1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Per Entity Load Tracking (PELT)
4 *
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 *
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 *
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 *
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 *
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 *
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
22 *
23 * Move PELT related code from fair.c into this pelt.c file
24 * Author: Vincent Guittot <vincent.guittot@linaro.org>
25 */
26#include "pelt.h"
27
28/*
29 * Approximate:
30 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
31 */
32static u64 decay_load(u64 val, u64 n)
33{
34 unsigned int local_n;
35
36 if (unlikely(n > LOAD_AVG_PERIOD * 63))
37 return 0;
38
39 /* after bounds checking we can collapse to 32-bit */
40 local_n = n;
41
42 /*
43 * As y^PERIOD = 1/2, we can combine
44 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
45 * With a look-up table which covers y^n (n<PERIOD)
46 *
47 * To achieve constant time decay_load.
48 */
49 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
50 val >>= local_n / LOAD_AVG_PERIOD;
51 local_n %= LOAD_AVG_PERIOD;
52 }
53
54 val = mul_u64_u32_shr(a: val, mul: runnable_avg_yN_inv[local_n], shift: 32);
55 return val;
56}
57
58static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
59{
60 u32 c1, c2, c3 = d3; /* y^0 == 1 */
61
62 /*
63 * c1 = d1 y^p
64 */
65 c1 = decay_load(val: (u64)d1, n: periods);
66
67 /*
68 * p-1
69 * c2 = 1024 \Sum y^n
70 * n=1
71 *
72 * inf inf
73 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
74 * n=0 n=p
75 */
76 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, n: periods) - 1024;
77
78 return c1 + c2 + c3;
79}
80
81/*
82 * Accumulate the three separate parts of the sum; d1 the remainder
83 * of the last (incomplete) period, d2 the span of full periods and d3
84 * the remainder of the (incomplete) current period.
85 *
86 * d1 d2 d3
87 * ^ ^ ^
88 * | | |
89 * |<->|<----------------->|<--->|
90 * ... |---x---|------| ... |------|-----x (now)
91 *
92 * p-1
93 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
94 * n=1
95 *
96 * = u y^p + (Step 1)
97 *
98 * p-1
99 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
100 * n=1
101 */
102static __always_inline u32
103accumulate_sum(u64 delta, struct sched_avg *sa,
104 unsigned long load, unsigned long runnable, int running)
105{
106 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
107 u64 periods;
108
109 delta += sa->period_contrib;
110 periods = delta / 1024; /* A period is 1024us (~1ms) */
111
112 /*
113 * Step 1: decay old *_sum if we crossed period boundaries.
114 */
115 if (periods) {
116 sa->load_sum = decay_load(val: sa->load_sum, n: periods);
117 sa->runnable_sum =
118 decay_load(val: sa->runnable_sum, n: periods);
119 sa->util_sum = decay_load(val: (u64)(sa->util_sum), n: periods);
120
121 /*
122 * Step 2
123 */
124 delta %= 1024;
125 if (load) {
126 /*
127 * This relies on the:
128 *
129 * if (!load)
130 * runnable = running = 0;
131 *
132 * clause from ___update_load_sum(); this results in
133 * the below usage of @contrib to disappear entirely,
134 * so no point in calculating it.
135 */
136 contrib = __accumulate_pelt_segments(periods,
137 d1: 1024 - sa->period_contrib, d3: delta);
138 }
139 }
140 sa->period_contrib = delta;
141
142 if (load)
143 sa->load_sum += load * contrib;
144 if (runnable)
145 sa->runnable_sum += runnable * contrib << SCHED_CAPACITY_SHIFT;
146 if (running)
147 sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
148
149 return periods;
150}
151
152/*
153 * We can represent the historical contribution to runnable average as the
154 * coefficients of a geometric series. To do this we sub-divide our runnable
155 * history into segments of approximately 1ms (1024us); label the segment that
156 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
157 *
158 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
159 * p0 p1 p2
160 * (now) (~1ms ago) (~2ms ago)
161 *
162 * Let u_i denote the fraction of p_i that the entity was runnable.
163 *
164 * We then designate the fractions u_i as our co-efficients, yielding the
165 * following representation of historical load:
166 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
167 *
168 * We choose y based on the with of a reasonably scheduling period, fixing:
169 * y^32 = 0.5
170 *
171 * This means that the contribution to load ~32ms ago (u_32) will be weighted
172 * approximately half as much as the contribution to load within the last ms
173 * (u_0).
174 *
175 * When a period "rolls over" and we have new u_0`, multiplying the previous
176 * sum again by y is sufficient to update:
177 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
178 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
179 */
180static __always_inline int
181___update_load_sum(u64 now, struct sched_avg *sa,
182 unsigned long load, unsigned long runnable, int running)
183{
184 u64 delta;
185
186 delta = now - sa->last_update_time;
187 /*
188 * This should only happen when time goes backwards, which it
189 * unfortunately does during sched clock init when we swap over to TSC.
190 */
191 if ((s64)delta < 0) {
192 sa->last_update_time = now;
193 return 0;
194 }
195
196 /*
197 * Use 1024ns as the unit of measurement since it's a reasonable
198 * approximation of 1us and fast to compute.
199 */
200 delta >>= 10;
201 if (!delta)
202 return 0;
203
204 sa->last_update_time += delta << 10;
205
206 /*
207 * running is a subset of runnable (weight) so running can't be set if
208 * runnable is clear. But there are some corner cases where the current
209 * se has been already dequeued but cfs_rq->curr still points to it.
210 * This means that weight will be 0 but not running for a sched_entity
211 * but also for a cfs_rq if the latter becomes idle. As an example,
212 * this happens during sched_balance_newidle() which calls
213 * sched_balance_update_blocked_averages().
214 *
215 * Also see the comment in accumulate_sum().
216 */
217 if (!load)
218 runnable = running = 0;
219
220 /*
221 * Now we know we crossed measurement unit boundaries. The *_avg
222 * accrues by two steps:
223 *
224 * Step 1: accumulate *_sum since last_update_time. If we haven't
225 * crossed period boundaries, finish.
226 */
227 if (!accumulate_sum(delta, sa, load, runnable, running))
228 return 0;
229
230 return 1;
231}
232
233/*
234 * When syncing *_avg with *_sum, we must take into account the current
235 * position in the PELT segment otherwise the remaining part of the segment
236 * will be considered as idle time whereas it's not yet elapsed and this will
237 * generate unwanted oscillation in the range [1002..1024[.
238 *
239 * The max value of *_sum varies with the position in the time segment and is
240 * equals to :
241 *
242 * LOAD_AVG_MAX*y + sa->period_contrib
243 *
244 * which can be simplified into:
245 *
246 * LOAD_AVG_MAX - 1024 + sa->period_contrib
247 *
248 * because LOAD_AVG_MAX*y == LOAD_AVG_MAX-1024
249 *
250 * The same care must be taken when a sched entity is added, updated or
251 * removed from a cfs_rq and we need to update sched_avg. Scheduler entities
252 * and the cfs rq, to which they are attached, have the same position in the
253 * time segment because they use the same clock. This means that we can use
254 * the period_contrib of cfs_rq when updating the sched_avg of a sched_entity
255 * if it's more convenient.
256 */
257static __always_inline void
258___update_load_avg(struct sched_avg *sa, unsigned long load)
259{
260 u32 divider = get_pelt_divider(avg: sa);
261
262 /*
263 * Step 2: update *_avg.
264 */
265 sa->load_avg = div_u64(dividend: load * sa->load_sum, divisor: divider);
266 sa->runnable_avg = div_u64(dividend: sa->runnable_sum, divisor: divider);
267 WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
268}
269
270/*
271 * sched_entity:
272 *
273 * task:
274 * se_weight() = se->load.weight
275 * se_runnable() = !!on_rq
276 *
277 * group: [ see update_cfs_group() ]
278 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
279 * se_runnable() = grq->h_nr_runnable
280 *
281 * runnable_sum = se_runnable() * runnable = grq->runnable_sum
282 * runnable_avg = runnable_sum
283 *
284 * load_sum := runnable
285 * load_avg = se_weight(se) * load_sum
286 *
287 * cfq_rq:
288 *
289 * runnable_sum = \Sum se->avg.runnable_sum
290 * runnable_avg = \Sum se->avg.runnable_avg
291 *
292 * load_sum = \Sum se_weight(se) * se->avg.load_sum
293 * load_avg = \Sum se->avg.load_avg
294 */
295
296int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
297{
298 if (___update_load_sum(now, sa: &se->avg, load: 0, runnable: 0, running: 0)) {
299 ___update_load_avg(sa: &se->avg, load: se_weight(se));
300 trace_pelt_se_tp(se);
301 return 1;
302 }
303
304 return 0;
305}
306
307int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
308{
309 if (___update_load_sum(now, sa: &se->avg, load: !!se->on_rq, runnable: se_runnable(se),
310 running: cfs_rq->curr == se)) {
311
312 ___update_load_avg(sa: &se->avg, load: se_weight(se));
313 cfs_se_util_change(avg: &se->avg);
314 trace_pelt_se_tp(se);
315 return 1;
316 }
317
318 return 0;
319}
320
321int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
322{
323 if (___update_load_sum(now, sa: &cfs_rq->avg,
324 scale_load_down(cfs_rq->load.weight),
325 runnable: cfs_rq->h_nr_runnable,
326 running: cfs_rq->curr != NULL)) {
327
328 ___update_load_avg(sa: &cfs_rq->avg, load: 1);
329 trace_pelt_cfs_tp(cfs_rq);
330 return 1;
331 }
332
333 return 0;
334}
335
336/*
337 * rt_rq:
338 *
339 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
340 * util_sum = cpu_scale * load_sum
341 * runnable_sum = util_sum
342 *
343 * load_avg and runnable_avg are not supported and meaningless.
344 *
345 */
346
347int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
348{
349 if (___update_load_sum(now, sa: &rq->avg_rt,
350 load: running,
351 runnable: running,
352 running)) {
353
354 ___update_load_avg(sa: &rq->avg_rt, load: 1);
355 trace_pelt_rt_tp(rq);
356 return 1;
357 }
358
359 return 0;
360}
361
362/*
363 * dl_rq:
364 *
365 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
366 * util_sum = cpu_scale * load_sum
367 * runnable_sum = util_sum
368 *
369 * load_avg and runnable_avg are not supported and meaningless.
370 *
371 */
372
373int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
374{
375 if (___update_load_sum(now, sa: &rq->avg_dl,
376 load: running,
377 runnable: running,
378 running)) {
379
380 ___update_load_avg(sa: &rq->avg_dl, load: 1);
381 trace_pelt_dl_tp(rq);
382 return 1;
383 }
384
385 return 0;
386}
387
388#ifdef CONFIG_SCHED_HW_PRESSURE
389/*
390 * hardware:
391 *
392 * load_sum = \Sum se->avg.load_sum but se->avg.load_sum is not tracked
393 *
394 * util_avg and runnable_load_avg are not supported and meaningless.
395 *
396 * Unlike rt/dl utilization tracking that track time spent by a cpu
397 * running a rt/dl task through util_avg, the average HW pressure is
398 * tracked through load_avg. This is because HW pressure signal is
399 * time weighted "delta" capacity unlike util_avg which is binary.
400 * "delta capacity" = actual capacity -
401 * capped capacity a cpu due to a HW event.
402 */
403
404int update_hw_load_avg(u64 now, struct rq *rq, u64 capacity)
405{
406 if (___update_load_sum(now, &rq->avg_hw,
407 capacity,
408 capacity,
409 capacity)) {
410 ___update_load_avg(&rq->avg_hw, 1);
411 trace_pelt_hw_tp(rq);
412 return 1;
413 }
414
415 return 0;
416}
417#endif /* CONFIG_SCHED_HW_PRESSURE */
418
419#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
420/*
421 * IRQ:
422 *
423 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
424 * util_sum = cpu_scale * load_sum
425 * runnable_sum = util_sum
426 *
427 * load_avg and runnable_avg are not supported and meaningless.
428 *
429 */
430
431int update_irq_load_avg(struct rq *rq, u64 running)
432{
433 int ret = 0;
434
435 /*
436 * We can't use clock_pelt because IRQ time is not accounted in
437 * clock_task. Instead we directly scale the running time to
438 * reflect the real amount of computation
439 */
440 running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
441 running = cap_scale(running, arch_scale_cpu_capacity(cpu_of(rq)));
442
443 /*
444 * We know the time that has been used by interrupt since last update
445 * but we don't when. Let be pessimistic and assume that interrupt has
446 * happened just before the update. This is not so far from reality
447 * because interrupt will most probably wake up task and trig an update
448 * of rq clock during which the metric is updated.
449 * We start to decay with normal context time and then we add the
450 * interrupt context time.
451 * We can safely remove running from rq->clock because
452 * rq->clock += delta with delta >= running
453 */
454 ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
455 0,
456 0,
457 0);
458 ret += ___update_load_sum(rq->clock, &rq->avg_irq,
459 1,
460 1,
461 1);
462
463 if (ret) {
464 ___update_load_avg(&rq->avg_irq, 1);
465 trace_pelt_irq_tp(rq);
466 }
467
468 return ret;
469}
470#endif /* CONFIG_HAVE_SCHED_AVG_IRQ */
471
472/*
473 * Load avg and utiliztion metrics need to be updated periodically and before
474 * consumption. This function updates the metrics for all subsystems except for
475 * the fair class. @rq must be locked and have its clock updated.
476 */
477bool update_other_load_avgs(struct rq *rq)
478{
479 u64 now = rq_clock_pelt(rq);
480 const struct sched_class *curr_class = rq->donor->sched_class;
481 unsigned long hw_pressure = arch_scale_hw_pressure(cpu: cpu_of(rq));
482
483 lockdep_assert_rq_held(rq);
484
485 /* hw_pressure doesn't care about invariance */
486 return update_rt_rq_load_avg(now, rq, running: curr_class == &rt_sched_class) |
487 update_dl_rq_load_avg(now, rq, running: curr_class == &dl_sched_class) |
488 update_hw_load_avg(now: rq_clock_task(rq), rq, capacity: hw_pressure) |
489 update_irq_load_avg(rq, running: 0);
490}
491