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>
33 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
35 static u64 decay_load(u64 val, u64 n)
39 if (unlikely(n > LOAD_AVG_PERIOD * 63))
42 /* after bounds checking we can collapse to 32-bit */
46 * As y^PERIOD = 1/2, we can combine
47 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
48 * With a look-up table which covers y^n (n<PERIOD)
50 * To achieve constant time decay_load.
52 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
53 val >>= local_n / LOAD_AVG_PERIOD;
54 local_n %= LOAD_AVG_PERIOD;
57 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
61 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
63 u32 c1, c2, c3 = d3; /* y^0 == 1 */
68 c1 = decay_load((u64)d1, periods);
76 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
79 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
85 * Accumulate the three separate parts of the sum; d1 the remainder
86 * of the last (incomplete) period, d2 the span of full periods and d3
87 * the remainder of the (incomplete) current period.
92 * |<->|<----------------->|<--->|
93 * ... |---x---|------| ... |------|-----x (now)
96 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
102 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
105 static __always_inline u32
106 accumulate_sum(u64 delta, struct sched_avg *sa,
107 unsigned long load, unsigned long runnable, int running)
109 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
112 delta += sa->period_contrib;
113 periods = delta / 1024; /* A period is 1024us (~1ms) */
116 * Step 1: decay old *_sum if we crossed period boundaries.
119 sa->load_sum = decay_load(sa->load_sum, periods);
121 decay_load(sa->runnable_sum, periods);
122 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
130 * This relies on the:
133 * runnable = running = 0;
135 * clause from ___update_load_sum(); this results in
136 * the below usage of @contrib to disappear entirely,
137 * so no point in calculating it.
139 contrib = __accumulate_pelt_segments(periods,
140 1024 - sa->period_contrib, delta);
143 sa->period_contrib = delta;
146 sa->load_sum += load * contrib;
148 sa->runnable_sum += runnable * contrib << SCHED_CAPACITY_SHIFT;
150 sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
156 * We can represent the historical contribution to runnable average as the
157 * coefficients of a geometric series. To do this we sub-divide our runnable
158 * history into segments of approximately 1ms (1024us); label the segment that
159 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
161 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
163 * (now) (~1ms ago) (~2ms ago)
165 * Let u_i denote the fraction of p_i that the entity was runnable.
167 * We then designate the fractions u_i as our co-efficients, yielding the
168 * following representation of historical load:
169 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
171 * We choose y based on the with of a reasonably scheduling period, fixing:
174 * This means that the contribution to load ~32ms ago (u_32) will be weighted
175 * approximately half as much as the contribution to load within the last ms
178 * When a period "rolls over" and we have new u_0`, multiplying the previous
179 * sum again by y is sufficient to update:
180 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
181 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
183 static __always_inline int
184 ___update_load_sum(u64 now, struct sched_avg *sa,
185 unsigned long load, unsigned long runnable, int running)
189 delta = now - sa->last_update_time;
191 * This should only happen when time goes backwards, which it
192 * unfortunately does during sched clock init when we swap over to TSC.
194 if ((s64)delta < 0) {
195 sa->last_update_time = now;
200 * Use 1024ns as the unit of measurement since it's a reasonable
201 * approximation of 1us and fast to compute.
207 sa->last_update_time += delta << 10;
210 * running is a subset of runnable (weight) so running can't be set if
211 * runnable is clear. But there are some corner cases where the current
212 * se has been already dequeued but cfs_rq->curr still points to it.
213 * This means that weight will be 0 but not running for a sched_entity
214 * but also for a cfs_rq if the latter becomes idle. As an example,
215 * this happens during idle_balance() which calls
216 * update_blocked_averages().
218 * Also see the comment in accumulate_sum().
221 runnable = running = 0;
224 * Now we know we crossed measurement unit boundaries. The *_avg
225 * accrues by two steps:
227 * Step 1: accumulate *_sum since last_update_time. If we haven't
228 * crossed period boundaries, finish.
230 if (!accumulate_sum(delta, sa, load, runnable, running))
237 * When syncing *_avg with *_sum, we must take into account the current
238 * position in the PELT segment otherwise the remaining part of the segment
239 * will be considered as idle time whereas it's not yet elapsed and this will
240 * generate unwanted oscillation in the range [1002..1024[.
242 * The max value of *_sum varies with the position in the time segment and is
245 * LOAD_AVG_MAX*y + sa->period_contrib
247 * which can be simplified into:
249 * LOAD_AVG_MAX - 1024 + sa->period_contrib
251 * because LOAD_AVG_MAX*y == LOAD_AVG_MAX-1024
253 * The same care must be taken when a sched entity is added, updated or
254 * removed from a cfs_rq and we need to update sched_avg. Scheduler entities
255 * and the cfs rq, to which they are attached, have the same position in the
256 * time segment because they use the same clock. This means that we can use
257 * the period_contrib of cfs_rq when updating the sched_avg of a sched_entity
258 * if it's more convenient.
260 static __always_inline void
261 ___update_load_avg(struct sched_avg *sa, unsigned long load)
263 u32 divider = get_pelt_divider(sa);
266 * Step 2: update *_avg.
268 sa->load_avg = div_u64(load * sa->load_sum, divider);
269 sa->runnable_avg = div_u64(sa->runnable_sum, divider);
270 WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
277 * se_weight() = se->load.weight
278 * se_runnable() = !!on_rq
280 * group: [ see update_cfs_group() ]
281 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
282 * se_runnable() = grq->h_nr_running
284 * runnable_sum = se_runnable() * runnable = grq->runnable_sum
285 * runnable_avg = runnable_sum
287 * load_sum := runnable
288 * load_avg = se_weight(se) * load_sum
292 * runnable_sum = \Sum se->avg.runnable_sum
293 * runnable_avg = \Sum se->avg.runnable_avg
295 * load_sum = \Sum se_weight(se) * se->avg.load_sum
296 * load_avg = \Sum se->avg.load_avg
299 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
301 if (___update_load_sum(now, &se->avg, 0, 0, 0)) {
302 ___update_load_avg(&se->avg, se_weight(se));
303 trace_pelt_se_tp(se);
310 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
312 if (___update_load_sum(now, &se->avg, !!se->on_rq, se_runnable(se),
313 cfs_rq->curr == se)) {
315 ___update_load_avg(&se->avg, se_weight(se));
316 cfs_se_util_change(&se->avg);
317 trace_pelt_se_tp(se);
324 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
326 if (___update_load_sum(now, &cfs_rq->avg,
327 scale_load_down(cfs_rq->load.weight),
328 cfs_rq->h_nr_running,
329 cfs_rq->curr != NULL)) {
331 ___update_load_avg(&cfs_rq->avg, 1);
332 trace_pelt_cfs_tp(cfs_rq);
342 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
343 * util_sum = cpu_scale * load_sum
344 * runnable_sum = util_sum
346 * load_avg and runnable_avg are not supported and meaningless.
350 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
352 if (___update_load_sum(now, &rq->avg_rt,
357 ___update_load_avg(&rq->avg_rt, 1);
358 trace_pelt_rt_tp(rq);
368 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
369 * util_sum = cpu_scale * load_sum
370 * runnable_sum = util_sum
372 * load_avg and runnable_avg are not supported and meaningless.
376 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
378 if (___update_load_sum(now, &rq->avg_dl,
383 ___update_load_avg(&rq->avg_dl, 1);
384 trace_pelt_dl_tp(rq);
391 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
395 * load_sum = \Sum se->avg.load_sum but se->avg.load_sum is not tracked
397 * util_avg and runnable_load_avg are not supported and meaningless.
399 * Unlike rt/dl utilization tracking that track time spent by a cpu
400 * running a rt/dl task through util_avg, the average thermal pressure is
401 * tracked through load_avg. This is because thermal pressure signal is
402 * time weighted "delta" capacity unlike util_avg which is binary.
403 * "delta capacity" = actual capacity -
404 * capped capacity a cpu due to a thermal event.
407 int update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
409 if (___update_load_sum(now, &rq->avg_thermal,
413 ___update_load_avg(&rq->avg_thermal, 1);
414 trace_pelt_thermal_tp(rq);
422 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
426 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
427 * util_sum = cpu_scale * load_sum
428 * runnable_sum = util_sum
430 * load_avg and runnable_avg are not supported and meaningless.
434 int update_irq_load_avg(struct rq *rq, u64 running)
439 * We can't use clock_pelt because irq time is not accounted in
440 * clock_task. Instead we directly scale the running time to
441 * reflect the real amount of computation
443 running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
444 running = cap_scale(running, arch_scale_cpu_capacity(cpu_of(rq)));
447 * We know the time that has been used by interrupt since last update
448 * but we don't when. Let be pessimistic and assume that interrupt has
449 * happened just before the update. This is not so far from reality
450 * because interrupt will most probably wake up task and trig an update
451 * of rq clock during which the metric is updated.
452 * We start to decay with normal context time and then we add the
453 * interrupt context time.
454 * We can safely remove running from rq->clock because
455 * rq->clock += delta with delta >= running
457 ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
461 ret += ___update_load_sum(rq->clock, &rq->avg_irq,
467 ___update_load_avg(&rq->avg_irq, 1);
468 trace_pelt_irq_tp(rq);