xref: /openbmc/linux/mm/damon/core.c (revision 7ae5c03a)
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_is_registered_ops(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_is_registered_ops() - Check if a given damon_operations is registered.
44  * @id:	Id of the damon_operations to check if registered.
45  *
46  * Return: true if the ops is set, false otherwise.
47  */
48 bool damon_is_registered_ops(enum damon_ops_id id)
49 {
50 	bool registered;
51 
52 	if (id >= NR_DAMON_OPS)
53 		return false;
54 	mutex_lock(&damon_ops_lock);
55 	registered = __damon_is_registered_ops(id);
56 	mutex_unlock(&damon_ops_lock);
57 	return registered;
58 }
59 
60 /**
61  * damon_register_ops() - Register a monitoring operations set to DAMON.
62  * @ops:	monitoring operations set to register.
63  *
64  * This function registers a monitoring operations set of valid &struct
65  * damon_operations->id so that others can find and use them later.
66  *
67  * Return: 0 on success, negative error code otherwise.
68  */
69 int damon_register_ops(struct damon_operations *ops)
70 {
71 	int err = 0;
72 
73 	if (ops->id >= NR_DAMON_OPS)
74 		return -EINVAL;
75 	mutex_lock(&damon_ops_lock);
76 	/* Fail for already registered ops */
77 	if (__damon_is_registered_ops(ops->id)) {
78 		err = -EINVAL;
79 		goto out;
80 	}
81 	damon_registered_ops[ops->id] = *ops;
82 out:
83 	mutex_unlock(&damon_ops_lock);
84 	return err;
85 }
86 
87 /**
88  * damon_select_ops() - Select a monitoring operations to use with the context.
89  * @ctx:	monitoring context to use the operations.
90  * @id:		id of the registered monitoring operations to select.
91  *
92  * This function finds registered monitoring operations set of @id and make
93  * @ctx to use it.
94  *
95  * Return: 0 on success, negative error code otherwise.
96  */
97 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
98 {
99 	int err = 0;
100 
101 	if (id >= NR_DAMON_OPS)
102 		return -EINVAL;
103 
104 	mutex_lock(&damon_ops_lock);
105 	if (!__damon_is_registered_ops(id))
106 		err = -EINVAL;
107 	else
108 		ctx->ops = damon_registered_ops[id];
109 	mutex_unlock(&damon_ops_lock);
110 	return err;
111 }
112 
113 /*
114  * Construct a damon_region struct
115  *
116  * Returns the pointer to the new struct if success, or NULL otherwise
117  */
118 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
119 {
120 	struct damon_region *region;
121 
122 	region = kmalloc(sizeof(*region), GFP_KERNEL);
123 	if (!region)
124 		return NULL;
125 
126 	region->ar.start = start;
127 	region->ar.end = end;
128 	region->nr_accesses = 0;
129 	INIT_LIST_HEAD(&region->list);
130 
131 	region->age = 0;
132 	region->last_nr_accesses = 0;
133 
134 	return region;
135 }
136 
137 void damon_add_region(struct damon_region *r, struct damon_target *t)
138 {
139 	list_add_tail(&r->list, &t->regions_list);
140 	t->nr_regions++;
141 }
142 
143 static void damon_del_region(struct damon_region *r, struct damon_target *t)
144 {
145 	list_del(&r->list);
146 	t->nr_regions--;
147 }
148 
149 static void damon_free_region(struct damon_region *r)
150 {
151 	kfree(r);
152 }
153 
154 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
155 {
156 	damon_del_region(r, t);
157 	damon_free_region(r);
158 }
159 
160 /*
161  * Check whether a region is intersecting an address range
162  *
163  * Returns true if it is.
164  */
165 static bool damon_intersect(struct damon_region *r,
166 		struct damon_addr_range *re)
167 {
168 	return !(r->ar.end <= re->start || re->end <= r->ar.start);
169 }
170 
171 /*
172  * damon_set_regions() - Set regions of a target for given address ranges.
173  * @t:		the given target.
174  * @ranges:	array of new monitoring target ranges.
175  * @nr_ranges:	length of @ranges.
176  *
177  * This function adds new regions to, or modify existing regions of a
178  * monitoring target to fit in specific ranges.
179  *
180  * Return: 0 if success, or negative error code otherwise.
181  */
182 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
183 		unsigned int nr_ranges)
184 {
185 	struct damon_region *r, *next;
186 	unsigned int i;
187 
188 	/* Remove regions which are not in the new ranges */
189 	damon_for_each_region_safe(r, next, t) {
190 		for (i = 0; i < nr_ranges; i++) {
191 			if (damon_intersect(r, &ranges[i]))
192 				break;
193 		}
194 		if (i == nr_ranges)
195 			damon_destroy_region(r, t);
196 	}
197 
198 	/* Add new regions or resize existing regions to fit in the ranges */
199 	for (i = 0; i < nr_ranges; i++) {
200 		struct damon_region *first = NULL, *last, *newr;
201 		struct damon_addr_range *range;
202 
203 		range = &ranges[i];
204 		/* Get the first/last regions intersecting with the range */
205 		damon_for_each_region(r, t) {
206 			if (damon_intersect(r, range)) {
207 				if (!first)
208 					first = r;
209 				last = r;
210 			}
211 			if (r->ar.start >= range->end)
212 				break;
213 		}
214 		if (!first) {
215 			/* no region intersects with this range */
216 			newr = damon_new_region(
217 					ALIGN_DOWN(range->start,
218 						DAMON_MIN_REGION),
219 					ALIGN(range->end, DAMON_MIN_REGION));
220 			if (!newr)
221 				return -ENOMEM;
222 			damon_insert_region(newr, damon_prev_region(r), r, t);
223 		} else {
224 			/* resize intersecting regions to fit in this range */
225 			first->ar.start = ALIGN_DOWN(range->start,
226 					DAMON_MIN_REGION);
227 			last->ar.end = ALIGN(range->end, DAMON_MIN_REGION);
228 		}
229 	}
230 	return 0;
231 }
232 
233 struct damos *damon_new_scheme(
234 		unsigned long min_sz_region, unsigned long max_sz_region,
235 		unsigned int min_nr_accesses, unsigned int max_nr_accesses,
236 		unsigned int min_age_region, unsigned int max_age_region,
237 		enum damos_action action, struct damos_quota *quota,
238 		struct damos_watermarks *wmarks)
239 {
240 	struct damos *scheme;
241 
242 	scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
243 	if (!scheme)
244 		return NULL;
245 	scheme->min_sz_region = min_sz_region;
246 	scheme->max_sz_region = max_sz_region;
247 	scheme->min_nr_accesses = min_nr_accesses;
248 	scheme->max_nr_accesses = max_nr_accesses;
249 	scheme->min_age_region = min_age_region;
250 	scheme->max_age_region = max_age_region;
251 	scheme->action = action;
252 	scheme->stat = (struct damos_stat){};
253 	INIT_LIST_HEAD(&scheme->list);
254 
255 	scheme->quota.ms = quota->ms;
256 	scheme->quota.sz = quota->sz;
257 	scheme->quota.reset_interval = quota->reset_interval;
258 	scheme->quota.weight_sz = quota->weight_sz;
259 	scheme->quota.weight_nr_accesses = quota->weight_nr_accesses;
260 	scheme->quota.weight_age = quota->weight_age;
261 	scheme->quota.total_charged_sz = 0;
262 	scheme->quota.total_charged_ns = 0;
263 	scheme->quota.esz = 0;
264 	scheme->quota.charged_sz = 0;
265 	scheme->quota.charged_from = 0;
266 	scheme->quota.charge_target_from = NULL;
267 	scheme->quota.charge_addr_from = 0;
268 
269 	scheme->wmarks.metric = wmarks->metric;
270 	scheme->wmarks.interval = wmarks->interval;
271 	scheme->wmarks.high = wmarks->high;
272 	scheme->wmarks.mid = wmarks->mid;
273 	scheme->wmarks.low = wmarks->low;
274 	scheme->wmarks.activated = true;
275 
276 	return scheme;
277 }
278 
279 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
280 {
281 	list_add_tail(&s->list, &ctx->schemes);
282 }
283 
284 static void damon_del_scheme(struct damos *s)
285 {
286 	list_del(&s->list);
287 }
288 
289 static void damon_free_scheme(struct damos *s)
290 {
291 	kfree(s);
292 }
293 
294 void damon_destroy_scheme(struct damos *s)
295 {
296 	damon_del_scheme(s);
297 	damon_free_scheme(s);
298 }
299 
300 /*
301  * Construct a damon_target struct
302  *
303  * Returns the pointer to the new struct if success, or NULL otherwise
304  */
305 struct damon_target *damon_new_target(void)
306 {
307 	struct damon_target *t;
308 
309 	t = kmalloc(sizeof(*t), GFP_KERNEL);
310 	if (!t)
311 		return NULL;
312 
313 	t->pid = NULL;
314 	t->nr_regions = 0;
315 	INIT_LIST_HEAD(&t->regions_list);
316 
317 	return t;
318 }
319 
320 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
321 {
322 	list_add_tail(&t->list, &ctx->adaptive_targets);
323 }
324 
325 bool damon_targets_empty(struct damon_ctx *ctx)
326 {
327 	return list_empty(&ctx->adaptive_targets);
328 }
329 
330 static void damon_del_target(struct damon_target *t)
331 {
332 	list_del(&t->list);
333 }
334 
335 void damon_free_target(struct damon_target *t)
336 {
337 	struct damon_region *r, *next;
338 
339 	damon_for_each_region_safe(r, next, t)
340 		damon_free_region(r);
341 	kfree(t);
342 }
343 
344 void damon_destroy_target(struct damon_target *t)
345 {
346 	damon_del_target(t);
347 	damon_free_target(t);
348 }
349 
350 unsigned int damon_nr_regions(struct damon_target *t)
351 {
352 	return t->nr_regions;
353 }
354 
355 struct damon_ctx *damon_new_ctx(void)
356 {
357 	struct damon_ctx *ctx;
358 
359 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
360 	if (!ctx)
361 		return NULL;
362 
363 	ctx->sample_interval = 5 * 1000;
364 	ctx->aggr_interval = 100 * 1000;
365 	ctx->ops_update_interval = 60 * 1000 * 1000;
366 
367 	ktime_get_coarse_ts64(&ctx->last_aggregation);
368 	ctx->last_ops_update = ctx->last_aggregation;
369 
370 	mutex_init(&ctx->kdamond_lock);
371 
372 	ctx->min_nr_regions = 10;
373 	ctx->max_nr_regions = 1000;
374 
375 	INIT_LIST_HEAD(&ctx->adaptive_targets);
376 	INIT_LIST_HEAD(&ctx->schemes);
377 
378 	return ctx;
379 }
380 
381 static void damon_destroy_targets(struct damon_ctx *ctx)
382 {
383 	struct damon_target *t, *next_t;
384 
385 	if (ctx->ops.cleanup) {
386 		ctx->ops.cleanup(ctx);
387 		return;
388 	}
389 
390 	damon_for_each_target_safe(t, next_t, ctx)
391 		damon_destroy_target(t);
392 }
393 
394 void damon_destroy_ctx(struct damon_ctx *ctx)
395 {
396 	struct damos *s, *next_s;
397 
398 	damon_destroy_targets(ctx);
399 
400 	damon_for_each_scheme_safe(s, next_s, ctx)
401 		damon_destroy_scheme(s);
402 
403 	kfree(ctx);
404 }
405 
406 /**
407  * damon_set_attrs() - Set attributes for the monitoring.
408  * @ctx:		monitoring context
409  * @sample_int:		time interval between samplings
410  * @aggr_int:		time interval between aggregations
411  * @ops_upd_int:	time interval between monitoring operations updates
412  * @min_nr_reg:		minimal number of regions
413  * @max_nr_reg:		maximum number of regions
414  *
415  * This function should not be called while the kdamond is running.
416  * Every time interval is in micro-seconds.
417  *
418  * Return: 0 on success, negative error code otherwise.
419  */
420 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
421 		    unsigned long aggr_int, unsigned long ops_upd_int,
422 		    unsigned long min_nr_reg, unsigned long max_nr_reg)
423 {
424 	if (min_nr_reg < 3)
425 		return -EINVAL;
426 	if (min_nr_reg > max_nr_reg)
427 		return -EINVAL;
428 
429 	ctx->sample_interval = sample_int;
430 	ctx->aggr_interval = aggr_int;
431 	ctx->ops_update_interval = ops_upd_int;
432 	ctx->min_nr_regions = min_nr_reg;
433 	ctx->max_nr_regions = max_nr_reg;
434 
435 	return 0;
436 }
437 
438 /**
439  * damon_set_schemes() - Set data access monitoring based operation schemes.
440  * @ctx:	monitoring context
441  * @schemes:	array of the schemes
442  * @nr_schemes:	number of entries in @schemes
443  *
444  * This function should not be called while the kdamond of the context is
445  * running.
446  *
447  * Return: 0 if success, or negative error code otherwise.
448  */
449 int damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
450 			ssize_t nr_schemes)
451 {
452 	struct damos *s, *next;
453 	ssize_t i;
454 
455 	damon_for_each_scheme_safe(s, next, ctx)
456 		damon_destroy_scheme(s);
457 	for (i = 0; i < nr_schemes; i++)
458 		damon_add_scheme(ctx, schemes[i]);
459 	return 0;
460 }
461 
462 /**
463  * damon_nr_running_ctxs() - Return number of currently running contexts.
464  */
465 int damon_nr_running_ctxs(void)
466 {
467 	int nr_ctxs;
468 
469 	mutex_lock(&damon_lock);
470 	nr_ctxs = nr_running_ctxs;
471 	mutex_unlock(&damon_lock);
472 
473 	return nr_ctxs;
474 }
475 
476 /* Returns the size upper limit for each monitoring region */
477 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
478 {
479 	struct damon_target *t;
480 	struct damon_region *r;
481 	unsigned long sz = 0;
482 
483 	damon_for_each_target(t, ctx) {
484 		damon_for_each_region(r, t)
485 			sz += r->ar.end - r->ar.start;
486 	}
487 
488 	if (ctx->min_nr_regions)
489 		sz /= ctx->min_nr_regions;
490 	if (sz < DAMON_MIN_REGION)
491 		sz = DAMON_MIN_REGION;
492 
493 	return sz;
494 }
495 
496 static int kdamond_fn(void *data);
497 
498 /*
499  * __damon_start() - Starts monitoring with given context.
500  * @ctx:	monitoring context
501  *
502  * This function should be called while damon_lock is hold.
503  *
504  * Return: 0 on success, negative error code otherwise.
505  */
506 static int __damon_start(struct damon_ctx *ctx)
507 {
508 	int err = -EBUSY;
509 
510 	mutex_lock(&ctx->kdamond_lock);
511 	if (!ctx->kdamond) {
512 		err = 0;
513 		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
514 				nr_running_ctxs);
515 		if (IS_ERR(ctx->kdamond)) {
516 			err = PTR_ERR(ctx->kdamond);
517 			ctx->kdamond = NULL;
518 		}
519 	}
520 	mutex_unlock(&ctx->kdamond_lock);
521 
522 	return err;
523 }
524 
525 /**
526  * damon_start() - Starts the monitorings for a given group of contexts.
527  * @ctxs:	an array of the pointers for contexts to start monitoring
528  * @nr_ctxs:	size of @ctxs
529  * @exclusive:	exclusiveness of this contexts group
530  *
531  * This function starts a group of monitoring threads for a group of monitoring
532  * contexts.  One thread per each context is created and run in parallel.  The
533  * caller should handle synchronization between the threads by itself.  If
534  * @exclusive is true and a group of threads that created by other
535  * 'damon_start()' call is currently running, this function does nothing but
536  * returns -EBUSY.
537  *
538  * Return: 0 on success, negative error code otherwise.
539  */
540 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
541 {
542 	int i;
543 	int err = 0;
544 
545 	mutex_lock(&damon_lock);
546 	if ((exclusive && nr_running_ctxs) ||
547 			(!exclusive && running_exclusive_ctxs)) {
548 		mutex_unlock(&damon_lock);
549 		return -EBUSY;
550 	}
551 
552 	for (i = 0; i < nr_ctxs; i++) {
553 		err = __damon_start(ctxs[i]);
554 		if (err)
555 			break;
556 		nr_running_ctxs++;
557 	}
558 	if (exclusive && nr_running_ctxs)
559 		running_exclusive_ctxs = true;
560 	mutex_unlock(&damon_lock);
561 
562 	return err;
563 }
564 
565 /*
566  * __damon_stop() - Stops monitoring of a given context.
567  * @ctx:	monitoring context
568  *
569  * Return: 0 on success, negative error code otherwise.
570  */
571 static int __damon_stop(struct damon_ctx *ctx)
572 {
573 	struct task_struct *tsk;
574 
575 	mutex_lock(&ctx->kdamond_lock);
576 	tsk = ctx->kdamond;
577 	if (tsk) {
578 		get_task_struct(tsk);
579 		mutex_unlock(&ctx->kdamond_lock);
580 		kthread_stop(tsk);
581 		put_task_struct(tsk);
582 		return 0;
583 	}
584 	mutex_unlock(&ctx->kdamond_lock);
585 
586 	return -EPERM;
587 }
588 
589 /**
590  * damon_stop() - Stops the monitorings for a given group of contexts.
591  * @ctxs:	an array of the pointers for contexts to stop monitoring
592  * @nr_ctxs:	size of @ctxs
593  *
594  * Return: 0 on success, negative error code otherwise.
595  */
596 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
597 {
598 	int i, err = 0;
599 
600 	for (i = 0; i < nr_ctxs; i++) {
601 		/* nr_running_ctxs is decremented in kdamond_fn */
602 		err = __damon_stop(ctxs[i]);
603 		if (err)
604 			break;
605 	}
606 	return err;
607 }
608 
609 /*
610  * damon_check_reset_time_interval() - Check if a time interval is elapsed.
611  * @baseline:	the time to check whether the interval has elapsed since
612  * @interval:	the time interval (microseconds)
613  *
614  * See whether the given time interval has passed since the given baseline
615  * time.  If so, it also updates the baseline to current time for next check.
616  *
617  * Return:	true if the time interval has passed, or false otherwise.
618  */
619 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
620 		unsigned long interval)
621 {
622 	struct timespec64 now;
623 
624 	ktime_get_coarse_ts64(&now);
625 	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
626 			interval * 1000)
627 		return false;
628 	*baseline = now;
629 	return true;
630 }
631 
632 /*
633  * Check whether it is time to flush the aggregated information
634  */
635 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
636 {
637 	return damon_check_reset_time_interval(&ctx->last_aggregation,
638 			ctx->aggr_interval);
639 }
640 
641 /*
642  * Reset the aggregated monitoring results ('nr_accesses' of each region).
643  */
644 static void kdamond_reset_aggregated(struct damon_ctx *c)
645 {
646 	struct damon_target *t;
647 	unsigned int ti = 0;	/* target's index */
648 
649 	damon_for_each_target(t, c) {
650 		struct damon_region *r;
651 
652 		damon_for_each_region(r, t) {
653 			trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
654 			r->last_nr_accesses = r->nr_accesses;
655 			r->nr_accesses = 0;
656 		}
657 		ti++;
658 	}
659 }
660 
661 static void damon_split_region_at(struct damon_ctx *ctx,
662 		struct damon_target *t, struct damon_region *r,
663 		unsigned long sz_r);
664 
665 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
666 {
667 	unsigned long sz;
668 
669 	sz = r->ar.end - r->ar.start;
670 	return s->min_sz_region <= sz && sz <= s->max_sz_region &&
671 		s->min_nr_accesses <= r->nr_accesses &&
672 		r->nr_accesses <= s->max_nr_accesses &&
673 		s->min_age_region <= r->age && r->age <= s->max_age_region;
674 }
675 
676 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
677 		struct damon_region *r, struct damos *s)
678 {
679 	bool ret = __damos_valid_target(r, s);
680 
681 	if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
682 		return ret;
683 
684 	return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
685 }
686 
687 static void damon_do_apply_schemes(struct damon_ctx *c,
688 				   struct damon_target *t,
689 				   struct damon_region *r)
690 {
691 	struct damos *s;
692 
693 	damon_for_each_scheme(s, c) {
694 		struct damos_quota *quota = &s->quota;
695 		unsigned long sz = r->ar.end - r->ar.start;
696 		struct timespec64 begin, end;
697 		unsigned long sz_applied = 0;
698 
699 		if (!s->wmarks.activated)
700 			continue;
701 
702 		/* Check the quota */
703 		if (quota->esz && quota->charged_sz >= quota->esz)
704 			continue;
705 
706 		/* Skip previously charged regions */
707 		if (quota->charge_target_from) {
708 			if (t != quota->charge_target_from)
709 				continue;
710 			if (r == damon_last_region(t)) {
711 				quota->charge_target_from = NULL;
712 				quota->charge_addr_from = 0;
713 				continue;
714 			}
715 			if (quota->charge_addr_from &&
716 					r->ar.end <= quota->charge_addr_from)
717 				continue;
718 
719 			if (quota->charge_addr_from && r->ar.start <
720 					quota->charge_addr_from) {
721 				sz = ALIGN_DOWN(quota->charge_addr_from -
722 						r->ar.start, DAMON_MIN_REGION);
723 				if (!sz) {
724 					if (r->ar.end - r->ar.start <=
725 							DAMON_MIN_REGION)
726 						continue;
727 					sz = DAMON_MIN_REGION;
728 				}
729 				damon_split_region_at(c, t, r, sz);
730 				r = damon_next_region(r);
731 				sz = r->ar.end - r->ar.start;
732 			}
733 			quota->charge_target_from = NULL;
734 			quota->charge_addr_from = 0;
735 		}
736 
737 		if (!damos_valid_target(c, t, r, s))
738 			continue;
739 
740 		/* Apply the scheme */
741 		if (c->ops.apply_scheme) {
742 			if (quota->esz &&
743 					quota->charged_sz + sz > quota->esz) {
744 				sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
745 						DAMON_MIN_REGION);
746 				if (!sz)
747 					goto update_stat;
748 				damon_split_region_at(c, t, r, sz);
749 			}
750 			ktime_get_coarse_ts64(&begin);
751 			sz_applied = c->ops.apply_scheme(c, t, r, s);
752 			ktime_get_coarse_ts64(&end);
753 			quota->total_charged_ns += timespec64_to_ns(&end) -
754 				timespec64_to_ns(&begin);
755 			quota->charged_sz += sz;
756 			if (quota->esz && quota->charged_sz >= quota->esz) {
757 				quota->charge_target_from = t;
758 				quota->charge_addr_from = r->ar.end + 1;
759 			}
760 		}
761 		if (s->action != DAMOS_STAT)
762 			r->age = 0;
763 
764 update_stat:
765 		s->stat.nr_tried++;
766 		s->stat.sz_tried += sz;
767 		if (sz_applied)
768 			s->stat.nr_applied++;
769 		s->stat.sz_applied += sz_applied;
770 	}
771 }
772 
773 /* Shouldn't be called if quota->ms and quota->sz are zero */
774 static void damos_set_effective_quota(struct damos_quota *quota)
775 {
776 	unsigned long throughput;
777 	unsigned long esz;
778 
779 	if (!quota->ms) {
780 		quota->esz = quota->sz;
781 		return;
782 	}
783 
784 	if (quota->total_charged_ns)
785 		throughput = quota->total_charged_sz * 1000000 /
786 			quota->total_charged_ns;
787 	else
788 		throughput = PAGE_SIZE * 1024;
789 	esz = throughput * quota->ms;
790 
791 	if (quota->sz && quota->sz < esz)
792 		esz = quota->sz;
793 	quota->esz = esz;
794 }
795 
796 static void kdamond_apply_schemes(struct damon_ctx *c)
797 {
798 	struct damon_target *t;
799 	struct damon_region *r, *next_r;
800 	struct damos *s;
801 
802 	damon_for_each_scheme(s, c) {
803 		struct damos_quota *quota = &s->quota;
804 		unsigned long cumulated_sz;
805 		unsigned int score, max_score = 0;
806 
807 		if (!s->wmarks.activated)
808 			continue;
809 
810 		if (!quota->ms && !quota->sz)
811 			continue;
812 
813 		/* New charge window starts */
814 		if (time_after_eq(jiffies, quota->charged_from +
815 					msecs_to_jiffies(
816 						quota->reset_interval))) {
817 			if (quota->esz && quota->charged_sz >= quota->esz)
818 				s->stat.qt_exceeds++;
819 			quota->total_charged_sz += quota->charged_sz;
820 			quota->charged_from = jiffies;
821 			quota->charged_sz = 0;
822 			damos_set_effective_quota(quota);
823 		}
824 
825 		if (!c->ops.get_scheme_score)
826 			continue;
827 
828 		/* Fill up the score histogram */
829 		memset(quota->histogram, 0, sizeof(quota->histogram));
830 		damon_for_each_target(t, c) {
831 			damon_for_each_region(r, t) {
832 				if (!__damos_valid_target(r, s))
833 					continue;
834 				score = c->ops.get_scheme_score(
835 						c, t, r, s);
836 				quota->histogram[score] +=
837 					r->ar.end - r->ar.start;
838 				if (score > max_score)
839 					max_score = score;
840 			}
841 		}
842 
843 		/* Set the min score limit */
844 		for (cumulated_sz = 0, score = max_score; ; score--) {
845 			cumulated_sz += quota->histogram[score];
846 			if (cumulated_sz >= quota->esz || !score)
847 				break;
848 		}
849 		quota->min_score = score;
850 	}
851 
852 	damon_for_each_target(t, c) {
853 		damon_for_each_region_safe(r, next_r, t)
854 			damon_do_apply_schemes(c, t, r);
855 	}
856 }
857 
858 static inline unsigned long sz_damon_region(struct damon_region *r)
859 {
860 	return r->ar.end - r->ar.start;
861 }
862 
863 /*
864  * Merge two adjacent regions into one region
865  */
866 static void damon_merge_two_regions(struct damon_target *t,
867 		struct damon_region *l, struct damon_region *r)
868 {
869 	unsigned long sz_l = sz_damon_region(l), sz_r = sz_damon_region(r);
870 
871 	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
872 			(sz_l + sz_r);
873 	l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
874 	l->ar.end = r->ar.end;
875 	damon_destroy_region(r, t);
876 }
877 
878 /*
879  * Merge adjacent regions having similar access frequencies
880  *
881  * t		target affected by this merge operation
882  * thres	'->nr_accesses' diff threshold for the merge
883  * sz_limit	size upper limit of each region
884  */
885 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
886 				   unsigned long sz_limit)
887 {
888 	struct damon_region *r, *prev = NULL, *next;
889 
890 	damon_for_each_region_safe(r, next, t) {
891 		if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
892 			r->age = 0;
893 		else
894 			r->age++;
895 
896 		if (prev && prev->ar.end == r->ar.start &&
897 		    abs(prev->nr_accesses - r->nr_accesses) <= thres &&
898 		    sz_damon_region(prev) + sz_damon_region(r) <= sz_limit)
899 			damon_merge_two_regions(t, prev, r);
900 		else
901 			prev = r;
902 	}
903 }
904 
905 /*
906  * Merge adjacent regions having similar access frequencies
907  *
908  * threshold	'->nr_accesses' diff threshold for the merge
909  * sz_limit	size upper limit of each region
910  *
911  * This function merges monitoring target regions which are adjacent and their
912  * access frequencies are similar.  This is for minimizing the monitoring
913  * overhead under the dynamically changeable access pattern.  If a merge was
914  * unnecessarily made, later 'kdamond_split_regions()' will revert it.
915  */
916 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
917 				  unsigned long sz_limit)
918 {
919 	struct damon_target *t;
920 
921 	damon_for_each_target(t, c)
922 		damon_merge_regions_of(t, threshold, sz_limit);
923 }
924 
925 /*
926  * Split a region in two
927  *
928  * r		the region to be split
929  * sz_r		size of the first sub-region that will be made
930  */
931 static void damon_split_region_at(struct damon_ctx *ctx,
932 		struct damon_target *t, struct damon_region *r,
933 		unsigned long sz_r)
934 {
935 	struct damon_region *new;
936 
937 	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
938 	if (!new)
939 		return;
940 
941 	r->ar.end = new->ar.start;
942 
943 	new->age = r->age;
944 	new->last_nr_accesses = r->last_nr_accesses;
945 
946 	damon_insert_region(new, r, damon_next_region(r), t);
947 }
948 
949 /* Split every region in the given target into 'nr_subs' regions */
950 static void damon_split_regions_of(struct damon_ctx *ctx,
951 				     struct damon_target *t, int nr_subs)
952 {
953 	struct damon_region *r, *next;
954 	unsigned long sz_region, sz_sub = 0;
955 	int i;
956 
957 	damon_for_each_region_safe(r, next, t) {
958 		sz_region = r->ar.end - r->ar.start;
959 
960 		for (i = 0; i < nr_subs - 1 &&
961 				sz_region > 2 * DAMON_MIN_REGION; i++) {
962 			/*
963 			 * Randomly select size of left sub-region to be at
964 			 * least 10 percent and at most 90% of original region
965 			 */
966 			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
967 					sz_region / 10, DAMON_MIN_REGION);
968 			/* Do not allow blank region */
969 			if (sz_sub == 0 || sz_sub >= sz_region)
970 				continue;
971 
972 			damon_split_region_at(ctx, t, r, sz_sub);
973 			sz_region = sz_sub;
974 		}
975 	}
976 }
977 
978 /*
979  * Split every target region into randomly-sized small regions
980  *
981  * This function splits every target region into random-sized small regions if
982  * current total number of the regions is equal or smaller than half of the
983  * user-specified maximum number of regions.  This is for maximizing the
984  * monitoring accuracy under the dynamically changeable access patterns.  If a
985  * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
986  * it.
987  */
988 static void kdamond_split_regions(struct damon_ctx *ctx)
989 {
990 	struct damon_target *t;
991 	unsigned int nr_regions = 0;
992 	static unsigned int last_nr_regions;
993 	int nr_subregions = 2;
994 
995 	damon_for_each_target(t, ctx)
996 		nr_regions += damon_nr_regions(t);
997 
998 	if (nr_regions > ctx->max_nr_regions / 2)
999 		return;
1000 
1001 	/* Maybe the middle of the region has different access frequency */
1002 	if (last_nr_regions == nr_regions &&
1003 			nr_regions < ctx->max_nr_regions / 3)
1004 		nr_subregions = 3;
1005 
1006 	damon_for_each_target(t, ctx)
1007 		damon_split_regions_of(ctx, t, nr_subregions);
1008 
1009 	last_nr_regions = nr_regions;
1010 }
1011 
1012 /*
1013  * Check whether it is time to check and apply the operations-related data
1014  * structures.
1015  *
1016  * Returns true if it is.
1017  */
1018 static bool kdamond_need_update_operations(struct damon_ctx *ctx)
1019 {
1020 	return damon_check_reset_time_interval(&ctx->last_ops_update,
1021 			ctx->ops_update_interval);
1022 }
1023 
1024 /*
1025  * Check whether current monitoring should be stopped
1026  *
1027  * The monitoring is stopped when either the user requested to stop, or all
1028  * monitoring targets are invalid.
1029  *
1030  * Returns true if need to stop current monitoring.
1031  */
1032 static bool kdamond_need_stop(struct damon_ctx *ctx)
1033 {
1034 	struct damon_target *t;
1035 
1036 	if (kthread_should_stop())
1037 		return true;
1038 
1039 	if (!ctx->ops.target_valid)
1040 		return false;
1041 
1042 	damon_for_each_target(t, ctx) {
1043 		if (ctx->ops.target_valid(t))
1044 			return false;
1045 	}
1046 
1047 	return true;
1048 }
1049 
1050 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
1051 {
1052 	struct sysinfo i;
1053 
1054 	switch (metric) {
1055 	case DAMOS_WMARK_FREE_MEM_RATE:
1056 		si_meminfo(&i);
1057 		return i.freeram * 1000 / i.totalram;
1058 	default:
1059 		break;
1060 	}
1061 	return -EINVAL;
1062 }
1063 
1064 /*
1065  * Returns zero if the scheme is active.  Else, returns time to wait for next
1066  * watermark check in micro-seconds.
1067  */
1068 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1069 {
1070 	unsigned long metric;
1071 
1072 	if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
1073 		return 0;
1074 
1075 	metric = damos_wmark_metric_value(scheme->wmarks.metric);
1076 	/* higher than high watermark or lower than low watermark */
1077 	if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1078 		if (scheme->wmarks.activated)
1079 			pr_debug("deactivate a scheme (%d) for %s wmark\n",
1080 					scheme->action,
1081 					metric > scheme->wmarks.high ?
1082 					"high" : "low");
1083 		scheme->wmarks.activated = false;
1084 		return scheme->wmarks.interval;
1085 	}
1086 
1087 	/* inactive and higher than middle watermark */
1088 	if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1089 			!scheme->wmarks.activated)
1090 		return scheme->wmarks.interval;
1091 
1092 	if (!scheme->wmarks.activated)
1093 		pr_debug("activate a scheme (%d)\n", scheme->action);
1094 	scheme->wmarks.activated = true;
1095 	return 0;
1096 }
1097 
1098 static void kdamond_usleep(unsigned long usecs)
1099 {
1100 	/* See Documentation/timers/timers-howto.rst for the thresholds */
1101 	if (usecs > 20 * USEC_PER_MSEC)
1102 		schedule_timeout_idle(usecs_to_jiffies(usecs));
1103 	else
1104 		usleep_idle_range(usecs, usecs + 1);
1105 }
1106 
1107 /* Returns negative error code if it's not activated but should return */
1108 static int kdamond_wait_activation(struct damon_ctx *ctx)
1109 {
1110 	struct damos *s;
1111 	unsigned long wait_time;
1112 	unsigned long min_wait_time = 0;
1113 	bool init_wait_time = false;
1114 
1115 	while (!kdamond_need_stop(ctx)) {
1116 		damon_for_each_scheme(s, ctx) {
1117 			wait_time = damos_wmark_wait_us(s);
1118 			if (!init_wait_time || wait_time < min_wait_time) {
1119 				init_wait_time = true;
1120 				min_wait_time = wait_time;
1121 			}
1122 		}
1123 		if (!min_wait_time)
1124 			return 0;
1125 
1126 		kdamond_usleep(min_wait_time);
1127 
1128 		if (ctx->callback.after_wmarks_check &&
1129 				ctx->callback.after_wmarks_check(ctx))
1130 			break;
1131 	}
1132 	return -EBUSY;
1133 }
1134 
1135 /*
1136  * The monitoring daemon that runs as a kernel thread
1137  */
1138 static int kdamond_fn(void *data)
1139 {
1140 	struct damon_ctx *ctx = data;
1141 	struct damon_target *t;
1142 	struct damon_region *r, *next;
1143 	unsigned int max_nr_accesses = 0;
1144 	unsigned long sz_limit = 0;
1145 	bool done = false;
1146 
1147 	pr_debug("kdamond (%d) starts\n", current->pid);
1148 
1149 	if (ctx->ops.init)
1150 		ctx->ops.init(ctx);
1151 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1152 		done = true;
1153 
1154 	sz_limit = damon_region_sz_limit(ctx);
1155 
1156 	while (!kdamond_need_stop(ctx) && !done) {
1157 		if (kdamond_wait_activation(ctx)) {
1158 			done = true;
1159 			continue;
1160 		}
1161 
1162 		if (ctx->ops.prepare_access_checks)
1163 			ctx->ops.prepare_access_checks(ctx);
1164 		if (ctx->callback.after_sampling &&
1165 				ctx->callback.after_sampling(ctx)) {
1166 			done = true;
1167 			continue;
1168 		}
1169 
1170 		kdamond_usleep(ctx->sample_interval);
1171 
1172 		if (ctx->ops.check_accesses)
1173 			max_nr_accesses = ctx->ops.check_accesses(ctx);
1174 
1175 		if (kdamond_aggregate_interval_passed(ctx)) {
1176 			kdamond_merge_regions(ctx,
1177 					max_nr_accesses / 10,
1178 					sz_limit);
1179 			if (ctx->callback.after_aggregation &&
1180 					ctx->callback.after_aggregation(ctx)) {
1181 				done = true;
1182 				continue;
1183 			}
1184 			kdamond_apply_schemes(ctx);
1185 			kdamond_reset_aggregated(ctx);
1186 			kdamond_split_regions(ctx);
1187 			if (ctx->ops.reset_aggregated)
1188 				ctx->ops.reset_aggregated(ctx);
1189 		}
1190 
1191 		if (kdamond_need_update_operations(ctx)) {
1192 			if (ctx->ops.update)
1193 				ctx->ops.update(ctx);
1194 			sz_limit = damon_region_sz_limit(ctx);
1195 		}
1196 	}
1197 	damon_for_each_target(t, ctx) {
1198 		damon_for_each_region_safe(r, next, t)
1199 			damon_destroy_region(r, t);
1200 	}
1201 
1202 	if (ctx->callback.before_terminate)
1203 		ctx->callback.before_terminate(ctx);
1204 	if (ctx->ops.cleanup)
1205 		ctx->ops.cleanup(ctx);
1206 
1207 	pr_debug("kdamond (%d) finishes\n", current->pid);
1208 	mutex_lock(&ctx->kdamond_lock);
1209 	ctx->kdamond = NULL;
1210 	mutex_unlock(&ctx->kdamond_lock);
1211 
1212 	mutex_lock(&damon_lock);
1213 	nr_running_ctxs--;
1214 	if (!nr_running_ctxs && running_exclusive_ctxs)
1215 		running_exclusive_ctxs = false;
1216 	mutex_unlock(&damon_lock);
1217 
1218 	return 0;
1219 }
1220 
1221 #include "core-test.h"
1222