1 /* SPDX-License-Identifier: GPL-2.0 */
5 * Author: SeongJae Park <sjpark@amazon.de>
11 #include <linux/mutex.h>
12 #include <linux/time64.h>
13 #include <linux/types.h>
14 #include <linux/random.h>
16 /* Minimal region size. Every damon_region is aligned by this. */
17 #define DAMON_MIN_REGION PAGE_SIZE
18 /* Max priority score for DAMON-based operation schemes */
19 #define DAMOS_MAX_SCORE (99)
21 /* Get a random number in [l, r) */
22 static inline unsigned long damon_rand(unsigned long l, unsigned long r)
24 return l + prandom_u32_max(r - l);
28 * struct damon_addr_range - Represents an address region of [@start, @end).
29 * @start: Start address of the region (inclusive).
30 * @end: End address of the region (exclusive).
32 struct damon_addr_range {
38 * struct damon_region - Represents a monitoring target region.
39 * @ar: The address range of the region.
40 * @sampling_addr: Address of the sample for the next access check.
41 * @nr_accesses: Access frequency of this region.
42 * @list: List head for siblings.
43 * @age: Age of this region.
45 * @age is initially zero, increased for each aggregation interval, and reset
46 * to zero again if the access frequency is significantly changed. If two
47 * regions are merged into a new region, both @nr_accesses and @age of the new
48 * region are set as region size-weighted average of those of the two regions.
51 struct damon_addr_range ar;
52 unsigned long sampling_addr;
53 unsigned int nr_accesses;
54 struct list_head list;
57 /* private: Internal value for age calculation. */
58 unsigned int last_nr_accesses;
62 * struct damon_target - Represents a monitoring target.
63 * @id: Unique identifier for this target.
64 * @nr_regions: Number of monitoring target regions of this target.
65 * @regions_list: Head of the monitoring target regions of this target.
66 * @list: List head for siblings.
68 * Each monitoring context could have multiple targets. For example, a context
69 * for virtual memory address spaces could have multiple target processes. The
70 * @id of each target should be unique among the targets of the context. For
71 * example, in the virtual address monitoring context, it could be a pidfd or
72 * an address of an mm_struct.
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_STAT: Do nothing but count the stat.
98 DAMOS_STAT, /* Do nothing but only record the stat */
102 * struct damos_quota - Controls the aggressiveness of the given scheme.
103 * @ms: Maximum milliseconds that the scheme can use.
104 * @sz: Maximum bytes of memory that the action can be applied.
105 * @reset_interval: Charge reset interval in milliseconds.
107 * @weight_sz: Weight of the region's size for prioritization.
108 * @weight_nr_accesses: Weight of the region's nr_accesses for prioritization.
109 * @weight_age: Weight of the region's age for prioritization.
111 * To avoid consuming too much CPU time or IO resources for applying the
112 * &struct damos->action to large memory, DAMON allows users to set time and/or
113 * size quotas. The quotas can be set by writing non-zero values to &ms and
114 * &sz, respectively. If the time quota is set, DAMON tries to use only up to
115 * &ms milliseconds within &reset_interval for applying the action. If the
116 * size quota is set, DAMON tries to apply the action only up to &sz bytes
117 * within &reset_interval.
119 * Internally, the time quota is transformed to a size quota using estimated
120 * throughput of the scheme's action. DAMON then compares it against &sz and
121 * uses smaller one as the effective quota.
123 * For selecting regions within the quota, DAMON prioritizes current scheme's
124 * target memory regions using the &struct damon_primitive->get_scheme_score.
125 * You could customize the prioritization logic by setting &weight_sz,
126 * &weight_nr_accesses, and &weight_age, because monitoring primitives are
127 * encouraged to respect those.
132 unsigned long reset_interval;
134 unsigned int weight_sz;
135 unsigned int weight_nr_accesses;
136 unsigned int weight_age;
139 /* For throughput estimation */
140 unsigned long total_charged_sz;
141 unsigned long total_charged_ns;
143 unsigned long esz; /* Effective size quota in bytes */
145 /* For charging the quota */
146 unsigned long charged_sz;
147 unsigned long charged_from;
148 struct damon_target *charge_target_from;
149 unsigned long charge_addr_from;
151 /* For prioritization */
152 unsigned long histogram[DAMOS_MAX_SCORE + 1];
153 unsigned int min_score;
157 * enum damos_wmark_metric - Represents the watermark metric.
159 * @DAMOS_WMARK_NONE: Ignore the watermarks of the given scheme.
160 * @DAMOS_WMARK_FREE_MEM_RATE: Free memory rate of the system in [0,1000].
162 enum damos_wmark_metric {
164 DAMOS_WMARK_FREE_MEM_RATE,
168 * struct damos_watermarks - Controls when a given scheme should be activated.
169 * @metric: Metric for the watermarks.
170 * @interval: Watermarks check time interval in microseconds.
171 * @high: High watermark.
172 * @mid: Middle watermark.
173 * @low: Low watermark.
175 * If &metric is &DAMOS_WMARK_NONE, the scheme is always active. Being active
176 * means DAMON does monitoring and applying the action of the scheme to
177 * appropriate memory regions. Else, DAMON checks &metric of the system for at
178 * least every &interval microseconds and works as below.
180 * If &metric is higher than &high, the scheme is inactivated. If &metric is
181 * between &mid and &low, the scheme is activated. If &metric is lower than
182 * &low, the scheme is inactivated.
184 struct damos_watermarks {
185 enum damos_wmark_metric metric;
186 unsigned long interval;
196 * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
197 * @min_sz_region: Minimum size of target regions.
198 * @max_sz_region: Maximum size of target regions.
199 * @min_nr_accesses: Minimum ``->nr_accesses`` of target regions.
200 * @max_nr_accesses: Maximum ``->nr_accesses`` of target regions.
201 * @min_age_region: Minimum age of target regions.
202 * @max_age_region: Maximum age of target regions.
203 * @action: &damo_action to be applied to the target regions.
204 * @quota: Control the aggressiveness of this scheme.
205 * @wmarks: Watermarks for automated (in)activation of this scheme.
206 * @stat_count: Total number of regions that this scheme is applied.
207 * @stat_sz: Total size of regions that this scheme is applied.
208 * @list: List head for siblings.
210 * For each aggregation interval, DAMON finds regions which fit in the
211 * condition (&min_sz_region, &max_sz_region, &min_nr_accesses,
212 * &max_nr_accesses, &min_age_region, &max_age_region) and applies &action to
213 * those. To avoid consuming too much CPU time or IO resources for the
214 * &action, "a is used.
216 * To do the work only when needed, schemes can be activated for specific
217 * system situations using &wmarks. If all schemes that registered to the
218 * monitoring context are inactive, DAMON stops monitoring either, and just
219 * repeatedly checks the watermarks.
221 * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
222 * stops monitoring and just repeatedly checks the watermarks.
224 * After applying the &action to each region, &stat_count and &stat_sz is
225 * updated to reflect the number of regions and total size of regions that the
226 * &action is applied.
229 unsigned long min_sz_region;
230 unsigned long max_sz_region;
231 unsigned int min_nr_accesses;
232 unsigned int max_nr_accesses;
233 unsigned int min_age_region;
234 unsigned int max_age_region;
235 enum damos_action action;
236 struct damos_quota quota;
237 struct damos_watermarks wmarks;
238 unsigned long stat_count;
239 unsigned long stat_sz;
240 struct list_head list;
246 * struct damon_primitive - Monitoring primitives for given use cases.
248 * @init: Initialize primitive-internal data structures.
249 * @update: Update primitive-internal data structures.
250 * @prepare_access_checks: Prepare next access check of target regions.
251 * @check_accesses: Check the accesses to target regions.
252 * @reset_aggregated: Reset aggregated accesses monitoring results.
253 * @get_scheme_score: Get the score of a region for a scheme.
254 * @apply_scheme: Apply a DAMON-based operation scheme.
255 * @target_valid: Determine if the target is valid.
256 * @cleanup: Clean up the context.
258 * DAMON can be extended for various address spaces and usages. For this,
259 * users should register the low level primitives for their target address
260 * space and usecase via the &damon_ctx.primitive. Then, the monitoring thread
261 * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
262 * the monitoring, @update after each &damon_ctx.primitive_update_interval, and
263 * @check_accesses, @target_valid and @prepare_access_checks after each
264 * &damon_ctx.sample_interval. Finally, @reset_aggregated is called after each
265 * &damon_ctx.aggr_interval.
267 * @init should initialize primitive-internal data structures. For example,
268 * this could be used to construct proper monitoring target regions and link
269 * those to @damon_ctx.adaptive_targets.
270 * @update should update the primitive-internal data structures. For example,
271 * this could be used to update monitoring target regions for current status.
272 * @prepare_access_checks should manipulate the monitoring regions to be
273 * prepared for the next access check.
274 * @check_accesses should check the accesses to each region that made after the
275 * last preparation and update the number of observed accesses of each region.
276 * It should also return max number of observed accesses that made as a result
277 * of its update. The value will be used for regions adjustment threshold.
278 * @reset_aggregated should reset the access monitoring results that aggregated
279 * by @check_accesses.
280 * @get_scheme_score should return the priority score of a region for a scheme
281 * as an integer in [0, &DAMOS_MAX_SCORE].
282 * @apply_scheme is called from @kdamond when a region for user provided
283 * DAMON-based operation scheme is found. It should apply the scheme's action
284 * to the region. This is not used for &DAMON_ARBITRARY_TARGET case.
285 * @target_valid should check whether the target is still valid for the
287 * @cleanup is called from @kdamond just before its termination.
289 struct damon_primitive {
290 void (*init)(struct damon_ctx *context);
291 void (*update)(struct damon_ctx *context);
292 void (*prepare_access_checks)(struct damon_ctx *context);
293 unsigned int (*check_accesses)(struct damon_ctx *context);
294 void (*reset_aggregated)(struct damon_ctx *context);
295 int (*get_scheme_score)(struct damon_ctx *context,
296 struct damon_target *t, struct damon_region *r,
297 struct damos *scheme);
298 int (*apply_scheme)(struct damon_ctx *context, struct damon_target *t,
299 struct damon_region *r, struct damos *scheme);
300 bool (*target_valid)(void *target);
301 void (*cleanup)(struct damon_ctx *context);
305 * struct damon_callback - Monitoring events notification callbacks.
307 * @before_start: Called before starting the monitoring.
308 * @after_sampling: Called after each sampling.
309 * @after_aggregation: Called after each aggregation.
310 * @before_terminate: Called before terminating the monitoring.
311 * @private: User private data.
313 * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
314 * @before_terminate just before starting and finishing the monitoring,
315 * respectively. Therefore, those are good places for installing and cleaning
318 * The monitoring thread calls @after_sampling and @after_aggregation for each
319 * of the sampling intervals and aggregation intervals, respectively.
320 * Therefore, users can safely access the monitoring results without additional
321 * protection. For the reason, users are recommended to use these callback for
322 * the accesses to the results.
324 * If any callback returns non-zero, monitoring stops.
326 struct damon_callback {
329 int (*before_start)(struct damon_ctx *context);
330 int (*after_sampling)(struct damon_ctx *context);
331 int (*after_aggregation)(struct damon_ctx *context);
332 void (*before_terminate)(struct damon_ctx *context);
336 * struct damon_ctx - Represents a context for each monitoring. This is the
337 * main interface that allows users to set the attributes and get the results
340 * @sample_interval: The time between access samplings.
341 * @aggr_interval: The time between monitor results aggregations.
342 * @primitive_update_interval: The time between monitoring primitive updates.
344 * For each @sample_interval, DAMON checks whether each region is accessed or
345 * not. It aggregates and keeps the access information (number of accesses to
346 * each region) for @aggr_interval time. DAMON also checks whether the target
347 * memory regions need update (e.g., by ``mmap()`` calls from the application,
348 * in case of virtual memory monitoring) and applies the changes for each
349 * @primitive_update_interval. All time intervals are in micro-seconds.
350 * Please refer to &struct damon_primitive and &struct damon_callback for more
353 * @kdamond: Kernel thread who does the monitoring.
354 * @kdamond_stop: Notifies whether kdamond should stop.
355 * @kdamond_lock: Mutex for the synchronizations with @kdamond.
357 * For each monitoring context, one kernel thread for the monitoring is
358 * created. The pointer to the thread is stored in @kdamond.
360 * Once started, the monitoring thread runs until explicitly required to be
361 * terminated or every monitoring target is invalid. The validity of the
362 * targets is checked via the &damon_primitive.target_valid of @primitive. The
363 * termination can also be explicitly requested by writing non-zero to
364 * @kdamond_stop. The thread sets @kdamond to NULL when it terminates.
365 * Therefore, users can know whether the monitoring is ongoing or terminated by
366 * reading @kdamond. Reads and writes to @kdamond and @kdamond_stop from
367 * outside of the monitoring thread must be protected by @kdamond_lock.
369 * Note that the monitoring thread protects only @kdamond and @kdamond_stop via
370 * @kdamond_lock. Accesses to other fields must be protected by themselves.
372 * @primitive: Set of monitoring primitives for given use cases.
373 * @callback: Set of callbacks for monitoring events notifications.
375 * @min_nr_regions: The minimum number of adaptive monitoring regions.
376 * @max_nr_regions: The maximum number of adaptive monitoring regions.
377 * @adaptive_targets: Head of monitoring targets (&damon_target) list.
378 * @schemes: Head of schemes (&damos) list.
381 unsigned long sample_interval;
382 unsigned long aggr_interval;
383 unsigned long primitive_update_interval;
385 /* private: internal use only */
386 struct timespec64 last_aggregation;
387 struct timespec64 last_primitive_update;
390 struct task_struct *kdamond;
391 struct mutex kdamond_lock;
393 struct damon_primitive primitive;
394 struct damon_callback callback;
396 unsigned long min_nr_regions;
397 unsigned long max_nr_regions;
398 struct list_head adaptive_targets;
399 struct list_head schemes;
402 static inline struct damon_region *damon_next_region(struct damon_region *r)
404 return container_of(r->list.next, struct damon_region, list);
407 static inline struct damon_region *damon_prev_region(struct damon_region *r)
409 return container_of(r->list.prev, struct damon_region, list);
412 static inline struct damon_region *damon_last_region(struct damon_target *t)
414 return list_last_entry(&t->regions_list, struct damon_region, list);
417 #define damon_for_each_region(r, t) \
418 list_for_each_entry(r, &t->regions_list, list)
420 #define damon_for_each_region_safe(r, next, t) \
421 list_for_each_entry_safe(r, next, &t->regions_list, list)
423 #define damon_for_each_target(t, ctx) \
424 list_for_each_entry(t, &(ctx)->adaptive_targets, list)
426 #define damon_for_each_target_safe(t, next, ctx) \
427 list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
429 #define damon_for_each_scheme(s, ctx) \
430 list_for_each_entry(s, &(ctx)->schemes, list)
432 #define damon_for_each_scheme_safe(s, next, ctx) \
433 list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
437 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
438 inline void damon_insert_region(struct damon_region *r,
439 struct damon_region *prev, struct damon_region *next,
440 struct damon_target *t);
441 void damon_add_region(struct damon_region *r, struct damon_target *t);
442 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
444 struct damos *damon_new_scheme(
445 unsigned long min_sz_region, unsigned long max_sz_region,
446 unsigned int min_nr_accesses, unsigned int max_nr_accesses,
447 unsigned int min_age_region, unsigned int max_age_region,
448 enum damos_action action, struct damos_quota *quota,
449 struct damos_watermarks *wmarks);
450 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
451 void damon_destroy_scheme(struct damos *s);
453 struct damon_target *damon_new_target(unsigned long id);
454 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
455 bool damon_targets_empty(struct damon_ctx *ctx);
456 void damon_free_target(struct damon_target *t);
457 void damon_destroy_target(struct damon_target *t);
458 unsigned int damon_nr_regions(struct damon_target *t);
460 struct damon_ctx *damon_new_ctx(void);
461 void damon_destroy_ctx(struct damon_ctx *ctx);
462 int damon_set_targets(struct damon_ctx *ctx,
463 unsigned long *ids, ssize_t nr_ids);
464 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
465 unsigned long aggr_int, unsigned long primitive_upd_int,
466 unsigned long min_nr_reg, unsigned long max_nr_reg);
467 int damon_set_schemes(struct damon_ctx *ctx,
468 struct damos **schemes, ssize_t nr_schemes);
469 int damon_nr_running_ctxs(void);
471 int damon_start(struct damon_ctx **ctxs, int nr_ctxs);
472 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
474 #endif /* CONFIG_DAMON */
476 #ifdef CONFIG_DAMON_VADDR
477 bool damon_va_target_valid(void *t);
478 void damon_va_set_primitives(struct damon_ctx *ctx);
479 #endif /* CONFIG_DAMON_VADDR */
481 #ifdef CONFIG_DAMON_PADDR
482 bool damon_pa_target_valid(void *t);
483 void damon_pa_set_primitives(struct damon_ctx *ctx);
484 #endif /* CONFIG_DAMON_PADDR */
486 #endif /* _DAMON_H */