• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 #ifdef CONFIG_SMP
2 #include <trace/hooks/sched.h>
3 #include "sched-pelt.h"
4 
5 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se);
6 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se);
7 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq);
8 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running);
9 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running);
10 bool update_other_load_avgs(struct rq *rq);
11 
12 #ifdef CONFIG_SCHED_HW_PRESSURE
13 int update_hw_load_avg(u64 now, struct rq *rq, u64 capacity);
14 
hw_load_avg(struct rq * rq)15 static inline u64 hw_load_avg(struct rq *rq)
16 {
17 	return READ_ONCE(rq->avg_hw.load_avg);
18 }
19 #else
20 static inline int
update_hw_load_avg(u64 now,struct rq * rq,u64 capacity)21 update_hw_load_avg(u64 now, struct rq *rq, u64 capacity)
22 {
23 	return 0;
24 }
25 
hw_load_avg(struct rq * rq)26 static inline u64 hw_load_avg(struct rq *rq)
27 {
28 	return 0;
29 }
30 #endif
31 
32 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
33 int update_irq_load_avg(struct rq *rq, u64 running);
34 #else
35 static inline int
update_irq_load_avg(struct rq * rq,u64 running)36 update_irq_load_avg(struct rq *rq, u64 running)
37 {
38 	return 0;
39 }
40 #endif
41 
42 #define PELT_MIN_DIVIDER	(LOAD_AVG_MAX - 1024)
43 
get_pelt_divider(struct sched_avg * avg)44 static inline u32 get_pelt_divider(struct sched_avg *avg)
45 {
46 	return PELT_MIN_DIVIDER + avg->period_contrib;
47 }
48 
cfs_se_util_change(struct sched_avg * avg)49 static inline void cfs_se_util_change(struct sched_avg *avg)
50 {
51 	unsigned int enqueued;
52 
53 	if (!sched_feat(UTIL_EST))
54 		return;
55 
56 	/* Avoid store if the flag has been already reset */
57 	enqueued = avg->util_est;
58 	if (!(enqueued & UTIL_AVG_UNCHANGED))
59 		return;
60 
61 	/* Reset flag to report util_avg has been updated */
62 	enqueued &= ~UTIL_AVG_UNCHANGED;
63 	WRITE_ONCE(avg->util_est, enqueued);
64 }
65 
rq_clock_task_mult(struct rq * rq)66 static inline u64 rq_clock_task_mult(struct rq *rq)
67 {
68 	lockdep_assert_rq_held(rq);
69 	assert_clock_updated(rq);
70 
71 	return rq->clock_task_mult;
72 }
73 
rq_clock_pelt(struct rq * rq)74 static inline u64 rq_clock_pelt(struct rq *rq)
75 {
76 	lockdep_assert_rq_held(rq);
77 	assert_clock_updated(rq);
78 
79 	return rq->clock_pelt - rq->lost_idle_time;
80 }
81 
82 /* The rq is idle, we can sync to clock_task */
_update_idle_rq_clock_pelt(struct rq * rq)83 static inline void _update_idle_rq_clock_pelt(struct rq *rq)
84 {
85 	rq->clock_pelt = rq_clock_task_mult(rq);
86 
87 	u64_u32_store(rq->clock_idle, rq_clock(rq));
88 	/* Paired with smp_rmb in migrate_se_pelt_lag() */
89 	smp_wmb();
90 	u64_u32_store(rq->clock_pelt_idle, rq_clock_pelt(rq));
91 }
92 
93 /*
94  * The clock_pelt scales the time to reflect the effective amount of
95  * computation done during the running delta time but then sync back to
96  * clock_task when rq is idle.
97  *
98  *
99  * absolute time   | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
100  * @ max capacity  ------******---------------******---------------
101  * @ half capacity ------************---------************---------
102  * clock pelt      | 1| 2|    3|    4| 7| 8| 9|   10|   11|14|15|16
103  *
104  */
update_rq_clock_pelt(struct rq * rq,s64 delta)105 static inline void update_rq_clock_pelt(struct rq *rq, s64 delta)
106 {
107 	int ret = 0;
108 
109 	trace_android_rvh_update_rq_clock_pelt(rq, delta, &ret);
110 	if (ret)
111 		return;
112 
113 	if (unlikely(is_idle_task(rq->curr))) {
114 		_update_idle_rq_clock_pelt(rq);
115 		return;
116 	}
117 
118 	/*
119 	 * When a rq runs at a lower compute capacity, it will need
120 	 * more time to do the same amount of work than at max
121 	 * capacity. In order to be invariant, we scale the delta to
122 	 * reflect how much work has been really done.
123 	 * Running longer results in stealing idle time that will
124 	 * disturb the load signal compared to max capacity. This
125 	 * stolen idle time will be automatically reflected when the
126 	 * rq will be idle and the clock will be synced with
127 	 * rq_clock_task.
128 	 */
129 
130 	/*
131 	 * Scale the elapsed time to reflect the real amount of
132 	 * computation
133 	 */
134 	delta = cap_scale(delta, arch_scale_cpu_capacity(cpu_of(rq)));
135 	delta = cap_scale(delta, arch_scale_freq_capacity(cpu_of(rq)));
136 
137 	rq->clock_pelt += delta;
138 }
139 
140 extern unsigned int sched_pelt_lshift;
141 
142 /*
143  * absolute time   |1      |2      |3      |4      |5      |6      |
144  * @ mult = 1      --------****************--------****************-
145  * @ mult = 2      --------********----------------********---------
146  * @ mult = 4      --------****--------------------****-------------
147  * clock task mult
148  * @ mult = 2      |   |   |2  |3  |   |   |   |   |5  |6  |   |   |
149  * @ mult = 4      | | | | |2|3| | | | | | | | | | |5|6| | | | | | |
150  *
151  */
update_rq_clock_task_mult(struct rq * rq,s64 delta)152 static inline void update_rq_clock_task_mult(struct rq *rq, s64 delta)
153 {
154 	delta <<= READ_ONCE(sched_pelt_lshift);
155 
156 	rq->clock_task_mult += delta;
157 
158 	update_rq_clock_pelt(rq, delta);
159 }
160 
161 /*
162  * When rq becomes idle, we have to check if it has lost idle time
163  * because it was fully busy. A rq is fully used when the /Sum util_sum
164  * is greater or equal to:
165  * (LOAD_AVG_MAX - 1024 + rq->cfs.avg.period_contrib) << SCHED_CAPACITY_SHIFT;
166  * For optimization and computing rounding purpose, we don't take into account
167  * the position in the current window (period_contrib) and we use the higher
168  * bound of util_sum to decide.
169  */
update_idle_rq_clock_pelt(struct rq * rq)170 static inline void update_idle_rq_clock_pelt(struct rq *rq)
171 {
172 	u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT) - LOAD_AVG_MAX;
173 	u32 util_sum = rq->cfs.avg.util_sum;
174 	util_sum += rq->avg_rt.util_sum;
175 	util_sum += rq->avg_dl.util_sum;
176 
177 	/*
178 	 * Reflecting stolen time makes sense only if the idle
179 	 * phase would be present at max capacity. As soon as the
180 	 * utilization of a rq has reached the maximum value, it is
181 	 * considered as an always running rq without idle time to
182 	 * steal. This potential idle time is considered as lost in
183 	 * this case. We keep track of this lost idle time compare to
184 	 * rq's clock_task.
185 	 */
186 	if (util_sum >= divider)
187 		rq->lost_idle_time += rq_clock_task_mult(rq) - rq->clock_pelt;
188 
189 	_update_idle_rq_clock_pelt(rq);
190 }
191 
192 #ifdef CONFIG_CFS_BANDWIDTH
update_idle_cfs_rq_clock_pelt(struct cfs_rq * cfs_rq)193 static inline void update_idle_cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
194 {
195 	u64 throttled;
196 
197 	if (unlikely(cfs_rq->throttle_count))
198 		throttled = U64_MAX;
199 	else
200 		throttled = cfs_rq->throttled_clock_pelt_time;
201 
202 	u64_u32_store(cfs_rq->throttled_pelt_idle, throttled);
203 }
204 
205 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
cfs_rq_clock_pelt(struct cfs_rq * cfs_rq)206 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
207 {
208 	if (unlikely(cfs_rq->throttle_count))
209 		return cfs_rq->throttled_clock_pelt - cfs_rq->throttled_clock_pelt_time;
210 
211 	return rq_clock_pelt(rq_of(cfs_rq)) - cfs_rq->throttled_clock_pelt_time;
212 }
213 #else
update_idle_cfs_rq_clock_pelt(struct cfs_rq * cfs_rq)214 static inline void update_idle_cfs_rq_clock_pelt(struct cfs_rq *cfs_rq) { }
cfs_rq_clock_pelt(struct cfs_rq * cfs_rq)215 static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
216 {
217 	return rq_clock_pelt(rq_of(cfs_rq));
218 }
219 #endif
220 
221 #else
222 
223 static inline int
update_cfs_rq_load_avg(u64 now,struct cfs_rq * cfs_rq)224 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
225 {
226 	return 0;
227 }
228 
229 static inline int
update_rt_rq_load_avg(u64 now,struct rq * rq,int running)230 update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
231 {
232 	return 0;
233 }
234 
235 static inline int
update_dl_rq_load_avg(u64 now,struct rq * rq,int running)236 update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
237 {
238 	return 0;
239 }
240 
241 static inline int
update_hw_load_avg(u64 now,struct rq * rq,u64 capacity)242 update_hw_load_avg(u64 now, struct rq *rq, u64 capacity)
243 {
244 	return 0;
245 }
246 
hw_load_avg(struct rq * rq)247 static inline u64 hw_load_avg(struct rq *rq)
248 {
249 	return 0;
250 }
251 
252 static inline int
update_irq_load_avg(struct rq * rq,u64 running)253 update_irq_load_avg(struct rq *rq, u64 running)
254 {
255 	return 0;
256 }
257 
rq_clock_task_mult(struct rq * rq)258 static inline u64 rq_clock_task_mult(struct rq *rq)
259 {
260 	return rq_clock_task(rq);
261 }
262 
rq_clock_pelt(struct rq * rq)263 static inline u64 rq_clock_pelt(struct rq *rq)
264 {
265 	return rq_clock_task_mult(rq);
266 }
267 
268 static inline void
update_rq_clock_task_mult(struct rq * rq,s64 delta)269 update_rq_clock_task_mult(struct rq *rq, s64 delta) { }
270 
271 static inline void
update_idle_rq_clock_pelt(struct rq * rq)272 update_idle_rq_clock_pelt(struct rq *rq) { }
273 
update_idle_cfs_rq_clock_pelt(struct cfs_rq * cfs_rq)274 static inline void update_idle_cfs_rq_clock_pelt(struct cfs_rq *cfs_rq) { }
275 #endif
276 
277 
278