scheme->max_sz_region = max_sz_region;
89 	scheme->min_nr_accesses = min_nr_accesses;
90 	scheme->max_nr_accesses = max_nr_accesses;
91 	scheme->min_age_region = min_age_region;
92 	scheme->max_age_region = max_age_region;
93 	scheme->action = action;
94 	scheme->stat = (struct damos_stat){};
95 	INIT_LIST_HEAD(&scheme->list);
96 
97 	scheme->quota.ms = quota->ms;
98 	scheme->quota.sz = quota->sz;
99 	scheme->quota.reset_interval = quota->reset_interval;
100 	scheme->quota.weight_sz = quota->weight_sz;
101 	scheme->quota.weight_nr_accesses = quota->weight_nr_accesses;
102 	scheme->quota.weight_age = quota->weight_age;
103 	scheme->quota.total_charged_sz = 0;
104 	scheme->quota.total_charged_ns = 0;
105 	scheme->quota.esz = 0;
106 	scheme->quota.charged_sz = 0;
107 	scheme->quota.charged_from = 0;
108 	scheme->quota.charge_target_from = NULL;
109 	scheme->quota.charge_addr_from = 0;
110 
111 	scheme->wmarks.metric = wmarks->metric;
112 	scheme->wmarks.interval = wmarks->interval;
113 	scheme->wmarks.high = wmarks->high;
114 	scheme->wmarks.mid = wmarks->mid;
115 	scheme->wmarks.low = wmarks->low;
116 	scheme->wmarks.activated = true;
117 
118 	return scheme;
119 }
120 
121 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
122 {
123 	list_add_tail(&s->list, &ctx->schemes);
124 }
125 
126 static void damon_del_scheme(struct damos *s)
127 {
128 	list_del(&s->list);
129 }
130 
131 static void damon_free_scheme(struct damos *s)
132 {
133 	kfree(s);
134 }
135 
136 void damon_destroy_scheme(struct damos *s)
137 {
138 	damon_del_scheme(s);
139 	damon_free_scheme(s);
140 }
141 
142 /*
143  * Construct a damon_target struct
144  *
145  * Returns the pointer to the new struct if success, or NULL otherwise
146  */
147 struct damon_target *damon_new_target(unsigned long id)
148 {
149 	struct damon_target *t;
150 
151 	t = kmalloc(sizeof(*t), GFP_KERNEL);
152 	if (!t)
153 		return NULL;
154 
155 	t->id = id;
156 	t->nr_regions = 0;
157 	INIT_LIST_HEAD(&t->regions_list);
158 
159 	return t;
160 }
161 
162 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
163 {
164 	list_add_tail(&t->list, &ctx->adaptive_targets);
165 }
166 
167 bool damon_targets_empty(struct damon_ctx *ctx)
168 {
169 	return list_empty(&ctx->adaptive_targets);
170 }
171 
172 static void damon_del_target(struct damon_target *t)
173 {
174 	list_del(&t->list);
175 }
176 
177 void damon_free_target(struct damon_target *t)
178 {
179 	struct damon_region *r, *next;
180 
181 	damon_for_each_region_safe(r, next, t)
182 		damon_free_region(r);
183 	kfree(t);
184 }
185 
186 void damon_destroy_target(struct damon_target *t)
187 {
188 	damon_del_target(t);
189 	damon_free_target(t);
190 }
191 
192 unsigned int damon_nr_regions(struct damon_target *t)
193 {
194 	return t->nr_regions;
195 }
196 
197 struct damon_ctx *damon_new_ctx(void)
198 {
199 	struct damon_ctx *ctx;
200 
201 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
202 	if (!ctx)
203 		return NULL;
204 
205 	ctx->sample_interval = 5 * 1000;
206 	ctx->aggr_interval = 100 * 1000;
207 	ctx->primitive_update_interval = 60 * 1000 * 1000;
208 
209 	ktime_get_coarse_ts64(&ctx->last_aggregation);
210 	ctx->last_primitive_update = ctx->last_aggregation;
211 
212 	mutex_init(&ctx->kdamond_lock);
213 
214 	ctx->min_nr_regions = 10;
215 	ctx->max_nr_regions = 1000;
216 
217 	INIT_LIST_HEAD(&ctx->adaptive_targets);
218 	INIT_LIST_HEAD(&ctx->schemes);
219 
220 	return ctx;
221 }
222 
223 static void damon_destroy_targets(struct damon_ctx *ctx)
224 {
225 	struct damon_target *t, *next_t;
226 
227 	if (ctx->primitive.cleanup) {
228 		ctx->primitive.cleanup(ctx);
229 		return;
230 	}
231 
232 	damon_for_each_target_safe(t, next_t, ctx)
233 		damon_destroy_target(t);
234 }
235 
236 void damon_destroy_ctx(struct damon_ctx *ctx)
237 {
238 	struct damos *s, *next_s;
239 
240 	damon_destroy_targets(ctx);
241 
242 	damon_for_each_scheme_safe(s, next_s, ctx)
243 		damon_destroy_scheme(s);
244 
245 	kfree(ctx);
246 }
247 
248 /**
249  * damon_set_targets() - Set monitoring targets.
250  * @ctx:	monitoring context
251  * @ids:	array of target ids
252  * @nr_ids:	number of entries in @ids
253  *
254  * This function should not be called while the kdamond is running.
255  *
256  * Return: 0 on success, negative error code otherwise.
257  */
258 int damon_set_targets(struct damon_ctx *ctx,
259 		      unsigned long *ids, ssize_t nr_ids)
260 {
261 	ssize_t i;
262 	struct damon_target *t, *next;
263 
264 	damon_destroy_targets(ctx);
265 
266 	for (i = 0; i < nr_ids; i++) {
267 		t = damon_new_target(ids[i]);
268 		if (!t) {
269 			/* The caller should do cleanup of the ids itself */
270 			damon_for_each_target_safe(t, next, ctx)
271 				damon_destroy_target(t);
272 			return -ENOMEM;
273 		}
274 		damon_add_target(ctx, t);
275 	}
276 
277 	return 0;
278 }
279 
280 /**
281  * damon_set_attrs() - Set attributes for the monitoring.
282  * @ctx:		monitoring context
283  * @sample_int:		time interval between samplings
284  * @aggr_int:		time interval between aggregations
285  * @primitive_upd_int:	time interval between monitoring primitive updates
286  * @min_nr_reg:		minimal number of regions
287  * @max_nr_reg:		maximum number of regions
288  *
289  * This function should not be called while the kdamond is running.
290  * Every time interval is in micro-seconds.
291  *
292  * Return: 0 on success, negative error code otherwise.
293  */
294 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
295 		    unsigned long aggr_int, unsigned long primitive_upd_int,
296 		    unsigned long min_nr_reg, unsigned long max_nr_reg)
297 {
298 	if (min_nr_reg < 3)
299 		return -EINVAL;
300 	if (min_nr_reg > max_nr_reg)
301 		return -EINVAL;
302 
303 	ctx->sample_interval = sample_int;
304 	ctx->aggr_interval = aggr_int;
305 	ctx->primitive_update_interval = primitive_upd_int;
306 	ctx->min_nr_regions = min_nr_reg;
307 	ctx->max_nr_regions = max_nr_reg;
308 
309 	return 0;
310 }
311 
312 /**
313  * damon_set_schemes() - Set data access monitoring based operation schemes.
314  * @ctx:	monitoring context
315  * @schemes:	array of the schemes
316  * @nr_schemes:	number of entries in @schemes
317  *
318  * This function should not be called while the kdamond of the context is
319  * running.
320  *
321  * Return: 0 if success, or negative error code otherwise.
322  */
323 int damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
324 			ssize_t nr_schemes)
325 {
326 	struct damos *s, *next;
327 	ssize_t i;
328 
329 	damon_for_each_scheme_safe(s, next, ctx)
330 		damon_destroy_scheme(s);
331 	for (i = 0; i < nr_schemes; i++)
332 		damon_add_scheme(ctx, schemes[i]);
333 	return 0;
334 }
335 
336 /**
337  * damon_nr_running_ctxs() - Return number of currently running contexts.
338  */
339 int damon_nr_running_ctxs(void)
340 {
341 	int nr_ctxs;
342 
343 	mutex_lock(&damon_lock);
344 	nr_ctxs = nr_running_ctxs;
345 	mutex_unlock(&damon_lock);
346 
347 	return nr_ctxs;
348 }
349 
350 /* Returns the size upper limit for each monitoring region */
351 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
352 {
353 	struct damon_target *t;
354 	struct damon_region *r;
355 	unsigned long sz = 0;
356 
357 	damon_for_each_target(t, ctx) {
358 		damon_for_each_region(r, t)
359 			sz += r->ar.end - r->ar.start;
360 	}
361 
362 	if (ctx->min_nr_regions)
363 		sz /= ctx->min_nr_regions;
364 	if (sz < DAMON_MIN_REGION)
365 		sz = DAMON_MIN_REGION;
366 
367 	return sz;
368 }
369 
370 static int kdamond_fn(void *data);
371 
372 /*
373  * __damon_start() - Starts monitoring with given context.
374  * @ctx:	monitoring context
375  *
376  * This function should be called while damon_lock is hold.
377  *
378  * Return: 0 on success, negative error code otherwise.
379  */
380 static int __damon_start(struct damon_ctx *ctx)
381 {
382 	int err = -EBUSY;
383 
384 	mutex_lock(&ctx->kdamond_lock);
385 	if (!ctx->kdamond) {
386 		err = 0;
387 		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
388 				nr_running_ctxs);
389 		if (IS_ERR(ctx->kdamond)) {
390 			err = PTR_ERR(ctx->kdamond);
391 			ctx->kdamond = NULL;
392 		}
393 	}
394 	mutex_unlock(&ctx->kdamond_lock);
395 
396 	return err;
397 }
398 
399 /**
400  * damon_start() - Starts the monitorings for a given group of contexts.
401  * @ctxs:	an array of the pointers for contexts to start monitoring
402  * @nr_ctxs:	size of @ctxs
403  *
404  * This function starts a group of monitoring threads for a group of monitoring
405  * contexts.  One thread per each context is created and run in parallel.  The
406  * caller should handle synchronization between the threads by itself.  If a
407  * group of threads that created by other 'damon_start()' call is currently
408  * running, this function does nothing but returns -EBUSY.
409  *
410  * Return: 0 on success, negative error code otherwise.
411  */
412 int damon_start(struct damon_ctx **ctxs, int nr_ctxs)
413 {
414 	int i;
415 	int err = 0;
416 
417 	mutex_lock(&damon_lock);
418 	if (nr_running_ctxs) {
419 		mutex_unlock(&damon_lock);
420 		return -EBUSY;
421 	}
422 
423 	for (i = 0; i < nr_ctxs; i++) {
424 		err = __damon_start(ctxs[i]);
425 		if (err)
426 			break;
427 		nr_running_ctxs++;
428 	}
429 	mutex_unlock(&damon_lock);
430 
431 	return err;
432 }
433 
434 /*
435  * __damon_stop() - Stops monitoring of given context.
436  * @ctx:	monitoring context
437  *
438  * Return: 0 on success, negative error code otherwise.
439  */
440 static int __damon_stop(struct damon_ctx *ctx)
441 {
442 	struct task_struct *tsk;
443 
444 	mutex_lock(&ctx->kdamond_lock);
445 	tsk = ctx->kdamond;
446 	if (tsk) {
447 		get_task_struct(tsk);
448 		mutex_unlock(&ctx->kdamond_lock);
449 		kthread_stop(tsk);
450 		put_task_struct(tsk);
451 		return 0;
452 	}
453 	mutex_unlock(&ctx->kdamond_lock);
454 
455 	return -EPERM;
456 }
457 
458 /**
459  * damon_stop() - Stops the monitorings for a given group of contexts.
460  * @ctxs:	an array of the pointers for contexts to stop monitoring
461  * @nr_ctxs:	size of @ctxs
462  *
463  * Return: 0 on success, negative error code otherwise.
464  */
465 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
466 {
467 	int i, err = 0;
468 
469 	for (i = 0; i < nr_ctxs; i++) {
470 		/* nr_running_ctxs is decremented in kdamond_fn */
471 		err = __damon_stop(ctxs[i]);
472 		if (err)
473 			return err;
474 	}
475 
476 	return err;
477 }
478 
479 /*
480  * damon_check_reset_time_interval() - Check if a time interval is elapsed.
481  * @baseline:	the time to check whether the interval has elapsed since
482  * @interval:	the time interval (microseconds)
483  *
484  * See whether the given time interval has passed since the given baseline
485  * time.  If so, it also updates the baseline to current time for next check.
486  *
487  * Return:	true if the time interval has passed, or false otherwise.
488  */
489 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
490 		unsigned long interval)
491 {
492 	struct timespec64 now;
493 
494 	ktime_get_coarse_ts64(&now);
495 	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
496 			interval * 1000)
497 		return false;
498 	*baseline = now;
499 	return true;
500 }
501 
502 /*
503  * Check whether it is time to flush the aggregated information
504  */
505 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
506 {
507 	return damon_check_reset_time_interval(&ctx->last_aggregation,
508 			ctx->aggr_interval);
509 }
510 
511 /*
512  * Reset the aggregated monitoring results ('nr_accesses' of each region).
513  */
514 static void kdamond_reset_aggregated(struct damon_ctx *c)
515 {
516 	struct damon_target *t;
517 	unsigned int ti = 0;	/* target's index */
518 
519 	damon_for_each_target(t, c) {
520 		struct damon_region *r;
521 
522 		damon_for_each_region(r, t) {
523 			trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
524 			r->last_nr_accesses = r->nr_accesses;
525 			r->nr_accesses = 0;
526 		}
527 		ti++;
528 	}
529 }
530 
531 static void damon_split_region_at(struct damon_ctx *ctx,
532 		struct damon_target *t, struct damon_region *r,
533 		unsigned long sz_r);
534 
535 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
536 {
537 	unsigned long sz;
538 
539 	sz = r->ar.end - r->ar.start;
540 	return s->min_sz_region <= sz && sz <= s->max_sz_region &&
541 		s->min_nr_accesses <= r->nr_accesses &&
542 		r->nr_accesses <= s->max_nr_accesses &&
543 		s->min_age_region <= r->age && r->age <= s->max_age_region;
544 }
545 
546 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
547 		struct damon_region *r, struct damos *s)
548 {
549 	bool ret = __damos_valid_target(r, s);
550 
551 	if (!ret || !s->quota.esz || !c->primitive.get_scheme_score)
552 		return ret;
553 
554 	return c->primitive.get_scheme_score(c, t, r, s) >= s->quota.min_score;
555 }
556 
557 static void damon_do_apply_schemes(struct damon_ctx *c,
558 				   struct damon_target *t,
559 				   struct damon_region *r)
560 {
561 	struct damos *s;
562 
563 	damon_for_each_scheme(s, c) {
564 		struct damos_quota *quota = &s->quota;
565 		unsigned long sz = r->ar.end - r->ar.start;
566 		struct timespec64 begin, end;
567 		unsigned long sz_applied = 0;
568 
569 		if (!s->wmarks.activated)
570 			continue;
571 
572 		/* Check the quota */
573 		if (quota->esz && quota->charged_sz >= quota->esz)
574 			continue;
575 
576 		/* Skip previously charged regions */
577 		if (quota->charge_target_from) {
578 			if (t != quota->charge_target_from)
579 				continue;
580 			if (r == damon_last_region(t)) {
581 				quota->charge_target_from = NULL;
582 				quota->charge_addr_from = 0;
583 				continue;
584 			}
585 			if (quota->charge_addr_from &&
586 					r->ar.end <= quota->charge_addr_from)
587 				continue;
588 
589 			if (quota->charge_addr_from && r->ar.start <
590 					quota->charge_addr_from) {
591 				sz = ALIGN_DOWN(quota->charge_addr_from -
592 						r->ar.start, DAMON_MIN_REGION);
593 				if (!sz) {
594 					if (r->ar.end - r->ar.start <=
595 							DAMON_MIN_REGION)
596 						continue;
597 					sz = DAMON_MIN_REGION;
598 				}
599 				damon_split_region_at(c, t, r, sz);
600 				r = damon_next_region(r);
601 				sz = r->ar.end - r->ar.start;
602 			}
603 			quota->charge_target_from = NULL;
604 			quota->charge_addr_from = 0;
605 		}
606 
607 		if (!damos_valid_target(c, t, r, s))
608 			continue;
609 
610 		/* Apply the scheme */
611 		if (c->primitive.apply_scheme) {
612 			if (quota->esz &&
613 					quota->charged_sz + sz > quota->esz) {
614 				sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
615 						DAMON_MIN_REGION);
616 				if (!sz)
617 					goto update_stat;
618 				damon_split_region_at(c, t, r, sz);
619 			}
620 			ktime_get_coarse_ts64(&begin);
621 			sz_applied = c->primitive.apply_scheme(c, t, r, s);
622 			ktime_get_coarse_ts64(&end);
623 			quota->total_charged_ns += timespec64_to_ns(&end) -
624 				timespec64_to_ns(&begin);
625 			quota->charged_sz += sz;
626 			if (quota->esz && quota->charged_sz >= quota->esz) {
627 				quota->charge_target_from = t;
628 				quota->charge_addr_from = r->ar.end + 1;
629 			}
630 		}
631 		if (s->action != DAMOS_STAT)
632 			r->age = 0;
633 
634 update_stat:
635 		s->stat.nr_tried++;
636 		s->stat.sz_tried += sz;
637 		if (sz_applied)
638 			s->stat.nr_applied++;
639 		s->stat.sz_applied += sz_applied;
640 	}
641 }
642 
643 /* Shouldn't be called if quota->ms and quota->sz are zero */
644 static void damos_set_effective_quota(struct damos_quota *quota)
645 {
646 	unsigned long throughput;
647 	unsigned long esz;
648 
649 	if (!quota->ms) {
650 		quota->esz = quota->sz;
651 		return;
652 	}
653 
654 	if (quota->total_charged_ns)
655 		throughput = quota->total_charged_sz * 1000000 /
656 			quota->total_charged_ns;
657 	else
658 		throughput = PAGE_SIZE * 1024;
659 	esz = throughput * quota->ms;
660 
661 	if (quota->sz && quota->sz < esz)
662 		esz = quota->sz;
663 	quota->esz = esz;
664 }
665 
666 static void kdamond_apply_schemes(struct damon_ctx *c)
667 {
668 	struct damon_target *t;
669 	struct damon_region *r, *next_r;
670 	struct damos *s;
671 
672 	damon_for_each_scheme(s, c) {
673 		struct damos_quota *quota = &s->quota;
674 		unsigned long cumulated_sz;
675 		unsigned int score, max_score = 0;
676 
677 		if (!s->wmarks.activated)
678 			continue;
679 
680 		if (!quota->ms && !quota->sz)
681 			continue;
682 
683 		/* New charge window starts */
684 		if (time_after_eq(jiffies, quota->charged_from +
685 					msecs_to_jiffies(
686 						quota->reset_interval))) {
687 			if (quota->esz && quota->charged_sz >= quota->esz)
688 				s->stat.qt_exceeds++;
689 			quota->total_charged_sz += quota->charged_sz;
690 			quota->charged_from = jiffies;
691 			quota->charged_sz = 0;
692 			damos_set_effective_quota(quota);
693 		}
694 
695 		if (!c->primitive.get_scheme_score)
696 			continue;
697 
698 		/* Fill up the score histogram */
699 		memset(quota->histogram, 0, sizeof(quota->histogram));
700 		damon_for_each_target(t, c) {
701 			damon_for_each_region(r, t) {
702 				if (!__damos_valid_target(r, s))
703 					continue;
704 				score = c->primitive.get_scheme_score(
705 						c, t, r, s);
706 				quota->histogram[score] +=
707 					r->ar.end - r->ar.start;
708 				if (score > max_score)
709 					max_score = score;
710 			}
711 		}
712 
713 		/* Set the min score limit */
714 		for (cumulated_sz = 0, score = max_score; ; score--) {
715 			cumulated_sz += quota->histogram[score];
716 			if (cumulated_sz >= quota->esz || !score)
717 				break;
718 		}
719 		quota->min_score = score;
720 	}
721 
722 	damon_for_each_target(t, c) {
723 		damon_for_each_region_safe(r, next_r, t)
724 			damon_do_apply_schemes(c, t, r);
725 	}
726 }
727 
728 static inline unsigned long sz_damon_region(struct damon_region *r)
729 {
730 	return r->ar.end - r->ar.start;
731 }
732 
733 /*
734  * Merge two adjacent regions into one region
735  */
736 static void damon_merge_two_regions(struct damon_target *t,
737 		struct damon_region *l, struct damon_region *r)
738 {
739 	unsigned long sz_l = sz_damon_region(l), sz_r = sz_damon_region(r);
740 
741 	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
742 			(sz_l + sz_r);
743 	l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
744 	l->ar.end = r->ar.end;
745 	damon_destroy_region(r, t);
746 }
747 
748 /*
749  * Merge adjacent regions having similar access frequencies
750  *
751  * t		target affected by this merge operation
752  * thres	'->nr_accesses' diff threshold for the merge
753  * sz_limit	size upper limit of each region
754  */
755 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
756 				   unsigned long sz_limit)
757 {
758 	struct damon_region *r, *prev = NULL, *next;
759 
760 	damon_for_each_region_safe(r, next, t) {
761 		if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
762 			r->age = 0;
763 		else
764 			r->age++;
765 
766 		if (prev && prev->ar.end == r->ar.start &&
767 		    abs(prev->nr_accesses - r->nr_accesses) <= thres &&
768 		    sz_damon_region(prev) + sz_damon_region(r) <= sz_limit)
769 			damon_merge_two_regions(t, prev, r);
770 		else
771 			prev = r;
772 	}
773 }
774 
775 /*
776  * Merge adjacent regions having similar access frequencies
777  *
778  * threshold	'->nr_accesses' diff threshold for the merge
779  * sz_limit	size upper limit of each region
780  *
781  * This function merges monitoring target regions which are adjacent and their
782  * access frequencies are similar.  This is for minimizing the monitoring
783  * overhead under the dynamically changeable access pattern.  If a merge was
784  * unnecessarily made, later 'kdamond_split_regions()' will revert it.
785  */
786 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
787 				  unsigned long sz_limit)
788 {
789 	struct damon_target *t;
790 
791 	damon_for_each_target(t, c)
792 		damon_merge_regions_of(t, threshold, sz_limit);
793 }
794 
795 /*
796  * Split a region in two
797  *
798  * r		the region to be split
799  * sz_r		size of the first sub-region that will be made
800  */
801 static void damon_split_region_at(struct damon_ctx *ctx,
802 		struct damon_target *t, struct damon_region *r,
803 		unsigned long sz_r)
804 {
805 	struct damon_region *new;
806 
807 	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
808 	if (!new)
809 		return;
810 
811 	r->ar.end = new->ar.start;
812 
813 	new->age = r->age;
814 	new->last_nr_accesses = r->last_nr_accesses;
815 
816 	damon_insert_region(new, r, damon_next_region(r), t);
817 }
818 
819 /* Split every region in the given target into 'nr_subs' regions */
820 static void damon_split_regions_of(struct damon_ctx *ctx,
821 				     struct damon_target *t, int nr_subs)
822 {
823 	struct damon_region *r, *next;
824 	unsigned long sz_region, sz_sub = 0;
825 	int i;
826 
827 	damon_for_each_region_safe(r, next, t) {
828 		sz_region = r->ar.end - r->ar.start;
829 
830 		for (i = 0; i < nr_subs - 1 &&
831 				sz_region > 2 * DAMON_MIN_REGION; i++) {
832 			/*
833 			 * Randomly select size of left sub-region to be at
834 			 * least 10 percent and at most 90% of original region
835 			 */
836 			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
837 					sz_region / 10, DAMON_MIN_REGION);
838 			/* Do not allow blank region */
839 			if (sz_sub == 0 || sz_sub >= sz_region)
840 				continue;
841 
842 			damon_split_region_at(ctx, t, r, sz_sub);
843 			sz_region = sz_sub;
844 		}
845 	}
846 }
847 
848 /*
849  * Split every target region into randomly-sized small regions
850  *
851  * This function splits every target region into random-sized small regions if
852  * current total number of the regions is equal or smaller than half of the
853  * user-specified maximum number of regions.  This is for maximizing the
854  * monitoring accuracy under the dynamically changeable access patterns.  If a
855  * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
856  * it.
857  */
858 static void kdamond_split_regions(struct damon_ctx *ctx)
859 {
860 	struct damon_target *t;
861 	unsigned int nr_regions = 0;
862 	static unsigned int last_nr_regions;
863 	int nr_subregions = 2;
864 
865 	damon_for_each_target(t, ctx)
866 		nr_regions += damon_nr_regions(t);
867 
868 	if (nr_regions > ctx->max_nr_regions / 2)
869 		return;
870 
871 	/* Maybe the middle of the region has different access frequency */
872 	if (last_nr_regions == nr_regions &&
873 			nr_regions < ctx->max_nr_regions / 3)
874 		nr_subregions = 3;
875 
876 	damon_for_each_target(t, ctx)
877 		damon_split_regions_of(ctx, t, nr_subregions);
878 
879 	last_nr_regions = nr_regions;
880 }
881 
882 /*
883  * Check whether it is time to check and apply the target monitoring regions
884  *
885  * Returns true if it is.
886  */
887 static bool kdamond_need_update_primitive(struct damon_ctx *ctx)
888 {
889 	return damon_check_reset_time_interval(&ctx->last_primitive_update,
890 			ctx->primitive_update_interval);
891 }
892 
893 /*
894  * Check whether current monitoring should be stopped
895  *
896  * The monitoring is stopped when either the user requested to stop, or all
897  * monitoring targets are invalid.
898  *
899  * Returns true if need to stop current monitoring.
900  */
901 static bool kdamond_need_stop(struct damon_ctx *ctx)
902 {
903 	struct damon_target *t;
904 
905 	if (kthread_should_stop())
906 		return true;
907 
908 	if (!ctx->primitive.target_valid)
909 		return false;
910 
911 	damon_for_each_target(t, ctx) {
912 		if (ctx->primitive.target_valid(t))
913 			return false;
914 	}
915 
916 	return true;
917 }
918 
919 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
920 {
921 	struct sysinfo i;
922 
923 	switch (metric) {
924 	case DAMOS_WMARK_FREE_MEM_RATE:
925 		si_meminfo(&i);
926 		return i.freeram * 1000 / i.totalram;
927 	default:
928 		break;
929 	}
930 	return -EINVAL;
931 }
932 
933 /*
934  * Returns zero if the scheme is active.  Else, returns time to wait for next
935  * watermark check in micro-seconds.
936  */
937 static unsigned long damos_wmark_wait_us(struct damos *scheme)
938 {
939 	unsigned long metric;
940 
941 	if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
942 		return 0;
943 
944 	metric = damos_wmark_metric_value(scheme->wmarks.metric);
945 	/* higher than high watermark or lower than low watermark */
946 	if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
947 		if (scheme->wmarks.activated)
948 			pr_debug("deactivate a scheme (%d) for %s wmark\n",
949 					scheme->action,
950 					metric > scheme->wmarks.high ?
951 					"high" : "low");
952 		scheme->wmarks.activated = false;
953 		return scheme->wmarks.interval;
954 	}
955 
956 	/* inactive and higher than middle watermark */
957 	if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
958 			!scheme->wmarks.activated)
959 		return scheme->wmarks.interval;
960 
961 	if (!scheme->wmarks.activated)
962 		pr_debug("activate a scheme (%d)\n", scheme->action);
963 	scheme->wmarks.activated = true;
964 	return 0;
965 }
966 
967 static void kdamond_usleep(unsigned long usecs)
968 {
969 	/* See Documentation/timers/timers-howto.rst for the thresholds */
970 	if (usecs > 20 * USEC_PER_MSEC)
971 		schedule_timeout_idle(usecs_to_jiffies(usecs));
972 	else
973 		usleep_idle_range(usecs, usecs + 1);
974 }
975 
976 /* Returns negative error code if it's not activated but should return */
977 static int kdamond_wait_activation(struct damon_ctx *ctx)
978 {
979 	struct damos *s;
980 	unsigned long wait_time;
981 	unsigned long min_wait_time = 0;
982 
983 	while (!kdamond_need_stop(ctx)) {
984 		damon_for_each_scheme(s, ctx) {
985 			wait_time = damos_wmark_wait_us(s);
986 			if (!min_wait_time || wait_time < min_wait_time)
987 				min_wait_time = wait_time;
988 		}
989 		if (!min_wait_time)
990 			return 0;
991 
992 		kdamond_usleep(min_wait_time);
993 	}
994 	return -EBUSY;
995 }
996 
997 /*
998  * The monitoring daemon that runs as a kernel thread
999  */
1000 static int kdamond_fn(void *data)
1001 {
1002 	struct damon_ctx *ctx = (struct damon_ctx *)data;
1003 	struct damon_target *t;
1004 	struct damon_region *r, *next;
1005 	unsigned int max_nr_accesses = 0;
1006 	unsigned long sz_limit = 0;
1007 	bool done = false;
1008 
1009 	pr_debug("kdamond (%d) starts\n", current->pid);
1010 
1011 	if (ctx->primitive.init)
1012 		ctx->primitive.init(ctx);
1013 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1014 		done = true;
1015 
1016 	sz_limit = damon_region_sz_limit(ctx);
1017 
1018 	while (!kdamond_need_stop(ctx) && !done) {
1019 		if (kdamond_wait_activation(ctx))
1020 			continue;
1021 
1022 		if (ctx->primitive.prepare_access_checks)
1023 			ctx->primitive.prepare_access_checks(ctx);
1024 		if (ctx->callback.after_sampling &&
1025 				ctx->callback.after_sampling(ctx))
1026 			done = true;
1027 
1028 		kdamond_usleep(ctx->sample_interval);
1029 
1030 		if (ctx->primitive.check_accesses)
1031 			max_nr_accesses = ctx->primitive.check_accesses(ctx);
1032 
1033 		if (kdamond_aggregate_interval_passed(ctx)) {
1034 			kdamond_merge_regions(ctx,
1035 					max_nr_accesses / 10,
1036 					sz_limit);
1037 			if (ctx->callback.after_aggregation &&
1038 					ctx->callback.after_aggregation(ctx))
1039 				done = true;
1040 			kdamond_apply_schemes(ctx);
1041 			kdamond_reset_aggregated(ctx);
1042 			kdamond_split_regions(ctx);
1043 			if (ctx->primitive.reset_aggregated)
1044 				ctx->primitive.reset_aggregated(ctx);
1045 		}
1046 
1047 		if (kdamond_need_update_primitive(ctx)) {
1048 			if (ctx->primitive.update)
1049 				ctx->primitive.update(ctx);
1050 			sz_limit = damon_region_sz_limit(ctx);
1051 		}
1052 	}
1053 	damon_for_each_target(t, ctx) {
1054 		damon_for_each_region_safe(r, next, t)
1055 			damon_destroy_region(r, t);
1056 	}
1057 
1058 	if (ctx->callback.before_terminate)
1059 		ctx->callback.before_terminate(ctx);
1060 	if (ctx->primitive.cleanup)
1061 		ctx->primitive.cleanup(ctx);
1062 
1063 	pr_debug("kdamond (%d) finishes\n", current->pid);
1064 	mutex_lock(&ctx->kdamond_lock);
1065 	ctx->kdamond = NULL;
1066 	mutex_unlock(&ctx->kdamond_lock);
1067 
1068 	mutex_lock(&damon_lock);
1069 	nr_running_ctxs--;
1070 	mutex_unlock(&damon_lock);
1071 
1072 	return 0;
1073 }
1074 
1075 #include "core-test.h"
1076