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 * @pid: The PID of the virtual address space to monitor.
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 * @pid should be set for appropriate address space monitoring primitives
71 * including the virtual address spaces monitoring primitives.
75 unsigned int nr_regions;
76 struct list_head regions_list;
77 struct list_head list;
81 * enum damos_action - Represents an action of a Data Access Monitoring-based
84 * @DAMOS_WILLNEED: Call ``madvise()`` for the region with MADV_WILLNEED.
85 * @DAMOS_COLD: Call ``madvise()`` for the region with MADV_COLD.
86 * @DAMOS_PAGEOUT: Call ``madvise()`` for the region with MADV_PAGEOUT.
87 * @DAMOS_HUGEPAGE: Call ``madvise()`` for the region with MADV_HUGEPAGE.
88 * @DAMOS_NOHUGEPAGE: Call ``madvise()`` for the region with MADV_NOHUGEPAGE.
89 * @DAMOS_STAT: Do nothing but count the stat.
97 DAMOS_STAT, /* Do nothing but only record the stat */
101 * struct damos_quota - Controls the aggressiveness of the given scheme.
102 * @ms: Maximum milliseconds that the scheme can use.
103 * @sz: Maximum bytes of memory that the action can be applied.
104 * @reset_interval: Charge reset interval in milliseconds.
106 * @weight_sz: Weight of the region's size for prioritization.
107 * @weight_nr_accesses: Weight of the region's nr_accesses for prioritization.
108 * @weight_age: Weight of the region's age for prioritization.
110 * To avoid consuming too much CPU time or IO resources for applying the
111 * &struct damos->action to large memory, DAMON allows users to set time and/or
112 * size quotas. The quotas can be set by writing non-zero values to &ms and
113 * &sz, respectively. If the time quota is set, DAMON tries to use only up to
114 * &ms milliseconds within &reset_interval for applying the action. If the
115 * size quota is set, DAMON tries to apply the action only up to &sz bytes
116 * within &reset_interval.
118 * Internally, the time quota is transformed to a size quota using estimated
119 * throughput of the scheme's action. DAMON then compares it against &sz and
120 * uses smaller one as the effective quota.
122 * For selecting regions within the quota, DAMON prioritizes current scheme's
123 * target memory regions using the &struct damon_primitive->get_scheme_score.
124 * You could customize the prioritization logic by setting &weight_sz,
125 * &weight_nr_accesses, and &weight_age, because monitoring primitives are
126 * encouraged to respect those.
131 unsigned long reset_interval;
133 unsigned int weight_sz;
134 unsigned int weight_nr_accesses;
135 unsigned int weight_age;
138 /* For throughput estimation */
139 unsigned long total_charged_sz;
140 unsigned long total_charged_ns;
142 unsigned long esz; /* Effective size quota in bytes */
144 /* For charging the quota */
145 unsigned long charged_sz;
146 unsigned long charged_from;
147 struct damon_target *charge_target_from;
148 unsigned long charge_addr_from;
150 /* For prioritization */
151 unsigned long histogram[DAMOS_MAX_SCORE + 1];
152 unsigned int min_score;
156 * enum damos_wmark_metric - Represents the watermark metric.
158 * @DAMOS_WMARK_NONE: Ignore the watermarks of the given scheme.
159 * @DAMOS_WMARK_FREE_MEM_RATE: Free memory rate of the system in [0,1000].
161 enum damos_wmark_metric {
163 DAMOS_WMARK_FREE_MEM_RATE,
167 * struct damos_watermarks - Controls when a given scheme should be activated.
168 * @metric: Metric for the watermarks.
169 * @interval: Watermarks check time interval in microseconds.
170 * @high: High watermark.
171 * @mid: Middle watermark.
172 * @low: Low watermark.
174 * If &metric is &DAMOS_WMARK_NONE, the scheme is always active. Being active
175 * means DAMON does monitoring and applying the action of the scheme to
176 * appropriate memory regions. Else, DAMON checks &metric of the system for at
177 * least every &interval microseconds and works as below.
179 * If &metric is higher than &high, the scheme is inactivated. If &metric is
180 * between &mid and &low, the scheme is activated. If &metric is lower than
181 * &low, the scheme is inactivated.
183 struct damos_watermarks {
184 enum damos_wmark_metric metric;
185 unsigned long interval;
195 * struct damos_stat - Statistics on a given scheme.
196 * @nr_tried: Total number of regions that the scheme is tried to be applied.
197 * @sz_tried: Total size of regions that the scheme is tried to be applied.
198 * @nr_applied: Total number of regions that the scheme is applied.
199 * @sz_applied: Total size of regions that the scheme is applied.
200 * @qt_exceeds: Total number of times the quota of the scheme has exceeded.
203 unsigned long nr_tried;
204 unsigned long sz_tried;
205 unsigned long nr_applied;
206 unsigned long sz_applied;
207 unsigned long qt_exceeds;
211 * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
212 * @min_sz_region: Minimum size of target regions.
213 * @max_sz_region: Maximum size of target regions.
214 * @min_nr_accesses: Minimum ``->nr_accesses`` of target regions.
215 * @max_nr_accesses: Maximum ``->nr_accesses`` of target regions.
216 * @min_age_region: Minimum age of target regions.
217 * @max_age_region: Maximum age of target regions.
218 * @action: &damo_action to be applied to the target regions.
219 * @quota: Control the aggressiveness of this scheme.
220 * @wmarks: Watermarks for automated (in)activation of this scheme.
221 * @stat: Statistics of this scheme.
222 * @list: List head for siblings.
224 * For each aggregation interval, DAMON finds regions which fit in the
225 * condition (&min_sz_region, &max_sz_region, &min_nr_accesses,
226 * &max_nr_accesses, &min_age_region, &max_age_region) and applies &action to
227 * those. To avoid consuming too much CPU time or IO resources for the
228 * &action, "a is used.
230 * To do the work only when needed, schemes can be activated for specific
231 * system situations using &wmarks. If all schemes that registered to the
232 * monitoring context are inactive, DAMON stops monitoring either, and just
233 * repeatedly checks the watermarks.
235 * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
236 * stops monitoring and just repeatedly checks the watermarks.
238 * After applying the &action to each region, &stat_count and &stat_sz is
239 * updated to reflect the number of regions and total size of regions that the
240 * &action is applied.
243 unsigned long min_sz_region;
244 unsigned long max_sz_region;
245 unsigned int min_nr_accesses;
246 unsigned int max_nr_accesses;
247 unsigned int min_age_region;
248 unsigned int max_age_region;
249 enum damos_action action;
250 struct damos_quota quota;
251 struct damos_watermarks wmarks;
252 struct damos_stat stat;
253 struct list_head list;
259 * struct damon_primitive - Monitoring primitives for given use cases.
261 * @init: Initialize primitive-internal data structures.
262 * @update: Update primitive-internal data structures.
263 * @prepare_access_checks: Prepare next access check of target regions.
264 * @check_accesses: Check the accesses to target regions.
265 * @reset_aggregated: Reset aggregated accesses monitoring results.
266 * @get_scheme_score: Get the score of a region for a scheme.
267 * @apply_scheme: Apply a DAMON-based operation scheme.
268 * @target_valid: Determine if the target is valid.
269 * @cleanup: Clean up the context.
271 * DAMON can be extended for various address spaces and usages. For this,
272 * users should register the low level primitives for their target address
273 * space and usecase via the &damon_ctx.primitive. Then, the monitoring thread
274 * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
275 * the monitoring, @update after each &damon_ctx.primitive_update_interval, and
276 * @check_accesses, @target_valid and @prepare_access_checks after each
277 * &damon_ctx.sample_interval. Finally, @reset_aggregated is called after each
278 * &damon_ctx.aggr_interval.
280 * @init should initialize primitive-internal data structures. For example,
281 * this could be used to construct proper monitoring target regions and link
282 * those to @damon_ctx.adaptive_targets.
283 * @update should update the primitive-internal data structures. For example,
284 * this could be used to update monitoring target regions for current status.
285 * @prepare_access_checks should manipulate the monitoring regions to be
286 * prepared for the next access check.
287 * @check_accesses should check the accesses to each region that made after the
288 * last preparation and update the number of observed accesses of each region.
289 * It should also return max number of observed accesses that made as a result
290 * of its update. The value will be used for regions adjustment threshold.
291 * @reset_aggregated should reset the access monitoring results that aggregated
292 * by @check_accesses.
293 * @get_scheme_score should return the priority score of a region for a scheme
294 * as an integer in [0, &DAMOS_MAX_SCORE].
295 * @apply_scheme is called from @kdamond when a region for user provided
296 * DAMON-based operation scheme is found. It should apply the scheme's action
297 * to the region and return bytes of the region that the action is successfully
299 * @target_valid should check whether the target is still valid for the
301 * @cleanup is called from @kdamond just before its termination.
303 struct damon_primitive {
304 void (*init)(struct damon_ctx *context);
305 void (*update)(struct damon_ctx *context);
306 void (*prepare_access_checks)(struct damon_ctx *context);
307 unsigned int (*check_accesses)(struct damon_ctx *context);
308 void (*reset_aggregated)(struct damon_ctx *context);
309 int (*get_scheme_score)(struct damon_ctx *context,
310 struct damon_target *t, struct damon_region *r,
311 struct damos *scheme);
312 unsigned long (*apply_scheme)(struct damon_ctx *context,
313 struct damon_target *t, struct damon_region *r,
314 struct damos *scheme);
315 bool (*target_valid)(void *target);
316 void (*cleanup)(struct damon_ctx *context);
320 * struct damon_callback - Monitoring events notification callbacks.
322 * @before_start: Called before starting the monitoring.
323 * @after_sampling: Called after each sampling.
324 * @after_aggregation: Called after each aggregation.
325 * @before_terminate: Called before terminating the monitoring.
326 * @private: User private data.
328 * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
329 * @before_terminate just before starting and finishing the monitoring,
330 * respectively. Therefore, those are good places for installing and cleaning
333 * The monitoring thread calls @after_sampling and @after_aggregation for each
334 * of the sampling intervals and aggregation intervals, respectively.
335 * Therefore, users can safely access the monitoring results without additional
336 * protection. For the reason, users are recommended to use these callback for
337 * the accesses to the results.
339 * If any callback returns non-zero, monitoring stops.
341 struct damon_callback {
344 int (*before_start)(struct damon_ctx *context);
345 int (*after_sampling)(struct damon_ctx *context);
346 int (*after_aggregation)(struct damon_ctx *context);
347 void (*before_terminate)(struct damon_ctx *context);
351 * struct damon_ctx - Represents a context for each monitoring. This is the
352 * main interface that allows users to set the attributes and get the results
355 * @sample_interval: The time between access samplings.
356 * @aggr_interval: The time between monitor results aggregations.
357 * @primitive_update_interval: The time between monitoring primitive updates.
359 * For each @sample_interval, DAMON checks whether each region is accessed or
360 * not. It aggregates and keeps the access information (number of accesses to
361 * each region) for @aggr_interval time. DAMON also checks whether the target
362 * memory regions need update (e.g., by ``mmap()`` calls from the application,
363 * in case of virtual memory monitoring) and applies the changes for each
364 * @primitive_update_interval. All time intervals are in micro-seconds.
365 * Please refer to &struct damon_primitive and &struct damon_callback for more
368 * @kdamond: Kernel thread who does the monitoring.
369 * @kdamond_stop: Notifies whether kdamond should stop.
370 * @kdamond_lock: Mutex for the synchronizations with @kdamond.
372 * For each monitoring context, one kernel thread for the monitoring is
373 * created. The pointer to the thread is stored in @kdamond.
375 * Once started, the monitoring thread runs until explicitly required to be
376 * terminated or every monitoring target is invalid. The validity of the
377 * targets is checked via the &damon_primitive.target_valid of @primitive. The
378 * termination can also be explicitly requested by writing non-zero to
379 * @kdamond_stop. The thread sets @kdamond to NULL when it terminates.
380 * Therefore, users can know whether the monitoring is ongoing or terminated by
381 * reading @kdamond. Reads and writes to @kdamond and @kdamond_stop from
382 * outside of the monitoring thread must be protected by @kdamond_lock.
384 * Note that the monitoring thread protects only @kdamond and @kdamond_stop via
385 * @kdamond_lock. Accesses to other fields must be protected by themselves.
387 * @primitive: Set of monitoring primitives for given use cases.
388 * @callback: Set of callbacks for monitoring events notifications.
390 * @min_nr_regions: The minimum number of adaptive monitoring regions.
391 * @max_nr_regions: The maximum number of adaptive monitoring regions.
392 * @adaptive_targets: Head of monitoring targets (&damon_target) list.
393 * @schemes: Head of schemes (&damos) list.
396 unsigned long sample_interval;
397 unsigned long aggr_interval;
398 unsigned long primitive_update_interval;
400 /* private: internal use only */
401 struct timespec64 last_aggregation;
402 struct timespec64 last_primitive_update;
405 struct task_struct *kdamond;
406 struct mutex kdamond_lock;
408 struct damon_primitive primitive;
409 struct damon_callback callback;
411 unsigned long min_nr_regions;
412 unsigned long max_nr_regions;
413 struct list_head adaptive_targets;
414 struct list_head schemes;
417 static inline struct damon_region *damon_next_region(struct damon_region *r)
419 return container_of(r->list.next, struct damon_region, list);
422 static inline struct damon_region *damon_prev_region(struct damon_region *r)
424 return container_of(r->list.prev, struct damon_region, list);
427 static inline struct damon_region *damon_last_region(struct damon_target *t)
429 return list_last_entry(&t->regions_list, struct damon_region, list);
432 #define damon_for_each_region(r, t) \
433 list_for_each_entry(r, &t->regions_list, list)
435 #define damon_for_each_region_safe(r, next, t) \
436 list_for_each_entry_safe(r, next, &t->regions_list, list)
438 #define damon_for_each_target(t, ctx) \
439 list_for_each_entry(t, &(ctx)->adaptive_targets, list)
441 #define damon_for_each_target_safe(t, next, ctx) \
442 list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
444 #define damon_for_each_scheme(s, ctx) \
445 list_for_each_entry(s, &(ctx)->schemes, list)
447 #define damon_for_each_scheme_safe(s, next, ctx) \
448 list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
452 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
455 * Add a region between two other regions
457 static inline void damon_insert_region(struct damon_region *r,
458 struct damon_region *prev, struct damon_region *next,
459 struct damon_target *t)
461 __list_add(&r->list, &prev->list, &next->list);
465 void damon_add_region(struct damon_region *r, struct damon_target *t);
466 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
468 struct damos *damon_new_scheme(
469 unsigned long min_sz_region, unsigned long max_sz_region,
470 unsigned int min_nr_accesses, unsigned int max_nr_accesses,
471 unsigned int min_age_region, unsigned int max_age_region,
472 enum damos_action action, struct damos_quota *quota,
473 struct damos_watermarks *wmarks);
474 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
475 void damon_destroy_scheme(struct damos *s);
477 struct damon_target *damon_new_target(void);
478 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
479 bool damon_targets_empty(struct damon_ctx *ctx);
480 void damon_free_target(struct damon_target *t);
481 void damon_destroy_target(struct damon_target *t);
482 unsigned int damon_nr_regions(struct damon_target *t);
484 struct damon_ctx *damon_new_ctx(void);
485 void damon_destroy_ctx(struct damon_ctx *ctx);
486 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
487 unsigned long aggr_int, unsigned long primitive_upd_int,
488 unsigned long min_nr_reg, unsigned long max_nr_reg);
489 int damon_set_schemes(struct damon_ctx *ctx,
490 struct damos **schemes, ssize_t nr_schemes);
491 int damon_nr_running_ctxs(void);
493 int damon_start(struct damon_ctx **ctxs, int nr_ctxs);
494 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
496 #endif /* CONFIG_DAMON */
498 #ifdef CONFIG_DAMON_VADDR
499 bool damon_va_target_valid(void *t);
500 void damon_va_set_primitives(struct damon_ctx *ctx);
501 #endif /* CONFIG_DAMON_VADDR */
503 #ifdef CONFIG_DAMON_PADDR
504 bool damon_pa_target_valid(void *t);
505 void damon_pa_set_primitives(struct damon_ctx *ctx);
506 #endif /* CONFIG_DAMON_PADDR */
508 #endif /* _DAMON_H */