1 // SPDX-License-Identifier: GPL-2.0
3 * Per Entity Load Tracking
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 * Move PELT related code from fair.c into this pelt.c file
24 * Author: Vincent Guittot <vincent.guittot@linaro.org>
27 #include <linux/sched.h>
31 #include <trace/events/sched.h>
35 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
37 static u64
decay_load(u64 val
, u64 n
)
41 if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
44 /* after bounds checking we can collapse to 32-bit */
48 * As y^PERIOD = 1/2, we can combine
49 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
50 * With a look-up table which covers y^n (n<PERIOD)
52 * To achieve constant time decay_load.
54 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
55 val
>>= local_n
/ LOAD_AVG_PERIOD
;
56 local_n
%= LOAD_AVG_PERIOD
;
59 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
63 static u32
__accumulate_pelt_segments(u64 periods
, u32 d1
, u32 d3
)
65 u32 c1
, c2
, c3
= d3
; /* y^0 == 1 */
70 c1
= decay_load((u64
)d1
, periods
);
78 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
81 c2
= LOAD_AVG_MAX
- decay_load(LOAD_AVG_MAX
, periods
) - 1024;
86 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
89 * Accumulate the three separate parts of the sum; d1 the remainder
90 * of the last (incomplete) period, d2 the span of full periods and d3
91 * the remainder of the (incomplete) current period.
96 * |<->|<----------------->|<--->|
97 * ... |---x---|------| ... |------|-----x (now)
100 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
106 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
109 static __always_inline u32
110 accumulate_sum(u64 delta
, struct sched_avg
*sa
,
111 unsigned long load
, unsigned long runnable
, int running
)
113 u32 contrib
= (u32
)delta
; /* p == 0 -> delta < 1024 */
116 delta
+= sa
->period_contrib
;
117 periods
= delta
/ 1024; /* A period is 1024us (~1ms) */
120 * Step 1: decay old *_sum if we crossed period boundaries.
123 sa
->load_sum
= decay_load(sa
->load_sum
, periods
);
124 sa
->runnable_load_sum
=
125 decay_load(sa
->runnable_load_sum
, periods
);
126 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
);
134 * This relies on the:
137 * runnable = running = 0;
139 * clause from ___update_load_sum(); this results in
140 * the below usage of @contrib to dissapear entirely,
141 * so no point in calculating it.
143 contrib
= __accumulate_pelt_segments(periods
,
144 1024 - sa
->period_contrib
, delta
);
147 sa
->period_contrib
= delta
;
150 sa
->load_sum
+= load
* contrib
;
152 sa
->runnable_load_sum
+= runnable
* contrib
;
154 sa
->util_sum
+= contrib
<< SCHED_CAPACITY_SHIFT
;
160 * We can represent the historical contribution to runnable average as the
161 * coefficients of a geometric series. To do this we sub-divide our runnable
162 * history into segments of approximately 1ms (1024us); label the segment that
163 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
165 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
167 * (now) (~1ms ago) (~2ms ago)
169 * Let u_i denote the fraction of p_i that the entity was runnable.
171 * We then designate the fractions u_i as our co-efficients, yielding the
172 * following representation of historical load:
173 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
175 * We choose y based on the with of a reasonably scheduling period, fixing:
178 * This means that the contribution to load ~32ms ago (u_32) will be weighted
179 * approximately half as much as the contribution to load within the last ms
182 * When a period "rolls over" and we have new u_0`, multiplying the previous
183 * sum again by y is sufficient to update:
184 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
185 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
187 static __always_inline
int
188 ___update_load_sum(u64 now
, struct sched_avg
*sa
,
189 unsigned long load
, unsigned long runnable
, int running
)
193 delta
= now
- sa
->last_update_time
;
195 * This should only happen when time goes backwards, which it
196 * unfortunately does during sched clock init when we swap over to TSC.
198 if ((s64
)delta
< 0) {
199 sa
->last_update_time
= now
;
204 * Use 1024ns as the unit of measurement since it's a reasonable
205 * approximation of 1us and fast to compute.
211 sa
->last_update_time
+= delta
<< 10;
214 * running is a subset of runnable (weight) so running can't be set if
215 * runnable is clear. But there are some corner cases where the current
216 * se has been already dequeued but cfs_rq->curr still points to it.
217 * This means that weight will be 0 but not running for a sched_entity
218 * but also for a cfs_rq if the latter becomes idle. As an example,
219 * this happens during idle_balance() which calls
220 * update_blocked_averages().
222 * Also see the comment in accumulate_sum().
225 runnable
= running
= 0;
228 * Now we know we crossed measurement unit boundaries. The *_avg
229 * accrues by two steps:
231 * Step 1: accumulate *_sum since last_update_time. If we haven't
232 * crossed period boundaries, finish.
234 if (!accumulate_sum(delta
, sa
, load
, runnable
, running
))
240 static __always_inline
void
241 ___update_load_avg(struct sched_avg
*sa
, unsigned long load
, unsigned long runnable
)
243 u32 divider
= LOAD_AVG_MAX
- 1024 + sa
->period_contrib
;
246 * Step 2: update *_avg.
248 sa
->load_avg
= div_u64(load
* sa
->load_sum
, divider
);
249 sa
->runnable_load_avg
= div_u64(runnable
* sa
->runnable_load_sum
, divider
);
250 WRITE_ONCE(sa
->util_avg
, sa
->util_sum
/ divider
);
257 * se_runnable() == se_weight()
259 * group: [ see update_cfs_group() ]
260 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
261 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
263 * load_sum := runnable_sum
264 * load_avg = se_weight(se) * runnable_avg
266 * runnable_load_sum := runnable_sum
267 * runnable_load_avg = se_runnable(se) * runnable_avg
269 * XXX collapse load_sum and runnable_load_sum
273 * load_sum = \Sum se_weight(se) * se->avg.load_sum
274 * load_avg = \Sum se->avg.load_avg
276 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
277 * runnable_load_avg = \Sum se->avg.runable_load_avg
280 int __update_load_avg_blocked_se(u64 now
, struct sched_entity
*se
)
282 if (___update_load_sum(now
, &se
->avg
, 0, 0, 0)) {
283 ___update_load_avg(&se
->avg
, se_weight(se
), se_runnable(se
));
284 trace_pelt_se_tp(se
);
291 int __update_load_avg_se(u64 now
, struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
293 if (___update_load_sum(now
, &se
->avg
, !!se
->on_rq
, !!se
->on_rq
,
294 cfs_rq
->curr
== se
)) {
296 ___update_load_avg(&se
->avg
, se_weight(se
), se_runnable(se
));
297 cfs_se_util_change(&se
->avg
);
298 trace_pelt_se_tp(se
);
305 int __update_load_avg_cfs_rq(u64 now
, struct cfs_rq
*cfs_rq
)
307 if (___update_load_sum(now
, &cfs_rq
->avg
,
308 scale_load_down(cfs_rq
->load
.weight
),
309 scale_load_down(cfs_rq
->runnable_weight
),
310 cfs_rq
->curr
!= NULL
)) {
312 ___update_load_avg(&cfs_rq
->avg
, 1, 1);
313 trace_pelt_cfs_tp(cfs_rq
);
323 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
324 * util_sum = cpu_scale * load_sum
325 * runnable_load_sum = load_sum
327 * load_avg and runnable_load_avg are not supported and meaningless.
331 int update_rt_rq_load_avg(u64 now
, struct rq
*rq
, int running
)
333 if (___update_load_sum(now
, &rq
->avg_rt
,
338 ___update_load_avg(&rq
->avg_rt
, 1, 1);
339 trace_pelt_rt_tp(rq
);
349 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
350 * util_sum = cpu_scale * load_sum
351 * runnable_load_sum = load_sum
355 int update_dl_rq_load_avg(u64 now
, struct rq
*rq
, int running
)
357 if (___update_load_sum(now
, &rq
->avg_dl
,
362 ___update_load_avg(&rq
->avg_dl
, 1, 1);
363 trace_pelt_dl_tp(rq
);
370 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
374 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
375 * util_sum = cpu_scale * load_sum
376 * runnable_load_sum = load_sum
380 int update_irq_load_avg(struct rq
*rq
, u64 running
)
385 * We can't use clock_pelt because irq time is not accounted in
386 * clock_task. Instead we directly scale the running time to
387 * reflect the real amount of computation
389 running
= cap_scale(running
, arch_scale_freq_capacity(cpu_of(rq
)));
390 running
= cap_scale(running
, arch_scale_cpu_capacity(cpu_of(rq
)));
393 * We know the time that has been used by interrupt since last update
394 * but we don't when. Let be pessimistic and assume that interrupt has
395 * happened just before the update. This is not so far from reality
396 * because interrupt will most probably wake up task and trig an update
397 * of rq clock during which the metric is updated.
398 * We start to decay with normal context time and then we add the
399 * interrupt context time.
400 * We can safely remove running from rq->clock because
401 * rq->clock += delta with delta >= running
403 ret
= ___update_load_sum(rq
->clock
- running
, &rq
->avg_irq
,
407 ret
+= ___update_load_sum(rq
->clock
, &rq
->avg_irq
,
413 ___update_load_avg(&rq
->avg_irq
, 1, 1);
414 trace_pelt_irq_tp(rq
);