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 init_completion(&ctx->kdamond_started);
428 ctx->attrs.sample_interval = 5 * 1000;
429 ctx->attrs.aggr_interval = 100 * 1000;
430 ctx->attrs.ops_update_interval = 60 * 1000 * 1000;
432 ctx->passed_sample_intervals = 0;
433 /* These will be set from kdamond_init_intervals_sis() */
434 ctx->next_aggregation_sis = 0;
435 ctx->next_ops_update_sis = 0;
437 mutex_init(&ctx->kdamond_lock);
439 ctx->attrs.min_nr_regions = 10;
440 ctx->attrs.max_nr_regions = 1000;
442 INIT_LIST_HEAD(&ctx->adaptive_targets);
443 INIT_LIST_HEAD(&ctx->schemes);
448 static void damon_destroy_targets(struct damon_ctx *ctx)
450 struct damon_target *t, *next_t;
452 if (ctx->ops.cleanup) {
453 ctx->ops.cleanup(ctx);
457 damon_for_each_target_safe(t, next_t, ctx)
458 damon_destroy_target(t);
461 void damon_destroy_ctx(struct damon_ctx *ctx)
463 struct damos *s, *next_s;
465 damon_destroy_targets(ctx);
467 damon_for_each_scheme_safe(s, next_s, ctx)
468 damon_destroy_scheme(s);
473 static unsigned int damon_age_for_new_attrs(unsigned int age,
474 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
476 return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
479 /* convert access ratio in bp (per 10,000) to nr_accesses */
480 static unsigned int damon_accesses_bp_to_nr_accesses(
481 unsigned int accesses_bp, struct damon_attrs *attrs)
483 return accesses_bp * damon_max_nr_accesses(attrs) / 10000;
486 /* convert nr_accesses to access ratio in bp (per 10,000) */
487 static unsigned int damon_nr_accesses_to_accesses_bp(
488 unsigned int nr_accesses, struct damon_attrs *attrs)
490 return nr_accesses * 10000 / damon_max_nr_accesses(attrs);
493 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
494 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
496 return damon_accesses_bp_to_nr_accesses(
497 damon_nr_accesses_to_accesses_bp(
498 nr_accesses, old_attrs),
502 static void damon_update_monitoring_result(struct damon_region *r,
503 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
505 r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
506 old_attrs, new_attrs);
507 r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
511 * region->nr_accesses is the number of sampling intervals in the last
512 * aggregation interval that access to the region has found, and region->age is
513 * the number of aggregation intervals that its access pattern has maintained.
514 * For the reason, the real meaning of the two fields depend on current
515 * sampling interval and aggregation interval. This function updates
516 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
518 static void damon_update_monitoring_results(struct damon_ctx *ctx,
519 struct damon_attrs *new_attrs)
521 struct damon_attrs *old_attrs = &ctx->attrs;
522 struct damon_target *t;
523 struct damon_region *r;
525 /* if any interval is zero, simply forgive conversion */
526 if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
527 !new_attrs->sample_interval ||
528 !new_attrs->aggr_interval)
531 damon_for_each_target(t, ctx)
532 damon_for_each_region(r, t)
533 damon_update_monitoring_result(
534 r, old_attrs, new_attrs);
538 * damon_set_attrs() - Set attributes for the monitoring.
539 * @ctx: monitoring context
540 * @attrs: monitoring attributes
542 * This function should not be called while the kdamond is running.
543 * Every time interval is in micro-seconds.
545 * Return: 0 on success, negative error code otherwise.
547 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
549 unsigned long sample_interval = attrs->sample_interval ?
550 attrs->sample_interval : 1;
552 if (attrs->min_nr_regions < 3)
554 if (attrs->min_nr_regions > attrs->max_nr_regions)
556 if (attrs->sample_interval > attrs->aggr_interval)
559 ctx->next_aggregation_sis = ctx->passed_sample_intervals +
560 attrs->aggr_interval / sample_interval;
561 ctx->next_ops_update_sis = ctx->passed_sample_intervals +
562 attrs->ops_update_interval / sample_interval;
564 damon_update_monitoring_results(ctx, attrs);
570 * damon_set_schemes() - Set data access monitoring based operation schemes.
571 * @ctx: monitoring context
572 * @schemes: array of the schemes
573 * @nr_schemes: number of entries in @schemes
575 * This function should not be called while the kdamond of the context is
578 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
581 struct damos *s, *next;
584 damon_for_each_scheme_safe(s, next, ctx)
585 damon_destroy_scheme(s);
586 for (i = 0; i < nr_schemes; i++)
587 damon_add_scheme(ctx, schemes[i]);
591 * damon_nr_running_ctxs() - Return number of currently running contexts.
593 int damon_nr_running_ctxs(void)
597 mutex_lock(&damon_lock);
598 nr_ctxs = nr_running_ctxs;
599 mutex_unlock(&damon_lock);
604 /* Returns the size upper limit for each monitoring region */
605 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
607 struct damon_target *t;
608 struct damon_region *r;
609 unsigned long sz = 0;
611 damon_for_each_target(t, ctx) {
612 damon_for_each_region(r, t)
613 sz += damon_sz_region(r);
616 if (ctx->attrs.min_nr_regions)
617 sz /= ctx->attrs.min_nr_regions;
618 if (sz < DAMON_MIN_REGION)
619 sz = DAMON_MIN_REGION;
624 static int kdamond_fn(void *data);
627 * __damon_start() - Starts monitoring with given context.
628 * @ctx: monitoring context
630 * This function should be called while damon_lock is hold.
632 * Return: 0 on success, negative error code otherwise.
634 static int __damon_start(struct damon_ctx *ctx)
638 mutex_lock(&ctx->kdamond_lock);
641 reinit_completion(&ctx->kdamond_started);
642 ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
644 if (IS_ERR(ctx->kdamond)) {
645 err = PTR_ERR(ctx->kdamond);
648 wait_for_completion(&ctx->kdamond_started);
651 mutex_unlock(&ctx->kdamond_lock);
657 * damon_start() - Starts the monitorings for a given group of contexts.
658 * @ctxs: an array of the pointers for contexts to start monitoring
659 * @nr_ctxs: size of @ctxs
660 * @exclusive: exclusiveness of this contexts group
662 * This function starts a group of monitoring threads for a group of monitoring
663 * contexts. One thread per each context is created and run in parallel. The
664 * caller should handle synchronization between the threads by itself. If
665 * @exclusive is true and a group of threads that created by other
666 * 'damon_start()' call is currently running, this function does nothing but
669 * Return: 0 on success, negative error code otherwise.
671 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
676 mutex_lock(&damon_lock);
677 if ((exclusive && nr_running_ctxs) ||
678 (!exclusive && running_exclusive_ctxs)) {
679 mutex_unlock(&damon_lock);
683 for (i = 0; i < nr_ctxs; i++) {
684 err = __damon_start(ctxs[i]);
689 if (exclusive && nr_running_ctxs)
690 running_exclusive_ctxs = true;
691 mutex_unlock(&damon_lock);
697 * __damon_stop() - Stops monitoring of a given context.
698 * @ctx: monitoring context
700 * Return: 0 on success, negative error code otherwise.
702 static int __damon_stop(struct damon_ctx *ctx)
704 struct task_struct *tsk;
706 mutex_lock(&ctx->kdamond_lock);
709 get_task_struct(tsk);
710 mutex_unlock(&ctx->kdamond_lock);
712 put_task_struct(tsk);
715 mutex_unlock(&ctx->kdamond_lock);
721 * damon_stop() - Stops the monitorings for a given group of contexts.
722 * @ctxs: an array of the pointers for contexts to stop monitoring
723 * @nr_ctxs: size of @ctxs
725 * Return: 0 on success, negative error code otherwise.
727 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
731 for (i = 0; i < nr_ctxs; i++) {
732 /* nr_running_ctxs is decremented in kdamond_fn */
733 err = __damon_stop(ctxs[i]);
741 * Reset the aggregated monitoring results ('nr_accesses' of each region).
743 static void kdamond_reset_aggregated(struct damon_ctx *c)
745 struct damon_target *t;
746 unsigned int ti = 0; /* target's index */
748 damon_for_each_target(t, c) {
749 struct damon_region *r;
751 damon_for_each_region(r, t) {
752 trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
753 r->last_nr_accesses = r->nr_accesses;
760 static void damon_split_region_at(struct damon_target *t,
761 struct damon_region *r, unsigned long sz_r);
763 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
767 sz = damon_sz_region(r);
768 return s->pattern.min_sz_region <= sz &&
769 sz <= s->pattern.max_sz_region &&
770 s->pattern.min_nr_accesses <= r->nr_accesses &&
771 r->nr_accesses <= s->pattern.max_nr_accesses &&
772 s->pattern.min_age_region <= r->age &&
773 r->age <= s->pattern.max_age_region;
776 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
777 struct damon_region *r, struct damos *s)
779 bool ret = __damos_valid_target(r, s);
781 if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
784 return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
788 * damos_skip_charged_region() - Check if the given region or starting part of
789 * it is already charged for the DAMOS quota.
790 * @t: The target of the region.
791 * @rp: The pointer to the region.
792 * @s: The scheme to be applied.
794 * If a quota of a scheme has exceeded in a quota charge window, the scheme's
795 * action would applied to only a part of the target access pattern fulfilling
796 * regions. To avoid applying the scheme action to only already applied
797 * regions, DAMON skips applying the scheme action to the regions that charged
798 * in the previous charge window.
800 * This function checks if a given region should be skipped or not for the
801 * reason. If only the starting part of the region has previously charged,
802 * this function splits the region into two so that the second one covers the
803 * area that not charged in the previous charge widnow and saves the second
804 * region in *rp and returns false, so that the caller can apply DAMON action
807 * Return: true if the region should be entirely skipped, false otherwise.
809 static bool damos_skip_charged_region(struct damon_target *t,
810 struct damon_region **rp, struct damos *s)
812 struct damon_region *r = *rp;
813 struct damos_quota *quota = &s->quota;
814 unsigned long sz_to_skip;
816 /* Skip previously charged regions */
817 if (quota->charge_target_from) {
818 if (t != quota->charge_target_from)
820 if (r == damon_last_region(t)) {
821 quota->charge_target_from = NULL;
822 quota->charge_addr_from = 0;
825 if (quota->charge_addr_from &&
826 r->ar.end <= quota->charge_addr_from)
829 if (quota->charge_addr_from && r->ar.start <
830 quota->charge_addr_from) {
831 sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
832 r->ar.start, DAMON_MIN_REGION);
834 if (damon_sz_region(r) <= DAMON_MIN_REGION)
836 sz_to_skip = DAMON_MIN_REGION;
838 damon_split_region_at(t, r, sz_to_skip);
839 r = damon_next_region(r);
842 quota->charge_target_from = NULL;
843 quota->charge_addr_from = 0;
848 static void damos_update_stat(struct damos *s,
849 unsigned long sz_tried, unsigned long sz_applied)
852 s->stat.sz_tried += sz_tried;
854 s->stat.nr_applied++;
855 s->stat.sz_applied += sz_applied;
858 static bool __damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
859 struct damon_region *r, struct damos_filter *filter)
861 bool matched = false;
862 struct damon_target *ti;
864 unsigned long start, end;
866 switch (filter->type) {
867 case DAMOS_FILTER_TYPE_TARGET:
868 damon_for_each_target(ti, ctx) {
873 matched = target_idx == filter->target_idx;
875 case DAMOS_FILTER_TYPE_ADDR:
876 start = ALIGN_DOWN(filter->addr_range.start, DAMON_MIN_REGION);
877 end = ALIGN_DOWN(filter->addr_range.end, DAMON_MIN_REGION);
879 /* inside the range */
880 if (start <= r->ar.start && r->ar.end <= end) {
884 /* outside of the range */
885 if (r->ar.end <= start || end <= r->ar.start) {
889 /* start before the range and overlap */
890 if (r->ar.start < start) {
891 damon_split_region_at(t, r, start - r->ar.start);
895 /* start inside the range */
896 damon_split_region_at(t, r, end - r->ar.start);
903 return matched == filter->matching;
906 static bool damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
907 struct damon_region *r, struct damos *s)
909 struct damos_filter *filter;
911 damos_for_each_filter(filter, s) {
912 if (__damos_filter_out(ctx, t, r, filter))
918 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
919 struct damon_region *r, struct damos *s)
921 struct damos_quota *quota = &s->quota;
922 unsigned long sz = damon_sz_region(r);
923 struct timespec64 begin, end;
924 unsigned long sz_applied = 0;
927 if (c->ops.apply_scheme) {
928 if (quota->esz && quota->charged_sz + sz > quota->esz) {
929 sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
933 damon_split_region_at(t, r, sz);
935 if (damos_filter_out(c, t, r, s))
937 ktime_get_coarse_ts64(&begin);
938 if (c->callback.before_damos_apply)
939 err = c->callback.before_damos_apply(c, t, r, s);
941 sz_applied = c->ops.apply_scheme(c, t, r, s);
942 ktime_get_coarse_ts64(&end);
943 quota->total_charged_ns += timespec64_to_ns(&end) -
944 timespec64_to_ns(&begin);
945 quota->charged_sz += sz;
946 if (quota->esz && quota->charged_sz >= quota->esz) {
947 quota->charge_target_from = t;
948 quota->charge_addr_from = r->ar.end + 1;
951 if (s->action != DAMOS_STAT)
955 damos_update_stat(s, sz, sz_applied);
958 static void damon_do_apply_schemes(struct damon_ctx *c,
959 struct damon_target *t,
960 struct damon_region *r)
964 damon_for_each_scheme(s, c) {
965 struct damos_quota *quota = &s->quota;
967 if (!s->wmarks.activated)
970 /* Check the quota */
971 if (quota->esz && quota->charged_sz >= quota->esz)
974 if (damos_skip_charged_region(t, &r, s))
977 if (!damos_valid_target(c, t, r, s))
980 damos_apply_scheme(c, t, r, s);
984 /* Shouldn't be called if quota->ms and quota->sz are zero */
985 static void damos_set_effective_quota(struct damos_quota *quota)
987 unsigned long throughput;
991 quota->esz = quota->sz;
995 if (quota->total_charged_ns)
996 throughput = quota->total_charged_sz * 1000000 /
997 quota->total_charged_ns;
999 throughput = PAGE_SIZE * 1024;
1000 esz = throughput * quota->ms;
1002 if (quota->sz && quota->sz < esz)
1007 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
1009 struct damos_quota *quota = &s->quota;
1010 struct damon_target *t;
1011 struct damon_region *r;
1012 unsigned long cumulated_sz;
1013 unsigned int score, max_score = 0;
1015 if (!quota->ms && !quota->sz)
1018 /* New charge window starts */
1019 if (time_after_eq(jiffies, quota->charged_from +
1020 msecs_to_jiffies(quota->reset_interval))) {
1021 if (quota->esz && quota->charged_sz >= quota->esz)
1022 s->stat.qt_exceeds++;
1023 quota->total_charged_sz += quota->charged_sz;
1024 quota->charged_from = jiffies;
1025 quota->charged_sz = 0;
1026 damos_set_effective_quota(quota);
1029 if (!c->ops.get_scheme_score)
1032 /* Fill up the score histogram */
1033 memset(quota->histogram, 0, sizeof(quota->histogram));
1034 damon_for_each_target(t, c) {
1035 damon_for_each_region(r, t) {
1036 if (!__damos_valid_target(r, s))
1038 score = c->ops.get_scheme_score(c, t, r, s);
1039 quota->histogram[score] += damon_sz_region(r);
1040 if (score > max_score)
1045 /* Set the min score limit */
1046 for (cumulated_sz = 0, score = max_score; ; score--) {
1047 cumulated_sz += quota->histogram[score];
1048 if (cumulated_sz >= quota->esz || !score)
1051 quota->min_score = score;
1054 static void kdamond_apply_schemes(struct damon_ctx *c)
1056 struct damon_target *t;
1057 struct damon_region *r, *next_r;
1060 damon_for_each_scheme(s, c) {
1061 if (!s->wmarks.activated)
1064 damos_adjust_quota(c, s);
1067 damon_for_each_target(t, c) {
1068 damon_for_each_region_safe(r, next_r, t)
1069 damon_do_apply_schemes(c, t, r);
1074 * Merge two adjacent regions into one region
1076 static void damon_merge_two_regions(struct damon_target *t,
1077 struct damon_region *l, struct damon_region *r)
1079 unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1081 l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1083 l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1084 l->ar.end = r->ar.end;
1085 damon_destroy_region(r, t);
1089 * Merge adjacent regions having similar access frequencies
1091 * t target affected by this merge operation
1092 * thres '->nr_accesses' diff threshold for the merge
1093 * sz_limit size upper limit of each region
1095 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1096 unsigned long sz_limit)
1098 struct damon_region *r, *prev = NULL, *next;
1100 damon_for_each_region_safe(r, next, t) {
1101 if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1106 if (prev && prev->ar.end == r->ar.start &&
1107 abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1108 damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1109 damon_merge_two_regions(t, prev, r);
1116 * Merge adjacent regions having similar access frequencies
1118 * threshold '->nr_accesses' diff threshold for the merge
1119 * sz_limit size upper limit of each region
1121 * This function merges monitoring target regions which are adjacent and their
1122 * access frequencies are similar. This is for minimizing the monitoring
1123 * overhead under the dynamically changeable access pattern. If a merge was
1124 * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1126 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1127 unsigned long sz_limit)
1129 struct damon_target *t;
1131 damon_for_each_target(t, c)
1132 damon_merge_regions_of(t, threshold, sz_limit);
1136 * Split a region in two
1138 * r the region to be split
1139 * sz_r size of the first sub-region that will be made
1141 static void damon_split_region_at(struct damon_target *t,
1142 struct damon_region *r, unsigned long sz_r)
1144 struct damon_region *new;
1146 new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1150 r->ar.end = new->ar.start;
1153 new->last_nr_accesses = r->last_nr_accesses;
1155 damon_insert_region(new, r, damon_next_region(r), t);
1158 /* Split every region in the given target into 'nr_subs' regions */
1159 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1161 struct damon_region *r, *next;
1162 unsigned long sz_region, sz_sub = 0;
1165 damon_for_each_region_safe(r, next, t) {
1166 sz_region = damon_sz_region(r);
1168 for (i = 0; i < nr_subs - 1 &&
1169 sz_region > 2 * DAMON_MIN_REGION; i++) {
1171 * Randomly select size of left sub-region to be at
1172 * least 10 percent and at most 90% of original region
1174 sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1175 sz_region / 10, DAMON_MIN_REGION);
1176 /* Do not allow blank region */
1177 if (sz_sub == 0 || sz_sub >= sz_region)
1180 damon_split_region_at(t, r, sz_sub);
1187 * Split every target region into randomly-sized small regions
1189 * This function splits every target region into random-sized small regions if
1190 * current total number of the regions is equal or smaller than half of the
1191 * user-specified maximum number of regions. This is for maximizing the
1192 * monitoring accuracy under the dynamically changeable access patterns. If a
1193 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1196 static void kdamond_split_regions(struct damon_ctx *ctx)
1198 struct damon_target *t;
1199 unsigned int nr_regions = 0;
1200 static unsigned int last_nr_regions;
1201 int nr_subregions = 2;
1203 damon_for_each_target(t, ctx)
1204 nr_regions += damon_nr_regions(t);
1206 if (nr_regions > ctx->attrs.max_nr_regions / 2)
1209 /* Maybe the middle of the region has different access frequency */
1210 if (last_nr_regions == nr_regions &&
1211 nr_regions < ctx->attrs.max_nr_regions / 3)
1214 damon_for_each_target(t, ctx)
1215 damon_split_regions_of(t, nr_subregions);
1217 last_nr_regions = nr_regions;
1221 * Check whether current monitoring should be stopped
1223 * The monitoring is stopped when either the user requested to stop, or all
1224 * monitoring targets are invalid.
1226 * Returns true if need to stop current monitoring.
1228 static bool kdamond_need_stop(struct damon_ctx *ctx)
1230 struct damon_target *t;
1232 if (kthread_should_stop())
1235 if (!ctx->ops.target_valid)
1238 damon_for_each_target(t, ctx) {
1239 if (ctx->ops.target_valid(t))
1246 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
1251 case DAMOS_WMARK_FREE_MEM_RATE:
1253 return i.freeram * 1000 / i.totalram;
1261 * Returns zero if the scheme is active. Else, returns time to wait for next
1262 * watermark check in micro-seconds.
1264 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1266 unsigned long metric;
1268 if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
1271 metric = damos_wmark_metric_value(scheme->wmarks.metric);
1272 /* higher than high watermark or lower than low watermark */
1273 if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1274 if (scheme->wmarks.activated)
1275 pr_debug("deactivate a scheme (%d) for %s wmark\n",
1277 metric > scheme->wmarks.high ?
1279 scheme->wmarks.activated = false;
1280 return scheme->wmarks.interval;
1283 /* inactive and higher than middle watermark */
1284 if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1285 !scheme->wmarks.activated)
1286 return scheme->wmarks.interval;
1288 if (!scheme->wmarks.activated)
1289 pr_debug("activate a scheme (%d)\n", scheme->action);
1290 scheme->wmarks.activated = true;
1294 static void kdamond_usleep(unsigned long usecs)
1296 /* See Documentation/timers/timers-howto.rst for the thresholds */
1297 if (usecs > 20 * USEC_PER_MSEC)
1298 schedule_timeout_idle(usecs_to_jiffies(usecs));
1300 usleep_idle_range(usecs, usecs + 1);
1303 /* Returns negative error code if it's not activated but should return */
1304 static int kdamond_wait_activation(struct damon_ctx *ctx)
1307 unsigned long wait_time;
1308 unsigned long min_wait_time = 0;
1309 bool init_wait_time = false;
1311 while (!kdamond_need_stop(ctx)) {
1312 damon_for_each_scheme(s, ctx) {
1313 wait_time = damos_wmark_wait_us(s);
1314 if (!init_wait_time || wait_time < min_wait_time) {
1315 init_wait_time = true;
1316 min_wait_time = wait_time;
1322 kdamond_usleep(min_wait_time);
1324 if (ctx->callback.after_wmarks_check &&
1325 ctx->callback.after_wmarks_check(ctx))
1331 static void kdamond_init_intervals_sis(struct damon_ctx *ctx)
1333 unsigned long sample_interval = ctx->attrs.sample_interval ?
1334 ctx->attrs.sample_interval : 1;
1336 ctx->passed_sample_intervals = 0;
1337 ctx->next_aggregation_sis = ctx->attrs.aggr_interval / sample_interval;
1338 ctx->next_ops_update_sis = ctx->attrs.ops_update_interval /
1343 * The monitoring daemon that runs as a kernel thread
1345 static int kdamond_fn(void *data)
1347 struct damon_ctx *ctx = data;
1348 struct damon_target *t;
1349 struct damon_region *r, *next;
1350 unsigned int max_nr_accesses = 0;
1351 unsigned long sz_limit = 0;
1353 pr_debug("kdamond (%d) starts\n", current->pid);
1355 complete(&ctx->kdamond_started);
1356 kdamond_init_intervals_sis(ctx);
1360 if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1363 sz_limit = damon_region_sz_limit(ctx);
1365 while (!kdamond_need_stop(ctx)) {
1367 * ctx->attrs and ctx->next_{aggregation,ops_update}_sis could
1368 * be changed from after_wmarks_check() or after_aggregation()
1369 * callbacks. Read the values here, and use those for this
1370 * iteration. That is, damon_set_attrs() updated new values
1371 * are respected from next iteration.
1373 unsigned long next_aggregation_sis = ctx->next_aggregation_sis;
1374 unsigned long next_ops_update_sis = ctx->next_ops_update_sis;
1375 unsigned long sample_interval = ctx->attrs.sample_interval;
1377 if (kdamond_wait_activation(ctx))
1380 if (ctx->ops.prepare_access_checks)
1381 ctx->ops.prepare_access_checks(ctx);
1382 if (ctx->callback.after_sampling &&
1383 ctx->callback.after_sampling(ctx))
1386 kdamond_usleep(sample_interval);
1387 ctx->passed_sample_intervals++;
1389 if (ctx->ops.check_accesses)
1390 max_nr_accesses = ctx->ops.check_accesses(ctx);
1392 sample_interval = ctx->attrs.sample_interval ?
1393 ctx->attrs.sample_interval : 1;
1394 if (ctx->passed_sample_intervals == next_aggregation_sis) {
1395 ctx->next_aggregation_sis = next_aggregation_sis +
1396 ctx->attrs.aggr_interval / sample_interval;
1397 kdamond_merge_regions(ctx,
1398 max_nr_accesses / 10,
1400 if (ctx->callback.after_aggregation &&
1401 ctx->callback.after_aggregation(ctx))
1403 if (!list_empty(&ctx->schemes))
1404 kdamond_apply_schemes(ctx);
1405 kdamond_reset_aggregated(ctx);
1406 kdamond_split_regions(ctx);
1407 if (ctx->ops.reset_aggregated)
1408 ctx->ops.reset_aggregated(ctx);
1411 if (ctx->passed_sample_intervals == next_ops_update_sis) {
1412 ctx->next_ops_update_sis = next_ops_update_sis +
1413 ctx->attrs.ops_update_interval /
1415 if (ctx->ops.update)
1416 ctx->ops.update(ctx);
1417 sz_limit = damon_region_sz_limit(ctx);
1421 damon_for_each_target(t, ctx) {
1422 damon_for_each_region_safe(r, next, t)
1423 damon_destroy_region(r, t);
1426 if (ctx->callback.before_terminate)
1427 ctx->callback.before_terminate(ctx);
1428 if (ctx->ops.cleanup)
1429 ctx->ops.cleanup(ctx);
1431 pr_debug("kdamond (%d) finishes\n", current->pid);
1432 mutex_lock(&ctx->kdamond_lock);
1433 ctx->kdamond = NULL;
1434 mutex_unlock(&ctx->kdamond_lock);
1436 mutex_lock(&damon_lock);
1438 if (!nr_running_ctxs && running_exclusive_ctxs)
1439 running_exclusive_ctxs = false;
1440 mutex_unlock(&damon_lock);
1446 * struct damon_system_ram_region - System RAM resource address region of
1448 * @start: Start address of the region (inclusive).
1449 * @end: End address of the region (exclusive).
1451 struct damon_system_ram_region {
1452 unsigned long start;
1456 static int walk_system_ram(struct resource *res, void *arg)
1458 struct damon_system_ram_region *a = arg;
1460 if (a->end - a->start < resource_size(res)) {
1461 a->start = res->start;
1468 * Find biggest 'System RAM' resource and store its start and end address in
1469 * @start and @end, respectively. If no System RAM is found, returns false.
1471 static bool damon_find_biggest_system_ram(unsigned long *start,
1475 struct damon_system_ram_region arg = {};
1477 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
1478 if (arg.end <= arg.start)
1487 * damon_set_region_biggest_system_ram_default() - Set the region of the given
1488 * monitoring target as requested, or biggest 'System RAM'.
1489 * @t: The monitoring target to set the region.
1490 * @start: The pointer to the start address of the region.
1491 * @end: The pointer to the end address of the region.
1493 * This function sets the region of @t as requested by @start and @end. If the
1494 * values of @start and @end are zero, however, this function finds the biggest
1495 * 'System RAM' resource and sets the region to cover the resource. In the
1496 * latter case, this function saves the start and end addresses of the resource
1497 * in @start and @end, respectively.
1499 * Return: 0 on success, negative error code otherwise.
1501 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
1502 unsigned long *start, unsigned long *end)
1504 struct damon_addr_range addr_range;
1509 if (!*start && !*end &&
1510 !damon_find_biggest_system_ram(start, end))
1513 addr_range.start = *start;
1514 addr_range.end = *end;
1515 return damon_set_regions(t, &addr_range, 1);
1518 static int __init damon_init(void)
1520 damon_region_cache = KMEM_CACHE(damon_region, 0);
1521 if (unlikely(!damon_region_cache)) {
1522 pr_err("creating damon_region_cache fails\n");
1529 subsys_initcall(damon_init);
1531 #include "core-test.h"