1 // SPDX-License-Identifier: GPL-2.0
5 * Author: SeongJae Park <sjpark@amazon.de>
8 #define pr_fmt(fmt) "damon: " fmt
10 #include <linux/damon.h>
11 #include <linux/delay.h>
12 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
17 #define CREATE_TRACE_POINTS
18 #include <trace/events/damon.h>
20 #ifdef CONFIG_DAMON_KUNIT_TEST
21 #undef DAMON_MIN_REGION
22 #define DAMON_MIN_REGION 1
25 static DEFINE_MUTEX(damon_lock);
26 static int nr_running_ctxs;
27 static bool running_exclusive_ctxs;
29 static DEFINE_MUTEX(damon_ops_lock);
30 static struct damon_operations damon_registered_ops[NR_DAMON_OPS];
32 static struct kmem_cache *damon_region_cache __ro_after_init;
34 /* Should be called under damon_ops_lock with id smaller than NR_DAMON_OPS */
35 static bool __damon_is_registered_ops(enum damon_ops_id id)
37 struct damon_operations empty_ops = {};
39 if (!memcmp(&empty_ops, &damon_registered_ops[id], sizeof(empty_ops)))
45 * damon_is_registered_ops() - Check if a given damon_operations is registered.
46 * @id: Id of the damon_operations to check if registered.
48 * Return: true if the ops is set, false otherwise.
50 bool damon_is_registered_ops(enum damon_ops_id id)
54 if (id >= NR_DAMON_OPS)
56 mutex_lock(&damon_ops_lock);
57 registered = __damon_is_registered_ops(id);
58 mutex_unlock(&damon_ops_lock);
63 * damon_register_ops() - Register a monitoring operations set to DAMON.
64 * @ops: monitoring operations set to register.
66 * This function registers a monitoring operations set of valid &struct
67 * damon_operations->id so that others can find and use them later.
69 * Return: 0 on success, negative error code otherwise.
71 int damon_register_ops(struct damon_operations *ops)
75 if (ops->id >= NR_DAMON_OPS)
77 mutex_lock(&damon_ops_lock);
78 /* Fail for already registered ops */
79 if (__damon_is_registered_ops(ops->id)) {
83 damon_registered_ops[ops->id] = *ops;
85 mutex_unlock(&damon_ops_lock);
90 * damon_select_ops() - Select a monitoring operations to use with the context.
91 * @ctx: monitoring context to use the operations.
92 * @id: id of the registered monitoring operations to select.
94 * This function finds registered monitoring operations set of @id and make
97 * Return: 0 on success, negative error code otherwise.
99 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
103 if (id >= NR_DAMON_OPS)
106 mutex_lock(&damon_ops_lock);
107 if (!__damon_is_registered_ops(id))
110 ctx->ops = damon_registered_ops[id];
111 mutex_unlock(&damon_ops_lock);
116 * Construct a damon_region struct
118 * Returns the pointer to the new struct if success, or NULL otherwise
120 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
122 struct damon_region *region;
124 region = kmem_cache_alloc(damon_region_cache, GFP_KERNEL);
128 region->ar.start = start;
129 region->ar.end = end;
130 region->nr_accesses = 0;
131 INIT_LIST_HEAD(®ion->list);
134 region->last_nr_accesses = 0;
139 void damon_add_region(struct damon_region *r, struct damon_target *t)
141 list_add_tail(&r->list, &t->regions_list);
145 static void damon_del_region(struct damon_region *r, struct damon_target *t)
151 static void damon_free_region(struct damon_region *r)
153 kmem_cache_free(damon_region_cache, r);
156 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
158 damon_del_region(r, t);
159 damon_free_region(r);
163 * Check whether a region is intersecting an address range
165 * Returns true if it is.
167 static bool damon_intersect(struct damon_region *r,
168 struct damon_addr_range *re)
170 return !(r->ar.end <= re->start || re->end <= r->ar.start);
174 * Fill holes in regions with new regions.
176 static int damon_fill_regions_holes(struct damon_region *first,
177 struct damon_region *last, struct damon_target *t)
179 struct damon_region *r = first;
181 damon_for_each_region_from(r, t) {
182 struct damon_region *next, *newr;
186 next = damon_next_region(r);
187 if (r->ar.end != next->ar.start) {
188 newr = damon_new_region(r->ar.end, next->ar.start);
191 damon_insert_region(newr, r, next, t);
198 * damon_set_regions() - Set regions of a target for given address ranges.
199 * @t: the given target.
200 * @ranges: array of new monitoring target ranges.
201 * @nr_ranges: length of @ranges.
203 * This function adds new regions to, or modify existing regions of a
204 * monitoring target to fit in specific ranges.
206 * Return: 0 if success, or negative error code otherwise.
208 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
209 unsigned int nr_ranges)
211 struct damon_region *r, *next;
215 /* Remove regions which are not in the new ranges */
216 damon_for_each_region_safe(r, next, t) {
217 for (i = 0; i < nr_ranges; i++) {
218 if (damon_intersect(r, &ranges[i]))
222 damon_destroy_region(r, t);
225 r = damon_first_region(t);
226 /* Add new regions or resize existing regions to fit in the ranges */
227 for (i = 0; i < nr_ranges; i++) {
228 struct damon_region *first = NULL, *last, *newr;
229 struct damon_addr_range *range;
232 /* Get the first/last regions intersecting with the range */
233 damon_for_each_region_from(r, t) {
234 if (damon_intersect(r, range)) {
239 if (r->ar.start >= range->end)
243 /* no region intersects with this range */
244 newr = damon_new_region(
245 ALIGN_DOWN(range->start,
247 ALIGN(range->end, DAMON_MIN_REGION));
250 damon_insert_region(newr, damon_prev_region(r), r, t);
252 /* resize intersecting regions to fit in this range */
253 first->ar.start = ALIGN_DOWN(range->start,
255 last->ar.end = ALIGN(range->end, DAMON_MIN_REGION);
257 /* fill possible holes in the range */
258 err = damon_fill_regions_holes(first, last, t);
266 struct damos_filter *damos_new_filter(enum damos_filter_type type,
269 struct damos_filter *filter;
271 filter = kmalloc(sizeof(*filter), GFP_KERNEL);
275 filter->matching = matching;
276 INIT_LIST_HEAD(&filter->list);
280 void damos_add_filter(struct damos *s, struct damos_filter *f)
282 list_add_tail(&f->list, &s->filters);
285 static void damos_del_filter(struct damos_filter *f)
290 static void damos_free_filter(struct damos_filter *f)
295 void damos_destroy_filter(struct damos_filter *f)
298 damos_free_filter(f);
301 /* initialize private fields of damos_quota and return the pointer */
302 static struct damos_quota *damos_quota_init_priv(struct damos_quota *quota)
304 quota->total_charged_sz = 0;
305 quota->total_charged_ns = 0;
307 quota->charged_sz = 0;
308 quota->charged_from = 0;
309 quota->charge_target_from = NULL;
310 quota->charge_addr_from = 0;
314 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
315 enum damos_action action, struct damos_quota *quota,
316 struct damos_watermarks *wmarks)
318 struct damos *scheme;
320 scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
323 scheme->pattern = *pattern;
324 scheme->action = action;
325 INIT_LIST_HEAD(&scheme->filters);
326 scheme->stat = (struct damos_stat){};
327 INIT_LIST_HEAD(&scheme->list);
329 scheme->quota = *(damos_quota_init_priv(quota));
331 scheme->wmarks = *wmarks;
332 scheme->wmarks.activated = true;
337 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
339 list_add_tail(&s->list, &ctx->schemes);
342 static void damon_del_scheme(struct damos *s)
347 static void damon_free_scheme(struct damos *s)
352 void damon_destroy_scheme(struct damos *s)
354 struct damos_filter *f, *next;
356 damos_for_each_filter_safe(f, next, s)
357 damos_destroy_filter(f);
359 damon_free_scheme(s);
363 * Construct a damon_target struct
365 * Returns the pointer to the new struct if success, or NULL otherwise
367 struct damon_target *damon_new_target(void)
369 struct damon_target *t;
371 t = kmalloc(sizeof(*t), GFP_KERNEL);
377 INIT_LIST_HEAD(&t->regions_list);
378 INIT_LIST_HEAD(&t->list);
383 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
385 list_add_tail(&t->list, &ctx->adaptive_targets);
388 bool damon_targets_empty(struct damon_ctx *ctx)
390 return list_empty(&ctx->adaptive_targets);
393 static void damon_del_target(struct damon_target *t)
398 void damon_free_target(struct damon_target *t)
400 struct damon_region *r, *next;
402 damon_for_each_region_safe(r, next, t)
403 damon_free_region(r);
407 void damon_destroy_target(struct damon_target *t)
410 damon_free_target(t);
413 unsigned int damon_nr_regions(struct damon_target *t)
415 return t->nr_regions;
418 struct damon_ctx *damon_new_ctx(void)
420 struct damon_ctx *ctx;
422 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
426 ctx->attrs.sample_interval = 5 * 1000;
427 ctx->attrs.aggr_interval = 100 * 1000;
428 ctx->attrs.ops_update_interval = 60 * 1000 * 1000;
430 ktime_get_coarse_ts64(&ctx->last_aggregation);
431 ctx->last_ops_update = ctx->last_aggregation;
433 mutex_init(&ctx->kdamond_lock);
435 ctx->attrs.min_nr_regions = 10;
436 ctx->attrs.max_nr_regions = 1000;
438 INIT_LIST_HEAD(&ctx->adaptive_targets);
439 INIT_LIST_HEAD(&ctx->schemes);
444 static void damon_destroy_targets(struct damon_ctx *ctx)
446 struct damon_target *t, *next_t;
448 if (ctx->ops.cleanup) {
449 ctx->ops.cleanup(ctx);
453 damon_for_each_target_safe(t, next_t, ctx)
454 damon_destroy_target(t);
457 void damon_destroy_ctx(struct damon_ctx *ctx)
459 struct damos *s, *next_s;
461 damon_destroy_targets(ctx);
463 damon_for_each_scheme_safe(s, next_s, ctx)
464 damon_destroy_scheme(s);
469 static unsigned int damon_age_for_new_attrs(unsigned int age,
470 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
472 return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
475 /* convert access ratio in bp (per 10,000) to nr_accesses */
476 static unsigned int damon_accesses_bp_to_nr_accesses(
477 unsigned int accesses_bp, struct damon_attrs *attrs)
479 return accesses_bp * damon_max_nr_accesses(attrs) / 10000;
482 /* convert nr_accesses to access ratio in bp (per 10,000) */
483 static unsigned int damon_nr_accesses_to_accesses_bp(
484 unsigned int nr_accesses, struct damon_attrs *attrs)
486 return nr_accesses * 10000 / damon_max_nr_accesses(attrs);
489 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
490 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
492 return damon_accesses_bp_to_nr_accesses(
493 damon_nr_accesses_to_accesses_bp(
494 nr_accesses, old_attrs),
498 static void damon_update_monitoring_result(struct damon_region *r,
499 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
501 r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
502 old_attrs, new_attrs);
503 r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
507 * region->nr_accesses is the number of sampling intervals in the last
508 * aggregation interval that access to the region has found, and region->age is
509 * the number of aggregation intervals that its access pattern has maintained.
510 * For the reason, the real meaning of the two fields depend on current
511 * sampling interval and aggregation interval. This function updates
512 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
514 static void damon_update_monitoring_results(struct damon_ctx *ctx,
515 struct damon_attrs *new_attrs)
517 struct damon_attrs *old_attrs = &ctx->attrs;
518 struct damon_target *t;
519 struct damon_region *r;
521 /* if any interval is zero, simply forgive conversion */
522 if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
523 !new_attrs->sample_interval ||
524 !new_attrs->aggr_interval)
527 damon_for_each_target(t, ctx)
528 damon_for_each_region(r, t)
529 damon_update_monitoring_result(
530 r, old_attrs, new_attrs);
534 * damon_set_attrs() - Set attributes for the monitoring.
535 * @ctx: monitoring context
536 * @attrs: monitoring attributes
538 * This function should not be called while the kdamond is running.
539 * Every time interval is in micro-seconds.
541 * Return: 0 on success, negative error code otherwise.
543 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
545 if (attrs->min_nr_regions < 3)
547 if (attrs->min_nr_regions > attrs->max_nr_regions)
549 if (attrs->sample_interval > attrs->aggr_interval)
552 damon_update_monitoring_results(ctx, attrs);
558 * damon_set_schemes() - Set data access monitoring based operation schemes.
559 * @ctx: monitoring context
560 * @schemes: array of the schemes
561 * @nr_schemes: number of entries in @schemes
563 * This function should not be called while the kdamond of the context is
566 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
569 struct damos *s, *next;
572 damon_for_each_scheme_safe(s, next, ctx)
573 damon_destroy_scheme(s);
574 for (i = 0; i < nr_schemes; i++)
575 damon_add_scheme(ctx, schemes[i]);
579 * damon_nr_running_ctxs() - Return number of currently running contexts.
581 int damon_nr_running_ctxs(void)
585 mutex_lock(&damon_lock);
586 nr_ctxs = nr_running_ctxs;
587 mutex_unlock(&damon_lock);
592 /* Returns the size upper limit for each monitoring region */
593 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
595 struct damon_target *t;
596 struct damon_region *r;
597 unsigned long sz = 0;
599 damon_for_each_target(t, ctx) {
600 damon_for_each_region(r, t)
601 sz += damon_sz_region(r);
604 if (ctx->attrs.min_nr_regions)
605 sz /= ctx->attrs.min_nr_regions;
606 if (sz < DAMON_MIN_REGION)
607 sz = DAMON_MIN_REGION;
612 static int kdamond_fn(void *data);
615 * __damon_start() - Starts monitoring with given context.
616 * @ctx: monitoring context
618 * This function should be called while damon_lock is hold.
620 * Return: 0 on success, negative error code otherwise.
622 static int __damon_start(struct damon_ctx *ctx)
626 mutex_lock(&ctx->kdamond_lock);
629 ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
631 if (IS_ERR(ctx->kdamond)) {
632 err = PTR_ERR(ctx->kdamond);
636 mutex_unlock(&ctx->kdamond_lock);
642 * damon_start() - Starts the monitorings for a given group of contexts.
643 * @ctxs: an array of the pointers for contexts to start monitoring
644 * @nr_ctxs: size of @ctxs
645 * @exclusive: exclusiveness of this contexts group
647 * This function starts a group of monitoring threads for a group of monitoring
648 * contexts. One thread per each context is created and run in parallel. The
649 * caller should handle synchronization between the threads by itself. If
650 * @exclusive is true and a group of threads that created by other
651 * 'damon_start()' call is currently running, this function does nothing but
654 * Return: 0 on success, negative error code otherwise.
656 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
661 mutex_lock(&damon_lock);
662 if ((exclusive && nr_running_ctxs) ||
663 (!exclusive && running_exclusive_ctxs)) {
664 mutex_unlock(&damon_lock);
668 for (i = 0; i < nr_ctxs; i++) {
669 err = __damon_start(ctxs[i]);
674 if (exclusive && nr_running_ctxs)
675 running_exclusive_ctxs = true;
676 mutex_unlock(&damon_lock);
682 * __damon_stop() - Stops monitoring of a given context.
683 * @ctx: monitoring context
685 * Return: 0 on success, negative error code otherwise.
687 static int __damon_stop(struct damon_ctx *ctx)
689 struct task_struct *tsk;
691 mutex_lock(&ctx->kdamond_lock);
694 get_task_struct(tsk);
695 mutex_unlock(&ctx->kdamond_lock);
697 put_task_struct(tsk);
700 mutex_unlock(&ctx->kdamond_lock);
706 * damon_stop() - Stops the monitorings for a given group of contexts.
707 * @ctxs: an array of the pointers for contexts to stop monitoring
708 * @nr_ctxs: size of @ctxs
710 * Return: 0 on success, negative error code otherwise.
712 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
716 for (i = 0; i < nr_ctxs; i++) {
717 /* nr_running_ctxs is decremented in kdamond_fn */
718 err = __damon_stop(ctxs[i]);
726 * damon_check_reset_time_interval() - Check if a time interval is elapsed.
727 * @baseline: the time to check whether the interval has elapsed since
728 * @interval: the time interval (microseconds)
730 * See whether the given time interval has passed since the given baseline
731 * time. If so, it also updates the baseline to current time for next check.
733 * Return: true if the time interval has passed, or false otherwise.
735 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
736 unsigned long interval)
738 struct timespec64 now;
740 ktime_get_coarse_ts64(&now);
741 if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
749 * Check whether it is time to flush the aggregated information
751 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
753 return damon_check_reset_time_interval(&ctx->last_aggregation,
754 ctx->attrs.aggr_interval);
758 * Reset the aggregated monitoring results ('nr_accesses' of each region).
760 static void kdamond_reset_aggregated(struct damon_ctx *c)
762 struct damon_target *t;
763 unsigned int ti = 0; /* target's index */
765 damon_for_each_target(t, c) {
766 struct damon_region *r;
768 damon_for_each_region(r, t) {
769 trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
770 r->last_nr_accesses = r->nr_accesses;
777 static void damon_split_region_at(struct damon_target *t,
778 struct damon_region *r, unsigned long sz_r);
780 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
784 sz = damon_sz_region(r);
785 return s->pattern.min_sz_region <= sz &&
786 sz <= s->pattern.max_sz_region &&
787 s->pattern.min_nr_accesses <= r->nr_accesses &&
788 r->nr_accesses <= s->pattern.max_nr_accesses &&
789 s->pattern.min_age_region <= r->age &&
790 r->age <= s->pattern.max_age_region;
793 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
794 struct damon_region *r, struct damos *s)
796 bool ret = __damos_valid_target(r, s);
798 if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
801 return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
805 * damos_skip_charged_region() - Check if the given region or starting part of
806 * it is already charged for the DAMOS quota.
807 * @t: The target of the region.
808 * @rp: The pointer to the region.
809 * @s: The scheme to be applied.
811 * If a quota of a scheme has exceeded in a quota charge window, the scheme's
812 * action would applied to only a part of the target access pattern fulfilling
813 * regions. To avoid applying the scheme action to only already applied
814 * regions, DAMON skips applying the scheme action to the regions that charged
815 * in the previous charge window.
817 * This function checks if a given region should be skipped or not for the
818 * reason. If only the starting part of the region has previously charged,
819 * this function splits the region into two so that the second one covers the
820 * area that not charged in the previous charge widnow and saves the second
821 * region in *rp and returns false, so that the caller can apply DAMON action
824 * Return: true if the region should be entirely skipped, false otherwise.
826 static bool damos_skip_charged_region(struct damon_target *t,
827 struct damon_region **rp, struct damos *s)
829 struct damon_region *r = *rp;
830 struct damos_quota *quota = &s->quota;
831 unsigned long sz_to_skip;
833 /* Skip previously charged regions */
834 if (quota->charge_target_from) {
835 if (t != quota->charge_target_from)
837 if (r == damon_last_region(t)) {
838 quota->charge_target_from = NULL;
839 quota->charge_addr_from = 0;
842 if (quota->charge_addr_from &&
843 r->ar.end <= quota->charge_addr_from)
846 if (quota->charge_addr_from && r->ar.start <
847 quota->charge_addr_from) {
848 sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
849 r->ar.start, DAMON_MIN_REGION);
851 if (damon_sz_region(r) <= DAMON_MIN_REGION)
853 sz_to_skip = DAMON_MIN_REGION;
855 damon_split_region_at(t, r, sz_to_skip);
856 r = damon_next_region(r);
859 quota->charge_target_from = NULL;
860 quota->charge_addr_from = 0;
865 static void damos_update_stat(struct damos *s,
866 unsigned long sz_tried, unsigned long sz_applied)
869 s->stat.sz_tried += sz_tried;
871 s->stat.nr_applied++;
872 s->stat.sz_applied += sz_applied;
875 static bool __damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
876 struct damon_region *r, struct damos_filter *filter)
878 bool matched = false;
879 struct damon_target *ti;
881 unsigned long start, end;
883 switch (filter->type) {
884 case DAMOS_FILTER_TYPE_TARGET:
885 damon_for_each_target(ti, ctx) {
890 matched = target_idx == filter->target_idx;
892 case DAMOS_FILTER_TYPE_ADDR:
893 start = ALIGN_DOWN(filter->addr_range.start, DAMON_MIN_REGION);
894 end = ALIGN_DOWN(filter->addr_range.end, DAMON_MIN_REGION);
896 /* inside the range */
897 if (start <= r->ar.start && r->ar.end <= end) {
901 /* outside of the range */
902 if (r->ar.end <= start || end <= r->ar.start) {
906 /* start before the range and overlap */
907 if (r->ar.start < start) {
908 damon_split_region_at(t, r, start - r->ar.start);
912 /* start inside the range */
913 damon_split_region_at(t, r, end - r->ar.start);
920 return matched == filter->matching;
923 static bool damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
924 struct damon_region *r, struct damos *s)
926 struct damos_filter *filter;
928 damos_for_each_filter(filter, s) {
929 if (__damos_filter_out(ctx, t, r, filter))
935 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
936 struct damon_region *r, struct damos *s)
938 struct damos_quota *quota = &s->quota;
939 unsigned long sz = damon_sz_region(r);
940 struct timespec64 begin, end;
941 unsigned long sz_applied = 0;
944 if (c->ops.apply_scheme) {
945 if (quota->esz && quota->charged_sz + sz > quota->esz) {
946 sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
950 damon_split_region_at(t, r, sz);
952 if (damos_filter_out(c, t, r, s))
954 ktime_get_coarse_ts64(&begin);
955 if (c->callback.before_damos_apply)
956 err = c->callback.before_damos_apply(c, t, r, s);
958 sz_applied = c->ops.apply_scheme(c, t, r, s);
959 ktime_get_coarse_ts64(&end);
960 quota->total_charged_ns += timespec64_to_ns(&end) -
961 timespec64_to_ns(&begin);
962 quota->charged_sz += sz;
963 if (quota->esz && quota->charged_sz >= quota->esz) {
964 quota->charge_target_from = t;
965 quota->charge_addr_from = r->ar.end + 1;
968 if (s->action != DAMOS_STAT)
972 damos_update_stat(s, sz, sz_applied);
975 static void damon_do_apply_schemes(struct damon_ctx *c,
976 struct damon_target *t,
977 struct damon_region *r)
981 damon_for_each_scheme(s, c) {
982 struct damos_quota *quota = &s->quota;
984 if (!s->wmarks.activated)
987 /* Check the quota */
988 if (quota->esz && quota->charged_sz >= quota->esz)
991 if (damos_skip_charged_region(t, &r, s))
994 if (!damos_valid_target(c, t, r, s))
997 damos_apply_scheme(c, t, r, s);
1001 /* Shouldn't be called if quota->ms and quota->sz are zero */
1002 static void damos_set_effective_quota(struct damos_quota *quota)
1004 unsigned long throughput;
1008 quota->esz = quota->sz;
1012 if (quota->total_charged_ns)
1013 throughput = quota->total_charged_sz * 1000000 /
1014 quota->total_charged_ns;
1016 throughput = PAGE_SIZE * 1024;
1017 esz = throughput * quota->ms;
1019 if (quota->sz && quota->sz < esz)
1024 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
1026 struct damos_quota *quota = &s->quota;
1027 struct damon_target *t;
1028 struct damon_region *r;
1029 unsigned long cumulated_sz;
1030 unsigned int score, max_score = 0;
1032 if (!quota->ms && !quota->sz)
1035 /* New charge window starts */
1036 if (time_after_eq(jiffies, quota->charged_from +
1037 msecs_to_jiffies(quota->reset_interval))) {
1038 if (quota->esz && quota->charged_sz >= quota->esz)
1039 s->stat.qt_exceeds++;
1040 quota->total_charged_sz += quota->charged_sz;
1041 quota->charged_from = jiffies;
1042 quota->charged_sz = 0;
1043 damos_set_effective_quota(quota);
1046 if (!c->ops.get_scheme_score)
1049 /* Fill up the score histogram */
1050 memset(quota->histogram, 0, sizeof(quota->histogram));
1051 damon_for_each_target(t, c) {
1052 damon_for_each_region(r, t) {
1053 if (!__damos_valid_target(r, s))
1055 score = c->ops.get_scheme_score(c, t, r, s);
1056 quota->histogram[score] += damon_sz_region(r);
1057 if (score > max_score)
1062 /* Set the min score limit */
1063 for (cumulated_sz = 0, score = max_score; ; score--) {
1064 cumulated_sz += quota->histogram[score];
1065 if (cumulated_sz >= quota->esz || !score)
1068 quota->min_score = score;
1071 static void kdamond_apply_schemes(struct damon_ctx *c)
1073 struct damon_target *t;
1074 struct damon_region *r, *next_r;
1077 damon_for_each_scheme(s, c) {
1078 if (!s->wmarks.activated)
1081 damos_adjust_quota(c, s);
1084 damon_for_each_target(t, c) {
1085 damon_for_each_region_safe(r, next_r, t)
1086 damon_do_apply_schemes(c, t, r);
1091 * Merge two adjacent regions into one region
1093 static void damon_merge_two_regions(struct damon_target *t,
1094 struct damon_region *l, struct damon_region *r)
1096 unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1098 l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1100 l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1101 l->ar.end = r->ar.end;
1102 damon_destroy_region(r, t);
1106 * Merge adjacent regions having similar access frequencies
1108 * t target affected by this merge operation
1109 * thres '->nr_accesses' diff threshold for the merge
1110 * sz_limit size upper limit of each region
1112 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1113 unsigned long sz_limit)
1115 struct damon_region *r, *prev = NULL, *next;
1117 damon_for_each_region_safe(r, next, t) {
1118 if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1123 if (prev && prev->ar.end == r->ar.start &&
1124 abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1125 damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1126 damon_merge_two_regions(t, prev, r);
1133 * Merge adjacent regions having similar access frequencies
1135 * threshold '->nr_accesses' diff threshold for the merge
1136 * sz_limit size upper limit of each region
1138 * This function merges monitoring target regions which are adjacent and their
1139 * access frequencies are similar. This is for minimizing the monitoring
1140 * overhead under the dynamically changeable access pattern. If a merge was
1141 * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1143 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1144 unsigned long sz_limit)
1146 struct damon_target *t;
1148 damon_for_each_target(t, c)
1149 damon_merge_regions_of(t, threshold, sz_limit);
1153 * Split a region in two
1155 * r the region to be split
1156 * sz_r size of the first sub-region that will be made
1158 static void damon_split_region_at(struct damon_target *t,
1159 struct damon_region *r, unsigned long sz_r)
1161 struct damon_region *new;
1163 new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1167 r->ar.end = new->ar.start;
1170 new->last_nr_accesses = r->last_nr_accesses;
1172 damon_insert_region(new, r, damon_next_region(r), t);
1175 /* Split every region in the given target into 'nr_subs' regions */
1176 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1178 struct damon_region *r, *next;
1179 unsigned long sz_region, sz_sub = 0;
1182 damon_for_each_region_safe(r, next, t) {
1183 sz_region = damon_sz_region(r);
1185 for (i = 0; i < nr_subs - 1 &&
1186 sz_region > 2 * DAMON_MIN_REGION; i++) {
1188 * Randomly select size of left sub-region to be at
1189 * least 10 percent and at most 90% of original region
1191 sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1192 sz_region / 10, DAMON_MIN_REGION);
1193 /* Do not allow blank region */
1194 if (sz_sub == 0 || sz_sub >= sz_region)
1197 damon_split_region_at(t, r, sz_sub);
1204 * Split every target region into randomly-sized small regions
1206 * This function splits every target region into random-sized small regions if
1207 * current total number of the regions is equal or smaller than half of the
1208 * user-specified maximum number of regions. This is for maximizing the
1209 * monitoring accuracy under the dynamically changeable access patterns. If a
1210 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1213 static void kdamond_split_regions(struct damon_ctx *ctx)
1215 struct damon_target *t;
1216 unsigned int nr_regions = 0;
1217 static unsigned int last_nr_regions;
1218 int nr_subregions = 2;
1220 damon_for_each_target(t, ctx)
1221 nr_regions += damon_nr_regions(t);
1223 if (nr_regions > ctx->attrs.max_nr_regions / 2)
1226 /* Maybe the middle of the region has different access frequency */
1227 if (last_nr_regions == nr_regions &&
1228 nr_regions < ctx->attrs.max_nr_regions / 3)
1231 damon_for_each_target(t, ctx)
1232 damon_split_regions_of(t, nr_subregions);
1234 last_nr_regions = nr_regions;
1238 * Check whether it is time to check and apply the operations-related data
1241 * Returns true if it is.
1243 static bool kdamond_need_update_operations(struct damon_ctx *ctx)
1245 return damon_check_reset_time_interval(&ctx->last_ops_update,
1246 ctx->attrs.ops_update_interval);
1250 * Check whether current monitoring should be stopped
1252 * The monitoring is stopped when either the user requested to stop, or all
1253 * monitoring targets are invalid.
1255 * Returns true if need to stop current monitoring.
1257 static bool kdamond_need_stop(struct damon_ctx *ctx)
1259 struct damon_target *t;
1261 if (kthread_should_stop())
1264 if (!ctx->ops.target_valid)
1267 damon_for_each_target(t, ctx) {
1268 if (ctx->ops.target_valid(t))
1275 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
1280 case DAMOS_WMARK_FREE_MEM_RATE:
1282 return i.freeram * 1000 / i.totalram;
1290 * Returns zero if the scheme is active. Else, returns time to wait for next
1291 * watermark check in micro-seconds.
1293 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1295 unsigned long metric;
1297 if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
1300 metric = damos_wmark_metric_value(scheme->wmarks.metric);
1301 /* higher than high watermark or lower than low watermark */
1302 if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1303 if (scheme->wmarks.activated)
1304 pr_debug("deactivate a scheme (%d) for %s wmark\n",
1306 metric > scheme->wmarks.high ?
1308 scheme->wmarks.activated = false;
1309 return scheme->wmarks.interval;
1312 /* inactive and higher than middle watermark */
1313 if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1314 !scheme->wmarks.activated)
1315 return scheme->wmarks.interval;
1317 if (!scheme->wmarks.activated)
1318 pr_debug("activate a scheme (%d)\n", scheme->action);
1319 scheme->wmarks.activated = true;
1323 static void kdamond_usleep(unsigned long usecs)
1325 /* See Documentation/timers/timers-howto.rst for the thresholds */
1326 if (usecs > 20 * USEC_PER_MSEC)
1327 schedule_timeout_idle(usecs_to_jiffies(usecs));
1329 usleep_idle_range(usecs, usecs + 1);
1332 /* Returns negative error code if it's not activated but should return */
1333 static int kdamond_wait_activation(struct damon_ctx *ctx)
1336 unsigned long wait_time;
1337 unsigned long min_wait_time = 0;
1338 bool init_wait_time = false;
1340 while (!kdamond_need_stop(ctx)) {
1341 damon_for_each_scheme(s, ctx) {
1342 wait_time = damos_wmark_wait_us(s);
1343 if (!init_wait_time || wait_time < min_wait_time) {
1344 init_wait_time = true;
1345 min_wait_time = wait_time;
1351 kdamond_usleep(min_wait_time);
1353 if (ctx->callback.after_wmarks_check &&
1354 ctx->callback.after_wmarks_check(ctx))
1361 * The monitoring daemon that runs as a kernel thread
1363 static int kdamond_fn(void *data)
1365 struct damon_ctx *ctx = data;
1366 struct damon_target *t;
1367 struct damon_region *r, *next;
1368 unsigned int max_nr_accesses = 0;
1369 unsigned long sz_limit = 0;
1371 pr_debug("kdamond (%d) starts\n", current->pid);
1375 if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1378 sz_limit = damon_region_sz_limit(ctx);
1380 while (!kdamond_need_stop(ctx)) {
1381 if (kdamond_wait_activation(ctx))
1384 if (ctx->ops.prepare_access_checks)
1385 ctx->ops.prepare_access_checks(ctx);
1386 if (ctx->callback.after_sampling &&
1387 ctx->callback.after_sampling(ctx))
1390 kdamond_usleep(ctx->attrs.sample_interval);
1392 if (ctx->ops.check_accesses)
1393 max_nr_accesses = ctx->ops.check_accesses(ctx);
1395 if (kdamond_aggregate_interval_passed(ctx)) {
1396 kdamond_merge_regions(ctx,
1397 max_nr_accesses / 10,
1399 if (ctx->callback.after_aggregation &&
1400 ctx->callback.after_aggregation(ctx))
1402 if (!list_empty(&ctx->schemes))
1403 kdamond_apply_schemes(ctx);
1404 kdamond_reset_aggregated(ctx);
1405 kdamond_split_regions(ctx);
1406 if (ctx->ops.reset_aggregated)
1407 ctx->ops.reset_aggregated(ctx);
1410 if (kdamond_need_update_operations(ctx)) {
1411 if (ctx->ops.update)
1412 ctx->ops.update(ctx);
1413 sz_limit = damon_region_sz_limit(ctx);
1417 damon_for_each_target(t, ctx) {
1418 damon_for_each_region_safe(r, next, t)
1419 damon_destroy_region(r, t);
1422 if (ctx->callback.before_terminate)
1423 ctx->callback.before_terminate(ctx);
1424 if (ctx->ops.cleanup)
1425 ctx->ops.cleanup(ctx);
1427 pr_debug("kdamond (%d) finishes\n", current->pid);
1428 mutex_lock(&ctx->kdamond_lock);
1429 ctx->kdamond = NULL;
1430 mutex_unlock(&ctx->kdamond_lock);
1432 mutex_lock(&damon_lock);
1434 if (!nr_running_ctxs && running_exclusive_ctxs)
1435 running_exclusive_ctxs = false;
1436 mutex_unlock(&damon_lock);
1442 * struct damon_system_ram_region - System RAM resource address region of
1444 * @start: Start address of the region (inclusive).
1445 * @end: End address of the region (exclusive).
1447 struct damon_system_ram_region {
1448 unsigned long start;
1452 static int walk_system_ram(struct resource *res, void *arg)
1454 struct damon_system_ram_region *a = arg;
1456 if (a->end - a->start < resource_size(res)) {
1457 a->start = res->start;
1464 * Find biggest 'System RAM' resource and store its start and end address in
1465 * @start and @end, respectively. If no System RAM is found, returns false.
1467 static bool damon_find_biggest_system_ram(unsigned long *start,
1471 struct damon_system_ram_region arg = {};
1473 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
1474 if (arg.end <= arg.start)
1483 * damon_set_region_biggest_system_ram_default() - Set the region of the given
1484 * monitoring target as requested, or biggest 'System RAM'.
1485 * @t: The monitoring target to set the region.
1486 * @start: The pointer to the start address of the region.
1487 * @end: The pointer to the end address of the region.
1489 * This function sets the region of @t as requested by @start and @end. If the
1490 * values of @start and @end are zero, however, this function finds the biggest
1491 * 'System RAM' resource and sets the region to cover the resource. In the
1492 * latter case, this function saves the start and end addresses of the resource
1493 * in @start and @end, respectively.
1495 * Return: 0 on success, negative error code otherwise.
1497 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
1498 unsigned long *start, unsigned long *end)
1500 struct damon_addr_range addr_range;
1505 if (!*start && !*end &&
1506 !damon_find_biggest_system_ram(start, end))
1509 addr_range.start = *start;
1510 addr_range.end = *end;
1511 return damon_set_regions(t, &addr_range, 1);
1514 static int __init damon_init(void)
1516 damon_region_cache = KMEM_CACHE(damon_region, 0);
1517 if (unlikely(!damon_region_cache)) {
1518 pr_err("creating damon_region_cache fails\n");
1525 subsys_initcall(damon_init);
1527 #include "core-test.h"