1 /* SPDX-License-Identifier: GPL-2.0 */
5 * Author: SeongJae Park <sjpark@amazon.de>
11 #include <linux/memcontrol.h>
12 #include <linux/mutex.h>
13 #include <linux/time64.h>
14 #include <linux/types.h>
15 #include <linux/random.h>
17 /* Minimal region size. Every damon_region is aligned by this. */
18 #define DAMON_MIN_REGION PAGE_SIZE
19 /* Max priority score for DAMON-based operation schemes */
20 #define DAMOS_MAX_SCORE (99)
22 /* Get a random number in [l, r) */
23 static inline unsigned long damon_rand(unsigned long l, unsigned long r)
25 return l + get_random_u32_below(r - l);
29 * struct damon_addr_range - Represents an address region of [@start, @end).
30 * @start: Start address of the region (inclusive).
31 * @end: End address of the region (exclusive).
33 struct damon_addr_range {
39 * struct damon_region - Represents a monitoring target region.
40 * @ar: The address range of the region.
41 * @sampling_addr: Address of the sample for the next access check.
42 * @nr_accesses: Access frequency of this region.
43 * @list: List head for siblings.
44 * @age: Age of this region.
46 * @age is initially zero, increased for each aggregation interval, and reset
47 * to zero again if the access frequency is significantly changed. If two
48 * regions are merged into a new region, both @nr_accesses and @age of the new
49 * region are set as region size-weighted average of those of the two regions.
52 struct damon_addr_range ar;
53 unsigned long sampling_addr;
54 unsigned int nr_accesses;
55 struct list_head list;
58 /* private: Internal value for age calculation. */
59 unsigned int last_nr_accesses;
63 * struct damon_target - Represents a monitoring target.
64 * @pid: The PID of the virtual address space to monitor.
65 * @nr_regions: Number of monitoring target regions of this target.
66 * @regions_list: Head of the monitoring target regions of this target.
67 * @list: List head for siblings.
69 * Each monitoring context could have multiple targets. For example, a context
70 * for virtual memory address spaces could have multiple target processes. The
71 * @pid should be set for appropriate &struct damon_operations including the
72 * virtual address spaces monitoring operations.
76 unsigned int nr_regions;
77 struct list_head regions_list;
78 struct list_head list;
82 * enum damos_action - Represents an action of a Data Access Monitoring-based
85 * @DAMOS_WILLNEED: Call ``madvise()`` for the region with MADV_WILLNEED.
86 * @DAMOS_COLD: Call ``madvise()`` for the region with MADV_COLD.
87 * @DAMOS_PAGEOUT: Call ``madvise()`` for the region with MADV_PAGEOUT.
88 * @DAMOS_HUGEPAGE: Call ``madvise()`` for the region with MADV_HUGEPAGE.
89 * @DAMOS_NOHUGEPAGE: Call ``madvise()`` for the region with MADV_NOHUGEPAGE.
90 * @DAMOS_LRU_PRIO: Prioritize the region on its LRU lists.
91 * @DAMOS_LRU_DEPRIO: Deprioritize the region on its LRU lists.
92 * @DAMOS_STAT: Do nothing but count the stat.
93 * @NR_DAMOS_ACTIONS: Total number of DAMOS actions
95 * The support of each action is up to running &struct damon_operations.
96 * &enum DAMON_OPS_VADDR and &enum DAMON_OPS_FVADDR supports all actions except
97 * &enum DAMOS_LRU_PRIO and &enum DAMOS_LRU_DEPRIO. &enum DAMON_OPS_PADDR
98 * supports only &enum DAMOS_PAGEOUT, &enum DAMOS_LRU_PRIO, &enum
99 * DAMOS_LRU_DEPRIO, and &DAMOS_STAT.
109 DAMOS_STAT, /* Do nothing but only record the stat */
114 * struct damos_quota - Controls the aggressiveness of the given scheme.
115 * @ms: Maximum milliseconds that the scheme can use.
116 * @sz: Maximum bytes of memory that the action can be applied.
117 * @reset_interval: Charge reset interval in milliseconds.
119 * @weight_sz: Weight of the region's size for prioritization.
120 * @weight_nr_accesses: Weight of the region's nr_accesses for prioritization.
121 * @weight_age: Weight of the region's age for prioritization.
123 * To avoid consuming too much CPU time or IO resources for applying the
124 * &struct damos->action to large memory, DAMON allows users to set time and/or
125 * size quotas. The quotas can be set by writing non-zero values to &ms and
126 * &sz, respectively. If the time quota is set, DAMON tries to use only up to
127 * &ms milliseconds within &reset_interval for applying the action. If the
128 * size quota is set, DAMON tries to apply the action only up to &sz bytes
129 * within &reset_interval.
131 * Internally, the time quota is transformed to a size quota using estimated
132 * throughput of the scheme's action. DAMON then compares it against &sz and
133 * uses smaller one as the effective quota.
135 * For selecting regions within the quota, DAMON prioritizes current scheme's
136 * target memory regions using the &struct damon_operations->get_scheme_score.
137 * You could customize the prioritization logic by setting &weight_sz,
138 * &weight_nr_accesses, and &weight_age, because monitoring operations are
139 * encouraged to respect those.
144 unsigned long reset_interval;
146 unsigned int weight_sz;
147 unsigned int weight_nr_accesses;
148 unsigned int weight_age;
151 /* For throughput estimation */
152 unsigned long total_charged_sz;
153 unsigned long total_charged_ns;
155 unsigned long esz; /* Effective size quota in bytes */
157 /* For charging the quota */
158 unsigned long charged_sz;
159 unsigned long charged_from;
160 struct damon_target *charge_target_from;
161 unsigned long charge_addr_from;
163 /* For prioritization */
164 unsigned long histogram[DAMOS_MAX_SCORE + 1];
165 unsigned int min_score;
169 * enum damos_wmark_metric - Represents the watermark metric.
171 * @DAMOS_WMARK_NONE: Ignore the watermarks of the given scheme.
172 * @DAMOS_WMARK_FREE_MEM_RATE: Free memory rate of the system in [0,1000].
173 * @NR_DAMOS_WMARK_METRICS: Total number of DAMOS watermark metrics
175 enum damos_wmark_metric {
177 DAMOS_WMARK_FREE_MEM_RATE,
178 NR_DAMOS_WMARK_METRICS,
182 * struct damos_watermarks - Controls when a given scheme should be activated.
183 * @metric: Metric for the watermarks.
184 * @interval: Watermarks check time interval in microseconds.
185 * @high: High watermark.
186 * @mid: Middle watermark.
187 * @low: Low watermark.
189 * If &metric is &DAMOS_WMARK_NONE, the scheme is always active. Being active
190 * means DAMON does monitoring and applying the action of the scheme to
191 * appropriate memory regions. Else, DAMON checks &metric of the system for at
192 * least every &interval microseconds and works as below.
194 * If &metric is higher than &high, the scheme is inactivated. If &metric is
195 * between &mid and &low, the scheme is activated. If &metric is lower than
196 * &low, the scheme is inactivated.
198 struct damos_watermarks {
199 enum damos_wmark_metric metric;
200 unsigned long interval;
210 * struct damos_stat - Statistics on a given scheme.
211 * @nr_tried: Total number of regions that the scheme is tried to be applied.
212 * @sz_tried: Total size of regions that the scheme is tried to be applied.
213 * @nr_applied: Total number of regions that the scheme is applied.
214 * @sz_applied: Total size of regions that the scheme is applied.
215 * @qt_exceeds: Total number of times the quota of the scheme has exceeded.
218 unsigned long nr_tried;
219 unsigned long sz_tried;
220 unsigned long nr_applied;
221 unsigned long sz_applied;
222 unsigned long qt_exceeds;
226 * enum damos_filter_type - Type of memory for &struct damos_filter
227 * @DAMOS_FILTER_TYPE_ANON: Anonymous pages.
228 * @DAMOS_FILTER_TYPE_MEMCG: Specific memcg's pages.
229 * @DAMOS_FILTER_TYPE_ADDR: Address range.
230 * @DAMOS_FILTER_TYPE_TARGET: Data Access Monitoring target.
231 * @NR_DAMOS_FILTER_TYPES: Number of filter types.
233 * The anon pages type and memcg type filters are handled by underlying
234 * &struct damon_operations as a part of scheme action trying, and therefore
235 * accounted as 'tried'. In contrast, other types are handled by core layer
236 * before trying of the action and therefore not accounted as 'tried'.
238 * The support of the filters that handled by &struct damon_operations depend
239 * on the running &struct damon_operations.
240 * &enum DAMON_OPS_PADDR supports both anon pages type and memcg type filters,
241 * while &enum DAMON_OPS_VADDR and &enum DAMON_OPS_FVADDR don't support any of
244 enum damos_filter_type {
245 DAMOS_FILTER_TYPE_ANON,
246 DAMOS_FILTER_TYPE_MEMCG,
247 DAMOS_FILTER_TYPE_ADDR,
248 DAMOS_FILTER_TYPE_TARGET,
249 NR_DAMOS_FILTER_TYPES,
253 * struct damos_filter - DAMOS action target memory filter.
254 * @type: Type of the page.
255 * @matching: If the matching page should filtered out or in.
256 * @memcg_id: Memcg id of the question if @type is DAMOS_FILTER_MEMCG.
257 * @addr_range: Address range if @type is DAMOS_FILTER_TYPE_ADDR.
258 * @target_idx: Index of the &struct damon_target of
259 * &damon_ctx->adaptive_targets if @type is
260 * DAMOS_FILTER_TYPE_TARGET.
261 * @list: List head for siblings.
263 * Before applying the &damos->action to a memory region, DAMOS checks if each
264 * page of the region matches to this and avoid applying the action if so.
265 * Support of each filter type depends on the running &struct damon_operations
266 * and the type. Refer to &enum damos_filter_type for more detai.
268 struct damos_filter {
269 enum damos_filter_type type;
272 unsigned short memcg_id;
273 struct damon_addr_range addr_range;
276 struct list_head list;
280 * struct damos_access_pattern - Target access pattern of the given scheme.
281 * @min_sz_region: Minimum size of target regions.
282 * @max_sz_region: Maximum size of target regions.
283 * @min_nr_accesses: Minimum ``->nr_accesses`` of target regions.
284 * @max_nr_accesses: Maximum ``->nr_accesses`` of target regions.
285 * @min_age_region: Minimum age of target regions.
286 * @max_age_region: Maximum age of target regions.
288 struct damos_access_pattern {
289 unsigned long min_sz_region;
290 unsigned long max_sz_region;
291 unsigned int min_nr_accesses;
292 unsigned int max_nr_accesses;
293 unsigned int min_age_region;
294 unsigned int max_age_region;
298 * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
299 * @pattern: Access pattern of target regions.
300 * @action: &damo_action to be applied to the target regions.
301 * @quota: Control the aggressiveness of this scheme.
302 * @wmarks: Watermarks for automated (in)activation of this scheme.
303 * @filters: Additional set of &struct damos_filter for &action.
304 * @stat: Statistics of this scheme.
305 * @list: List head for siblings.
307 * For each aggregation interval, DAMON finds regions which fit in the
308 * &pattern and applies &action to those. To avoid consuming too much
309 * CPU time or IO resources for the &action, "a is used.
311 * To do the work only when needed, schemes can be activated for specific
312 * system situations using &wmarks. If all schemes that registered to the
313 * monitoring context are inactive, DAMON stops monitoring either, and just
314 * repeatedly checks the watermarks.
316 * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
317 * stops monitoring and just repeatedly checks the watermarks.
319 * Before applying the &action to a memory region, &struct damon_operations
320 * implementation could check pages of the region and skip &action to respect
323 * After applying the &action to each region, &stat_count and &stat_sz is
324 * updated to reflect the number of regions and total size of regions that the
325 * &action is applied.
328 struct damos_access_pattern pattern;
329 enum damos_action action;
330 struct damos_quota quota;
331 struct damos_watermarks wmarks;
332 struct list_head filters;
333 struct damos_stat stat;
334 struct list_head list;
338 * enum damon_ops_id - Identifier for each monitoring operations implementation
340 * @DAMON_OPS_VADDR: Monitoring operations for virtual address spaces
341 * @DAMON_OPS_FVADDR: Monitoring operations for only fixed ranges of virtual
343 * @DAMON_OPS_PADDR: Monitoring operations for the physical address space
344 * @NR_DAMON_OPS: Number of monitoring operations implementations
356 * struct damon_operations - Monitoring operations for given use cases.
358 * @id: Identifier of this operations set.
359 * @init: Initialize operations-related data structures.
360 * @update: Update operations-related data structures.
361 * @prepare_access_checks: Prepare next access check of target regions.
362 * @check_accesses: Check the accesses to target regions.
363 * @reset_aggregated: Reset aggregated accesses monitoring results.
364 * @get_scheme_score: Get the score of a region for a scheme.
365 * @apply_scheme: Apply a DAMON-based operation scheme.
366 * @target_valid: Determine if the target is valid.
367 * @cleanup: Clean up the context.
369 * DAMON can be extended for various address spaces and usages. For this,
370 * users should register the low level operations for their target address
371 * space and usecase via the &damon_ctx.ops. Then, the monitoring thread
372 * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
373 * the monitoring, @update after each &damon_attrs.ops_update_interval, and
374 * @check_accesses, @target_valid and @prepare_access_checks after each
375 * &damon_attrs.sample_interval. Finally, @reset_aggregated is called after
376 * each &damon_attrs.aggr_interval.
378 * Each &struct damon_operations instance having valid @id can be registered
379 * via damon_register_ops() and selected by damon_select_ops() later.
380 * @init should initialize operations-related data structures. For example,
381 * this could be used to construct proper monitoring target regions and link
382 * those to @damon_ctx.adaptive_targets.
383 * @update should update the operations-related data structures. For example,
384 * this could be used to update monitoring target regions for current status.
385 * @prepare_access_checks should manipulate the monitoring regions to be
386 * prepared for the next access check.
387 * @check_accesses should check the accesses to each region that made after the
388 * last preparation and update the number of observed accesses of each region.
389 * It should also return max number of observed accesses that made as a result
390 * of its update. The value will be used for regions adjustment threshold.
391 * @reset_aggregated should reset the access monitoring results that aggregated
392 * by @check_accesses.
393 * @get_scheme_score should return the priority score of a region for a scheme
394 * as an integer in [0, &DAMOS_MAX_SCORE].
395 * @apply_scheme is called from @kdamond when a region for user provided
396 * DAMON-based operation scheme is found. It should apply the scheme's action
397 * to the region and return bytes of the region that the action is successfully
399 * @target_valid should check whether the target is still valid for the
401 * @cleanup is called from @kdamond just before its termination.
403 struct damon_operations {
404 enum damon_ops_id id;
405 void (*init)(struct damon_ctx *context);
406 void (*update)(struct damon_ctx *context);
407 void (*prepare_access_checks)(struct damon_ctx *context);
408 unsigned int (*check_accesses)(struct damon_ctx *context);
409 void (*reset_aggregated)(struct damon_ctx *context);
410 int (*get_scheme_score)(struct damon_ctx *context,
411 struct damon_target *t, struct damon_region *r,
412 struct damos *scheme);
413 unsigned long (*apply_scheme)(struct damon_ctx *context,
414 struct damon_target *t, struct damon_region *r,
415 struct damos *scheme);
416 bool (*target_valid)(struct damon_target *t);
417 void (*cleanup)(struct damon_ctx *context);
421 * struct damon_callback - Monitoring events notification callbacks.
423 * @before_start: Called before starting the monitoring.
424 * @after_wmarks_check: Called after each schemes' watermarks check.
425 * @after_sampling: Called after each sampling.
426 * @after_aggregation: Called after each aggregation.
427 * @before_damos_apply: Called before applying DAMOS action.
428 * @before_terminate: Called before terminating the monitoring.
429 * @private: User private data.
431 * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
432 * @before_terminate just before starting and finishing the monitoring,
433 * respectively. Therefore, those are good places for installing and cleaning
436 * The monitoring thread calls @after_wmarks_check after each DAMON-based
437 * operation schemes' watermarks check. If users need to make changes to the
438 * attributes of the monitoring context while it's deactivated due to the
439 * watermarks, this is the good place to do.
441 * The monitoring thread calls @after_sampling and @after_aggregation for each
442 * of the sampling intervals and aggregation intervals, respectively.
443 * Therefore, users can safely access the monitoring results without additional
444 * protection. For the reason, users are recommended to use these callback for
445 * the accesses to the results.
447 * If any callback returns non-zero, monitoring stops.
449 struct damon_callback {
452 int (*before_start)(struct damon_ctx *context);
453 int (*after_wmarks_check)(struct damon_ctx *context);
454 int (*after_sampling)(struct damon_ctx *context);
455 int (*after_aggregation)(struct damon_ctx *context);
456 int (*before_damos_apply)(struct damon_ctx *context,
457 struct damon_target *target,
458 struct damon_region *region,
459 struct damos *scheme);
460 void (*before_terminate)(struct damon_ctx *context);
464 * struct damon_attrs - Monitoring attributes for accuracy/overhead control.
466 * @sample_interval: The time between access samplings.
467 * @aggr_interval: The time between monitor results aggregations.
468 * @ops_update_interval: The time between monitoring operations updates.
469 * @min_nr_regions: The minimum number of adaptive monitoring
471 * @max_nr_regions: The maximum number of adaptive monitoring
474 * For each @sample_interval, DAMON checks whether each region is accessed or
475 * not. It aggregates and keeps the access information (number of accesses to
476 * each region) for @aggr_interval time. DAMON also checks whether the target
477 * memory regions need update (e.g., by ``mmap()`` calls from the application,
478 * in case of virtual memory monitoring) and applies the changes for each
479 * @ops_update_interval. All time intervals are in micro-seconds.
480 * Please refer to &struct damon_operations and &struct damon_callback for more
484 unsigned long sample_interval;
485 unsigned long aggr_interval;
486 unsigned long ops_update_interval;
487 unsigned long min_nr_regions;
488 unsigned long max_nr_regions;
492 * struct damon_ctx - Represents a context for each monitoring. This is the
493 * main interface that allows users to set the attributes and get the results
496 * @attrs: Monitoring attributes for accuracy/overhead control.
497 * @kdamond: Kernel thread who does the monitoring.
498 * @kdamond_lock: Mutex for the synchronizations with @kdamond.
500 * For each monitoring context, one kernel thread for the monitoring is
501 * created. The pointer to the thread is stored in @kdamond.
503 * Once started, the monitoring thread runs until explicitly required to be
504 * terminated or every monitoring target is invalid. The validity of the
505 * targets is checked via the &damon_operations.target_valid of @ops. The
506 * termination can also be explicitly requested by calling damon_stop().
507 * The thread sets @kdamond to NULL when it terminates. Therefore, users can
508 * know whether the monitoring is ongoing or terminated by reading @kdamond.
509 * Reads and writes to @kdamond from outside of the monitoring thread must
510 * be protected by @kdamond_lock.
512 * Note that the monitoring thread protects only @kdamond via @kdamond_lock.
513 * Accesses to other fields must be protected by themselves.
515 * @ops: Set of monitoring operations for given use cases.
516 * @callback: Set of callbacks for monitoring events notifications.
518 * @adaptive_targets: Head of monitoring targets (&damon_target) list.
519 * @schemes: Head of schemes (&damos) list.
522 struct damon_attrs attrs;
524 /* private: internal use only */
525 struct timespec64 last_aggregation;
526 struct timespec64 last_ops_update;
529 struct task_struct *kdamond;
530 struct mutex kdamond_lock;
532 struct damon_operations ops;
533 struct damon_callback callback;
535 struct list_head adaptive_targets;
536 struct list_head schemes;
539 static inline struct damon_region *damon_next_region(struct damon_region *r)
541 return container_of(r->list.next, struct damon_region, list);
544 static inline struct damon_region *damon_prev_region(struct damon_region *r)
546 return container_of(r->list.prev, struct damon_region, list);
549 static inline struct damon_region *damon_last_region(struct damon_target *t)
551 return list_last_entry(&t->regions_list, struct damon_region, list);
554 static inline struct damon_region *damon_first_region(struct damon_target *t)
556 return list_first_entry(&t->regions_list, struct damon_region, list);
559 static inline unsigned long damon_sz_region(struct damon_region *r)
561 return r->ar.end - r->ar.start;
565 #define damon_for_each_region(r, t) \
566 list_for_each_entry(r, &t->regions_list, list)
568 #define damon_for_each_region_from(r, t) \
569 list_for_each_entry_from(r, &t->regions_list, list)
571 #define damon_for_each_region_safe(r, next, t) \
572 list_for_each_entry_safe(r, next, &t->regions_list, list)
574 #define damon_for_each_target(t, ctx) \
575 list_for_each_entry(t, &(ctx)->adaptive_targets, list)
577 #define damon_for_each_target_safe(t, next, ctx) \
578 list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
580 #define damon_for_each_scheme(s, ctx) \
581 list_for_each_entry(s, &(ctx)->schemes, list)
583 #define damon_for_each_scheme_safe(s, next, ctx) \
584 list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
586 #define damos_for_each_filter(f, scheme) \
587 list_for_each_entry(f, &(scheme)->filters, list)
589 #define damos_for_each_filter_safe(f, next, scheme) \
590 list_for_each_entry_safe(f, next, &(scheme)->filters, list)
594 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
597 * Add a region between two other regions
599 static inline void damon_insert_region(struct damon_region *r,
600 struct damon_region *prev, struct damon_region *next,
601 struct damon_target *t)
603 __list_add(&r->list, &prev->list, &next->list);
607 void damon_add_region(struct damon_region *r, struct damon_target *t);
608 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
609 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
610 unsigned int nr_ranges);
612 struct damos_filter *damos_new_filter(enum damos_filter_type type,
614 void damos_add_filter(struct damos *s, struct damos_filter *f);
615 void damos_destroy_filter(struct damos_filter *f);
617 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
618 enum damos_action action, struct damos_quota *quota,
619 struct damos_watermarks *wmarks);
620 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
621 void damon_destroy_scheme(struct damos *s);
623 struct damon_target *damon_new_target(void);
624 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
625 bool damon_targets_empty(struct damon_ctx *ctx);
626 void damon_free_target(struct damon_target *t);
627 void damon_destroy_target(struct damon_target *t);
628 unsigned int damon_nr_regions(struct damon_target *t);
630 struct damon_ctx *damon_new_ctx(void);
631 void damon_destroy_ctx(struct damon_ctx *ctx);
632 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs);
633 void damon_set_schemes(struct damon_ctx *ctx,
634 struct damos **schemes, ssize_t nr_schemes);
635 int damon_nr_running_ctxs(void);
636 bool damon_is_registered_ops(enum damon_ops_id id);
637 int damon_register_ops(struct damon_operations *ops);
638 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id);
640 static inline bool damon_target_has_pid(const struct damon_ctx *ctx)
642 return ctx->ops.id == DAMON_OPS_VADDR || ctx->ops.id == DAMON_OPS_FVADDR;
645 static inline unsigned int damon_max_nr_accesses(const struct damon_attrs *attrs)
647 /* {aggr,sample}_interval are unsigned long, hence could overflow */
648 return min(attrs->aggr_interval / attrs->sample_interval,
649 (unsigned long)UINT_MAX);
653 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive);
654 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
656 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
657 unsigned long *start, unsigned long *end);
659 #endif /* CONFIG_DAMON */
661 #endif /* _DAMON_H */