xref: /linux/kernel/sched/pelt.h (revision a4eb44a6435d6d8f9e642407a4a06f65eb90ca04)
1 #ifdef CONFIG_SMP
2 #include "sched-pelt.h"
3 
4 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se);
5 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se);
6 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq);
7 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running);
8 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running);
9 
10 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
11 int update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity);
12 
13 static inline u64 thermal_load_avg(struct rq *rq)
14 {
15 	return READ_ONCE(rq->avg_thermal.load_avg);
16 }
17 #else
18 static inline int
19 update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
20 {
21 	return 0;
22 }
23 
24 static inline u64 thermal_load_avg(struct rq *rq)
25 {
26 	return 0;
27 }
28 #endif
29 
30 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
31 int update_irq_load_avg(struct rq *rq, u64 running);
32 #else
33 static inline int
34 update_irq_load_avg(struct rq *rq, u64 running)
35 {
36 	return 0;
37 }
38 #endif
39 
40 #define PELT_MIN_DIVIDER	(LOAD_AVG_MAX - 1024)
41 
42 static inline u32 get_pelt_divider(struct sched_avg *avg)
43 {
44 	return PELT_MIN_DIVIDER + avg->period_contrib;
45 }
46 
47 static inline void cfs_se_util_change(struct sched_avg *avg)
48 {
49 	unsigned int enqueued;
50 
51 	if (!sched_feat(UTIL_EST))
52 		return;
53 
54 	/* Avoid store if the flag has been already reset */
55 	enqueued = avg->util_est.enqueued;
56 	if (!(enqueued & UTIL_AVG_UNCHANGED))
57 		return;
58 
59 	/* Reset flag to report util_avg has been updated */
60 	enqueued &= ~UTIL_AVG_UNCHANGED;
61 	WRITE_ONCE(avg->util_est.enqueued, enqueued);
62 }
63 
64 /*
65  * The clock_pelt scales the time to reflect the effective amount of
66  * computation done during the running delta time but then sync back to
67  * clock_task when rq is idle.
68  *
69  *
70  * absolute time   | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
71  * @ max capacity  ------******---------------******---------------
72  * @ half capacity ------************---------************---------
73  * clock pelt      | 1| 2|    3|    4| 7| 8| 9|   10|   11|14|15|16
74  *
75  */
76 static inline void update_rq_clock_pelt(struct rq *rq, s64 delta)
77 {
78 	if (unlikely(is_idle_task(rq->curr))) {
79 		/* The rq is idle, we can sync to clock_task */
80 		rq->clock_pelt  = rq_clock_task(rq);
81 		return;
82 	}
83 
84 	/*
85 	 * When a rq runs at a lower compute capacity, it will need
86 	 * more time to do the same amount of work than at max
87 	 * capacity. In order to be invariant, we scale the delta to
88 	 * reflect how much work has been really done.
89 	 * Running longer results in stealing idle time that will
90 	 * disturb the load signal compared to max capacity. This
91 	 * stolen idle time will be automatically reflected when the
92 	 * rq will be idle and the clock will be synced with
93 	 * rq_clock_task.
94 	 */
95 
96 	/*
97 	 * Scale the elapsed time to reflect the real amount of
98 	 * computation
99 	 */
100 	delta = cap_scale(delta, arch_scale_cpu_capacity(cpu_of(rq)));
101 	delta = cap_scale(delta, arch_scale_freq_capacity(cpu_of(rq)));
102 
103 	rq->clock_pelt += delta;
104 }
105 
106 /*
107  * When rq becomes idle, we have to check if it has lost idle time
108  * because it was fully busy. A rq is fully used when the /Sum util_sum
109  * is greater or equal to:
110  * (LOAD_AVG_MAX - 1024 + rq->cfs.avg.period_contrib) << SCHED_CAPACITY_SHIFT;
111  * For optimization and computing rounding purpose, we don't take into account
112  * the position in the current window (period_contrib) and we use the higher
113  * bound of util_sum to decide.
114  */
115 static inline void update_idle_rq_clock_pelt(struct rq *rq)
116 {
117 	u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT) - LOAD_AVG_MAX;
118 	u32 util_sum = rq->cfs.avg.util_sum;
119 	util_sum += rq->avg_rt.util_sum;
120 	util_sum += rq->avg_dl.util_sum;
121 
122 	/*
123 	 * Reflecting stolen time makes sense only if the idle
124 	 * phase would be present at max capacity. As soon as the
125 	 * utilization of a rq has reached the maximum value, it is
126 	 * considered as an always running rq without idle time to
127 	 * steal. This potential idle time is considered as lost in
128 	 * this case. We keep track of this lost idle time compare to
129 	 * rq's clock_task.
130 	 */
131 	if (util_sum >= divider)
132 		rq->lost_idle_time += rq_clock_task(rq) - rq->clock_pelt;
133 }
134 
135 static inline u64 rq_clock_pelt(struct rq *rq)
136 {
137 	lockdep_assert_rq_held(rq);
138 	assert_clock_updated(rq);
139 
140 	return rq->clock_pelt - rq->lost_idle_time;
141 }
142 
143 #ifdef CONFIG_CFS_BANDWIDTH
144 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
145 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
146 {
147 	if (unlikely(cfs_rq->throttle_count))
148 		return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
149 
150 	return rq_clock_pelt(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
151 }
152 #else
153 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
154 {
155 	return rq_clock_pelt(rq_of(cfs_rq));
156 }
157 #endif
158 
159 #else
160 
161 static inline int
162 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
163 {
164 	return 0;
165 }
166 
167 static inline int
168 update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
169 {
170 	return 0;
171 }
172 
173 static inline int
174 update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
175 {
176 	return 0;
177 }
178 
179 static inline int
180 update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
181 {
182 	return 0;
183 }
184 
185 static inline u64 thermal_load_avg(struct rq *rq)
186 {
187 	return 0;
188 }
189 
190 static inline int
191 update_irq_load_avg(struct rq *rq, u64 running)
192 {
193 	return 0;
194 }
195 
196 static inline u64 rq_clock_pelt(struct rq *rq)
197 {
198 	return rq_clock_task(rq);
199 }
200 
201 static inline void
202 update_rq_clock_pelt(struct rq *rq, s64 delta) { }
203 
204 static inline void
205 update_idle_rq_clock_pelt(struct rq *rq) { }
206 
207 #endif
208 
209 
210