xref: /xnu-11215.61.5/osfmk/kern/sched_clutch.h (revision 4f1223e81cd707a65cc109d0b8ad6653699da3c4)
1 /*
2  * Copyright (c) 2018 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 
29 #ifndef _KERN_SCHED_CLUTCH_H_
30 #define _KERN_SCHED_CLUTCH_H_
31 
32 #include <kern/sched.h>
33 #include <kern/priority_queue.h>
34 #include <kern/bits.h>
35 
36 #if !SCHED_TEST_HARNESS
37 
38 #include <machine/atomic.h>
39 #include <kern/thread_group.h>
40 
41 #endif /* !SCHED_TEST_HARNESS */
42 
43 #if CONFIG_SCHED_CLUTCH
44 
45 /*
46  * For the current implementation, bound threads are not managed
47  * in the clutch hierarchy. This helper macro is used to indicate
48  * if the thread should be in the hierarchy.
49  */
50 #define SCHED_CLUTCH_THREAD_ELIGIBLE(thread)    ((thread->bound_processor) == PROCESSOR_NULL)
51 
52 #if CONFIG_SCHED_EDGE
53 #define SCHED_CLUTCH_THREAD_CLUSTER_BOUND(thread)       (thread->th_bound_cluster_id != THREAD_BOUND_CLUSTER_NONE)
54 #define SCHED_CLUTCH_THREAD_CLUSTER_BOUND_SOFT(thread)  ((thread->sched_flags & TH_SFLAG_BOUND_SOFT) != 0)
55 
56 #else /* CONFIG_SCHED_EDGE */
57 #define SCHED_CLUTCH_THREAD_CLUSTER_BOUND(thread)       (0)
58 #define SCHED_CLUTCH_THREAD_CLUSTER_BOUND_SOFT(thread)  (0)
59 #endif /* CONFIG_SCHED_EDGE */
60 
61 /*
62  * Clutch Bucket Runqueue Structure.
63  */
64 struct sched_clutch_bucket_runq {
65 	int                     scbrq_highq;
66 	int                     scbrq_count;
67 	bitmap_t                scbrq_bitmap[BITMAP_LEN(NRQS_MAX)];
68 	circle_queue_head_t     scbrq_queues[NRQS_MAX];
69 };
70 typedef struct sched_clutch_bucket_runq *sched_clutch_bucket_runq_t;
71 
72 /*
73  *
74  * Clutch hierarchy locking protocol
75  *
76  * The scheduler clutch hierarchy is protected by a combination of
77  * atomics and pset lock.
78  * - All fields protected by the pset lock are annotated with (P)
79  * - All fields updated using atomics are annotated with (A)
80  * - All fields that are unprotected and are not updated after
81  *   initialization are annotated with (I)
82  */
83 
84 /*
85  * struct sched_clutch_root_bucket
86  *
87  * A clutch_root_bucket represents all threads across all thread groups
88  * that are in the same scheduler bucket (FG/IN/...). The clutch_root_bucket
89  * is selected for execution by the root level bucket selection algorithm
90  * which bases the decision on the clutch_root_bucket's deadline (EDF). The
91  * deadline for a root bucket is calculated based on its runnable timestamp
92  * and the worst-case-execution-latency values specied in sched_clutch_root_bucket_wcel[]
93  */
94 struct sched_clutch_root_bucket {
95 	/* (I) sched bucket represented by this root bucket */
96 	uint8_t                         scrb_bucket;
97 	/* (I) Indicates the root bucket represents cluster bound threads */
98 	bool                            scrb_bound;
99 	/* (P) Indicates if the root bucket is in starvation avoidance mode */
100 	bool                            scrb_starvation_avoidance;
101 
102 	union {
103 		/* (P) priority queue for all unbound clutch buckets in this sched bucket */
104 		struct sched_clutch_bucket_runq scrb_clutch_buckets;
105 		/* (P) Runqueue for all bound threads part of this root bucket */
106 		struct run_queue                scrb_bound_thread_runq;
107 	};
108 	/* (P) priority queue entry to use for enqueueing root bucket into root prioq */
109 	struct priority_queue_entry_deadline scrb_pqlink;
110 	/* (P) warped deadline for root bucket */
111 	uint64_t                        scrb_warped_deadline;
112 	/* (P) warp remaining for root bucket */
113 	uint64_t                        scrb_warp_remaining;
114 	/* (P) timestamp for the start of the starvation avoidance window */
115 	uint64_t                        scrb_starvation_ts;
116 };
117 typedef struct sched_clutch_root_bucket *sched_clutch_root_bucket_t;
118 
119 /*
120  * struct sched_clutch_root
121  *
122  * A clutch_root represents the root of the hierarchy. It maintains a
123  * priority queue of all runnable root buckets. The clutch_root also
124  * maintains the information about the last clutch_root_bucket scheduled
125  * in order to implement bucket level quantum. The bucket level quantums
126  * allow low priority buckets to get a "fair" chance of using the CPU even
127  * if they contain a bunch of short executing threads. The bucket quantums
128  * are configured using sched_clutch_root_bucket_quantum[]
129  */
130 struct sched_clutch_root {
131 	/* (P) root level priority; represents the highest runnable thread in the hierarchy */
132 	int16_t                         scr_priority;
133 	/* (P) total number of runnable threads in the hierarchy */
134 	uint16_t                        scr_thr_count;
135 	/* (P) root level urgency; represents the urgency of the whole hierarchy for pre-emption purposes */
136 	int16_t                         scr_urgency;
137 #if CONFIG_SCHED_EDGE
138 	/* (P) runnable shared resource load enqueued in this cluster/root hierarchy */
139 	uint16_t                        scr_shared_rsrc_load_runnable[CLUSTER_SHARED_RSRC_TYPE_COUNT];
140 #endif /* CONFIG_SCHED_EDGE */
141 
142 	uint32_t                        scr_cluster_id;
143 	/* (I) processor set this hierarchy belongs to */
144 	processor_set_t                 scr_pset;
145 	/*
146 	 * (P) list of all runnable clutch buckets across the system;
147 	 * allows easy iteration in the sched tick based timesharing code
148 	 */
149 	queue_head_t                    scr_clutch_buckets;
150 
151 	/*
152 	 * (P) priority queue of all runnable foreign buckets in this hierarchy;
153 	 * used for tracking thread groups which need to be migrated when
154 	 * psets are available or rebalancing threads on CPU idle.
155 	 */
156 	struct priority_queue_sched_max scr_foreign_buckets;
157 
158 	/* Root level bucket management */
159 
160 	/* (P) bitmap of all runnable unbounded root buckets */
161 	bitmap_t                        scr_unbound_runnable_bitmap[BITMAP_LEN(TH_BUCKET_SCHED_MAX)];
162 	/* (P) bitmap of all runnable unbounded root buckets which have warps remaining */
163 	bitmap_t                        scr_unbound_warp_available[BITMAP_LEN(TH_BUCKET_SCHED_MAX)];
164 	/* (P) bitmap of all runnable bounded root buckets */
165 	bitmap_t                        scr_bound_runnable_bitmap[BITMAP_LEN(TH_BUCKET_SCHED_MAX)];
166 	/* (P) bitmap of all runnable bounded root buckets which have warps remaining */
167 	bitmap_t                        scr_bound_warp_available[BITMAP_LEN(TH_BUCKET_SCHED_MAX)];
168 
169 	/* (P) priority queue of all runnable unbounded root buckets in deadline order */
170 	struct priority_queue_deadline_min scr_unbound_root_buckets;
171 	/* (P) priority queue of all bounded root buckets in deadline order */
172 	struct priority_queue_deadline_min scr_bound_root_buckets;
173 
174 	/* (P) cumulative run counts at each bucket for load average calculation */
175 	uint16_t _Atomic                scr_cumulative_run_count[TH_BUCKET_SCHED_MAX];
176 
177 	/* (P) storage for all unbound clutch_root_buckets */
178 	struct sched_clutch_root_bucket scr_unbound_buckets[TH_BUCKET_SCHED_MAX];
179 	/* (P) storage for all bound clutch_root_buckets */
180 	struct sched_clutch_root_bucket scr_bound_buckets[TH_BUCKET_SCHED_MAX];
181 };
182 typedef struct sched_clutch_root *sched_clutch_root_t;
183 
184 /* forward declaration for sched_clutch */
185 struct sched_clutch;
186 
187 /*
188  * sched_clutch_bucket_cpu_data_t
189  *
190  * Used for maintaining clutch bucket used and blocked time. The
191  * values are used for calculating the interactivity score for the
192  * clutch bucket.
193  */
194 #define CLUTCH_CPU_DATA_MAX             (UINT64_MAX)
195 typedef uint64_t                        clutch_cpu_data_t;
196 typedef unsigned __int128               clutch_cpu_data_wide_t;
197 
198 typedef union sched_clutch_bucket_cpu_data {
199 	struct {
200 		/* Clutch bucket CPU used across all threads */
201 		clutch_cpu_data_t       scbcd_cpu_used;
202 		/* Clutch bucket voluntary blocked time */
203 		clutch_cpu_data_t       scbcd_cpu_blocked;
204 	} cpu_data;
205 	clutch_cpu_data_wide_t          scbcd_cpu_data_packed;
206 } sched_clutch_bucket_cpu_data_t;
207 
208 /*
209  * struct sched_clutch_bucket
210  *
211  * A sched_clutch_bucket represents the set of threads for a thread
212  * group at a particular scheduling bucket in a specific cluster.
213  * It maintains information about the CPU usage & blocking behavior
214  * of all threads part of the clutch_bucket. It inherits the timeshare
215  * values from the clutch_bucket_group for decay and timesharing among
216  * threads in the clutch.
217  *
218  * Since the clutch bucket is a per thread group per-QoS entity it is
219  * important to keep its size small and the structure well aligned.
220  */
221 struct sched_clutch_bucket {
222 #if CONFIG_SCHED_EDGE
223 	/* (P) flag to indicate if the bucket is a foreign bucket */
224 	bool                            scb_foreign;
225 #endif /* CONFIG_SCHED_EDGE */
226 	/* (I) bucket for the clutch_bucket */
227 	uint8_t                         scb_bucket;
228 	/* (P) priority of the clutch bucket */
229 	uint8_t                         scb_priority;
230 	/* (P) number of threads in this clutch_bucket; should match runq.count */
231 	uint16_t                        scb_thr_count;
232 
233 	/* Pointer to the clutch bucket group this clutch bucket belongs to */
234 	struct sched_clutch_bucket_group *scb_group;
235 	/* (A) pointer to the root of the hierarchy this bucket is in */
236 	struct sched_clutch_root        *scb_root;
237 	/* (P) priority queue of threads based on their promoted/base priority */
238 	struct priority_queue_sched_max scb_clutchpri_prioq;
239 	/* (P) runq of threads in clutch_bucket */
240 	struct priority_queue_sched_stable_max scb_thread_runq;
241 
242 	/* (P) linkage for all clutch_buckets in a root bucket; used for tick operations */
243 	queue_chain_t                   scb_listlink;
244 	/* (P) linkage for clutch_bucket in root_bucket runqueue */
245 	queue_chain_t                   scb_runqlink;
246 	/* (P) queue of threads for timesharing purposes */
247 	queue_head_t                    scb_thread_timeshare_queue;
248 #if CONFIG_SCHED_EDGE
249 	/* (P) linkage for all "foreign" clutch buckets in the root clutch */
250 	struct priority_queue_entry_sched     scb_foreignlink;
251 #endif /* CONFIG_SCHED_EDGE */
252 };
253 typedef struct sched_clutch_bucket *sched_clutch_bucket_t;
254 
255 /*
256  * sched_clutch_counter_time_t
257  *
258  * Holds thread counts and a timestamp (typically for a clutch bucket group).
259  * Used to allow atomic updates to these fields.
260  */
261 typedef union sched_clutch_counter_time {
262 	struct {
263 		uint64_t                scct_count;
264 		uint64_t                scct_timestamp;
265 	};
266 	unsigned __int128               scct_packed;
267 } __attribute__((aligned(16))) sched_clutch_counter_time_t;
268 
269 /*
270  * struct sched_clutch_bucket_group
271  *
272  * It represents all the threads for a thread group at a particular
273  * QoS/Scheduling bucket. This structure also maintains the timesharing
274  * properties that are used for decay calculation for all threads in the
275  * thread group at the specific scheduling bucket.
276  */
277 struct sched_clutch_bucket_group {
278 	/* (I) bucket for the clutch_bucket_group */
279 	uint8_t                         scbg_bucket;
280 	/* (A) sched tick when the clutch bucket group load/shifts were updated */
281 	uint32_t _Atomic                scbg_timeshare_tick;
282 	/* (A) priority shifts for threads in the clutch_bucket_group */
283 	uint32_t _Atomic                scbg_pri_shift;
284 	/* (A) preferred cluster ID for clutch bucket */
285 	uint32_t _Atomic                scbg_preferred_cluster;
286 	/* (A) cluster ID for AMP rebalancing */
287 	uint32_t                        scbg_amp_rebalance_last_chosen;
288 	/* (I) clutch to which this clutch bucket_group belongs */
289 	struct sched_clutch             *scbg_clutch;
290 	/* (A) holds blocked timestamp and runnable/running count */
291 	sched_clutch_counter_time_t     scbg_blocked_data;
292 	/* (P/A depending on scheduler) holds pending timestamp and thread count */
293 	sched_clutch_counter_time_t     scbg_pending_data;
294 	/* (P/A depending on scheduler) holds interactivity timestamp and score */
295 	sched_clutch_counter_time_t     scbg_interactivity_data;
296 	/* (A) CPU usage information for the clutch bucket group */
297 	sched_clutch_bucket_cpu_data_t  scbg_cpu_data;
298 	/* Storage for all clutch buckets for a thread group at scbg_bucket */
299 	struct sched_clutch_bucket      *scbg_clutch_buckets;
300 };
301 typedef struct sched_clutch_bucket_group *sched_clutch_bucket_group_t;
302 
303 
304 /*
305  * struct sched_clutch
306  *
307  * A sched_clutch is a 1:1 mapping to a thread group. It maintains the
308  * storage for all clutch buckets for this thread group and some properties
309  * of the thread group (such as flags etc.)
310  */
311 struct sched_clutch {
312 	/*
313 	 * (A) number of runnable threads in sched_clutch; needs to be atomic
314 	 * to support cross cluster sched_clutch migrations.
315 	 */
316 	uint16_t _Atomic                sc_thr_count;
317 	/*
318 	 * Grouping specific parameters. Currently the implementation only
319 	 * supports thread_group based grouping.
320 	 */
321 	union {
322 		/* (I) Pointer to thread group */
323 		struct thread_group     *sc_tg;
324 	};
325 	/* (I) storage for all clutch_buckets for this clutch */
326 	struct sched_clutch_bucket_group sc_clutch_groups[TH_BUCKET_SCHED_MAX];
327 };
328 typedef struct sched_clutch *sched_clutch_t;
329 
330 
331 /* Clutch lifecycle management */
332 void sched_clutch_init_with_thread_group(sched_clutch_t, struct thread_group *);
333 void sched_clutch_destroy(sched_clutch_t);
334 
335 /* Clutch thread membership management */
336 void sched_clutch_thread_clutch_update(thread_t, sched_clutch_t, sched_clutch_t);
337 uint32_t sched_edge_thread_preferred_cluster(thread_t);
338 
339 /* Clutch timesharing stats management */
340 uint32_t sched_clutch_thread_run_bucket_incr(thread_t, sched_bucket_t);
341 uint32_t sched_clutch_thread_run_bucket_decr(thread_t, sched_bucket_t);
342 void sched_clutch_cpu_usage_update(thread_t, uint64_t);
343 uint32_t sched_clutch_thread_pri_shift(thread_t, sched_bucket_t);
344 
345 /* Clutch properties accessors */
346 uint32_t sched_clutch_root_count(sched_clutch_root_t);
347 
348 /* Grouping specific external routines */
349 extern sched_clutch_t sched_clutch_for_thread(thread_t);
350 extern sched_clutch_t sched_clutch_for_thread_group(struct thread_group *);
351 
352 #if CONFIG_SCHED_EDGE
353 
354 /*
355  * Getter and Setter for Edge configuration. Used by CLPC to affect thread migration behavior.
356  */
357 void sched_edge_matrix_get(sched_clutch_edge *edge_matrix, bool *edge_request_bitmap, uint64_t flags, uint64_t matrix_order);
358 void sched_edge_matrix_set(sched_clutch_edge *edge_matrix, bool *edge_changes_bitmap, uint64_t flags, uint64_t matrix_order);
359 void sched_edge_tg_preferred_cluster_change(struct thread_group *tg, uint32_t *tg_bucket_preferred_cluster, sched_perfcontrol_preferred_cluster_options_t options);
360 
361 uint16_t sched_edge_cluster_cumulative_count(sched_clutch_root_t root_clutch, sched_bucket_t bucket);
362 uint16_t sched_edge_shared_rsrc_runnable_load(sched_clutch_root_t root_clutch, cluster_shared_rsrc_type_t load_type);
363 
364 #endif /* CONFIG_SCHED_EDGE */
365 
366 #endif /* CONFIG_SCHED_CLUTCH */
367 
368 #endif /* _KERN_SCHED_CLUTCH_H_ */
369