xref: /openbmc/linux/kernel/sched/pelt.c (revision e7bae9bb)
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 #include <linux/sched.h>
28 #include "sched.h"
29 #include "pelt.h"
30 
31 /*
32  * Approximate:
33  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
34  */
35 static u64 decay_load(u64 val, u64 n)
36 {
37 	unsigned int local_n;
38 
39 	if (unlikely(n > LOAD_AVG_PERIOD * 63))
40 		return 0;
41 
42 	/* after bounds checking we can collapse to 32-bit */
43 	local_n = n;
44 
45 	/*
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)
49 	 *
50 	 * To achieve constant time decay_load.
51 	 */
52 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
53 		val >>= local_n / LOAD_AVG_PERIOD;
54 		local_n %= LOAD_AVG_PERIOD;
55 	}
56 
57 	val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
58 	return val;
59 }
60 
61 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
62 {
63 	u32 c1, c2, c3 = d3; /* y^0 == 1 */
64 
65 	/*
66 	 * c1 = d1 y^p
67 	 */
68 	c1 = decay_load((u64)d1, periods);
69 
70 	/*
71 	 *            p-1
72 	 * c2 = 1024 \Sum y^n
73 	 *            n=1
74 	 *
75 	 *              inf        inf
76 	 *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
77 	 *              n=0        n=p
78 	 */
79 	c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
80 
81 	return c1 + c2 + c3;
82 }
83 
84 /*
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.
88  *
89  *           d1          d2           d3
90  *           ^           ^            ^
91  *           |           |            |
92  *         |<->|<----------------->|<--->|
93  * ... |---x---|------| ... |------|-----x (now)
94  *
95  *                           p-1
96  * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
97  *                           n=1
98  *
99  *    = u y^p +					(Step 1)
100  *
101  *                     p-1
102  *      d1 y^p + 1024 \Sum y^n + d3 y^0		(Step 2)
103  *                     n=1
104  */
105 static __always_inline u32
106 accumulate_sum(u64 delta, struct sched_avg *sa,
107 	       unsigned long load, unsigned long runnable, int running)
108 {
109 	u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
110 	u64 periods;
111 
112 	delta += sa->period_contrib;
113 	periods = delta / 1024; /* A period is 1024us (~1ms) */
114 
115 	/*
116 	 * Step 1: decay old *_sum if we crossed period boundaries.
117 	 */
118 	if (periods) {
119 		sa->load_sum = decay_load(sa->load_sum, periods);
120 		sa->runnable_sum =
121 			decay_load(sa->runnable_sum, periods);
122 		sa->util_sum = decay_load((u64)(sa->util_sum), periods);
123 
124 		/*
125 		 * Step 2
126 		 */
127 		delta %= 1024;
128 		if (load) {
129 			/*
130 			 * This relies on the:
131 			 *
132 			 * if (!load)
133 			 *	runnable = running = 0;
134 			 *
135 			 * clause from ___update_load_sum(); this results in
136 			 * the below usage of @contrib to dissapear entirely,
137 			 * so no point in calculating it.
138 			 */
139 			contrib = __accumulate_pelt_segments(periods,
140 					1024 - sa->period_contrib, delta);
141 		}
142 	}
143 	sa->period_contrib = delta;
144 
145 	if (load)
146 		sa->load_sum += load * contrib;
147 	if (runnable)
148 		sa->runnable_sum += runnable * contrib << SCHED_CAPACITY_SHIFT;
149 	if (running)
150 		sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
151 
152 	return periods;
153 }
154 
155 /*
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.
160  *
161  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
162  *      p0            p1           p2
163  *     (now)       (~1ms ago)  (~2ms ago)
164  *
165  * Let u_i denote the fraction of p_i that the entity was runnable.
166  *
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 + ...
170  *
171  * We choose y based on the with of a reasonably scheduling period, fixing:
172  *   y^32 = 0.5
173  *
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
176  * (u_0).
177  *
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}]
182  */
183 static __always_inline int
184 ___update_load_sum(u64 now, struct sched_avg *sa,
185 		  unsigned long load, unsigned long runnable, int running)
186 {
187 	u64 delta;
188 
189 	delta = now - sa->last_update_time;
190 	/*
191 	 * This should only happen when time goes backwards, which it
192 	 * unfortunately does during sched clock init when we swap over to TSC.
193 	 */
194 	if ((s64)delta < 0) {
195 		sa->last_update_time = now;
196 		return 0;
197 	}
198 
199 	/*
200 	 * Use 1024ns as the unit of measurement since it's a reasonable
201 	 * approximation of 1us and fast to compute.
202 	 */
203 	delta >>= 10;
204 	if (!delta)
205 		return 0;
206 
207 	sa->last_update_time += delta << 10;
208 
209 	/*
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().
217 	 *
218 	 * Also see the comment in accumulate_sum().
219 	 */
220 	if (!load)
221 		runnable = running = 0;
222 
223 	/*
224 	 * Now we know we crossed measurement unit boundaries. The *_avg
225 	 * accrues by two steps:
226 	 *
227 	 * Step 1: accumulate *_sum since last_update_time. If we haven't
228 	 * crossed period boundaries, finish.
229 	 */
230 	if (!accumulate_sum(delta, sa, load, runnable, running))
231 		return 0;
232 
233 	return 1;
234 }
235 
236 /*
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[.
241  *
242  * The max value of *_sum varies with the position in the time segment and is
243  * equals to :
244  *
245  *   LOAD_AVG_MAX*y + sa->period_contrib
246  *
247  * which can be simplified into:
248  *
249  *   LOAD_AVG_MAX - 1024 + sa->period_contrib
250  *
251  * because LOAD_AVG_MAX*y == LOAD_AVG_MAX-1024
252  *
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.
259  */
260 static __always_inline void
261 ___update_load_avg(struct sched_avg *sa, unsigned long load)
262 {
263 	u32 divider = get_pelt_divider(sa);
264 
265 	/*
266 	 * Step 2: update *_avg.
267 	 */
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);
271 }
272 
273 /*
274  * sched_entity:
275  *
276  *   task:
277  *     se_weight()   = se->load.weight
278  *     se_runnable() = !!on_rq
279  *
280  *   group: [ see update_cfs_group() ]
281  *     se_weight()   = tg->weight * grq->load_avg / tg->load_avg
282  *     se_runnable() = grq->h_nr_running
283  *
284  *   runnable_sum = se_runnable() * runnable = grq->runnable_sum
285  *   runnable_avg = runnable_sum
286  *
287  *   load_sum := runnable
288  *   load_avg = se_weight(se) * load_sum
289  *
290  * cfq_rq:
291  *
292  *   runnable_sum = \Sum se->avg.runnable_sum
293  *   runnable_avg = \Sum se->avg.runnable_avg
294  *
295  *   load_sum = \Sum se_weight(se) * se->avg.load_sum
296  *   load_avg = \Sum se->avg.load_avg
297  */
298 
299 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
300 {
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);
304 		return 1;
305 	}
306 
307 	return 0;
308 }
309 
310 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
311 {
312 	if (___update_load_sum(now, &se->avg, !!se->on_rq, se_runnable(se),
313 				cfs_rq->curr == se)) {
314 
315 		___update_load_avg(&se->avg, se_weight(se));
316 		cfs_se_util_change(&se->avg);
317 		trace_pelt_se_tp(se);
318 		return 1;
319 	}
320 
321 	return 0;
322 }
323 
324 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
325 {
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)) {
330 
331 		___update_load_avg(&cfs_rq->avg, 1);
332 		trace_pelt_cfs_tp(cfs_rq);
333 		return 1;
334 	}
335 
336 	return 0;
337 }
338 
339 /*
340  * rt_rq:
341  *
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
345  *
346  *   load_avg and runnable_avg are not supported and meaningless.
347  *
348  */
349 
350 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
351 {
352 	if (___update_load_sum(now, &rq->avg_rt,
353 				running,
354 				running,
355 				running)) {
356 
357 		___update_load_avg(&rq->avg_rt, 1);
358 		trace_pelt_rt_tp(rq);
359 		return 1;
360 	}
361 
362 	return 0;
363 }
364 
365 /*
366  * dl_rq:
367  *
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
371  *
372  *   load_avg and runnable_avg are not supported and meaningless.
373  *
374  */
375 
376 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
377 {
378 	if (___update_load_sum(now, &rq->avg_dl,
379 				running,
380 				running,
381 				running)) {
382 
383 		___update_load_avg(&rq->avg_dl, 1);
384 		trace_pelt_dl_tp(rq);
385 		return 1;
386 	}
387 
388 	return 0;
389 }
390 
391 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
392 /*
393  * thermal:
394  *
395  *   load_sum = \Sum se->avg.load_sum but se->avg.load_sum is not tracked
396  *
397  *   util_avg and runnable_load_avg are not supported and meaningless.
398  *
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.
405  */
406 
407 int update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
408 {
409 	if (___update_load_sum(now, &rq->avg_thermal,
410 			       capacity,
411 			       capacity,
412 			       capacity)) {
413 		___update_load_avg(&rq->avg_thermal, 1);
414 		trace_pelt_thermal_tp(rq);
415 		return 1;
416 	}
417 
418 	return 0;
419 }
420 #endif
421 
422 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
423 /*
424  * irq:
425  *
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
429  *
430  *   load_avg and runnable_avg are not supported and meaningless.
431  *
432  */
433 
434 int update_irq_load_avg(struct rq *rq, u64 running)
435 {
436 	int ret = 0;
437 
438 	/*
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
442 	 */
443 	running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
444 	running = cap_scale(running, arch_scale_cpu_capacity(cpu_of(rq)));
445 
446 	/*
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
456 	 */
457 	ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
458 				0,
459 				0,
460 				0);
461 	ret += ___update_load_sum(rq->clock, &rq->avg_irq,
462 				1,
463 				1,
464 				1);
465 
466 	if (ret) {
467 		___update_load_avg(&rq->avg_irq, 1);
468 		trace_pelt_irq_tp(rq);
469 	}
470 
471 	return ret;
472 }
473 #endif
474