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) */
damon_rand(unsigned long l,unsigned long 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  * @NR_DAMOS_ACTIONS:	Total number of DAMOS actions
91  */
92 enum damos_action {
93 	DAMOS_WILLNEED,
94 	DAMOS_COLD,
95 	DAMOS_PAGEOUT,
96 	DAMOS_HUGEPAGE,
97 	DAMOS_NOHUGEPAGE,
98 	DAMOS_STAT,		/* Do nothing but only record the stat */
99 	NR_DAMOS_ACTIONS,
100 };
101 
102 /**
103  * struct damos_quota - Controls the aggressiveness of the given scheme.
104  * @ms:			Maximum milliseconds that the scheme can use.
105  * @sz:			Maximum bytes of memory that the action can be applied.
106  * @reset_interval:	Charge reset interval in milliseconds.
107  *
108  * @weight_sz:		Weight of the region's size for prioritization.
109  * @weight_nr_accesses:	Weight of the region's nr_accesses for prioritization.
110  * @weight_age:		Weight of the region's age for prioritization.
111  *
112  * To avoid consuming too much CPU time or IO resources for applying the
113  * &struct damos->action to large memory, DAMON allows users to set time and/or
114  * size quotas.  The quotas can be set by writing non-zero values to &ms and
115  * &sz, respectively.  If the time quota is set, DAMON tries to use only up to
116  * &ms milliseconds within &reset_interval for applying the action.  If the
117  * size quota is set, DAMON tries to apply the action only up to &sz bytes
118  * within &reset_interval.
119  *
120  * Internally, the time quota is transformed to a size quota using estimated
121  * throughput of the scheme's action.  DAMON then compares it against &sz and
122  * uses smaller one as the effective quota.
123  *
124  * For selecting regions within the quota, DAMON prioritizes current scheme's
125  * target memory regions using the &struct damon_operations->get_scheme_score.
126  * You could customize the prioritization logic by setting &weight_sz,
127  * &weight_nr_accesses, and &weight_age, because monitoring operations are
128  * encouraged to respect those.
129  */
130 struct damos_quota {
131 	unsigned long ms;
132 	unsigned long sz;
133 	unsigned long reset_interval;
134 
135 	unsigned int weight_sz;
136 	unsigned int weight_nr_accesses;
137 	unsigned int weight_age;
138 
139 /* private: */
140 	/* For throughput estimation */
141 	unsigned long total_charged_sz;
142 	unsigned long total_charged_ns;
143 
144 	unsigned long esz;	/* Effective size quota in bytes */
145 
146 	/* For charging the quota */
147 	unsigned long charged_sz;
148 	unsigned long charged_from;
149 	struct damon_target *charge_target_from;
150 	unsigned long charge_addr_from;
151 
152 	/* For prioritization */
153 	unsigned long histogram[DAMOS_MAX_SCORE + 1];
154 	unsigned int min_score;
155 };
156 
157 /**
158  * enum damos_wmark_metric - Represents the watermark metric.
159  *
160  * @DAMOS_WMARK_NONE:		Ignore the watermarks of the given scheme.
161  * @DAMOS_WMARK_FREE_MEM_RATE:	Free memory rate of the system in [0,1000].
162  * @NR_DAMOS_WMARK_METRICS:	Total number of DAMOS watermark metrics
163  */
164 enum damos_wmark_metric {
165 	DAMOS_WMARK_NONE,
166 	DAMOS_WMARK_FREE_MEM_RATE,
167 	NR_DAMOS_WMARK_METRICS,
168 };
169 
170 /**
171  * struct damos_watermarks - Controls when a given scheme should be activated.
172  * @metric:	Metric for the watermarks.
173  * @interval:	Watermarks check time interval in microseconds.
174  * @high:	High watermark.
175  * @mid:	Middle watermark.
176  * @low:	Low watermark.
177  *
178  * If &metric is &DAMOS_WMARK_NONE, the scheme is always active.  Being active
179  * means DAMON does monitoring and applying the action of the scheme to
180  * appropriate memory regions.  Else, DAMON checks &metric of the system for at
181  * least every &interval microseconds and works as below.
182  *
183  * If &metric is higher than &high, the scheme is inactivated.  If &metric is
184  * between &mid and &low, the scheme is activated.  If &metric is lower than
185  * &low, the scheme is inactivated.
186  */
187 struct damos_watermarks {
188 	enum damos_wmark_metric metric;
189 	unsigned long interval;
190 	unsigned long high;
191 	unsigned long mid;
192 	unsigned long low;
193 
194 /* private: */
195 	bool activated;
196 };
197 
198 /**
199  * struct damos_stat - Statistics on a given scheme.
200  * @nr_tried:	Total number of regions that the scheme is tried to be applied.
201  * @sz_tried:	Total size of regions that the scheme is tried to be applied.
202  * @nr_applied:	Total number of regions that the scheme is applied.
203  * @sz_applied:	Total size of regions that the scheme is applied.
204  * @qt_exceeds: Total number of times the quota of the scheme has exceeded.
205  */
206 struct damos_stat {
207 	unsigned long nr_tried;
208 	unsigned long sz_tried;
209 	unsigned long nr_applied;
210 	unsigned long sz_applied;
211 	unsigned long qt_exceeds;
212 };
213 
214 /**
215  * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
216  * @min_sz_region:	Minimum size of target regions.
217  * @max_sz_region:	Maximum size of target regions.
218  * @min_nr_accesses:	Minimum ``->nr_accesses`` of target regions.
219  * @max_nr_accesses:	Maximum ``->nr_accesses`` of target regions.
220  * @min_age_region:	Minimum age of target regions.
221  * @max_age_region:	Maximum age of target regions.
222  * @action:		&damo_action to be applied to the target regions.
223  * @quota:		Control the aggressiveness of this scheme.
224  * @wmarks:		Watermarks for automated (in)activation of this scheme.
225  * @stat:		Statistics of this scheme.
226  * @list:		List head for siblings.
227  *
228  * For each aggregation interval, DAMON finds regions which fit in the
229  * condition (&min_sz_region, &max_sz_region, &min_nr_accesses,
230  * &max_nr_accesses, &min_age_region, &max_age_region) and applies &action to
231  * those.  To avoid consuming too much CPU time or IO resources for the
232  * &action, &quota is used.
233  *
234  * To do the work only when needed, schemes can be activated for specific
235  * system situations using &wmarks.  If all schemes that registered to the
236  * monitoring context are inactive, DAMON stops monitoring either, and just
237  * repeatedly checks the watermarks.
238  *
239  * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
240  * stops monitoring and just repeatedly checks the watermarks.
241  *
242  * After applying the &action to each region, &stat_count and &stat_sz is
243  * updated to reflect the number of regions and total size of regions that the
244  * &action is applied.
245  */
246 struct damos {
247 	unsigned long min_sz_region;
248 	unsigned long max_sz_region;
249 	unsigned int min_nr_accesses;
250 	unsigned int max_nr_accesses;
251 	unsigned int min_age_region;
252 	unsigned int max_age_region;
253 	enum damos_action action;
254 	struct damos_quota quota;
255 	struct damos_watermarks wmarks;
256 	struct damos_stat stat;
257 	struct list_head list;
258 };
259 
260 /**
261  * enum damon_ops_id - Identifier for each monitoring operations implementation
262  *
263  * @DAMON_OPS_VADDR:	Monitoring operations for virtual address spaces
264  * @DAMON_OPS_FVADDR:	Monitoring operations for only fixed ranges of virtual
265  *			address spaces
266  * @DAMON_OPS_PADDR:	Monitoring operations for the physical address space
267  * @NR_DAMON_OPS:	Number of monitoring operations implementations
268  */
269 enum damon_ops_id {
270 	DAMON_OPS_VADDR,
271 	DAMON_OPS_FVADDR,
272 	DAMON_OPS_PADDR,
273 	NR_DAMON_OPS,
274 };
275 
276 struct damon_ctx;
277 
278 /**
279  * struct damon_operations - Monitoring operations for given use cases.
280  *
281  * @id:				Identifier of this operations set.
282  * @init:			Initialize operations-related data structures.
283  * @update:			Update operations-related data structures.
284  * @prepare_access_checks:	Prepare next access check of target regions.
285  * @check_accesses:		Check the accesses to target regions.
286  * @reset_aggregated:		Reset aggregated accesses monitoring results.
287  * @get_scheme_score:		Get the score of a region for a scheme.
288  * @apply_scheme:		Apply a DAMON-based operation scheme.
289  * @target_valid:		Determine if the target is valid.
290  * @cleanup:			Clean up the context.
291  *
292  * DAMON can be extended for various address spaces and usages.  For this,
293  * users should register the low level operations for their target address
294  * space and usecase via the &damon_ctx.ops.  Then, the monitoring thread
295  * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
296  * the monitoring, @update after each &damon_ctx.ops_update_interval, and
297  * @check_accesses, @target_valid and @prepare_access_checks after each
298  * &damon_ctx.sample_interval.  Finally, @reset_aggregated is called after each
299  * &damon_ctx.aggr_interval.
300  *
301  * Each &struct damon_operations instance having valid @id can be registered
302  * via damon_register_ops() and selected by damon_select_ops() later.
303  * @init should initialize operations-related data structures.  For example,
304  * this could be used to construct proper monitoring target regions and link
305  * those to @damon_ctx.adaptive_targets.
306  * @update should update the operations-related data structures.  For example,
307  * this could be used to update monitoring target regions for current status.
308  * @prepare_access_checks should manipulate the monitoring regions to be
309  * prepared for the next access check.
310  * @check_accesses should check the accesses to each region that made after the
311  * last preparation and update the number of observed accesses of each region.
312  * It should also return max number of observed accesses that made as a result
313  * of its update.  The value will be used for regions adjustment threshold.
314  * @reset_aggregated should reset the access monitoring results that aggregated
315  * by @check_accesses.
316  * @get_scheme_score should return the priority score of a region for a scheme
317  * as an integer in [0, &DAMOS_MAX_SCORE].
318  * @apply_scheme is called from @kdamond when a region for user provided
319  * DAMON-based operation scheme is found.  It should apply the scheme's action
320  * to the region and return bytes of the region that the action is successfully
321  * applied.
322  * @target_valid should check whether the target is still valid for the
323  * monitoring.
324  * @cleanup is called from @kdamond just before its termination.
325  */
326 struct damon_operations {
327 	enum damon_ops_id id;
328 	void (*init)(struct damon_ctx *context);
329 	void (*update)(struct damon_ctx *context);
330 	void (*prepare_access_checks)(struct damon_ctx *context);
331 	unsigned int (*check_accesses)(struct damon_ctx *context);
332 	void (*reset_aggregated)(struct damon_ctx *context);
333 	int (*get_scheme_score)(struct damon_ctx *context,
334 			struct damon_target *t, struct damon_region *r,
335 			struct damos *scheme);
336 	unsigned long (*apply_scheme)(struct damon_ctx *context,
337 			struct damon_target *t, struct damon_region *r,
338 			struct damos *scheme);
339 	bool (*target_valid)(void *target);
340 	void (*cleanup)(struct damon_ctx *context);
341 };
342 
343 /**
344  * struct damon_callback - Monitoring events notification callbacks.
345  *
346  * @before_start:	Called before starting the monitoring.
347  * @after_wmarks_check:	Called after each schemes' watermarks check.
348  * @after_sampling:	Called after each sampling.
349  * @after_aggregation:	Called after each aggregation.
350  * @before_terminate:	Called before terminating the monitoring.
351  * @private:		User private data.
352  *
353  * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
354  * @before_terminate just before starting and finishing the monitoring,
355  * respectively.  Therefore, those are good places for installing and cleaning
356  * @private.
357  *
358  * The monitoring thread calls @after_wmarks_check after each DAMON-based
359  * operation schemes' watermarks check.  If users need to make changes to the
360  * attributes of the monitoring context while it's deactivated due to the
361  * watermarks, this is the good place to do.
362  *
363  * The monitoring thread calls @after_sampling and @after_aggregation for each
364  * of the sampling intervals and aggregation intervals, respectively.
365  * Therefore, users can safely access the monitoring results without additional
366  * protection.  For the reason, users are recommended to use these callback for
367  * the accesses to the results.
368  *
369  * If any callback returns non-zero, monitoring stops.
370  */
371 struct damon_callback {
372 	void *private;
373 
374 	int (*before_start)(struct damon_ctx *context);
375 	int (*after_wmarks_check)(struct damon_ctx *context);
376 	int (*after_sampling)(struct damon_ctx *context);
377 	int (*after_aggregation)(struct damon_ctx *context);
378 	void (*before_terminate)(struct damon_ctx *context);
379 };
380 
381 /**
382  * struct damon_ctx - Represents a context for each monitoring.  This is the
383  * main interface that allows users to set the attributes and get the results
384  * of the monitoring.
385  *
386  * @sample_interval:		The time between access samplings.
387  * @aggr_interval:		The time between monitor results aggregations.
388  * @ops_update_interval:	The time between monitoring operations updates.
389  *
390  * For each @sample_interval, DAMON checks whether each region is accessed or
391  * not.  It aggregates and keeps the access information (number of accesses to
392  * each region) for @aggr_interval time.  DAMON also checks whether the target
393  * memory regions need update (e.g., by ``mmap()`` calls from the application,
394  * in case of virtual memory monitoring) and applies the changes for each
395  * @ops_update_interval.  All time intervals are in micro-seconds.
396  * Please refer to &struct damon_operations and &struct damon_callback for more
397  * detail.
398  *
399  * @kdamond:		Kernel thread who does the monitoring.
400  * @kdamond_stop:	Notifies whether kdamond should stop.
401  * @kdamond_lock:	Mutex for the synchronizations with @kdamond.
402  *
403  * For each monitoring context, one kernel thread for the monitoring is
404  * created.  The pointer to the thread is stored in @kdamond.
405  *
406  * Once started, the monitoring thread runs until explicitly required to be
407  * terminated or every monitoring target is invalid.  The validity of the
408  * targets is checked via the &damon_operations.target_valid of @ops.  The
409  * termination can also be explicitly requested by writing non-zero to
410  * @kdamond_stop.  The thread sets @kdamond to NULL when it terminates.
411  * Therefore, users can know whether the monitoring is ongoing or terminated by
412  * reading @kdamond.  Reads and writes to @kdamond and @kdamond_stop from
413  * outside of the monitoring thread must be protected by @kdamond_lock.
414  *
415  * Note that the monitoring thread protects only @kdamond and @kdamond_stop via
416  * @kdamond_lock.  Accesses to other fields must be protected by themselves.
417  *
418  * @ops:	Set of monitoring operations for given use cases.
419  * @callback:	Set of callbacks for monitoring events notifications.
420  *
421  * @min_nr_regions:	The minimum number of adaptive monitoring regions.
422  * @max_nr_regions:	The maximum number of adaptive monitoring regions.
423  * @adaptive_targets:	Head of monitoring targets (&damon_target) list.
424  * @schemes:		Head of schemes (&damos) list.
425  */
426 struct damon_ctx {
427 	unsigned long sample_interval;
428 	unsigned long aggr_interval;
429 	unsigned long ops_update_interval;
430 
431 /* private: internal use only */
432 	struct timespec64 last_aggregation;
433 	struct timespec64 last_ops_update;
434 
435 /* public: */
436 	struct task_struct *kdamond;
437 	struct mutex kdamond_lock;
438 
439 	struct damon_operations ops;
440 	struct damon_callback callback;
441 
442 	unsigned long min_nr_regions;
443 	unsigned long max_nr_regions;
444 	struct list_head adaptive_targets;
445 	struct list_head schemes;
446 };
447 
damon_next_region(struct damon_region * r)448 static inline struct damon_region *damon_next_region(struct damon_region *r)
449 {
450 	return container_of(r->list.next, struct damon_region, list);
451 }
452 
damon_prev_region(struct damon_region * r)453 static inline struct damon_region *damon_prev_region(struct damon_region *r)
454 {
455 	return container_of(r->list.prev, struct damon_region, list);
456 }
457 
damon_last_region(struct damon_target * t)458 static inline struct damon_region *damon_last_region(struct damon_target *t)
459 {
460 	return list_last_entry(&t->regions_list, struct damon_region, list);
461 }
462 
463 #define damon_for_each_region(r, t) \
464 	list_for_each_entry(r, &t->regions_list, list)
465 
466 #define damon_for_each_region_safe(r, next, t) \
467 	list_for_each_entry_safe(r, next, &t->regions_list, list)
468 
469 #define damon_for_each_target(t, ctx) \
470 	list_for_each_entry(t, &(ctx)->adaptive_targets, list)
471 
472 #define damon_for_each_target_safe(t, next, ctx)	\
473 	list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
474 
475 #define damon_for_each_scheme(s, ctx) \
476 	list_for_each_entry(s, &(ctx)->schemes, list)
477 
478 #define damon_for_each_scheme_safe(s, next, ctx) \
479 	list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
480 
481 #ifdef CONFIG_DAMON
482 
483 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
484 
485 /*
486  * Add a region between two other regions
487  */
damon_insert_region(struct damon_region * r,struct damon_region * prev,struct damon_region * next,struct damon_target * t)488 static inline void damon_insert_region(struct damon_region *r,
489 		struct damon_region *prev, struct damon_region *next,
490 		struct damon_target *t)
491 {
492 	__list_add(&r->list, &prev->list, &next->list);
493 	t->nr_regions++;
494 }
495 
496 void damon_add_region(struct damon_region *r, struct damon_target *t);
497 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
498 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
499 		unsigned int nr_ranges);
500 
501 struct damos *damon_new_scheme(
502 		unsigned long min_sz_region, unsigned long max_sz_region,
503 		unsigned int min_nr_accesses, unsigned int max_nr_accesses,
504 		unsigned int min_age_region, unsigned int max_age_region,
505 		enum damos_action action, struct damos_quota *quota,
506 		struct damos_watermarks *wmarks);
507 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
508 void damon_destroy_scheme(struct damos *s);
509 
510 struct damon_target *damon_new_target(void);
511 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
512 bool damon_targets_empty(struct damon_ctx *ctx);
513 void damon_free_target(struct damon_target *t);
514 void damon_destroy_target(struct damon_target *t);
515 unsigned int damon_nr_regions(struct damon_target *t);
516 
517 struct damon_ctx *damon_new_ctx(void);
518 void damon_destroy_ctx(struct damon_ctx *ctx);
519 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
520 		unsigned long aggr_int, unsigned long ops_upd_int,
521 		unsigned long min_nr_reg, unsigned long max_nr_reg);
522 int damon_set_schemes(struct damon_ctx *ctx,
523 			struct damos **schemes, ssize_t nr_schemes);
524 int damon_nr_running_ctxs(void);
525 bool damon_is_registered_ops(enum damon_ops_id id);
526 int damon_register_ops(struct damon_operations *ops);
527 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id);
528 
529 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive);
530 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
531 
532 #endif	/* CONFIG_DAMON */
533 
534 #endif	/* _DAMON_H */
535