xref: /openbmc/linux/kernel/sched/pelt.h (revision 81de3bf3)
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_HAVE_SCHED_AVG_IRQ
11 int update_irq_load_avg(struct rq *rq, u64 running);
12 #else
13 static inline int
14 update_irq_load_avg(struct rq *rq, u64 running)
15 {
16 	return 0;
17 }
18 #endif
19 
20 /*
21  * When a task is dequeued, its estimated utilization should not be update if
22  * its util_avg has not been updated at least once.
23  * This flag is used to synchronize util_avg updates with util_est updates.
24  * We map this information into the LSB bit of the utilization saved at
25  * dequeue time (i.e. util_est.dequeued).
26  */
27 #define UTIL_AVG_UNCHANGED 0x1
28 
29 static inline void cfs_se_util_change(struct sched_avg *avg)
30 {
31 	unsigned int enqueued;
32 
33 	if (!sched_feat(UTIL_EST))
34 		return;
35 
36 	/* Avoid store if the flag has been already set */
37 	enqueued = avg->util_est.enqueued;
38 	if (!(enqueued & UTIL_AVG_UNCHANGED))
39 		return;
40 
41 	/* Reset flag to report util_avg has been updated */
42 	enqueued &= ~UTIL_AVG_UNCHANGED;
43 	WRITE_ONCE(avg->util_est.enqueued, enqueued);
44 }
45 
46 /*
47  * The clock_pelt scales the time to reflect the effective amount of
48  * computation done during the running delta time but then sync back to
49  * clock_task when rq is idle.
50  *
51  *
52  * absolute time   | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
53  * @ max capacity  ------******---------------******---------------
54  * @ half capacity ------************---------************---------
55  * clock pelt      | 1| 2|    3|    4| 7| 8| 9|   10|   11|14|15|16
56  *
57  */
58 static inline void update_rq_clock_pelt(struct rq *rq, s64 delta)
59 {
60 	if (unlikely(is_idle_task(rq->curr))) {
61 		/* The rq is idle, we can sync to clock_task */
62 		rq->clock_pelt  = rq_clock_task(rq);
63 		return;
64 	}
65 
66 	/*
67 	 * When a rq runs at a lower compute capacity, it will need
68 	 * more time to do the same amount of work than at max
69 	 * capacity. In order to be invariant, we scale the delta to
70 	 * reflect how much work has been really done.
71 	 * Running longer results in stealing idle time that will
72 	 * disturb the load signal compared to max capacity. This
73 	 * stolen idle time will be automatically reflected when the
74 	 * rq will be idle and the clock will be synced with
75 	 * rq_clock_task.
76 	 */
77 
78 	/*
79 	 * Scale the elapsed time to reflect the real amount of
80 	 * computation
81 	 */
82 	delta = cap_scale(delta, arch_scale_cpu_capacity(cpu_of(rq)));
83 	delta = cap_scale(delta, arch_scale_freq_capacity(cpu_of(rq)));
84 
85 	rq->clock_pelt += delta;
86 }
87 
88 /*
89  * When rq becomes idle, we have to check if it has lost idle time
90  * because it was fully busy. A rq is fully used when the /Sum util_sum
91  * is greater or equal to:
92  * (LOAD_AVG_MAX - 1024 + rq->cfs.avg.period_contrib) << SCHED_CAPACITY_SHIFT;
93  * For optimization and computing rounding purpose, we don't take into account
94  * the position in the current window (period_contrib) and we use the higher
95  * bound of util_sum to decide.
96  */
97 static inline void update_idle_rq_clock_pelt(struct rq *rq)
98 {
99 	u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT) - LOAD_AVG_MAX;
100 	u32 util_sum = rq->cfs.avg.util_sum;
101 	util_sum += rq->avg_rt.util_sum;
102 	util_sum += rq->avg_dl.util_sum;
103 
104 	/*
105 	 * Reflecting stolen time makes sense only if the idle
106 	 * phase would be present at max capacity. As soon as the
107 	 * utilization of a rq has reached the maximum value, it is
108 	 * considered as an always runnig rq without idle time to
109 	 * steal. This potential idle time is considered as lost in
110 	 * this case. We keep track of this lost idle time compare to
111 	 * rq's clock_task.
112 	 */
113 	if (util_sum >= divider)
114 		rq->lost_idle_time += rq_clock_task(rq) - rq->clock_pelt;
115 }
116 
117 static inline u64 rq_clock_pelt(struct rq *rq)
118 {
119 	lockdep_assert_held(&rq->lock);
120 	assert_clock_updated(rq);
121 
122 	return rq->clock_pelt - rq->lost_idle_time;
123 }
124 
125 #ifdef CONFIG_CFS_BANDWIDTH
126 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
127 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
128 {
129 	if (unlikely(cfs_rq->throttle_count))
130 		return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
131 
132 	return rq_clock_pelt(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
133 }
134 #else
135 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
136 {
137 	return rq_clock_pelt(rq_of(cfs_rq));
138 }
139 #endif
140 
141 #else
142 
143 static inline int
144 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
145 {
146 	return 0;
147 }
148 
149 static inline int
150 update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
151 {
152 	return 0;
153 }
154 
155 static inline int
156 update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
157 {
158 	return 0;
159 }
160 
161 static inline int
162 update_irq_load_avg(struct rq *rq, u64 running)
163 {
164 	return 0;
165 }
166 
167 static inline u64 rq_clock_pelt(struct rq *rq)
168 {
169 	return rq_clock_task(rq);
170 }
171 
172 static inline void
173 update_rq_clock_pelt(struct rq *rq, s64 delta) { }
174 
175 static inline void
176 update_idle_rq_clock_pelt(struct rq *rq) { }
177 
178 #endif
179 
180 
181