1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* 3 * DAMON api 4 * 5 * Author: SeongJae Park <sjpark@amazon.de> 6 */ 7 8 #ifndef _DAMON_H_ 9 #define _DAMON_H_ 10 11 #include <linux/mutex.h> 12 #include <linux/time64.h> 13 #include <linux/types.h> 14 #include <linux/random.h> 15 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) 20 21 /* Get a random number in [l, r) */ 22 static inline unsigned long damon_rand(unsigned long l, unsigned long r) 23 { 24 return l + prandom_u32_max(r - l); 25 } 26 27 /** 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). 31 */ 32 struct damon_addr_range { 33 unsigned long start; 34 unsigned long end; 35 }; 36 37 /** 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. 44 * 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. 49 */ 50 struct damon_region { 51 struct damon_addr_range ar; 52 unsigned long sampling_addr; 53 unsigned int nr_accesses; 54 struct list_head list; 55 56 unsigned int age; 57 /* private: Internal value for age calculation. */ 58 unsigned int last_nr_accesses; 59 }; 60 61 /** 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. 67 * 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 &struct damon_operations including the 71 * virtual address spaces monitoring operations. 72 */ 73 struct damon_target { 74 struct pid *pid; 75 unsigned int nr_regions; 76 struct list_head regions_list; 77 struct list_head list; 78 }; 79 80 /** 81 * enum damos_action - Represents an action of a Data Access Monitoring-based 82 * Operation Scheme. 83 * 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. 90 */ 91 enum damos_action { 92 DAMOS_WILLNEED, 93 DAMOS_COLD, 94 DAMOS_PAGEOUT, 95 DAMOS_HUGEPAGE, 96 DAMOS_NOHUGEPAGE, 97 DAMOS_STAT, /* Do nothing but only record the stat */ 98 }; 99 100 /** 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. 105 * 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. 109 * 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. 117 * 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. 121 * 122 * For selecting regions within the quota, DAMON prioritizes current scheme's 123 * target memory regions using the &struct damon_operations->get_scheme_score. 124 * You could customize the prioritization logic by setting &weight_sz, 125 * &weight_nr_accesses, and &weight_age, because monitoring operations are 126 * encouraged to respect those. 127 */ 128 struct damos_quota { 129 unsigned long ms; 130 unsigned long sz; 131 unsigned long reset_interval; 132 133 unsigned int weight_sz; 134 unsigned int weight_nr_accesses; 135 unsigned int weight_age; 136 137 /* private: */ 138 /* For throughput estimation */ 139 unsigned long total_charged_sz; 140 unsigned long total_charged_ns; 141 142 unsigned long esz; /* Effective size quota in bytes */ 143 144 /* 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; 149 150 /* For prioritization */ 151 unsigned long histogram[DAMOS_MAX_SCORE + 1]; 152 unsigned int min_score; 153 }; 154 155 /** 156 * enum damos_wmark_metric - Represents the watermark metric. 157 * 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]. 160 */ 161 enum damos_wmark_metric { 162 DAMOS_WMARK_NONE, 163 DAMOS_WMARK_FREE_MEM_RATE, 164 }; 165 166 /** 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. 173 * 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. 178 * 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. 182 */ 183 struct damos_watermarks { 184 enum damos_wmark_metric metric; 185 unsigned long interval; 186 unsigned long high; 187 unsigned long mid; 188 unsigned long low; 189 190 /* private: */ 191 bool activated; 192 }; 193 194 /** 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. 201 */ 202 struct damos_stat { 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; 208 }; 209 210 /** 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. 223 * 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. 229 * 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. 234 * 235 * If all schemes that registered to a &struct damon_ctx are inactive, DAMON 236 * stops monitoring and just repeatedly checks the watermarks. 237 * 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. 241 */ 242 struct damos { 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; 254 }; 255 256 /** 257 * enum damon_ops_id - Identifier for each monitoring operations implementation 258 * 259 * @DAMON_OPS_VADDR: Monitoring operations for virtual address spaces 260 * @DAMON_OPS_PADDR: Monitoring operations for the physical address space 261 */ 262 enum damon_ops_id { 263 DAMON_OPS_VADDR, 264 DAMON_OPS_PADDR, 265 NR_DAMON_OPS, 266 }; 267 268 struct damon_ctx; 269 270 /** 271 * struct damon_operations - Monitoring operations for given use cases. 272 * 273 * @id: Identifier of this operations set. 274 * @init: Initialize operations-related data structures. 275 * @update: Update operations-related data structures. 276 * @prepare_access_checks: Prepare next access check of target regions. 277 * @check_accesses: Check the accesses to target regions. 278 * @reset_aggregated: Reset aggregated accesses monitoring results. 279 * @get_scheme_score: Get the score of a region for a scheme. 280 * @apply_scheme: Apply a DAMON-based operation scheme. 281 * @target_valid: Determine if the target is valid. 282 * @cleanup: Clean up the context. 283 * 284 * DAMON can be extended for various address spaces and usages. For this, 285 * users should register the low level operations for their target address 286 * space and usecase via the &damon_ctx.ops. Then, the monitoring thread 287 * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting 288 * the monitoring, @update after each &damon_ctx.ops_update_interval, and 289 * @check_accesses, @target_valid and @prepare_access_checks after each 290 * &damon_ctx.sample_interval. Finally, @reset_aggregated is called after each 291 * &damon_ctx.aggr_interval. 292 * 293 * Each &struct damon_operations instance having valid @id can be registered 294 * via damon_register_ops() and selected by damon_select_ops() later. 295 * @init should initialize operations-related data structures. For example, 296 * this could be used to construct proper monitoring target regions and link 297 * those to @damon_ctx.adaptive_targets. 298 * @update should update the operations-related data structures. For example, 299 * this could be used to update monitoring target regions for current status. 300 * @prepare_access_checks should manipulate the monitoring regions to be 301 * prepared for the next access check. 302 * @check_accesses should check the accesses to each region that made after the 303 * last preparation and update the number of observed accesses of each region. 304 * It should also return max number of observed accesses that made as a result 305 * of its update. The value will be used for regions adjustment threshold. 306 * @reset_aggregated should reset the access monitoring results that aggregated 307 * by @check_accesses. 308 * @get_scheme_score should return the priority score of a region for a scheme 309 * as an integer in [0, &DAMOS_MAX_SCORE]. 310 * @apply_scheme is called from @kdamond when a region for user provided 311 * DAMON-based operation scheme is found. It should apply the scheme's action 312 * to the region and return bytes of the region that the action is successfully 313 * applied. 314 * @target_valid should check whether the target is still valid for the 315 * monitoring. 316 * @cleanup is called from @kdamond just before its termination. 317 */ 318 struct damon_operations { 319 enum damon_ops_id id; 320 void (*init)(struct damon_ctx *context); 321 void (*update)(struct damon_ctx *context); 322 void (*prepare_access_checks)(struct damon_ctx *context); 323 unsigned int (*check_accesses)(struct damon_ctx *context); 324 void (*reset_aggregated)(struct damon_ctx *context); 325 int (*get_scheme_score)(struct damon_ctx *context, 326 struct damon_target *t, struct damon_region *r, 327 struct damos *scheme); 328 unsigned long (*apply_scheme)(struct damon_ctx *context, 329 struct damon_target *t, struct damon_region *r, 330 struct damos *scheme); 331 bool (*target_valid)(void *target); 332 void (*cleanup)(struct damon_ctx *context); 333 }; 334 335 /** 336 * struct damon_callback - Monitoring events notification callbacks. 337 * 338 * @before_start: Called before starting the monitoring. 339 * @after_sampling: Called after each sampling. 340 * @after_aggregation: Called after each aggregation. 341 * @before_terminate: Called before terminating the monitoring. 342 * @private: User private data. 343 * 344 * The monitoring thread (&damon_ctx.kdamond) calls @before_start and 345 * @before_terminate just before starting and finishing the monitoring, 346 * respectively. Therefore, those are good places for installing and cleaning 347 * @private. 348 * 349 * The monitoring thread calls @after_sampling and @after_aggregation for each 350 * of the sampling intervals and aggregation intervals, respectively. 351 * Therefore, users can safely access the monitoring results without additional 352 * protection. For the reason, users are recommended to use these callback for 353 * the accesses to the results. 354 * 355 * If any callback returns non-zero, monitoring stops. 356 */ 357 struct damon_callback { 358 void *private; 359 360 int (*before_start)(struct damon_ctx *context); 361 int (*after_sampling)(struct damon_ctx *context); 362 int (*after_aggregation)(struct damon_ctx *context); 363 void (*before_terminate)(struct damon_ctx *context); 364 }; 365 366 /** 367 * struct damon_ctx - Represents a context for each monitoring. This is the 368 * main interface that allows users to set the attributes and get the results 369 * of the monitoring. 370 * 371 * @sample_interval: The time between access samplings. 372 * @aggr_interval: The time between monitor results aggregations. 373 * @ops_update_interval: The time between monitoring operations updates. 374 * 375 * For each @sample_interval, DAMON checks whether each region is accessed or 376 * not. It aggregates and keeps the access information (number of accesses to 377 * each region) for @aggr_interval time. DAMON also checks whether the target 378 * memory regions need update (e.g., by ``mmap()`` calls from the application, 379 * in case of virtual memory monitoring) and applies the changes for each 380 * @ops_update_interval. All time intervals are in micro-seconds. 381 * Please refer to &struct damon_operations and &struct damon_callback for more 382 * detail. 383 * 384 * @kdamond: Kernel thread who does the monitoring. 385 * @kdamond_stop: Notifies whether kdamond should stop. 386 * @kdamond_lock: Mutex for the synchronizations with @kdamond. 387 * 388 * For each monitoring context, one kernel thread for the monitoring is 389 * created. The pointer to the thread is stored in @kdamond. 390 * 391 * Once started, the monitoring thread runs until explicitly required to be 392 * terminated or every monitoring target is invalid. The validity of the 393 * targets is checked via the &damon_operations.target_valid of @ops. The 394 * termination can also be explicitly requested by writing non-zero to 395 * @kdamond_stop. The thread sets @kdamond to NULL when it terminates. 396 * Therefore, users can know whether the monitoring is ongoing or terminated by 397 * reading @kdamond. Reads and writes to @kdamond and @kdamond_stop from 398 * outside of the monitoring thread must be protected by @kdamond_lock. 399 * 400 * Note that the monitoring thread protects only @kdamond and @kdamond_stop via 401 * @kdamond_lock. Accesses to other fields must be protected by themselves. 402 * 403 * @ops: Set of monitoring operations for given use cases. 404 * @callback: Set of callbacks for monitoring events notifications. 405 * 406 * @min_nr_regions: The minimum number of adaptive monitoring regions. 407 * @max_nr_regions: The maximum number of adaptive monitoring regions. 408 * @adaptive_targets: Head of monitoring targets (&damon_target) list. 409 * @schemes: Head of schemes (&damos) list. 410 */ 411 struct damon_ctx { 412 unsigned long sample_interval; 413 unsigned long aggr_interval; 414 unsigned long ops_update_interval; 415 416 /* private: internal use only */ 417 struct timespec64 last_aggregation; 418 struct timespec64 last_ops_update; 419 420 /* public: */ 421 struct task_struct *kdamond; 422 struct mutex kdamond_lock; 423 424 struct damon_operations ops; 425 struct damon_callback callback; 426 427 unsigned long min_nr_regions; 428 unsigned long max_nr_regions; 429 struct list_head adaptive_targets; 430 struct list_head schemes; 431 }; 432 433 static inline struct damon_region *damon_next_region(struct damon_region *r) 434 { 435 return container_of(r->list.next, struct damon_region, list); 436 } 437 438 static inline struct damon_region *damon_prev_region(struct damon_region *r) 439 { 440 return container_of(r->list.prev, struct damon_region, list); 441 } 442 443 static inline struct damon_region *damon_last_region(struct damon_target *t) 444 { 445 return list_last_entry(&t->regions_list, struct damon_region, list); 446 } 447 448 #define damon_for_each_region(r, t) \ 449 list_for_each_entry(r, &t->regions_list, list) 450 451 #define damon_for_each_region_safe(r, next, t) \ 452 list_for_each_entry_safe(r, next, &t->regions_list, list) 453 454 #define damon_for_each_target(t, ctx) \ 455 list_for_each_entry(t, &(ctx)->adaptive_targets, list) 456 457 #define damon_for_each_target_safe(t, next, ctx) \ 458 list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list) 459 460 #define damon_for_each_scheme(s, ctx) \ 461 list_for_each_entry(s, &(ctx)->schemes, list) 462 463 #define damon_for_each_scheme_safe(s, next, ctx) \ 464 list_for_each_entry_safe(s, next, &(ctx)->schemes, list) 465 466 #ifdef CONFIG_DAMON 467 468 struct damon_region *damon_new_region(unsigned long start, unsigned long end); 469 470 /* 471 * Add a region between two other regions 472 */ 473 static inline void damon_insert_region(struct damon_region *r, 474 struct damon_region *prev, struct damon_region *next, 475 struct damon_target *t) 476 { 477 __list_add(&r->list, &prev->list, &next->list); 478 t->nr_regions++; 479 } 480 481 void damon_add_region(struct damon_region *r, struct damon_target *t); 482 void damon_destroy_region(struct damon_region *r, struct damon_target *t); 483 484 struct damos *damon_new_scheme( 485 unsigned long min_sz_region, unsigned long max_sz_region, 486 unsigned int min_nr_accesses, unsigned int max_nr_accesses, 487 unsigned int min_age_region, unsigned int max_age_region, 488 enum damos_action action, struct damos_quota *quota, 489 struct damos_watermarks *wmarks); 490 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s); 491 void damon_destroy_scheme(struct damos *s); 492 493 struct damon_target *damon_new_target(void); 494 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t); 495 bool damon_targets_empty(struct damon_ctx *ctx); 496 void damon_free_target(struct damon_target *t); 497 void damon_destroy_target(struct damon_target *t); 498 unsigned int damon_nr_regions(struct damon_target *t); 499 500 struct damon_ctx *damon_new_ctx(void); 501 void damon_destroy_ctx(struct damon_ctx *ctx); 502 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int, 503 unsigned long aggr_int, unsigned long ops_upd_int, 504 unsigned long min_nr_reg, unsigned long max_nr_reg); 505 int damon_set_schemes(struct damon_ctx *ctx, 506 struct damos **schemes, ssize_t nr_schemes); 507 int damon_nr_running_ctxs(void); 508 int damon_register_ops(struct damon_operations *ops); 509 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id); 510 511 int damon_start(struct damon_ctx **ctxs, int nr_ctxs); 512 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs); 513 514 #endif /* CONFIG_DAMON */ 515 516 #endif /* _DAMON_H */ 517