1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Data Access Monitor
4 *
5 * Author: SeongJae Park <sj@kernel.org>
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/psi.h>
15 #include <linux/slab.h>
16 #include <linux/string.h>
17
18 #define CREATE_TRACE_POINTS
19 #include <trace/events/damon.h>
20
21 #ifdef CONFIG_DAMON_KUNIT_TEST
22 #undef DAMON_MIN_REGION
23 #define DAMON_MIN_REGION 1
24 #endif
25
26 static DEFINE_MUTEX(damon_lock);
27 static int nr_running_ctxs;
28 static bool running_exclusive_ctxs;
29
30 static DEFINE_MUTEX(damon_ops_lock);
31 static struct damon_operations damon_registered_ops[NR_DAMON_OPS];
32
33 static struct kmem_cache *damon_region_cache __ro_after_init;
34
35 /* Should be called under damon_ops_lock with id smaller than NR_DAMON_OPS */
__damon_is_registered_ops(enum damon_ops_id id)36 static bool __damon_is_registered_ops(enum damon_ops_id id)
37 {
38 struct damon_operations empty_ops = {};
39
40 if (!memcmp(&empty_ops, &damon_registered_ops[id], sizeof(empty_ops)))
41 return false;
42 return true;
43 }
44
45 /**
46 * damon_is_registered_ops() - Check if a given damon_operations is registered.
47 * @id: Id of the damon_operations to check if registered.
48 *
49 * Return: true if the ops is set, false otherwise.
50 */
damon_is_registered_ops(enum damon_ops_id id)51 bool damon_is_registered_ops(enum damon_ops_id id)
52 {
53 bool registered;
54
55 if (id >= NR_DAMON_OPS)
56 return false;
57 mutex_lock(&damon_ops_lock);
58 registered = __damon_is_registered_ops(id);
59 mutex_unlock(&damon_ops_lock);
60 return registered;
61 }
62
63 /**
64 * damon_register_ops() - Register a monitoring operations set to DAMON.
65 * @ops: monitoring operations set to register.
66 *
67 * This function registers a monitoring operations set of valid &struct
68 * damon_operations->id so that others can find and use them later.
69 *
70 * Return: 0 on success, negative error code otherwise.
71 */
damon_register_ops(struct damon_operations * ops)72 int damon_register_ops(struct damon_operations *ops)
73 {
74 int err = 0;
75
76 if (ops->id >= NR_DAMON_OPS)
77 return -EINVAL;
78 mutex_lock(&damon_ops_lock);
79 /* Fail for already registered ops */
80 if (__damon_is_registered_ops(ops->id)) {
81 err = -EINVAL;
82 goto out;
83 }
84 damon_registered_ops[ops->id] = *ops;
85 out:
86 mutex_unlock(&damon_ops_lock);
87 return err;
88 }
89
90 /**
91 * damon_select_ops() - Select a monitoring operations to use with the context.
92 * @ctx: monitoring context to use the operations.
93 * @id: id of the registered monitoring operations to select.
94 *
95 * This function finds registered monitoring operations set of @id and make
96 * @ctx to use it.
97 *
98 * Return: 0 on success, negative error code otherwise.
99 */
damon_select_ops(struct damon_ctx * ctx,enum damon_ops_id id)100 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
101 {
102 int err = 0;
103
104 if (id >= NR_DAMON_OPS)
105 return -EINVAL;
106
107 mutex_lock(&damon_ops_lock);
108 if (!__damon_is_registered_ops(id))
109 err = -EINVAL;
110 else
111 ctx->ops = damon_registered_ops[id];
112 mutex_unlock(&damon_ops_lock);
113 return err;
114 }
115
116 /*
117 * Construct a damon_region struct
118 *
119 * Returns the pointer to the new struct if success, or NULL otherwise
120 */
damon_new_region(unsigned long start,unsigned long end)121 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
122 {
123 struct damon_region *region;
124
125 region = kmem_cache_alloc(damon_region_cache, GFP_KERNEL);
126 if (!region)
127 return NULL;
128
129 region->ar.start = start;
130 region->ar.end = end;
131 region->nr_accesses = 0;
132 region->nr_accesses_bp = 0;
133 INIT_LIST_HEAD(®ion->list);
134
135 region->age = 0;
136 region->last_nr_accesses = 0;
137
138 return region;
139 }
140
damon_add_region(struct damon_region * r,struct damon_target * t)141 void damon_add_region(struct damon_region *r, struct damon_target *t)
142 {
143 list_add_tail(&r->list, &t->regions_list);
144 t->nr_regions++;
145 }
146
damon_del_region(struct damon_region * r,struct damon_target * t)147 static void damon_del_region(struct damon_region *r, struct damon_target *t)
148 {
149 list_del(&r->list);
150 t->nr_regions--;
151 }
152
damon_free_region(struct damon_region * r)153 static void damon_free_region(struct damon_region *r)
154 {
155 kmem_cache_free(damon_region_cache, r);
156 }
157
damon_destroy_region(struct damon_region * r,struct damon_target * t)158 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
159 {
160 damon_del_region(r, t);
161 damon_free_region(r);
162 }
163
164 /*
165 * Check whether a region is intersecting an address range
166 *
167 * Returns true if it is.
168 */
damon_intersect(struct damon_region * r,struct damon_addr_range * re)169 static bool damon_intersect(struct damon_region *r,
170 struct damon_addr_range *re)
171 {
172 return !(r->ar.end <= re->start || re->end <= r->ar.start);
173 }
174
175 /*
176 * Fill holes in regions with new regions.
177 */
damon_fill_regions_holes(struct damon_region * first,struct damon_region * last,struct damon_target * t)178 static int damon_fill_regions_holes(struct damon_region *first,
179 struct damon_region *last, struct damon_target *t)
180 {
181 struct damon_region *r = first;
182
183 damon_for_each_region_from(r, t) {
184 struct damon_region *next, *newr;
185
186 if (r == last)
187 break;
188 next = damon_next_region(r);
189 if (r->ar.end != next->ar.start) {
190 newr = damon_new_region(r->ar.end, next->ar.start);
191 if (!newr)
192 return -ENOMEM;
193 damon_insert_region(newr, r, next, t);
194 }
195 }
196 return 0;
197 }
198
199 /*
200 * damon_set_regions() - Set regions of a target for given address ranges.
201 * @t: the given target.
202 * @ranges: array of new monitoring target ranges.
203 * @nr_ranges: length of @ranges.
204 *
205 * This function adds new regions to, or modify existing regions of a
206 * monitoring target to fit in specific ranges.
207 *
208 * Return: 0 if success, or negative error code otherwise.
209 */
damon_set_regions(struct damon_target * t,struct damon_addr_range * ranges,unsigned int nr_ranges)210 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
211 unsigned int nr_ranges)
212 {
213 struct damon_region *r, *next;
214 unsigned int i;
215 int err;
216
217 /* Remove regions which are not in the new ranges */
218 damon_for_each_region_safe(r, next, t) {
219 for (i = 0; i < nr_ranges; i++) {
220 if (damon_intersect(r, &ranges[i]))
221 break;
222 }
223 if (i == nr_ranges)
224 damon_destroy_region(r, t);
225 }
226
227 r = damon_first_region(t);
228 /* Add new regions or resize existing regions to fit in the ranges */
229 for (i = 0; i < nr_ranges; i++) {
230 struct damon_region *first = NULL, *last, *newr;
231 struct damon_addr_range *range;
232
233 range = &ranges[i];
234 /* Get the first/last regions intersecting with the range */
235 damon_for_each_region_from(r, t) {
236 if (damon_intersect(r, range)) {
237 if (!first)
238 first = r;
239 last = r;
240 }
241 if (r->ar.start >= range->end)
242 break;
243 }
244 if (!first) {
245 /* no region intersects with this range */
246 newr = damon_new_region(
247 ALIGN_DOWN(range->start,
248 DAMON_MIN_REGION),
249 ALIGN(range->end, DAMON_MIN_REGION));
250 if (!newr)
251 return -ENOMEM;
252 damon_insert_region(newr, damon_prev_region(r), r, t);
253 } else {
254 /* resize intersecting regions to fit in this range */
255 first->ar.start = ALIGN_DOWN(range->start,
256 DAMON_MIN_REGION);
257 last->ar.end = ALIGN(range->end, DAMON_MIN_REGION);
258
259 /* fill possible holes in the range */
260 err = damon_fill_regions_holes(first, last, t);
261 if (err)
262 return err;
263 }
264 }
265 return 0;
266 }
267
damos_new_filter(enum damos_filter_type type,bool matching)268 struct damos_filter *damos_new_filter(enum damos_filter_type type,
269 bool matching)
270 {
271 struct damos_filter *filter;
272
273 filter = kmalloc(sizeof(*filter), GFP_KERNEL);
274 if (!filter)
275 return NULL;
276 filter->type = type;
277 filter->matching = matching;
278 INIT_LIST_HEAD(&filter->list);
279 return filter;
280 }
281
damos_add_filter(struct damos * s,struct damos_filter * f)282 void damos_add_filter(struct damos *s, struct damos_filter *f)
283 {
284 list_add_tail(&f->list, &s->filters);
285 }
286
damos_del_filter(struct damos_filter * f)287 static void damos_del_filter(struct damos_filter *f)
288 {
289 list_del(&f->list);
290 }
291
damos_free_filter(struct damos_filter * f)292 static void damos_free_filter(struct damos_filter *f)
293 {
294 kfree(f);
295 }
296
damos_destroy_filter(struct damos_filter * f)297 void damos_destroy_filter(struct damos_filter *f)
298 {
299 damos_del_filter(f);
300 damos_free_filter(f);
301 }
302
damos_new_quota_goal(enum damos_quota_goal_metric metric,unsigned long target_value)303 struct damos_quota_goal *damos_new_quota_goal(
304 enum damos_quota_goal_metric metric,
305 unsigned long target_value)
306 {
307 struct damos_quota_goal *goal;
308
309 goal = kmalloc(sizeof(*goal), GFP_KERNEL);
310 if (!goal)
311 return NULL;
312 goal->metric = metric;
313 goal->target_value = target_value;
314 INIT_LIST_HEAD(&goal->list);
315 return goal;
316 }
317
damos_add_quota_goal(struct damos_quota * q,struct damos_quota_goal * g)318 void damos_add_quota_goal(struct damos_quota *q, struct damos_quota_goal *g)
319 {
320 list_add_tail(&g->list, &q->goals);
321 }
322
damos_del_quota_goal(struct damos_quota_goal * g)323 static void damos_del_quota_goal(struct damos_quota_goal *g)
324 {
325 list_del(&g->list);
326 }
327
damos_free_quota_goal(struct damos_quota_goal * g)328 static void damos_free_quota_goal(struct damos_quota_goal *g)
329 {
330 kfree(g);
331 }
332
damos_destroy_quota_goal(struct damos_quota_goal * g)333 void damos_destroy_quota_goal(struct damos_quota_goal *g)
334 {
335 damos_del_quota_goal(g);
336 damos_free_quota_goal(g);
337 }
338
339 /* initialize fields of @quota that normally API users wouldn't set */
damos_quota_init(struct damos_quota * quota)340 static struct damos_quota *damos_quota_init(struct damos_quota *quota)
341 {
342 quota->esz = 0;
343 quota->total_charged_sz = 0;
344 quota->total_charged_ns = 0;
345 quota->charged_sz = 0;
346 quota->charged_from = 0;
347 quota->charge_target_from = NULL;
348 quota->charge_addr_from = 0;
349 quota->esz_bp = 0;
350 return quota;
351 }
352
damon_new_scheme(struct damos_access_pattern * pattern,enum damos_action action,unsigned long apply_interval_us,struct damos_quota * quota,struct damos_watermarks * wmarks)353 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
354 enum damos_action action,
355 unsigned long apply_interval_us,
356 struct damos_quota *quota,
357 struct damos_watermarks *wmarks)
358 {
359 struct damos *scheme;
360
361 scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
362 if (!scheme)
363 return NULL;
364 scheme->pattern = *pattern;
365 scheme->action = action;
366 scheme->apply_interval_us = apply_interval_us;
367 /*
368 * next_apply_sis will be set when kdamond starts. While kdamond is
369 * running, it will also updated when it is added to the DAMON context,
370 * or damon_attrs are updated.
371 */
372 scheme->next_apply_sis = 0;
373 INIT_LIST_HEAD(&scheme->filters);
374 scheme->stat = (struct damos_stat){};
375 INIT_LIST_HEAD(&scheme->list);
376
377 scheme->quota = *(damos_quota_init(quota));
378 /* quota.goals should be separately set by caller */
379 INIT_LIST_HEAD(&scheme->quota.goals);
380
381 scheme->wmarks = *wmarks;
382 scheme->wmarks.activated = true;
383
384 return scheme;
385 }
386
damos_set_next_apply_sis(struct damos * s,struct damon_ctx * ctx)387 static void damos_set_next_apply_sis(struct damos *s, struct damon_ctx *ctx)
388 {
389 unsigned long sample_interval = ctx->attrs.sample_interval ?
390 ctx->attrs.sample_interval : 1;
391 unsigned long apply_interval = s->apply_interval_us ?
392 s->apply_interval_us : ctx->attrs.aggr_interval;
393
394 s->next_apply_sis = ctx->passed_sample_intervals +
395 apply_interval / sample_interval;
396 }
397
damon_add_scheme(struct damon_ctx * ctx,struct damos * s)398 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
399 {
400 list_add_tail(&s->list, &ctx->schemes);
401 damos_set_next_apply_sis(s, ctx);
402 }
403
damon_del_scheme(struct damos * s)404 static void damon_del_scheme(struct damos *s)
405 {
406 list_del(&s->list);
407 }
408
damon_free_scheme(struct damos * s)409 static void damon_free_scheme(struct damos *s)
410 {
411 kfree(s);
412 }
413
damon_destroy_scheme(struct damos * s)414 void damon_destroy_scheme(struct damos *s)
415 {
416 struct damos_quota_goal *g, *g_next;
417 struct damos_filter *f, *next;
418
419 damos_for_each_quota_goal_safe(g, g_next, &s->quota)
420 damos_destroy_quota_goal(g);
421
422 damos_for_each_filter_safe(f, next, s)
423 damos_destroy_filter(f);
424 damon_del_scheme(s);
425 damon_free_scheme(s);
426 }
427
428 /*
429 * Construct a damon_target struct
430 *
431 * Returns the pointer to the new struct if success, or NULL otherwise
432 */
damon_new_target(void)433 struct damon_target *damon_new_target(void)
434 {
435 struct damon_target *t;
436
437 t = kmalloc(sizeof(*t), GFP_KERNEL);
438 if (!t)
439 return NULL;
440
441 t->pid = NULL;
442 t->nr_regions = 0;
443 INIT_LIST_HEAD(&t->regions_list);
444 INIT_LIST_HEAD(&t->list);
445
446 return t;
447 }
448
damon_add_target(struct damon_ctx * ctx,struct damon_target * t)449 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
450 {
451 list_add_tail(&t->list, &ctx->adaptive_targets);
452 }
453
damon_targets_empty(struct damon_ctx * ctx)454 bool damon_targets_empty(struct damon_ctx *ctx)
455 {
456 return list_empty(&ctx->adaptive_targets);
457 }
458
damon_del_target(struct damon_target * t)459 static void damon_del_target(struct damon_target *t)
460 {
461 list_del(&t->list);
462 }
463
damon_free_target(struct damon_target * t)464 void damon_free_target(struct damon_target *t)
465 {
466 struct damon_region *r, *next;
467
468 damon_for_each_region_safe(r, next, t)
469 damon_free_region(r);
470 kfree(t);
471 }
472
damon_destroy_target(struct damon_target * t)473 void damon_destroy_target(struct damon_target *t)
474 {
475 damon_del_target(t);
476 damon_free_target(t);
477 }
478
damon_nr_regions(struct damon_target * t)479 unsigned int damon_nr_regions(struct damon_target *t)
480 {
481 return t->nr_regions;
482 }
483
damon_new_ctx(void)484 struct damon_ctx *damon_new_ctx(void)
485 {
486 struct damon_ctx *ctx;
487
488 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
489 if (!ctx)
490 return NULL;
491
492 init_completion(&ctx->kdamond_started);
493
494 ctx->attrs.sample_interval = 5 * 1000;
495 ctx->attrs.aggr_interval = 100 * 1000;
496 ctx->attrs.ops_update_interval = 60 * 1000 * 1000;
497
498 ctx->passed_sample_intervals = 0;
499 /* These will be set from kdamond_init_intervals_sis() */
500 ctx->next_aggregation_sis = 0;
501 ctx->next_ops_update_sis = 0;
502
503 mutex_init(&ctx->kdamond_lock);
504
505 ctx->attrs.min_nr_regions = 10;
506 ctx->attrs.max_nr_regions = 1000;
507
508 INIT_LIST_HEAD(&ctx->adaptive_targets);
509 INIT_LIST_HEAD(&ctx->schemes);
510
511 return ctx;
512 }
513
damon_destroy_targets(struct damon_ctx * ctx)514 static void damon_destroy_targets(struct damon_ctx *ctx)
515 {
516 struct damon_target *t, *next_t;
517
518 if (ctx->ops.cleanup) {
519 ctx->ops.cleanup(ctx);
520 return;
521 }
522
523 damon_for_each_target_safe(t, next_t, ctx)
524 damon_destroy_target(t);
525 }
526
damon_destroy_ctx(struct damon_ctx * ctx)527 void damon_destroy_ctx(struct damon_ctx *ctx)
528 {
529 struct damos *s, *next_s;
530
531 damon_destroy_targets(ctx);
532
533 damon_for_each_scheme_safe(s, next_s, ctx)
534 damon_destroy_scheme(s);
535
536 kfree(ctx);
537 }
538
damon_age_for_new_attrs(unsigned int age,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)539 static unsigned int damon_age_for_new_attrs(unsigned int age,
540 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
541 {
542 return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
543 }
544
545 /* convert access ratio in bp (per 10,000) to nr_accesses */
damon_accesses_bp_to_nr_accesses(unsigned int accesses_bp,struct damon_attrs * attrs)546 static unsigned int damon_accesses_bp_to_nr_accesses(
547 unsigned int accesses_bp, struct damon_attrs *attrs)
548 {
549 return accesses_bp * damon_max_nr_accesses(attrs) / 10000;
550 }
551
552 /* convert nr_accesses to access ratio in bp (per 10,000) */
damon_nr_accesses_to_accesses_bp(unsigned int nr_accesses,struct damon_attrs * attrs)553 static unsigned int damon_nr_accesses_to_accesses_bp(
554 unsigned int nr_accesses, struct damon_attrs *attrs)
555 {
556 return nr_accesses * 10000 / damon_max_nr_accesses(attrs);
557 }
558
damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)559 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
560 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
561 {
562 return damon_accesses_bp_to_nr_accesses(
563 damon_nr_accesses_to_accesses_bp(
564 nr_accesses, old_attrs),
565 new_attrs);
566 }
567
damon_update_monitoring_result(struct damon_region * r,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)568 static void damon_update_monitoring_result(struct damon_region *r,
569 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
570 {
571 r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
572 old_attrs, new_attrs);
573 r->nr_accesses_bp = r->nr_accesses * 10000;
574 r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
575 }
576
577 /*
578 * region->nr_accesses is the number of sampling intervals in the last
579 * aggregation interval that access to the region has found, and region->age is
580 * the number of aggregation intervals that its access pattern has maintained.
581 * For the reason, the real meaning of the two fields depend on current
582 * sampling interval and aggregation interval. This function updates
583 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
584 */
damon_update_monitoring_results(struct damon_ctx * ctx,struct damon_attrs * new_attrs)585 static void damon_update_monitoring_results(struct damon_ctx *ctx,
586 struct damon_attrs *new_attrs)
587 {
588 struct damon_attrs *old_attrs = &ctx->attrs;
589 struct damon_target *t;
590 struct damon_region *r;
591
592 /* if any interval is zero, simply forgive conversion */
593 if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
594 !new_attrs->sample_interval ||
595 !new_attrs->aggr_interval)
596 return;
597
598 damon_for_each_target(t, ctx)
599 damon_for_each_region(r, t)
600 damon_update_monitoring_result(
601 r, old_attrs, new_attrs);
602 }
603
604 /**
605 * damon_set_attrs() - Set attributes for the monitoring.
606 * @ctx: monitoring context
607 * @attrs: monitoring attributes
608 *
609 * This function should be called while the kdamond is not running, or an
610 * access check results aggregation is not ongoing (e.g., from
611 * &struct damon_callback->after_aggregation or
612 * &struct damon_callback->after_wmarks_check callbacks).
613 *
614 * Every time interval is in micro-seconds.
615 *
616 * Return: 0 on success, negative error code otherwise.
617 */
damon_set_attrs(struct damon_ctx * ctx,struct damon_attrs * attrs)618 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
619 {
620 unsigned long sample_interval = attrs->sample_interval ?
621 attrs->sample_interval : 1;
622 struct damos *s;
623
624 if (attrs->min_nr_regions < 3)
625 return -EINVAL;
626 if (attrs->min_nr_regions > attrs->max_nr_regions)
627 return -EINVAL;
628 if (attrs->sample_interval > attrs->aggr_interval)
629 return -EINVAL;
630
631 ctx->next_aggregation_sis = ctx->passed_sample_intervals +
632 attrs->aggr_interval / sample_interval;
633 ctx->next_ops_update_sis = ctx->passed_sample_intervals +
634 attrs->ops_update_interval / sample_interval;
635
636 damon_update_monitoring_results(ctx, attrs);
637 ctx->attrs = *attrs;
638
639 damon_for_each_scheme(s, ctx)
640 damos_set_next_apply_sis(s, ctx);
641
642 return 0;
643 }
644
645 /**
646 * damon_set_schemes() - Set data access monitoring based operation schemes.
647 * @ctx: monitoring context
648 * @schemes: array of the schemes
649 * @nr_schemes: number of entries in @schemes
650 *
651 * This function should not be called while the kdamond of the context is
652 * running.
653 */
damon_set_schemes(struct damon_ctx * ctx,struct damos ** schemes,ssize_t nr_schemes)654 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
655 ssize_t nr_schemes)
656 {
657 struct damos *s, *next;
658 ssize_t i;
659
660 damon_for_each_scheme_safe(s, next, ctx)
661 damon_destroy_scheme(s);
662 for (i = 0; i < nr_schemes; i++)
663 damon_add_scheme(ctx, schemes[i]);
664 }
665
666 /**
667 * damon_nr_running_ctxs() - Return number of currently running contexts.
668 */
damon_nr_running_ctxs(void)669 int damon_nr_running_ctxs(void)
670 {
671 int nr_ctxs;
672
673 mutex_lock(&damon_lock);
674 nr_ctxs = nr_running_ctxs;
675 mutex_unlock(&damon_lock);
676
677 return nr_ctxs;
678 }
679
680 /* Returns the size upper limit for each monitoring region */
damon_region_sz_limit(struct damon_ctx * ctx)681 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
682 {
683 struct damon_target *t;
684 struct damon_region *r;
685 unsigned long sz = 0;
686
687 damon_for_each_target(t, ctx) {
688 damon_for_each_region(r, t)
689 sz += damon_sz_region(r);
690 }
691
692 if (ctx->attrs.min_nr_regions)
693 sz /= ctx->attrs.min_nr_regions;
694 if (sz < DAMON_MIN_REGION)
695 sz = DAMON_MIN_REGION;
696
697 return sz;
698 }
699
700 static int kdamond_fn(void *data);
701
702 /*
703 * __damon_start() - Starts monitoring with given context.
704 * @ctx: monitoring context
705 *
706 * This function should be called while damon_lock is hold.
707 *
708 * Return: 0 on success, negative error code otherwise.
709 */
__damon_start(struct damon_ctx * ctx)710 static int __damon_start(struct damon_ctx *ctx)
711 {
712 int err = -EBUSY;
713
714 mutex_lock(&ctx->kdamond_lock);
715 if (!ctx->kdamond) {
716 err = 0;
717 reinit_completion(&ctx->kdamond_started);
718 ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
719 nr_running_ctxs);
720 if (IS_ERR(ctx->kdamond)) {
721 err = PTR_ERR(ctx->kdamond);
722 ctx->kdamond = NULL;
723 } else {
724 wait_for_completion(&ctx->kdamond_started);
725 }
726 }
727 mutex_unlock(&ctx->kdamond_lock);
728
729 return err;
730 }
731
732 /**
733 * damon_start() - Starts the monitorings for a given group of contexts.
734 * @ctxs: an array of the pointers for contexts to start monitoring
735 * @nr_ctxs: size of @ctxs
736 * @exclusive: exclusiveness of this contexts group
737 *
738 * This function starts a group of monitoring threads for a group of monitoring
739 * contexts. One thread per each context is created and run in parallel. The
740 * caller should handle synchronization between the threads by itself. If
741 * @exclusive is true and a group of threads that created by other
742 * 'damon_start()' call is currently running, this function does nothing but
743 * returns -EBUSY.
744 *
745 * Return: 0 on success, negative error code otherwise.
746 */
damon_start(struct damon_ctx ** ctxs,int nr_ctxs,bool exclusive)747 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
748 {
749 int i;
750 int err = 0;
751
752 mutex_lock(&damon_lock);
753 if ((exclusive && nr_running_ctxs) ||
754 (!exclusive && running_exclusive_ctxs)) {
755 mutex_unlock(&damon_lock);
756 return -EBUSY;
757 }
758
759 for (i = 0; i < nr_ctxs; i++) {
760 err = __damon_start(ctxs[i]);
761 if (err)
762 break;
763 nr_running_ctxs++;
764 }
765 if (exclusive && nr_running_ctxs)
766 running_exclusive_ctxs = true;
767 mutex_unlock(&damon_lock);
768
769 return err;
770 }
771
772 /*
773 * __damon_stop() - Stops monitoring of a given context.
774 * @ctx: monitoring context
775 *
776 * Return: 0 on success, negative error code otherwise.
777 */
__damon_stop(struct damon_ctx * ctx)778 static int __damon_stop(struct damon_ctx *ctx)
779 {
780 struct task_struct *tsk;
781
782 mutex_lock(&ctx->kdamond_lock);
783 tsk = ctx->kdamond;
784 if (tsk) {
785 get_task_struct(tsk);
786 mutex_unlock(&ctx->kdamond_lock);
787 kthread_stop_put(tsk);
788 return 0;
789 }
790 mutex_unlock(&ctx->kdamond_lock);
791
792 return -EPERM;
793 }
794
795 /**
796 * damon_stop() - Stops the monitorings for a given group of contexts.
797 * @ctxs: an array of the pointers for contexts to stop monitoring
798 * @nr_ctxs: size of @ctxs
799 *
800 * Return: 0 on success, negative error code otherwise.
801 */
damon_stop(struct damon_ctx ** ctxs,int nr_ctxs)802 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
803 {
804 int i, err = 0;
805
806 for (i = 0; i < nr_ctxs; i++) {
807 /* nr_running_ctxs is decremented in kdamond_fn */
808 err = __damon_stop(ctxs[i]);
809 if (err)
810 break;
811 }
812 return err;
813 }
814
815 /*
816 * Reset the aggregated monitoring results ('nr_accesses' of each region).
817 */
kdamond_reset_aggregated(struct damon_ctx * c)818 static void kdamond_reset_aggregated(struct damon_ctx *c)
819 {
820 struct damon_target *t;
821 unsigned int ti = 0; /* target's index */
822
823 damon_for_each_target(t, c) {
824 struct damon_region *r;
825
826 damon_for_each_region(r, t) {
827 trace_damon_aggregated(ti, r, damon_nr_regions(t));
828 r->last_nr_accesses = r->nr_accesses;
829 r->nr_accesses = 0;
830 }
831 ti++;
832 }
833 }
834
835 static void damon_split_region_at(struct damon_target *t,
836 struct damon_region *r, unsigned long sz_r);
837
__damos_valid_target(struct damon_region * r,struct damos * s)838 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
839 {
840 unsigned long sz;
841 unsigned int nr_accesses = r->nr_accesses_bp / 10000;
842
843 sz = damon_sz_region(r);
844 return s->pattern.min_sz_region <= sz &&
845 sz <= s->pattern.max_sz_region &&
846 s->pattern.min_nr_accesses <= nr_accesses &&
847 nr_accesses <= s->pattern.max_nr_accesses &&
848 s->pattern.min_age_region <= r->age &&
849 r->age <= s->pattern.max_age_region;
850 }
851
damos_valid_target(struct damon_ctx * c,struct damon_target * t,struct damon_region * r,struct damos * s)852 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
853 struct damon_region *r, struct damos *s)
854 {
855 bool ret = __damos_valid_target(r, s);
856
857 if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
858 return ret;
859
860 return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
861 }
862
863 /*
864 * damos_skip_charged_region() - Check if the given region or starting part of
865 * it is already charged for the DAMOS quota.
866 * @t: The target of the region.
867 * @rp: The pointer to the region.
868 * @s: The scheme to be applied.
869 *
870 * If a quota of a scheme has exceeded in a quota charge window, the scheme's
871 * action would applied to only a part of the target access pattern fulfilling
872 * regions. To avoid applying the scheme action to only already applied
873 * regions, DAMON skips applying the scheme action to the regions that charged
874 * in the previous charge window.
875 *
876 * This function checks if a given region should be skipped or not for the
877 * reason. If only the starting part of the region has previously charged,
878 * this function splits the region into two so that the second one covers the
879 * area that not charged in the previous charge widnow and saves the second
880 * region in *rp and returns false, so that the caller can apply DAMON action
881 * to the second one.
882 *
883 * Return: true if the region should be entirely skipped, false otherwise.
884 */
damos_skip_charged_region(struct damon_target * t,struct damon_region ** rp,struct damos * s)885 static bool damos_skip_charged_region(struct damon_target *t,
886 struct damon_region **rp, struct damos *s)
887 {
888 struct damon_region *r = *rp;
889 struct damos_quota *quota = &s->quota;
890 unsigned long sz_to_skip;
891
892 /* Skip previously charged regions */
893 if (quota->charge_target_from) {
894 if (t != quota->charge_target_from)
895 return true;
896 if (r == damon_last_region(t)) {
897 quota->charge_target_from = NULL;
898 quota->charge_addr_from = 0;
899 return true;
900 }
901 if (quota->charge_addr_from &&
902 r->ar.end <= quota->charge_addr_from)
903 return true;
904
905 if (quota->charge_addr_from && r->ar.start <
906 quota->charge_addr_from) {
907 sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
908 r->ar.start, DAMON_MIN_REGION);
909 if (!sz_to_skip) {
910 if (damon_sz_region(r) <= DAMON_MIN_REGION)
911 return true;
912 sz_to_skip = DAMON_MIN_REGION;
913 }
914 damon_split_region_at(t, r, sz_to_skip);
915 r = damon_next_region(r);
916 *rp = r;
917 }
918 quota->charge_target_from = NULL;
919 quota->charge_addr_from = 0;
920 }
921 return false;
922 }
923
damos_update_stat(struct damos * s,unsigned long sz_tried,unsigned long sz_applied)924 static void damos_update_stat(struct damos *s,
925 unsigned long sz_tried, unsigned long sz_applied)
926 {
927 s->stat.nr_tried++;
928 s->stat.sz_tried += sz_tried;
929 if (sz_applied)
930 s->stat.nr_applied++;
931 s->stat.sz_applied += sz_applied;
932 }
933
__damos_filter_out(struct damon_ctx * ctx,struct damon_target * t,struct damon_region * r,struct damos_filter * filter)934 static bool __damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
935 struct damon_region *r, struct damos_filter *filter)
936 {
937 bool matched = false;
938 struct damon_target *ti;
939 int target_idx = 0;
940 unsigned long start, end;
941
942 switch (filter->type) {
943 case DAMOS_FILTER_TYPE_TARGET:
944 damon_for_each_target(ti, ctx) {
945 if (ti == t)
946 break;
947 target_idx++;
948 }
949 matched = target_idx == filter->target_idx;
950 break;
951 case DAMOS_FILTER_TYPE_ADDR:
952 start = ALIGN_DOWN(filter->addr_range.start, DAMON_MIN_REGION);
953 end = ALIGN_DOWN(filter->addr_range.end, DAMON_MIN_REGION);
954
955 /* inside the range */
956 if (start <= r->ar.start && r->ar.end <= end) {
957 matched = true;
958 break;
959 }
960 /* outside of the range */
961 if (r->ar.end <= start || end <= r->ar.start) {
962 matched = false;
963 break;
964 }
965 /* start before the range and overlap */
966 if (r->ar.start < start) {
967 damon_split_region_at(t, r, start - r->ar.start);
968 matched = false;
969 break;
970 }
971 /* start inside the range */
972 damon_split_region_at(t, r, end - r->ar.start);
973 matched = true;
974 break;
975 default:
976 return false;
977 }
978
979 return matched == filter->matching;
980 }
981
damos_filter_out(struct damon_ctx * ctx,struct damon_target * t,struct damon_region * r,struct damos * s)982 static bool damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
983 struct damon_region *r, struct damos *s)
984 {
985 struct damos_filter *filter;
986
987 damos_for_each_filter(filter, s) {
988 if (__damos_filter_out(ctx, t, r, filter))
989 return true;
990 }
991 return false;
992 }
993
damos_apply_scheme(struct damon_ctx * c,struct damon_target * t,struct damon_region * r,struct damos * s)994 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
995 struct damon_region *r, struct damos *s)
996 {
997 struct damos_quota *quota = &s->quota;
998 unsigned long sz = damon_sz_region(r);
999 struct timespec64 begin, end;
1000 unsigned long sz_applied = 0;
1001 int err = 0;
1002 /*
1003 * We plan to support multiple context per kdamond, as DAMON sysfs
1004 * implies with 'nr_contexts' file. Nevertheless, only single context
1005 * per kdamond is supported for now. So, we can simply use '0' context
1006 * index here.
1007 */
1008 unsigned int cidx = 0;
1009 struct damos *siter; /* schemes iterator */
1010 unsigned int sidx = 0;
1011 struct damon_target *titer; /* targets iterator */
1012 unsigned int tidx = 0;
1013 bool do_trace = false;
1014
1015 /* get indices for trace_damos_before_apply() */
1016 if (trace_damos_before_apply_enabled()) {
1017 damon_for_each_scheme(siter, c) {
1018 if (siter == s)
1019 break;
1020 sidx++;
1021 }
1022 damon_for_each_target(titer, c) {
1023 if (titer == t)
1024 break;
1025 tidx++;
1026 }
1027 do_trace = true;
1028 }
1029
1030 if (c->ops.apply_scheme) {
1031 if (quota->esz && quota->charged_sz + sz > quota->esz) {
1032 sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
1033 DAMON_MIN_REGION);
1034 if (!sz)
1035 goto update_stat;
1036 damon_split_region_at(t, r, sz);
1037 }
1038 if (damos_filter_out(c, t, r, s))
1039 return;
1040 ktime_get_coarse_ts64(&begin);
1041 if (c->callback.before_damos_apply)
1042 err = c->callback.before_damos_apply(c, t, r, s);
1043 if (!err) {
1044 trace_damos_before_apply(cidx, sidx, tidx, r,
1045 damon_nr_regions(t), do_trace);
1046 sz_applied = c->ops.apply_scheme(c, t, r, s);
1047 }
1048 ktime_get_coarse_ts64(&end);
1049 quota->total_charged_ns += timespec64_to_ns(&end) -
1050 timespec64_to_ns(&begin);
1051 quota->charged_sz += sz;
1052 if (quota->esz && quota->charged_sz >= quota->esz) {
1053 quota->charge_target_from = t;
1054 quota->charge_addr_from = r->ar.end + 1;
1055 }
1056 }
1057 if (s->action != DAMOS_STAT)
1058 r->age = 0;
1059
1060 update_stat:
1061 damos_update_stat(s, sz, sz_applied);
1062 }
1063
damon_do_apply_schemes(struct damon_ctx * c,struct damon_target * t,struct damon_region * r)1064 static void damon_do_apply_schemes(struct damon_ctx *c,
1065 struct damon_target *t,
1066 struct damon_region *r)
1067 {
1068 struct damos *s;
1069
1070 damon_for_each_scheme(s, c) {
1071 struct damos_quota *quota = &s->quota;
1072
1073 if (c->passed_sample_intervals != s->next_apply_sis)
1074 continue;
1075
1076 if (!s->wmarks.activated)
1077 continue;
1078
1079 /* Check the quota */
1080 if (quota->esz && quota->charged_sz >= quota->esz)
1081 continue;
1082
1083 if (damos_skip_charged_region(t, &r, s))
1084 continue;
1085
1086 if (!damos_valid_target(c, t, r, s))
1087 continue;
1088
1089 damos_apply_scheme(c, t, r, s);
1090 }
1091 }
1092
1093 /*
1094 * damon_feed_loop_next_input() - get next input to achieve a target score.
1095 * @last_input The last input.
1096 * @score Current score that made with @last_input.
1097 *
1098 * Calculate next input to achieve the target score, based on the last input
1099 * and current score. Assuming the input and the score are positively
1100 * proportional, calculate how much compensation should be added to or
1101 * subtracted from the last input as a proportion of the last input. Avoid
1102 * next input always being zero by setting it non-zero always. In short form
1103 * (assuming support of float and signed calculations), the algorithm is as
1104 * below.
1105 *
1106 * next_input = max(last_input * ((goal - current) / goal + 1), 1)
1107 *
1108 * For simple implementation, we assume the target score is always 10,000. The
1109 * caller should adjust @score for this.
1110 *
1111 * Returns next input that assumed to achieve the target score.
1112 */
damon_feed_loop_next_input(unsigned long last_input,unsigned long score)1113 static unsigned long damon_feed_loop_next_input(unsigned long last_input,
1114 unsigned long score)
1115 {
1116 const unsigned long goal = 10000;
1117 unsigned long score_goal_diff = max(goal, score) - min(goal, score);
1118 unsigned long score_goal_diff_bp = score_goal_diff * 10000 / goal;
1119 unsigned long compensation = last_input * score_goal_diff_bp / 10000;
1120 /* Set minimum input as 10000 to avoid compensation be zero */
1121 const unsigned long min_input = 10000;
1122
1123 if (goal > score)
1124 return last_input + compensation;
1125 if (last_input > compensation + min_input)
1126 return last_input - compensation;
1127 return min_input;
1128 }
1129
1130 #ifdef CONFIG_PSI
1131
damos_get_some_mem_psi_total(void)1132 static u64 damos_get_some_mem_psi_total(void)
1133 {
1134 if (static_branch_likely(&psi_disabled))
1135 return 0;
1136 return div_u64(psi_system.total[PSI_AVGS][PSI_MEM * 2],
1137 NSEC_PER_USEC);
1138 }
1139
1140 #else /* CONFIG_PSI */
1141
damos_get_some_mem_psi_total(void)1142 static inline u64 damos_get_some_mem_psi_total(void)
1143 {
1144 return 0;
1145 };
1146
1147 #endif /* CONFIG_PSI */
1148
damos_set_quota_goal_current_value(struct damos_quota_goal * goal)1149 static void damos_set_quota_goal_current_value(struct damos_quota_goal *goal)
1150 {
1151 u64 now_psi_total;
1152
1153 switch (goal->metric) {
1154 case DAMOS_QUOTA_USER_INPUT:
1155 /* User should already set goal->current_value */
1156 break;
1157 case DAMOS_QUOTA_SOME_MEM_PSI_US:
1158 now_psi_total = damos_get_some_mem_psi_total();
1159 goal->current_value = now_psi_total - goal->last_psi_total;
1160 goal->last_psi_total = now_psi_total;
1161 break;
1162 default:
1163 break;
1164 }
1165 }
1166
1167 /* Return the highest score since it makes schemes least aggressive */
damos_quota_score(struct damos_quota * quota)1168 static unsigned long damos_quota_score(struct damos_quota *quota)
1169 {
1170 struct damos_quota_goal *goal;
1171 unsigned long highest_score = 0;
1172
1173 damos_for_each_quota_goal(goal, quota) {
1174 damos_set_quota_goal_current_value(goal);
1175 highest_score = max(highest_score,
1176 goal->current_value * 10000 /
1177 goal->target_value);
1178 }
1179
1180 return highest_score;
1181 }
1182
1183 /*
1184 * Called only if quota->ms, or quota->sz are set, or quota->goals is not empty
1185 */
damos_set_effective_quota(struct damos_quota * quota)1186 static void damos_set_effective_quota(struct damos_quota *quota)
1187 {
1188 unsigned long throughput;
1189 unsigned long esz;
1190
1191 if (!quota->ms && list_empty("a->goals)) {
1192 quota->esz = quota->sz;
1193 return;
1194 }
1195
1196 if (!list_empty("a->goals)) {
1197 unsigned long score = damos_quota_score(quota);
1198
1199 quota->esz_bp = damon_feed_loop_next_input(
1200 max(quota->esz_bp, 10000UL),
1201 score);
1202 esz = quota->esz_bp / 10000;
1203 }
1204
1205 if (quota->ms) {
1206 if (quota->total_charged_ns)
1207 throughput = quota->total_charged_sz * 1000000 /
1208 quota->total_charged_ns;
1209 else
1210 throughput = PAGE_SIZE * 1024;
1211 if (!list_empty("a->goals))
1212 esz = min(throughput * quota->ms, esz);
1213 else
1214 esz = throughput * quota->ms;
1215 }
1216
1217 if (quota->sz && quota->sz < esz)
1218 esz = quota->sz;
1219
1220 quota->esz = esz;
1221 }
1222
damos_adjust_quota(struct damon_ctx * c,struct damos * s)1223 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
1224 {
1225 struct damos_quota *quota = &s->quota;
1226 struct damon_target *t;
1227 struct damon_region *r;
1228 unsigned long cumulated_sz;
1229 unsigned int score, max_score = 0;
1230
1231 if (!quota->ms && !quota->sz && list_empty("a->goals))
1232 return;
1233
1234 /* New charge window starts */
1235 if (time_after_eq(jiffies, quota->charged_from +
1236 msecs_to_jiffies(quota->reset_interval))) {
1237 if (quota->esz && quota->charged_sz >= quota->esz)
1238 s->stat.qt_exceeds++;
1239 quota->total_charged_sz += quota->charged_sz;
1240 quota->charged_from = jiffies;
1241 quota->charged_sz = 0;
1242 damos_set_effective_quota(quota);
1243 }
1244
1245 if (!c->ops.get_scheme_score)
1246 return;
1247
1248 /* Fill up the score histogram */
1249 memset(quota->histogram, 0, sizeof(quota->histogram));
1250 damon_for_each_target(t, c) {
1251 damon_for_each_region(r, t) {
1252 if (!__damos_valid_target(r, s))
1253 continue;
1254 score = c->ops.get_scheme_score(c, t, r, s);
1255 quota->histogram[score] += damon_sz_region(r);
1256 if (score > max_score)
1257 max_score = score;
1258 }
1259 }
1260
1261 /* Set the min score limit */
1262 for (cumulated_sz = 0, score = max_score; ; score--) {
1263 cumulated_sz += quota->histogram[score];
1264 if (cumulated_sz >= quota->esz || !score)
1265 break;
1266 }
1267 quota->min_score = score;
1268 }
1269
kdamond_apply_schemes(struct damon_ctx * c)1270 static void kdamond_apply_schemes(struct damon_ctx *c)
1271 {
1272 struct damon_target *t;
1273 struct damon_region *r, *next_r;
1274 struct damos *s;
1275 unsigned long sample_interval = c->attrs.sample_interval ?
1276 c->attrs.sample_interval : 1;
1277 bool has_schemes_to_apply = false;
1278
1279 damon_for_each_scheme(s, c) {
1280 if (c->passed_sample_intervals != s->next_apply_sis)
1281 continue;
1282
1283 if (!s->wmarks.activated)
1284 continue;
1285
1286 has_schemes_to_apply = true;
1287
1288 damos_adjust_quota(c, s);
1289 }
1290
1291 if (!has_schemes_to_apply)
1292 return;
1293
1294 damon_for_each_target(t, c) {
1295 damon_for_each_region_safe(r, next_r, t)
1296 damon_do_apply_schemes(c, t, r);
1297 }
1298
1299 damon_for_each_scheme(s, c) {
1300 if (c->passed_sample_intervals != s->next_apply_sis)
1301 continue;
1302 s->next_apply_sis +=
1303 (s->apply_interval_us ? s->apply_interval_us :
1304 c->attrs.aggr_interval) / sample_interval;
1305 }
1306 }
1307
1308 /*
1309 * Merge two adjacent regions into one region
1310 */
damon_merge_two_regions(struct damon_target * t,struct damon_region * l,struct damon_region * r)1311 static void damon_merge_two_regions(struct damon_target *t,
1312 struct damon_region *l, struct damon_region *r)
1313 {
1314 unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1315
1316 l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1317 (sz_l + sz_r);
1318 l->nr_accesses_bp = l->nr_accesses * 10000;
1319 l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1320 l->ar.end = r->ar.end;
1321 damon_destroy_region(r, t);
1322 }
1323
1324 /*
1325 * Merge adjacent regions having similar access frequencies
1326 *
1327 * t target affected by this merge operation
1328 * thres '->nr_accesses' diff threshold for the merge
1329 * sz_limit size upper limit of each region
1330 */
damon_merge_regions_of(struct damon_target * t,unsigned int thres,unsigned long sz_limit)1331 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1332 unsigned long sz_limit)
1333 {
1334 struct damon_region *r, *prev = NULL, *next;
1335
1336 damon_for_each_region_safe(r, next, t) {
1337 if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1338 r->age = 0;
1339 else
1340 r->age++;
1341
1342 if (prev && prev->ar.end == r->ar.start &&
1343 abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1344 damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1345 damon_merge_two_regions(t, prev, r);
1346 else
1347 prev = r;
1348 }
1349 }
1350
1351 /*
1352 * Merge adjacent regions having similar access frequencies
1353 *
1354 * threshold '->nr_accesses' diff threshold for the merge
1355 * sz_limit size upper limit of each region
1356 *
1357 * This function merges monitoring target regions which are adjacent and their
1358 * access frequencies are similar. This is for minimizing the monitoring
1359 * overhead under the dynamically changeable access pattern. If a merge was
1360 * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1361 */
kdamond_merge_regions(struct damon_ctx * c,unsigned int threshold,unsigned long sz_limit)1362 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1363 unsigned long sz_limit)
1364 {
1365 struct damon_target *t;
1366
1367 damon_for_each_target(t, c)
1368 damon_merge_regions_of(t, threshold, sz_limit);
1369 }
1370
1371 /*
1372 * Split a region in two
1373 *
1374 * r the region to be split
1375 * sz_r size of the first sub-region that will be made
1376 */
damon_split_region_at(struct damon_target * t,struct damon_region * r,unsigned long sz_r)1377 static void damon_split_region_at(struct damon_target *t,
1378 struct damon_region *r, unsigned long sz_r)
1379 {
1380 struct damon_region *new;
1381
1382 new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1383 if (!new)
1384 return;
1385
1386 r->ar.end = new->ar.start;
1387
1388 new->age = r->age;
1389 new->last_nr_accesses = r->last_nr_accesses;
1390 new->nr_accesses_bp = r->nr_accesses_bp;
1391 new->nr_accesses = r->nr_accesses;
1392
1393 damon_insert_region(new, r, damon_next_region(r), t);
1394 }
1395
1396 /* Split every region in the given target into 'nr_subs' regions */
damon_split_regions_of(struct damon_target * t,int nr_subs)1397 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1398 {
1399 struct damon_region *r, *next;
1400 unsigned long sz_region, sz_sub = 0;
1401 int i;
1402
1403 damon_for_each_region_safe(r, next, t) {
1404 sz_region = damon_sz_region(r);
1405
1406 for (i = 0; i < nr_subs - 1 &&
1407 sz_region > 2 * DAMON_MIN_REGION; i++) {
1408 /*
1409 * Randomly select size of left sub-region to be at
1410 * least 10 percent and at most 90% of original region
1411 */
1412 sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1413 sz_region / 10, DAMON_MIN_REGION);
1414 /* Do not allow blank region */
1415 if (sz_sub == 0 || sz_sub >= sz_region)
1416 continue;
1417
1418 damon_split_region_at(t, r, sz_sub);
1419 sz_region = sz_sub;
1420 }
1421 }
1422 }
1423
1424 /*
1425 * Split every target region into randomly-sized small regions
1426 *
1427 * This function splits every target region into random-sized small regions if
1428 * current total number of the regions is equal or smaller than half of the
1429 * user-specified maximum number of regions. This is for maximizing the
1430 * monitoring accuracy under the dynamically changeable access patterns. If a
1431 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1432 * it.
1433 */
kdamond_split_regions(struct damon_ctx * ctx)1434 static void kdamond_split_regions(struct damon_ctx *ctx)
1435 {
1436 struct damon_target *t;
1437 unsigned int nr_regions = 0;
1438 static unsigned int last_nr_regions;
1439 int nr_subregions = 2;
1440
1441 damon_for_each_target(t, ctx)
1442 nr_regions += damon_nr_regions(t);
1443
1444 if (nr_regions > ctx->attrs.max_nr_regions / 2)
1445 return;
1446
1447 /* Maybe the middle of the region has different access frequency */
1448 if (last_nr_regions == nr_regions &&
1449 nr_regions < ctx->attrs.max_nr_regions / 3)
1450 nr_subregions = 3;
1451
1452 damon_for_each_target(t, ctx)
1453 damon_split_regions_of(t, nr_subregions);
1454
1455 last_nr_regions = nr_regions;
1456 }
1457
1458 /*
1459 * Check whether current monitoring should be stopped
1460 *
1461 * The monitoring is stopped when either the user requested to stop, or all
1462 * monitoring targets are invalid.
1463 *
1464 * Returns true if need to stop current monitoring.
1465 */
kdamond_need_stop(struct damon_ctx * ctx)1466 static bool kdamond_need_stop(struct damon_ctx *ctx)
1467 {
1468 struct damon_target *t;
1469
1470 if (kthread_should_stop())
1471 return true;
1472
1473 if (!ctx->ops.target_valid)
1474 return false;
1475
1476 damon_for_each_target(t, ctx) {
1477 if (ctx->ops.target_valid(t))
1478 return false;
1479 }
1480
1481 return true;
1482 }
1483
damos_get_wmark_metric_value(enum damos_wmark_metric metric,unsigned long * metric_value)1484 static int damos_get_wmark_metric_value(enum damos_wmark_metric metric,
1485 unsigned long *metric_value)
1486 {
1487 switch (metric) {
1488 case DAMOS_WMARK_FREE_MEM_RATE:
1489 *metric_value = global_zone_page_state(NR_FREE_PAGES) * 1000 /
1490 totalram_pages();
1491 return 0;
1492 default:
1493 break;
1494 }
1495 return -EINVAL;
1496 }
1497
1498 /*
1499 * Returns zero if the scheme is active. Else, returns time to wait for next
1500 * watermark check in micro-seconds.
1501 */
damos_wmark_wait_us(struct damos * scheme)1502 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1503 {
1504 unsigned long metric;
1505
1506 if (damos_get_wmark_metric_value(scheme->wmarks.metric, &metric))
1507 return 0;
1508
1509 /* higher than high watermark or lower than low watermark */
1510 if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1511 if (scheme->wmarks.activated)
1512 pr_debug("deactivate a scheme (%d) for %s wmark\n",
1513 scheme->action,
1514 metric > scheme->wmarks.high ?
1515 "high" : "low");
1516 scheme->wmarks.activated = false;
1517 return scheme->wmarks.interval;
1518 }
1519
1520 /* inactive and higher than middle watermark */
1521 if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1522 !scheme->wmarks.activated)
1523 return scheme->wmarks.interval;
1524
1525 if (!scheme->wmarks.activated)
1526 pr_debug("activate a scheme (%d)\n", scheme->action);
1527 scheme->wmarks.activated = true;
1528 return 0;
1529 }
1530
kdamond_usleep(unsigned long usecs)1531 static void kdamond_usleep(unsigned long usecs)
1532 {
1533 /* See Documentation/timers/timers-howto.rst for the thresholds */
1534 if (usecs > 20 * USEC_PER_MSEC)
1535 schedule_timeout_idle(usecs_to_jiffies(usecs));
1536 else
1537 usleep_idle_range(usecs, usecs + 1);
1538 }
1539
1540 /* Returns negative error code if it's not activated but should return */
kdamond_wait_activation(struct damon_ctx * ctx)1541 static int kdamond_wait_activation(struct damon_ctx *ctx)
1542 {
1543 struct damos *s;
1544 unsigned long wait_time;
1545 unsigned long min_wait_time = 0;
1546 bool init_wait_time = false;
1547
1548 while (!kdamond_need_stop(ctx)) {
1549 damon_for_each_scheme(s, ctx) {
1550 wait_time = damos_wmark_wait_us(s);
1551 if (!init_wait_time || wait_time < min_wait_time) {
1552 init_wait_time = true;
1553 min_wait_time = wait_time;
1554 }
1555 }
1556 if (!min_wait_time)
1557 return 0;
1558
1559 kdamond_usleep(min_wait_time);
1560
1561 if (ctx->callback.after_wmarks_check &&
1562 ctx->callback.after_wmarks_check(ctx))
1563 break;
1564 }
1565 return -EBUSY;
1566 }
1567
kdamond_init_intervals_sis(struct damon_ctx * ctx)1568 static void kdamond_init_intervals_sis(struct damon_ctx *ctx)
1569 {
1570 unsigned long sample_interval = ctx->attrs.sample_interval ?
1571 ctx->attrs.sample_interval : 1;
1572 unsigned long apply_interval;
1573 struct damos *scheme;
1574
1575 ctx->passed_sample_intervals = 0;
1576 ctx->next_aggregation_sis = ctx->attrs.aggr_interval / sample_interval;
1577 ctx->next_ops_update_sis = ctx->attrs.ops_update_interval /
1578 sample_interval;
1579
1580 damon_for_each_scheme(scheme, ctx) {
1581 apply_interval = scheme->apply_interval_us ?
1582 scheme->apply_interval_us : ctx->attrs.aggr_interval;
1583 scheme->next_apply_sis = apply_interval / sample_interval;
1584 }
1585 }
1586
1587 /*
1588 * The monitoring daemon that runs as a kernel thread
1589 */
kdamond_fn(void * data)1590 static int kdamond_fn(void *data)
1591 {
1592 struct damon_ctx *ctx = data;
1593 struct damon_target *t;
1594 struct damon_region *r, *next;
1595 unsigned int max_nr_accesses = 0;
1596 unsigned long sz_limit = 0;
1597
1598 pr_debug("kdamond (%d) starts\n", current->pid);
1599
1600 complete(&ctx->kdamond_started);
1601 kdamond_init_intervals_sis(ctx);
1602
1603 if (ctx->ops.init)
1604 ctx->ops.init(ctx);
1605 if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1606 goto done;
1607
1608 sz_limit = damon_region_sz_limit(ctx);
1609
1610 while (!kdamond_need_stop(ctx)) {
1611 /*
1612 * ctx->attrs and ctx->next_{aggregation,ops_update}_sis could
1613 * be changed from after_wmarks_check() or after_aggregation()
1614 * callbacks. Read the values here, and use those for this
1615 * iteration. That is, damon_set_attrs() updated new values
1616 * are respected from next iteration.
1617 */
1618 unsigned long next_aggregation_sis = ctx->next_aggregation_sis;
1619 unsigned long next_ops_update_sis = ctx->next_ops_update_sis;
1620 unsigned long sample_interval = ctx->attrs.sample_interval;
1621
1622 if (kdamond_wait_activation(ctx))
1623 break;
1624
1625 if (ctx->ops.prepare_access_checks)
1626 ctx->ops.prepare_access_checks(ctx);
1627 if (ctx->callback.after_sampling &&
1628 ctx->callback.after_sampling(ctx))
1629 break;
1630
1631 kdamond_usleep(sample_interval);
1632 ctx->passed_sample_intervals++;
1633
1634 if (ctx->ops.check_accesses)
1635 max_nr_accesses = ctx->ops.check_accesses(ctx);
1636
1637 if (ctx->passed_sample_intervals == next_aggregation_sis) {
1638 kdamond_merge_regions(ctx,
1639 max_nr_accesses / 10,
1640 sz_limit);
1641 if (ctx->callback.after_aggregation &&
1642 ctx->callback.after_aggregation(ctx))
1643 break;
1644 }
1645
1646 /*
1647 * do kdamond_apply_schemes() after kdamond_merge_regions() if
1648 * possible, to reduce overhead
1649 */
1650 if (!list_empty(&ctx->schemes))
1651 kdamond_apply_schemes(ctx);
1652
1653 sample_interval = ctx->attrs.sample_interval ?
1654 ctx->attrs.sample_interval : 1;
1655 if (ctx->passed_sample_intervals == next_aggregation_sis) {
1656 ctx->next_aggregation_sis = next_aggregation_sis +
1657 ctx->attrs.aggr_interval / sample_interval;
1658
1659 kdamond_reset_aggregated(ctx);
1660 kdamond_split_regions(ctx);
1661 if (ctx->ops.reset_aggregated)
1662 ctx->ops.reset_aggregated(ctx);
1663 }
1664
1665 if (ctx->passed_sample_intervals == next_ops_update_sis) {
1666 ctx->next_ops_update_sis = next_ops_update_sis +
1667 ctx->attrs.ops_update_interval /
1668 sample_interval;
1669 if (ctx->ops.update)
1670 ctx->ops.update(ctx);
1671 sz_limit = damon_region_sz_limit(ctx);
1672 }
1673 }
1674 done:
1675 damon_for_each_target(t, ctx) {
1676 damon_for_each_region_safe(r, next, t)
1677 damon_destroy_region(r, t);
1678 }
1679
1680 if (ctx->callback.before_terminate)
1681 ctx->callback.before_terminate(ctx);
1682 if (ctx->ops.cleanup)
1683 ctx->ops.cleanup(ctx);
1684
1685 pr_debug("kdamond (%d) finishes\n", current->pid);
1686 mutex_lock(&ctx->kdamond_lock);
1687 ctx->kdamond = NULL;
1688 mutex_unlock(&ctx->kdamond_lock);
1689
1690 mutex_lock(&damon_lock);
1691 nr_running_ctxs--;
1692 if (!nr_running_ctxs && running_exclusive_ctxs)
1693 running_exclusive_ctxs = false;
1694 mutex_unlock(&damon_lock);
1695
1696 return 0;
1697 }
1698
1699 /*
1700 * struct damon_system_ram_region - System RAM resource address region of
1701 * [@start, @end).
1702 * @start: Start address of the region (inclusive).
1703 * @end: End address of the region (exclusive).
1704 */
1705 struct damon_system_ram_region {
1706 unsigned long start;
1707 unsigned long end;
1708 };
1709
walk_system_ram(struct resource * res,void * arg)1710 static int walk_system_ram(struct resource *res, void *arg)
1711 {
1712 struct damon_system_ram_region *a = arg;
1713
1714 if (a->end - a->start < resource_size(res)) {
1715 a->start = res->start;
1716 a->end = res->end;
1717 }
1718 return 0;
1719 }
1720
1721 /*
1722 * Find biggest 'System RAM' resource and store its start and end address in
1723 * @start and @end, respectively. If no System RAM is found, returns false.
1724 */
damon_find_biggest_system_ram(unsigned long * start,unsigned long * end)1725 static bool damon_find_biggest_system_ram(unsigned long *start,
1726 unsigned long *end)
1727
1728 {
1729 struct damon_system_ram_region arg = {};
1730
1731 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
1732 if (arg.end <= arg.start)
1733 return false;
1734
1735 *start = arg.start;
1736 *end = arg.end;
1737 return true;
1738 }
1739
1740 /**
1741 * damon_set_region_biggest_system_ram_default() - Set the region of the given
1742 * monitoring target as requested, or biggest 'System RAM'.
1743 * @t: The monitoring target to set the region.
1744 * @start: The pointer to the start address of the region.
1745 * @end: The pointer to the end address of the region.
1746 *
1747 * This function sets the region of @t as requested by @start and @end. If the
1748 * values of @start and @end are zero, however, this function finds the biggest
1749 * 'System RAM' resource and sets the region to cover the resource. In the
1750 * latter case, this function saves the start and end addresses of the resource
1751 * in @start and @end, respectively.
1752 *
1753 * Return: 0 on success, negative error code otherwise.
1754 */
damon_set_region_biggest_system_ram_default(struct damon_target * t,unsigned long * start,unsigned long * end)1755 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
1756 unsigned long *start, unsigned long *end)
1757 {
1758 struct damon_addr_range addr_range;
1759
1760 if (*start > *end)
1761 return -EINVAL;
1762
1763 if (!*start && !*end &&
1764 !damon_find_biggest_system_ram(start, end))
1765 return -EINVAL;
1766
1767 addr_range.start = *start;
1768 addr_range.end = *end;
1769 return damon_set_regions(t, &addr_range, 1);
1770 }
1771
1772 /*
1773 * damon_moving_sum() - Calculate an inferred moving sum value.
1774 * @mvsum: Inferred sum of the last @len_window values.
1775 * @nomvsum: Non-moving sum of the last discrete @len_window window values.
1776 * @len_window: The number of last values to take care of.
1777 * @new_value: New value that will be added to the pseudo moving sum.
1778 *
1779 * Moving sum (moving average * window size) is good for handling noise, but
1780 * the cost of keeping past values can be high for arbitrary window size. This
1781 * function implements a lightweight pseudo moving sum function that doesn't
1782 * keep the past window values.
1783 *
1784 * It simply assumes there was no noise in the past, and get the no-noise
1785 * assumed past value to drop from @nomvsum and @len_window. @nomvsum is a
1786 * non-moving sum of the last window. For example, if @len_window is 10 and we
1787 * have 25 values, @nomvsum is the sum of the 11th to 20th values of the 25
1788 * values. Hence, this function simply drops @nomvsum / @len_window from
1789 * given @mvsum and add @new_value.
1790 *
1791 * For example, if @len_window is 10 and @nomvsum is 50, the last 10 values for
1792 * the last window could be vary, e.g., 0, 10, 0, 10, 0, 10, 0, 0, 0, 20. For
1793 * calculating next moving sum with a new value, we should drop 0 from 50 and
1794 * add the new value. However, this function assumes it got value 5 for each
1795 * of the last ten times. Based on the assumption, when the next value is
1796 * measured, it drops the assumed past value, 5 from the current sum, and add
1797 * the new value to get the updated pseduo-moving average.
1798 *
1799 * This means the value could have errors, but the errors will be disappeared
1800 * for every @len_window aligned calls. For example, if @len_window is 10, the
1801 * pseudo moving sum with 11th value to 19th value would have an error. But
1802 * the sum with 20th value will not have the error.
1803 *
1804 * Return: Pseudo-moving average after getting the @new_value.
1805 */
damon_moving_sum(unsigned int mvsum,unsigned int nomvsum,unsigned int len_window,unsigned int new_value)1806 static unsigned int damon_moving_sum(unsigned int mvsum, unsigned int nomvsum,
1807 unsigned int len_window, unsigned int new_value)
1808 {
1809 return mvsum - nomvsum / len_window + new_value;
1810 }
1811
1812 /**
1813 * damon_update_region_access_rate() - Update the access rate of a region.
1814 * @r: The DAMON region to update for its access check result.
1815 * @accessed: Whether the region has accessed during last sampling interval.
1816 * @attrs: The damon_attrs of the DAMON context.
1817 *
1818 * Update the access rate of a region with the region's last sampling interval
1819 * access check result.
1820 *
1821 * Usually this will be called by &damon_operations->check_accesses callback.
1822 */
damon_update_region_access_rate(struct damon_region * r,bool accessed,struct damon_attrs * attrs)1823 void damon_update_region_access_rate(struct damon_region *r, bool accessed,
1824 struct damon_attrs *attrs)
1825 {
1826 unsigned int len_window = 1;
1827
1828 /*
1829 * sample_interval can be zero, but cannot be larger than
1830 * aggr_interval, owing to validation of damon_set_attrs().
1831 */
1832 if (attrs->sample_interval)
1833 len_window = damon_max_nr_accesses(attrs);
1834 r->nr_accesses_bp = damon_moving_sum(r->nr_accesses_bp,
1835 r->last_nr_accesses * 10000, len_window,
1836 accessed ? 10000 : 0);
1837
1838 if (accessed)
1839 r->nr_accesses++;
1840 }
1841
damon_init(void)1842 static int __init damon_init(void)
1843 {
1844 damon_region_cache = KMEM_CACHE(damon_region, 0);
1845 if (unlikely(!damon_region_cache)) {
1846 pr_err("creating damon_region_cache fails\n");
1847 return -ENOMEM;
1848 }
1849
1850 return 0;
1851 }
1852
1853 subsys_initcall(damon_init);
1854
1855 #include "core-test.h"
1856