xref: /openbmc/linux/mm/damon/core.c (revision 4b33b5ff)
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 static bool running_exclusive_ctxs;
28 
29 static DEFINE_MUTEX(damon_ops_lock);
30 static struct damon_operations damon_registered_ops[NR_DAMON_OPS];
31 
32 /* Should be called under damon_ops_lock with id smaller than NR_DAMON_OPS */
33 static bool damon_registered_ops_id(enum damon_ops_id id)
34 {
35 	struct damon_operations empty_ops = {};
36 
37 	if (!memcmp(&empty_ops, &damon_registered_ops[id], sizeof(empty_ops)))
38 		return false;
39 	return true;
40 }
41 
42 /**
43  * damon_register_ops() - Register a monitoring operations set to DAMON.
44  * @ops:	monitoring operations set to register.
45  *
46  * This function registers a monitoring operations set of valid &struct
47  * damon_operations->id so that others can find and use them later.
48  *
49  * Return: 0 on success, negative error code otherwise.
50  */
51 int damon_register_ops(struct damon_operations *ops)
52 {
53 	int err = 0;
54 
55 	if (ops->id >= NR_DAMON_OPS)
56 		return -EINVAL;
57 	mutex_lock(&damon_ops_lock);
58 	/* Fail for already registered ops */
59 	if (damon_registered_ops_id(ops->id)) {
60 		err = -EINVAL;
61 		goto out;
62 	}
63 	damon_registered_ops[ops->id] = *ops;
64 out:
65 	mutex_unlock(&damon_ops_lock);
66 	return err;
67 }
68 
69 /**
70  * damon_select_ops() - Select a monitoring operations to use with the context.
71  * @ctx:	monitoring context to use the operations.
72  * @id:		id of the registered monitoring operations to select.
73  *
74  * This function finds registered monitoring operations set of @id and make
75  * @ctx to use it.
76  *
77  * Return: 0 on success, negative error code otherwise.
78  */
79 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
80 {
81 	int err = 0;
82 
83 	if (id >= NR_DAMON_OPS)
84 		return -EINVAL;
85 
86 	mutex_lock(&damon_ops_lock);
87 	if (!damon_registered_ops_id(id))
88 		err = -EINVAL;
89 	else
90 		ctx->ops = damon_registered_ops[id];
91 	mutex_unlock(&damon_ops_lock);
92 	return err;
93 }
94 
95 /*
96  * Construct a damon_region struct
97  *
98  * Returns the pointer to the new struct if success, or NULL otherwise
99  */
100 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
101 {
102 	struct damon_region *region;
103 
104 	region = kmalloc(sizeof(*region), GFP_KERNEL);
105 	if (!region)
106 		return NULL;
107 
108 	region->ar.start = start;
109 	region->ar.end = end;
110 	region->nr_accesses = 0;
111 	INIT_LIST_HEAD(&region->list);
112 
113 	region->age = 0;
114 	region->last_nr_accesses = 0;
115 
116 	return region;
117 }
118 
119 void damon_add_region(struct damon_region *r, struct damon_target *t)
120 {
121 	list_add_tail(&r->list, &t->regions_list);
122 	t->nr_regions++;
123 }
124 
125 static void damon_del_region(struct damon_region *r, struct damon_target *t)
126 {
127 	list_del(&r->list);
128 	t->nr_regions--;
129 }
130 
131 static void damon_free_region(struct damon_region *r)
132 {
133 	kfree(r);
134 }
135 
136 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
137 {
138 	damon_del_region(r, t);
139 	damon_free_region(r);
140 }
141 
142 struct damos *damon_new_scheme(
143 		unsigned long min_sz_region, unsigned long max_sz_region,
144 		unsigned int min_nr_accesses, unsigned int max_nr_accesses,
145 		unsigned int min_age_region, unsigned int max_age_region,
146 		enum damos_action action, struct damos_quota *quota,
147 		struct damos_watermarks *wmarks)
148 {
149 	struct damos *scheme;
150 
151 	scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
152 	if (!scheme)
153 		return NULL;
154 	scheme->min_sz_region = min_sz_region;
155 	scheme->max_sz_region = max_sz_region;
156 	scheme->min_nr_accesses = min_nr_accesses;
157 	scheme->max_nr_accesses = max_nr_accesses;
158 	scheme->min_age_region = min_age_region;
159 	scheme->max_age_region = max_age_region;
160 	scheme->action = action;
161 	scheme->stat = (struct damos_stat){};
162 	INIT_LIST_HEAD(&scheme->list);
163 
164 	scheme->quota.ms = quota->ms;
165 	scheme->quota.sz = quota->sz;
166 	scheme->quota.reset_interval = quota->reset_interval;
167 	scheme->quota.weight_sz = quota->weight_sz;
168 	scheme->quota.weight_nr_accesses = quota->weight_nr_accesses;
169 	scheme->quota.weight_age = quota->weight_age;
170 	scheme->quota.total_charged_sz = 0;
171 	scheme->quota.total_charged_ns = 0;
172 	scheme->quota.esz = 0;
173 	scheme->quota.charged_sz = 0;
174 	scheme->quota.charged_from = 0;
175 	scheme->quota.charge_target_from = NULL;
176 	scheme->quota.charge_addr_from = 0;
177 
178 	scheme->wmarks.metric = wmarks->metric;
179 	scheme->wmarks.interval = wmarks->interval;
180 	scheme->wmarks.high = wmarks->high;
181 	scheme->wmarks.mid = wmarks->mid;
182 	scheme->wmarks.low = wmarks->low;
183 	scheme->wmarks.activated = true;
184 
185 	return scheme;
186 }
187 
188 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
189 {
190 	list_add_tail(&s->list, &ctx->schemes);
191 }
192 
193 static void damon_del_scheme(struct damos *s)
194 {
195 	list_del(&s->list);
196 }
197 
198 static void damon_free_scheme(struct damos *s)
199 {
200 	kfree(s);
201 }
202 
203 void damon_destroy_scheme(struct damos *s)
204 {
205 	damon_del_scheme(s);
206 	damon_free_scheme(s);
207 }
208 
209 /*
210  * Construct a damon_target struct
211  *
212  * Returns the pointer to the new struct if success, or NULL otherwise
213  */
214 struct damon_target *damon_new_target(void)
215 {
216 	struct damon_target *t;
217 
218 	t = kmalloc(sizeof(*t), GFP_KERNEL);
219 	if (!t)
220 		return NULL;
221 
222 	t->pid = NULL;
223 	t->nr_regions = 0;
224 	INIT_LIST_HEAD(&t->regions_list);
225 
226 	return t;
227 }
228 
229 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
230 {
231 	list_add_tail(&t->list, &ctx->adaptive_targets);
232 }
233 
234 bool damon_targets_empty(struct damon_ctx *ctx)
235 {
236 	return list_empty(&ctx->adaptive_targets);
237 }
238 
239 static void damon_del_target(struct damon_target *t)
240 {
241 	list_del(&t->list);
242 }
243 
244 void damon_free_target(struct damon_target *t)
245 {
246 	struct damon_region *r, *next;
247 
248 	damon_for_each_region_safe(r, next, t)
249 		damon_free_region(r);
250 	kfree(t);
251 }
252 
253 void damon_destroy_target(struct damon_target *t)
254 {
255 	damon_del_target(t);
256 	damon_free_target(t);
257 }
258 
259 unsigned int damon_nr_regions(struct damon_target *t)
260 {
261 	return t->nr_regions;
262 }
263 
264 struct damon_ctx *damon_new_ctx(void)
265 {
266 	struct damon_ctx *ctx;
267 
268 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
269 	if (!ctx)
270 		return NULL;
271 
272 	ctx->sample_interval = 5 * 1000;
273 	ctx->aggr_interval = 100 * 1000;
274 	ctx->ops_update_interval = 60 * 1000 * 1000;
275 
276 	ktime_get_coarse_ts64(&ctx->last_aggregation);
277 	ctx->last_ops_update = ctx->last_aggregation;
278 
279 	mutex_init(&ctx->kdamond_lock);
280 
281 	ctx->min_nr_regions = 10;
282 	ctx->max_nr_regions = 1000;
283 
284 	INIT_LIST_HEAD(&ctx->adaptive_targets);
285 	INIT_LIST_HEAD(&ctx->schemes);
286 
287 	return ctx;
288 }
289 
290 static void damon_destroy_targets(struct damon_ctx *ctx)
291 {
292 	struct damon_target *t, *next_t;
293 
294 	if (ctx->ops.cleanup) {
295 		ctx->ops.cleanup(ctx);
296 		return;
297 	}
298 
299 	damon_for_each_target_safe(t, next_t, ctx)
300 		damon_destroy_target(t);
301 }
302 
303 void damon_destroy_ctx(struct damon_ctx *ctx)
304 {
305 	struct damos *s, *next_s;
306 
307 	damon_destroy_targets(ctx);
308 
309 	damon_for_each_scheme_safe(s, next_s, ctx)
310 		damon_destroy_scheme(s);
311 
312 	kfree(ctx);
313 }
314 
315 /**
316  * damon_set_attrs() - Set attributes for the monitoring.
317  * @ctx:		monitoring context
318  * @sample_int:		time interval between samplings
319  * @aggr_int:		time interval between aggregations
320  * @ops_upd_int:	time interval between monitoring operations updates
321  * @min_nr_reg:		minimal number of regions
322  * @max_nr_reg:		maximum number of regions
323  *
324  * This function should not be called while the kdamond is running.
325  * Every time interval is in micro-seconds.
326  *
327  * Return: 0 on success, negative error code otherwise.
328  */
329 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
330 		    unsigned long aggr_int, unsigned long ops_upd_int,
331 		    unsigned long min_nr_reg, unsigned long max_nr_reg)
332 {
333 	if (min_nr_reg < 3)
334 		return -EINVAL;
335 	if (min_nr_reg > max_nr_reg)
336 		return -EINVAL;
337 
338 	ctx->sample_interval = sample_int;
339 	ctx->aggr_interval = aggr_int;
340 	ctx->ops_update_interval = ops_upd_int;
341 	ctx->min_nr_regions = min_nr_reg;
342 	ctx->max_nr_regions = max_nr_reg;
343 
344 	return 0;
345 }
346 
347 /**
348  * damon_set_schemes() - Set data access monitoring based operation schemes.
349  * @ctx:	monitoring context
350  * @schemes:	array of the schemes
351  * @nr_schemes:	number of entries in @schemes
352  *
353  * This function should not be called while the kdamond of the context is
354  * running.
355  *
356  * Return: 0 if success, or negative error code otherwise.
357  */
358 int damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
359 			ssize_t nr_schemes)
360 {
361 	struct damos *s, *next;
362 	ssize_t i;
363 
364 	damon_for_each_scheme_safe(s, next, ctx)
365 		damon_destroy_scheme(s);
366 	for (i = 0; i < nr_schemes; i++)
367 		damon_add_scheme(ctx, schemes[i]);
368 	return 0;
369 }
370 
371 /**
372  * damon_nr_running_ctxs() - Return number of currently running contexts.
373  */
374 int damon_nr_running_ctxs(void)
375 {
376 	int nr_ctxs;
377 
378 	mutex_lock(&damon_lock);
379 	nr_ctxs = nr_running_ctxs;
380 	mutex_unlock(&damon_lock);
381 
382 	return nr_ctxs;
383 }
384 
385 /* Returns the size upper limit for each monitoring region */
386 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
387 {
388 	struct damon_target *t;
389 	struct damon_region *r;
390 	unsigned long sz = 0;
391 
392 	damon_for_each_target(t, ctx) {
393 		damon_for_each_region(r, t)
394 			sz += r->ar.end - r->ar.start;
395 	}
396 
397 	if (ctx->min_nr_regions)
398 		sz /= ctx->min_nr_regions;
399 	if (sz < DAMON_MIN_REGION)
400 		sz = DAMON_MIN_REGION;
401 
402 	return sz;
403 }
404 
405 static int kdamond_fn(void *data);
406 
407 /*
408  * __damon_start() - Starts monitoring with given context.
409  * @ctx:	monitoring context
410  *
411  * This function should be called while damon_lock is hold.
412  *
413  * Return: 0 on success, negative error code otherwise.
414  */
415 static int __damon_start(struct damon_ctx *ctx)
416 {
417 	int err = -EBUSY;
418 
419 	mutex_lock(&ctx->kdamond_lock);
420 	if (!ctx->kdamond) {
421 		err = 0;
422 		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
423 				nr_running_ctxs);
424 		if (IS_ERR(ctx->kdamond)) {
425 			err = PTR_ERR(ctx->kdamond);
426 			ctx->kdamond = NULL;
427 		}
428 	}
429 	mutex_unlock(&ctx->kdamond_lock);
430 
431 	return err;
432 }
433 
434 /**
435  * damon_start() - Starts the monitorings for a given group of contexts.
436  * @ctxs:	an array of the pointers for contexts to start monitoring
437  * @nr_ctxs:	size of @ctxs
438  * @exclusive:	exclusiveness of this contexts group
439  *
440  * This function starts a group of monitoring threads for a group of monitoring
441  * contexts.  One thread per each context is created and run in parallel.  The
442  * caller should handle synchronization between the threads by itself.  If
443  * @exclusive is true and a group of threads that created by other
444  * 'damon_start()' call is currently running, this function does nothing but
445  * returns -EBUSY.
446  *
447  * Return: 0 on success, negative error code otherwise.
448  */
449 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
450 {
451 	int i;
452 	int err = 0;
453 
454 	mutex_lock(&damon_lock);
455 	if ((exclusive && nr_running_ctxs) ||
456 			(!exclusive && running_exclusive_ctxs)) {
457 		mutex_unlock(&damon_lock);
458 		return -EBUSY;
459 	}
460 
461 	for (i = 0; i < nr_ctxs; i++) {
462 		err = __damon_start(ctxs[i]);
463 		if (err)
464 			break;
465 		nr_running_ctxs++;
466 	}
467 	if (exclusive && nr_running_ctxs)
468 		running_exclusive_ctxs = true;
469 	mutex_unlock(&damon_lock);
470 
471 	return err;
472 }
473 
474 /*
475  * __damon_stop() - Stops monitoring of a given context.
476  * @ctx:	monitoring context
477  *
478  * Return: 0 on success, negative error code otherwise.
479  */
480 static int __damon_stop(struct damon_ctx *ctx)
481 {
482 	struct task_struct *tsk;
483 
484 	mutex_lock(&ctx->kdamond_lock);
485 	tsk = ctx->kdamond;
486 	if (tsk) {
487 		get_task_struct(tsk);
488 		mutex_unlock(&ctx->kdamond_lock);
489 		kthread_stop(tsk);
490 		put_task_struct(tsk);
491 		return 0;
492 	}
493 	mutex_unlock(&ctx->kdamond_lock);
494 
495 	return -EPERM;
496 }
497 
498 /**
499  * damon_stop() - Stops the monitorings for a given group of contexts.
500  * @ctxs:	an array of the pointers for contexts to stop monitoring
501  * @nr_ctxs:	size of @ctxs
502  *
503  * Return: 0 on success, negative error code otherwise.
504  */
505 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
506 {
507 	int i, err = 0;
508 
509 	for (i = 0; i < nr_ctxs; i++) {
510 		/* nr_running_ctxs is decremented in kdamond_fn */
511 		err = __damon_stop(ctxs[i]);
512 		if (err)
513 			break;
514 	}
515 	return err;
516 }
517 
518 /*
519  * damon_check_reset_time_interval() - Check if a time interval is elapsed.
520  * @baseline:	the time to check whether the interval has elapsed since
521  * @interval:	the time interval (microseconds)
522  *
523  * See whether the given time interval has passed since the given baseline
524  * time.  If so, it also updates the baseline to current time for next check.
525  *
526  * Return:	true if the time interval has passed, or false otherwise.
527  */
528 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
529 		unsigned long interval)
530 {
531 	struct timespec64 now;
532 
533 	ktime_get_coarse_ts64(&now);
534 	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
535 			interval * 1000)
536 		return false;
537 	*baseline = now;
538 	return true;
539 }
540 
541 /*
542  * Check whether it is time to flush the aggregated information
543  */
544 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
545 {
546 	return damon_check_reset_time_interval(&ctx->last_aggregation,
547 			ctx->aggr_interval);
548 }
549 
550 /*
551  * Reset the aggregated monitoring results ('nr_accesses' of each region).
552  */
553 static void kdamond_reset_aggregated(struct damon_ctx *c)
554 {
555 	struct damon_target *t;
556 	unsigned int ti = 0;	/* target's index */
557 
558 	damon_for_each_target(t, c) {
559 		struct damon_region *r;
560 
561 		damon_for_each_region(r, t) {
562 			trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
563 			r->last_nr_accesses = r->nr_accesses;
564 			r->nr_accesses = 0;
565 		}
566 		ti++;
567 	}
568 }
569 
570 static void damon_split_region_at(struct damon_ctx *ctx,
571 		struct damon_target *t, struct damon_region *r,
572 		unsigned long sz_r);
573 
574 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
575 {
576 	unsigned long sz;
577 
578 	sz = r->ar.end - r->ar.start;
579 	return s->min_sz_region <= sz && sz <= s->max_sz_region &&
580 		s->min_nr_accesses <= r->nr_accesses &&
581 		r->nr_accesses <= s->max_nr_accesses &&
582 		s->min_age_region <= r->age && r->age <= s->max_age_region;
583 }
584 
585 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
586 		struct damon_region *r, struct damos *s)
587 {
588 	bool ret = __damos_valid_target(r, s);
589 
590 	if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
591 		return ret;
592 
593 	return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
594 }
595 
596 static void damon_do_apply_schemes(struct damon_ctx *c,
597 				   struct damon_target *t,
598 				   struct damon_region *r)
599 {
600 	struct damos *s;
601 
602 	damon_for_each_scheme(s, c) {
603 		struct damos_quota *quota = &s->quota;
604 		unsigned long sz = r->ar.end - r->ar.start;
605 		struct timespec64 begin, end;
606 		unsigned long sz_applied = 0;
607 
608 		if (!s->wmarks.activated)
609 			continue;
610 
611 		/* Check the quota */
612 		if (quota->esz && quota->charged_sz >= quota->esz)
613 			continue;
614 
615 		/* Skip previously charged regions */
616 		if (quota->charge_target_from) {
617 			if (t != quota->charge_target_from)
618 				continue;
619 			if (r == damon_last_region(t)) {
620 				quota->charge_target_from = NULL;
621 				quota->charge_addr_from = 0;
622 				continue;
623 			}
624 			if (quota->charge_addr_from &&
625 					r->ar.end <= quota->charge_addr_from)
626 				continue;
627 
628 			if (quota->charge_addr_from && r->ar.start <
629 					quota->charge_addr_from) {
630 				sz = ALIGN_DOWN(quota->charge_addr_from -
631 						r->ar.start, DAMON_MIN_REGION);
632 				if (!sz) {
633 					if (r->ar.end - r->ar.start <=
634 							DAMON_MIN_REGION)
635 						continue;
636 					sz = DAMON_MIN_REGION;
637 				}
638 				damon_split_region_at(c, t, r, sz);
639 				r = damon_next_region(r);
640 				sz = r->ar.end - r->ar.start;
641 			}
642 			quota->charge_target_from = NULL;
643 			quota->charge_addr_from = 0;
644 		}
645 
646 		if (!damos_valid_target(c, t, r, s))
647 			continue;
648 
649 		/* Apply the scheme */
650 		if (c->ops.apply_scheme) {
651 			if (quota->esz &&
652 					quota->charged_sz + sz > quota->esz) {
653 				sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
654 						DAMON_MIN_REGION);
655 				if (!sz)
656 					goto update_stat;
657 				damon_split_region_at(c, t, r, sz);
658 			}
659 			ktime_get_coarse_ts64(&begin);
660 			sz_applied = c->ops.apply_scheme(c, t, r, s);
661 			ktime_get_coarse_ts64(&end);
662 			quota->total_charged_ns += timespec64_to_ns(&end) -
663 				timespec64_to_ns(&begin);
664 			quota->charged_sz += sz;
665 			if (quota->esz && quota->charged_sz >= quota->esz) {
666 				quota->charge_target_from = t;
667 				quota->charge_addr_from = r->ar.end + 1;
668 			}
669 		}
670 		if (s->action != DAMOS_STAT)
671 			r->age = 0;
672 
673 update_stat:
674 		s->stat.nr_tried++;
675 		s->stat.sz_tried += sz;
676 		if (sz_applied)
677 			s->stat.nr_applied++;
678 		s->stat.sz_applied += sz_applied;
679 	}
680 }
681 
682 /* Shouldn't be called if quota->ms and quota->sz are zero */
683 static void damos_set_effective_quota(struct damos_quota *quota)
684 {
685 	unsigned long throughput;
686 	unsigned long esz;
687 
688 	if (!quota->ms) {
689 		quota->esz = quota->sz;
690 		return;
691 	}
692 
693 	if (quota->total_charged_ns)
694 		throughput = quota->total_charged_sz * 1000000 /
695 			quota->total_charged_ns;
696 	else
697 		throughput = PAGE_SIZE * 1024;
698 	esz = throughput * quota->ms;
699 
700 	if (quota->sz && quota->sz < esz)
701 		esz = quota->sz;
702 	quota->esz = esz;
703 }
704 
705 static void kdamond_apply_schemes(struct damon_ctx *c)
706 {
707 	struct damon_target *t;
708 	struct damon_region *r, *next_r;
709 	struct damos *s;
710 
711 	damon_for_each_scheme(s, c) {
712 		struct damos_quota *quota = &s->quota;
713 		unsigned long cumulated_sz;
714 		unsigned int score, max_score = 0;
715 
716 		if (!s->wmarks.activated)
717 			continue;
718 
719 		if (!quota->ms && !quota->sz)
720 			continue;
721 
722 		/* New charge window starts */
723 		if (time_after_eq(jiffies, quota->charged_from +
724 					msecs_to_jiffies(
725 						quota->reset_interval))) {
726 			if (quota->esz && quota->charged_sz >= quota->esz)
727 				s->stat.qt_exceeds++;
728 			quota->total_charged_sz += quota->charged_sz;
729 			quota->charged_from = jiffies;
730 			quota->charged_sz = 0;
731 			damos_set_effective_quota(quota);
732 		}
733 
734 		if (!c->ops.get_scheme_score)
735 			continue;
736 
737 		/* Fill up the score histogram */
738 		memset(quota->histogram, 0, sizeof(quota->histogram));
739 		damon_for_each_target(t, c) {
740 			damon_for_each_region(r, t) {
741 				if (!__damos_valid_target(r, s))
742 					continue;
743 				score = c->ops.get_scheme_score(
744 						c, t, r, s);
745 				quota->histogram[score] +=
746 					r->ar.end - r->ar.start;
747 				if (score > max_score)
748 					max_score = score;
749 			}
750 		}
751 
752 		/* Set the min score limit */
753 		for (cumulated_sz = 0, score = max_score; ; score--) {
754 			cumulated_sz += quota->histogram[score];
755 			if (cumulated_sz >= quota->esz || !score)
756 				break;
757 		}
758 		quota->min_score = score;
759 	}
760 
761 	damon_for_each_target(t, c) {
762 		damon_for_each_region_safe(r, next_r, t)
763 			damon_do_apply_schemes(c, t, r);
764 	}
765 }
766 
767 static inline unsigned long sz_damon_region(struct damon_region *r)
768 {
769 	return r->ar.end - r->ar.start;
770 }
771 
772 /*
773  * Merge two adjacent regions into one region
774  */
775 static void damon_merge_two_regions(struct damon_target *t,
776 		struct damon_region *l, struct damon_region *r)
777 {
778 	unsigned long sz_l = sz_damon_region(l), sz_r = sz_damon_region(r);
779 
780 	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
781 			(sz_l + sz_r);
782 	l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
783 	l->ar.end = r->ar.end;
784 	damon_destroy_region(r, t);
785 }
786 
787 /*
788  * Merge adjacent regions having similar access frequencies
789  *
790  * t		target affected by this merge operation
791  * thres	'->nr_accesses' diff threshold for the merge
792  * sz_limit	size upper limit of each region
793  */
794 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
795 				   unsigned long sz_limit)
796 {
797 	struct damon_region *r, *prev = NULL, *next;
798 
799 	damon_for_each_region_safe(r, next, t) {
800 		if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
801 			r->age = 0;
802 		else
803 			r->age++;
804 
805 		if (prev && prev->ar.end == r->ar.start &&
806 		    abs(prev->nr_accesses - r->nr_accesses) <= thres &&
807 		    sz_damon_region(prev) + sz_damon_region(r) <= sz_limit)
808 			damon_merge_two_regions(t, prev, r);
809 		else
810 			prev = r;
811 	}
812 }
813 
814 /*
815  * Merge adjacent regions having similar access frequencies
816  *
817  * threshold	'->nr_accesses' diff threshold for the merge
818  * sz_limit	size upper limit of each region
819  *
820  * This function merges monitoring target regions which are adjacent and their
821  * access frequencies are similar.  This is for minimizing the monitoring
822  * overhead under the dynamically changeable access pattern.  If a merge was
823  * unnecessarily made, later 'kdamond_split_regions()' will revert it.
824  */
825 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
826 				  unsigned long sz_limit)
827 {
828 	struct damon_target *t;
829 
830 	damon_for_each_target(t, c)
831 		damon_merge_regions_of(t, threshold, sz_limit);
832 }
833 
834 /*
835  * Split a region in two
836  *
837  * r		the region to be split
838  * sz_r		size of the first sub-region that will be made
839  */
840 static void damon_split_region_at(struct damon_ctx *ctx,
841 		struct damon_target *t, struct damon_region *r,
842 		unsigned long sz_r)
843 {
844 	struct damon_region *new;
845 
846 	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
847 	if (!new)
848 		return;
849 
850 	r->ar.end = new->ar.start;
851 
852 	new->age = r->age;
853 	new->last_nr_accesses = r->last_nr_accesses;
854 
855 	damon_insert_region(new, r, damon_next_region(r), t);
856 }
857 
858 /* Split every region in the given target into 'nr_subs' regions */
859 static void damon_split_regions_of(struct damon_ctx *ctx,
860 				     struct damon_target *t, int nr_subs)
861 {
862 	struct damon_region *r, *next;
863 	unsigned long sz_region, sz_sub = 0;
864 	int i;
865 
866 	damon_for_each_region_safe(r, next, t) {
867 		sz_region = r->ar.end - r->ar.start;
868 
869 		for (i = 0; i < nr_subs - 1 &&
870 				sz_region > 2 * DAMON_MIN_REGION; i++) {
871 			/*
872 			 * Randomly select size of left sub-region to be at
873 			 * least 10 percent and at most 90% of original region
874 			 */
875 			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
876 					sz_region / 10, DAMON_MIN_REGION);
877 			/* Do not allow blank region */
878 			if (sz_sub == 0 || sz_sub >= sz_region)
879 				continue;
880 
881 			damon_split_region_at(ctx, t, r, sz_sub);
882 			sz_region = sz_sub;
883 		}
884 	}
885 }
886 
887 /*
888  * Split every target region into randomly-sized small regions
889  *
890  * This function splits every target region into random-sized small regions if
891  * current total number of the regions is equal or smaller than half of the
892  * user-specified maximum number of regions.  This is for maximizing the
893  * monitoring accuracy under the dynamically changeable access patterns.  If a
894  * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
895  * it.
896  */
897 static void kdamond_split_regions(struct damon_ctx *ctx)
898 {
899 	struct damon_target *t;
900 	unsigned int nr_regions = 0;
901 	static unsigned int last_nr_regions;
902 	int nr_subregions = 2;
903 
904 	damon_for_each_target(t, ctx)
905 		nr_regions += damon_nr_regions(t);
906 
907 	if (nr_regions > ctx->max_nr_regions / 2)
908 		return;
909 
910 	/* Maybe the middle of the region has different access frequency */
911 	if (last_nr_regions == nr_regions &&
912 			nr_regions < ctx->max_nr_regions / 3)
913 		nr_subregions = 3;
914 
915 	damon_for_each_target(t, ctx)
916 		damon_split_regions_of(ctx, t, nr_subregions);
917 
918 	last_nr_regions = nr_regions;
919 }
920 
921 /*
922  * Check whether it is time to check and apply the operations-related data
923  * structures.
924  *
925  * Returns true if it is.
926  */
927 static bool kdamond_need_update_operations(struct damon_ctx *ctx)
928 {
929 	return damon_check_reset_time_interval(&ctx->last_ops_update,
930 			ctx->ops_update_interval);
931 }
932 
933 /*
934  * Check whether current monitoring should be stopped
935  *
936  * The monitoring is stopped when either the user requested to stop, or all
937  * monitoring targets are invalid.
938  *
939  * Returns true if need to stop current monitoring.
940  */
941 static bool kdamond_need_stop(struct damon_ctx *ctx)
942 {
943 	struct damon_target *t;
944 
945 	if (kthread_should_stop())
946 		return true;
947 
948 	if (!ctx->ops.target_valid)
949 		return false;
950 
951 	damon_for_each_target(t, ctx) {
952 		if (ctx->ops.target_valid(t))
953 			return false;
954 	}
955 
956 	return true;
957 }
958 
959 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
960 {
961 	struct sysinfo i;
962 
963 	switch (metric) {
964 	case DAMOS_WMARK_FREE_MEM_RATE:
965 		si_meminfo(&i);
966 		return i.freeram * 1000 / i.totalram;
967 	default:
968 		break;
969 	}
970 	return -EINVAL;
971 }
972 
973 /*
974  * Returns zero if the scheme is active.  Else, returns time to wait for next
975  * watermark check in micro-seconds.
976  */
977 static unsigned long damos_wmark_wait_us(struct damos *scheme)
978 {
979 	unsigned long metric;
980 
981 	if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
982 		return 0;
983 
984 	metric = damos_wmark_metric_value(scheme->wmarks.metric);
985 	/* higher than high watermark or lower than low watermark */
986 	if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
987 		if (scheme->wmarks.activated)
988 			pr_debug("deactivate a scheme (%d) for %s wmark\n",
989 					scheme->action,
990 					metric > scheme->wmarks.high ?
991 					"high" : "low");
992 		scheme->wmarks.activated = false;
993 		return scheme->wmarks.interval;
994 	}
995 
996 	/* inactive and higher than middle watermark */
997 	if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
998 			!scheme->wmarks.activated)
999 		return scheme->wmarks.interval;
1000 
1001 	if (!scheme->wmarks.activated)
1002 		pr_debug("activate a scheme (%d)\n", scheme->action);
1003 	scheme->wmarks.activated = true;
1004 	return 0;
1005 }
1006 
1007 static void kdamond_usleep(unsigned long usecs)
1008 {
1009 	/* See Documentation/timers/timers-howto.rst for the thresholds */
1010 	if (usecs > 20 * USEC_PER_MSEC)
1011 		schedule_timeout_idle(usecs_to_jiffies(usecs));
1012 	else
1013 		usleep_idle_range(usecs, usecs + 1);
1014 }
1015 
1016 /* Returns negative error code if it's not activated but should return */
1017 static int kdamond_wait_activation(struct damon_ctx *ctx)
1018 {
1019 	struct damos *s;
1020 	unsigned long wait_time;
1021 	unsigned long min_wait_time = 0;
1022 	bool init_wait_time = false;
1023 
1024 	while (!kdamond_need_stop(ctx)) {
1025 		damon_for_each_scheme(s, ctx) {
1026 			wait_time = damos_wmark_wait_us(s);
1027 			if (!init_wait_time || wait_time < min_wait_time) {
1028 				init_wait_time = true;
1029 				min_wait_time = wait_time;
1030 			}
1031 		}
1032 		if (!min_wait_time)
1033 			return 0;
1034 
1035 		kdamond_usleep(min_wait_time);
1036 	}
1037 	return -EBUSY;
1038 }
1039 
1040 /*
1041  * The monitoring daemon that runs as a kernel thread
1042  */
1043 static int kdamond_fn(void *data)
1044 {
1045 	struct damon_ctx *ctx = (struct damon_ctx *)data;
1046 	struct damon_target *t;
1047 	struct damon_region *r, *next;
1048 	unsigned int max_nr_accesses = 0;
1049 	unsigned long sz_limit = 0;
1050 	bool done = false;
1051 
1052 	pr_debug("kdamond (%d) starts\n", current->pid);
1053 
1054 	if (ctx->ops.init)
1055 		ctx->ops.init(ctx);
1056 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1057 		done = true;
1058 
1059 	sz_limit = damon_region_sz_limit(ctx);
1060 
1061 	while (!kdamond_need_stop(ctx) && !done) {
1062 		if (kdamond_wait_activation(ctx))
1063 			continue;
1064 
1065 		if (ctx->ops.prepare_access_checks)
1066 			ctx->ops.prepare_access_checks(ctx);
1067 		if (ctx->callback.after_sampling &&
1068 				ctx->callback.after_sampling(ctx))
1069 			done = true;
1070 
1071 		kdamond_usleep(ctx->sample_interval);
1072 
1073 		if (ctx->ops.check_accesses)
1074 			max_nr_accesses = ctx->ops.check_accesses(ctx);
1075 
1076 		if (kdamond_aggregate_interval_passed(ctx)) {
1077 			kdamond_merge_regions(ctx,
1078 					max_nr_accesses / 10,
1079 					sz_limit);
1080 			if (ctx->callback.after_aggregation &&
1081 					ctx->callback.after_aggregation(ctx))
1082 				done = true;
1083 			kdamond_apply_schemes(ctx);
1084 			kdamond_reset_aggregated(ctx);
1085 			kdamond_split_regions(ctx);
1086 			if (ctx->ops.reset_aggregated)
1087 				ctx->ops.reset_aggregated(ctx);
1088 		}
1089 
1090 		if (kdamond_need_update_operations(ctx)) {
1091 			if (ctx->ops.update)
1092 				ctx->ops.update(ctx);
1093 			sz_limit = damon_region_sz_limit(ctx);
1094 		}
1095 	}
1096 	damon_for_each_target(t, ctx) {
1097 		damon_for_each_region_safe(r, next, t)
1098 			damon_destroy_region(r, t);
1099 	}
1100 
1101 	if (ctx->callback.before_terminate)
1102 		ctx->callback.before_terminate(ctx);
1103 	if (ctx->ops.cleanup)
1104 		ctx->ops.cleanup(ctx);
1105 
1106 	pr_debug("kdamond (%d) finishes\n", current->pid);
1107 	mutex_lock(&ctx->kdamond_lock);
1108 	ctx->kdamond = NULL;
1109 	mutex_unlock(&ctx->kdamond_lock);
1110 
1111 	mutex_lock(&damon_lock);
1112 	nr_running_ctxs--;
1113 	if (!nr_running_ctxs && running_exclusive_ctxs)
1114 		running_exclusive_ctxs = false;
1115 	mutex_unlock(&damon_lock);
1116 
1117 	return 0;
1118 }
1119 
1120 #include "core-test.h"
1121