xref: /openbmc/linux/kernel/sched/pelt.c (revision f96eca43)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Per Entity Load Tracking
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 
27 /*
28  * Approximate:
29  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
30  */
decay_load(u64 val,u64 n)31 static u64 decay_load(u64 val, u64 n)
32 {
33 	unsigned int local_n;
34 
35 	if (unlikely(n > LOAD_AVG_PERIOD * 63))
36 		return 0;
37 
38 	/* after bounds checking we can collapse to 32-bit */
39 	local_n = n;
40 
41 	/*
42 	 * As y^PERIOD = 1/2, we can combine
43 	 *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
44 	 * With a look-up table which covers y^n (n<PERIOD)
45 	 *
46 	 * To achieve constant time decay_load.
47 	 */
48 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
49 		val >>= local_n / LOAD_AVG_PERIOD;
50 		local_n %= LOAD_AVG_PERIOD;
51 	}
52 
53 	val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
54 	return val;
55 }
56 
__accumulate_pelt_segments(u64 periods,u32 d1,u32 d3)57 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
58 {
59 	u32 c1, c2, c3 = d3; /* y^0 == 1 */
60 
61 	/*
62 	 * c1 = d1 y^p
63 	 */
64 	c1 = decay_load((u64)d1, periods);
65 
66 	/*
67 	 *            p-1
68 	 * c2 = 1024 \Sum y^n
69 	 *            n=1
70 	 *
71 	 *              inf        inf
72 	 *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
73 	 *              n=0        n=p
74 	 */
75 	c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
76 
77 	return c1 + c2 + c3;
78 }
79 
80 /*
81  * Accumulate the three separate parts of the sum; d1 the remainder
82  * of the last (incomplete) period, d2 the span of full periods and d3
83  * the remainder of the (incomplete) current period.
84  *
85  *           d1          d2           d3
86  *           ^           ^            ^
87  *           |           |            |
88  *         |<->|<----------------->|<--->|
89  * ... |---x---|------| ... |------|-----x (now)
90  *
91  *                           p-1
92  * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
93  *                           n=1
94  *
95  *    = u y^p +					(Step 1)
96  *
97  *                     p-1
98  *      d1 y^p + 1024 \Sum y^n + d3 y^0		(Step 2)
99  *                     n=1
100  */
101 static __always_inline u32
accumulate_sum(u64 delta,struct sched_avg * sa,unsigned long load,unsigned long runnable,int running)102 accumulate_sum(u64 delta, struct sched_avg *sa,
103 	       unsigned long load, unsigned long runnable, int running)
104 {
105 	u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
106 	u64 periods;
107 
108 	delta += sa->period_contrib;
109 	periods = delta / 1024; /* A period is 1024us (~1ms) */
110 
111 	/*
112 	 * Step 1: decay old *_sum if we crossed period boundaries.
113 	 */
114 	if (periods) {
115 		sa->load_sum = decay_load(sa->load_sum, periods);
116 		sa->runnable_sum =
117 			decay_load(sa->runnable_sum, periods);
118 		sa->util_sum = decay_load((u64)(sa->util_sum), periods);
119 
120 		/*
121 		 * Step 2
122 		 */
123 		delta %= 1024;
124 		if (load) {
125 			/*
126 			 * This relies on the:
127 			 *
128 			 * if (!load)
129 			 *	runnable = running = 0;
130 			 *
131 			 * clause from ___update_load_sum(); this results in
132 			 * the below usage of @contrib to disappear entirely,
133 			 * so no point in calculating it.
134 			 */
135 			contrib = __accumulate_pelt_segments(periods,
136 					1024 - sa->period_contrib, delta);
137 		}
138 	}
139 	sa->period_contrib = delta;
140 
141 	if (load)
142 		sa->load_sum += load * contrib;
143 	if (runnable)
144 		sa->runnable_sum += runnable * contrib << SCHED_CAPACITY_SHIFT;
145 	if (running)
146 		sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
147 
148 	return periods;
149 }
150 
151 /*
152  * We can represent the historical contribution to runnable average as the
153  * coefficients of a geometric series.  To do this we sub-divide our runnable
154  * history into segments of approximately 1ms (1024us); label the segment that
155  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
156  *
157  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
158  *      p0            p1           p2
159  *     (now)       (~1ms ago)  (~2ms ago)
160  *
161  * Let u_i denote the fraction of p_i that the entity was runnable.
162  *
163  * We then designate the fractions u_i as our co-efficients, yielding the
164  * following representation of historical load:
165  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
166  *
167  * We choose y based on the with of a reasonably scheduling period, fixing:
168  *   y^32 = 0.5
169  *
170  * This means that the contribution to load ~32ms ago (u_32) will be weighted
171  * approximately half as much as the contribution to load within the last ms
172  * (u_0).
173  *
174  * When a period "rolls over" and we have new u_0`, multiplying the previous
175  * sum again by y is sufficient to update:
176  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
177  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
178  */
179 static __always_inline int
___update_load_sum(u64 now,struct sched_avg * sa,unsigned long load,unsigned long runnable,int running)180 ___update_load_sum(u64 now, struct sched_avg *sa,
181 		  unsigned long load, unsigned long runnable, int running)
182 {
183 	u64 delta;
184 
185 	delta = now - sa->last_update_time;
186 	/*
187 	 * This should only happen when time goes backwards, which it
188 	 * unfortunately does during sched clock init when we swap over to TSC.
189 	 */
190 	if ((s64)delta < 0) {
191 		sa->last_update_time = now;
192 		return 0;
193 	}
194 
195 	/*
196 	 * Use 1024ns as the unit of measurement since it's a reasonable
197 	 * approximation of 1us and fast to compute.
198 	 */
199 	delta >>= 10;
200 	if (!delta)
201 		return 0;
202 
203 	sa->last_update_time += delta << 10;
204 
205 	/*
206 	 * running is a subset of runnable (weight) so running can't be set if
207 	 * runnable is clear. But there are some corner cases where the current
208 	 * se has been already dequeued but cfs_rq->curr still points to it.
209 	 * This means that weight will be 0 but not running for a sched_entity
210 	 * but also for a cfs_rq if the latter becomes idle. As an example,
211 	 * this happens during idle_balance() which calls
212 	 * update_blocked_averages().
213 	 *
214 	 * Also see the comment in accumulate_sum().
215 	 */
216 	if (!load)
217 		runnable = running = 0;
218 
219 	/*
220 	 * Now we know we crossed measurement unit boundaries. The *_avg
221 	 * accrues by two steps:
222 	 *
223 	 * Step 1: accumulate *_sum since last_update_time. If we haven't
224 	 * crossed period boundaries, finish.
225 	 */
226 	if (!accumulate_sum(delta, sa, load, runnable, running))
227 		return 0;
228 
229 	return 1;
230 }
231 
232 /*
233  * When syncing *_avg with *_sum, we must take into account the current
234  * position in the PELT segment otherwise the remaining part of the segment
235  * will be considered as idle time whereas it's not yet elapsed and this will
236  * generate unwanted oscillation in the range [1002..1024[.
237  *
238  * The max value of *_sum varies with the position in the time segment and is
239  * equals to :
240  *
241  *   LOAD_AVG_MAX*y + sa->period_contrib
242  *
243  * which can be simplified into:
244  *
245  *   LOAD_AVG_MAX - 1024 + sa->period_contrib
246  *
247  * because LOAD_AVG_MAX*y == LOAD_AVG_MAX-1024
248  *
249  * The same care must be taken when a sched entity is added, updated or
250  * removed from a cfs_rq and we need to update sched_avg. Scheduler entities
251  * and the cfs rq, to which they are attached, have the same position in the
252  * time segment because they use the same clock. This means that we can use
253  * the period_contrib of cfs_rq when updating the sched_avg of a sched_entity
254  * if it's more convenient.
255  */
256 static __always_inline void
___update_load_avg(struct sched_avg * sa,unsigned long load)257 ___update_load_avg(struct sched_avg *sa, unsigned long load)
258 {
259 	u32 divider = get_pelt_divider(sa);
260 
261 	/*
262 	 * Step 2: update *_avg.
263 	 */
264 	sa->load_avg = div_u64(load * sa->load_sum, divider);
265 	sa->runnable_avg = div_u64(sa->runnable_sum, divider);
266 	WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
267 }
268 
269 /*
270  * sched_entity:
271  *
272  *   task:
273  *     se_weight()   = se->load.weight
274  *     se_runnable() = !!on_rq
275  *
276  *   group: [ see update_cfs_group() ]
277  *     se_weight()   = tg->weight * grq->load_avg / tg->load_avg
278  *     se_runnable() = grq->h_nr_running
279  *
280  *   runnable_sum = se_runnable() * runnable = grq->runnable_sum
281  *   runnable_avg = runnable_sum
282  *
283  *   load_sum := runnable
284  *   load_avg = se_weight(se) * load_sum
285  *
286  * cfq_rq:
287  *
288  *   runnable_sum = \Sum se->avg.runnable_sum
289  *   runnable_avg = \Sum se->avg.runnable_avg
290  *
291  *   load_sum = \Sum se_weight(se) * se->avg.load_sum
292  *   load_avg = \Sum se->avg.load_avg
293  */
294 
__update_load_avg_blocked_se(u64 now,struct sched_entity * se)295 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
296 {
297 	if (___update_load_sum(now, &se->avg, 0, 0, 0)) {
298 		___update_load_avg(&se->avg, se_weight(se));
299 		trace_pelt_se_tp(se);
300 		return 1;
301 	}
302 
303 	return 0;
304 }
305 
__update_load_avg_se(u64 now,struct cfs_rq * cfs_rq,struct sched_entity * se)306 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
307 {
308 	if (___update_load_sum(now, &se->avg, !!se->on_rq, se_runnable(se),
309 				cfs_rq->curr == se)) {
310 
311 		___update_load_avg(&se->avg, se_weight(se));
312 		cfs_se_util_change(&se->avg);
313 		trace_pelt_se_tp(se);
314 		return 1;
315 	}
316 
317 	return 0;
318 }
319 
__update_load_avg_cfs_rq(u64 now,struct cfs_rq * cfs_rq)320 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
321 {
322 	if (___update_load_sum(now, &cfs_rq->avg,
323 				scale_load_down(cfs_rq->load.weight),
324 				cfs_rq->h_nr_running,
325 				cfs_rq->curr != NULL)) {
326 
327 		___update_load_avg(&cfs_rq->avg, 1);
328 		trace_pelt_cfs_tp(cfs_rq);
329 		return 1;
330 	}
331 
332 	return 0;
333 }
334 
335 /*
336  * rt_rq:
337  *
338  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
339  *   util_sum = cpu_scale * load_sum
340  *   runnable_sum = util_sum
341  *
342  *   load_avg and runnable_avg are not supported and meaningless.
343  *
344  */
345 
update_rt_rq_load_avg(u64 now,struct rq * rq,int running)346 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
347 {
348 	if (___update_load_sum(now, &rq->avg_rt,
349 				running,
350 				running,
351 				running)) {
352 
353 		___update_load_avg(&rq->avg_rt, 1);
354 		trace_pelt_rt_tp(rq);
355 		return 1;
356 	}
357 
358 	return 0;
359 }
360 
361 /*
362  * dl_rq:
363  *
364  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
365  *   util_sum = cpu_scale * load_sum
366  *   runnable_sum = util_sum
367  *
368  *   load_avg and runnable_avg are not supported and meaningless.
369  *
370  */
371 
update_dl_rq_load_avg(u64 now,struct rq * rq,int running)372 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
373 {
374 	if (___update_load_sum(now, &rq->avg_dl,
375 				running,
376 				running,
377 				running)) {
378 
379 		___update_load_avg(&rq->avg_dl, 1);
380 		trace_pelt_dl_tp(rq);
381 		return 1;
382 	}
383 
384 	return 0;
385 }
386 
387 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
388 /*
389  * thermal:
390  *
391  *   load_sum = \Sum se->avg.load_sum but se->avg.load_sum is not tracked
392  *
393  *   util_avg and runnable_load_avg are not supported and meaningless.
394  *
395  * Unlike rt/dl utilization tracking that track time spent by a cpu
396  * running a rt/dl task through util_avg, the average thermal pressure is
397  * tracked through load_avg. This is because thermal pressure signal is
398  * time weighted "delta" capacity unlike util_avg which is binary.
399  * "delta capacity" =  actual capacity  -
400  *			capped capacity a cpu due to a thermal event.
401  */
402 
update_thermal_load_avg(u64 now,struct rq * rq,u64 capacity)403 int update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
404 {
405 	if (___update_load_sum(now, &rq->avg_thermal,
406 			       capacity,
407 			       capacity,
408 			       capacity)) {
409 		___update_load_avg(&rq->avg_thermal, 1);
410 		trace_pelt_thermal_tp(rq);
411 		return 1;
412 	}
413 
414 	return 0;
415 }
416 #endif
417 
418 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
419 /*
420  * irq:
421  *
422  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
423  *   util_sum = cpu_scale * load_sum
424  *   runnable_sum = util_sum
425  *
426  *   load_avg and runnable_avg are not supported and meaningless.
427  *
428  */
429 
update_irq_load_avg(struct rq * rq,u64 running)430 int update_irq_load_avg(struct rq *rq, u64 running)
431 {
432 	int ret = 0;
433 
434 	/*
435 	 * We can't use clock_pelt because irq time is not accounted in
436 	 * clock_task. Instead we directly scale the running time to
437 	 * reflect the real amount of computation
438 	 */
439 	running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
440 	running = cap_scale(running, arch_scale_cpu_capacity(cpu_of(rq)));
441 
442 	/*
443 	 * We know the time that has been used by interrupt since last update
444 	 * but we don't when. Let be pessimistic and assume that interrupt has
445 	 * happened just before the update. This is not so far from reality
446 	 * because interrupt will most probably wake up task and trig an update
447 	 * of rq clock during which the metric is updated.
448 	 * We start to decay with normal context time and then we add the
449 	 * interrupt context time.
450 	 * We can safely remove running from rq->clock because
451 	 * rq->clock += delta with delta >= running
452 	 */
453 	ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
454 				0,
455 				0,
456 				0);
457 	ret += ___update_load_sum(rq->clock, &rq->avg_irq,
458 				1,
459 				1,
460 				1);
461 
462 	if (ret) {
463 		___update_load_avg(&rq->avg_irq, 1);
464 		trace_pelt_irq_tp(rq);
465 	}
466 
467 	return ret;
468 }
469 #endif
470