xref: /openbmc/linux/kernel/sched/pelt.h (revision aa0dc6a7)
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 static inline u32 get_pelt_divider(struct sched_avg *avg)
41 {
42 	return LOAD_AVG_MAX - 1024 + avg->period_contrib;
43 }
44 
45 static inline void cfs_se_util_change(struct sched_avg *avg)
46 {
47 	unsigned int enqueued;
48 
49 	if (!sched_feat(UTIL_EST))
50 		return;
51 
52 	/* Avoid store if the flag has been already reset */
53 	enqueued = avg->util_est.enqueued;
54 	if (!(enqueued & UTIL_AVG_UNCHANGED))
55 		return;
56 
57 	/* Reset flag to report util_avg has been updated */
58 	enqueued &= ~UTIL_AVG_UNCHANGED;
59 	WRITE_ONCE(avg->util_est.enqueued, enqueued);
60 }
61 
62 /*
63  * The clock_pelt scales the time to reflect the effective amount of
64  * computation done during the running delta time but then sync back to
65  * clock_task when rq is idle.
66  *
67  *
68  * absolute time   | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
69  * @ max capacity  ------******---------------******---------------
70  * @ half capacity ------************---------************---------
71  * clock pelt      | 1| 2|    3|    4| 7| 8| 9|   10|   11|14|15|16
72  *
73  */
74 static inline void update_rq_clock_pelt(struct rq *rq, s64 delta)
75 {
76 	if (unlikely(is_idle_task(rq->curr))) {
77 		/* The rq is idle, we can sync to clock_task */
78 		rq->clock_pelt  = rq_clock_task(rq);
79 		return;
80 	}
81 
82 	/*
83 	 * When a rq runs at a lower compute capacity, it will need
84 	 * more time to do the same amount of work than at max
85 	 * capacity. In order to be invariant, we scale the delta to
86 	 * reflect how much work has been really done.
87 	 * Running longer results in stealing idle time that will
88 	 * disturb the load signal compared to max capacity. This
89 	 * stolen idle time will be automatically reflected when the
90 	 * rq will be idle and the clock will be synced with
91 	 * rq_clock_task.
92 	 */
93 
94 	/*
95 	 * Scale the elapsed time to reflect the real amount of
96 	 * computation
97 	 */
98 	delta = cap_scale(delta, arch_scale_cpu_capacity(cpu_of(rq)));
99 	delta = cap_scale(delta, arch_scale_freq_capacity(cpu_of(rq)));
100 
101 	rq->clock_pelt += delta;
102 }
103 
104 /*
105  * When rq becomes idle, we have to check if it has lost idle time
106  * because it was fully busy. A rq is fully used when the /Sum util_sum
107  * is greater or equal to:
108  * (LOAD_AVG_MAX - 1024 + rq->cfs.avg.period_contrib) << SCHED_CAPACITY_SHIFT;
109  * For optimization and computing rounding purpose, we don't take into account
110  * the position in the current window (period_contrib) and we use the higher
111  * bound of util_sum to decide.
112  */
113 static inline void update_idle_rq_clock_pelt(struct rq *rq)
114 {
115 	u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT) - LOAD_AVG_MAX;
116 	u32 util_sum = rq->cfs.avg.util_sum;
117 	util_sum += rq->avg_rt.util_sum;
118 	util_sum += rq->avg_dl.util_sum;
119 
120 	/*
121 	 * Reflecting stolen time makes sense only if the idle
122 	 * phase would be present at max capacity. As soon as the
123 	 * utilization of a rq has reached the maximum value, it is
124 	 * considered as an always running rq without idle time to
125 	 * steal. This potential idle time is considered as lost in
126 	 * this case. We keep track of this lost idle time compare to
127 	 * rq's clock_task.
128 	 */
129 	if (util_sum >= divider)
130 		rq->lost_idle_time += rq_clock_task(rq) - rq->clock_pelt;
131 }
132 
133 static inline u64 rq_clock_pelt(struct rq *rq)
134 {
135 	lockdep_assert_rq_held(rq);
136 	assert_clock_updated(rq);
137 
138 	return rq->clock_pelt - rq->lost_idle_time;
139 }
140 
141 #ifdef CONFIG_CFS_BANDWIDTH
142 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
143 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
144 {
145 	if (unlikely(cfs_rq->throttle_count))
146 		return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
147 
148 	return rq_clock_pelt(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
149 }
150 #else
151 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
152 {
153 	return rq_clock_pelt(rq_of(cfs_rq));
154 }
155 #endif
156 
157 #else
158 
159 static inline int
160 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
161 {
162 	return 0;
163 }
164 
165 static inline int
166 update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
167 {
168 	return 0;
169 }
170 
171 static inline int
172 update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
173 {
174 	return 0;
175 }
176 
177 static inline int
178 update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
179 {
180 	return 0;
181 }
182 
183 static inline u64 thermal_load_avg(struct rq *rq)
184 {
185 	return 0;
186 }
187 
188 static inline int
189 update_irq_load_avg(struct rq *rq, u64 running)
190 {
191 	return 0;
192 }
193 
194 static inline u64 rq_clock_pelt(struct rq *rq)
195 {
196 	return rq_clock_task(rq);
197 }
198 
199 static inline void
200 update_rq_clock_pelt(struct rq *rq, s64 delta) { }
201 
202 static inline void
203 update_idle_rq_clock_pelt(struct rq *rq) { }
204 
205 #endif
206 
207 
208