xref: /openbmc/linux/kernel/sched/pelt.c (revision aad7ebb5)
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 #include <trace/events/sched.h>
32 
33 /*
34  * Approximate:
35  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
36  */
37 static u64 decay_load(u64 val, u64 n)
38 {
39 	unsigned int local_n;
40 
41 	if (unlikely(n > LOAD_AVG_PERIOD * 63))
42 		return 0;
43 
44 	/* after bounds checking we can collapse to 32-bit */
45 	local_n = n;
46 
47 	/*
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)
51 	 *
52 	 * To achieve constant time decay_load.
53 	 */
54 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
55 		val >>= local_n / LOAD_AVG_PERIOD;
56 		local_n %= LOAD_AVG_PERIOD;
57 	}
58 
59 	val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
60 	return val;
61 }
62 
63 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
64 {
65 	u32 c1, c2, c3 = d3; /* y^0 == 1 */
66 
67 	/*
68 	 * c1 = d1 y^p
69 	 */
70 	c1 = decay_load((u64)d1, periods);
71 
72 	/*
73 	 *            p-1
74 	 * c2 = 1024 \Sum y^n
75 	 *            n=1
76 	 *
77 	 *              inf        inf
78 	 *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
79 	 *              n=0        n=p
80 	 */
81 	c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
82 
83 	return c1 + c2 + c3;
84 }
85 
86 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
87 
88 /*
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.
92  *
93  *           d1          d2           d3
94  *           ^           ^            ^
95  *           |           |            |
96  *         |<->|<----------------->|<--->|
97  * ... |---x---|------| ... |------|-----x (now)
98  *
99  *                           p-1
100  * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
101  *                           n=1
102  *
103  *    = u y^p +					(Step 1)
104  *
105  *                     p-1
106  *      d1 y^p + 1024 \Sum y^n + d3 y^0		(Step 2)
107  *                     n=1
108  */
109 static __always_inline u32
110 accumulate_sum(u64 delta, struct sched_avg *sa,
111 	       unsigned long load, unsigned long runnable, int running)
112 {
113 	u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
114 	u64 periods;
115 
116 	delta += sa->period_contrib;
117 	periods = delta / 1024; /* A period is 1024us (~1ms) */
118 
119 	/*
120 	 * Step 1: decay old *_sum if we crossed period boundaries.
121 	 */
122 	if (periods) {
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);
127 
128 		/*
129 		 * Step 2
130 		 */
131 		delta %= 1024;
132 		contrib = __accumulate_pelt_segments(periods,
133 				1024 - sa->period_contrib, delta);
134 	}
135 	sa->period_contrib = delta;
136 
137 	if (load)
138 		sa->load_sum += load * contrib;
139 	if (runnable)
140 		sa->runnable_load_sum += runnable * contrib;
141 	if (running)
142 		sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
143 
144 	return periods;
145 }
146 
147 /*
148  * We can represent the historical contribution to runnable average as the
149  * coefficients of a geometric series.  To do this we sub-divide our runnable
150  * history into segments of approximately 1ms (1024us); label the segment that
151  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
152  *
153  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
154  *      p0            p1           p2
155  *     (now)       (~1ms ago)  (~2ms ago)
156  *
157  * Let u_i denote the fraction of p_i that the entity was runnable.
158  *
159  * We then designate the fractions u_i as our co-efficients, yielding the
160  * following representation of historical load:
161  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
162  *
163  * We choose y based on the with of a reasonably scheduling period, fixing:
164  *   y^32 = 0.5
165  *
166  * This means that the contribution to load ~32ms ago (u_32) will be weighted
167  * approximately half as much as the contribution to load within the last ms
168  * (u_0).
169  *
170  * When a period "rolls over" and we have new u_0`, multiplying the previous
171  * sum again by y is sufficient to update:
172  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
173  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
174  */
175 static __always_inline int
176 ___update_load_sum(u64 now, struct sched_avg *sa,
177 		  unsigned long load, unsigned long runnable, int running)
178 {
179 	u64 delta;
180 
181 	delta = now - sa->last_update_time;
182 	/*
183 	 * This should only happen when time goes backwards, which it
184 	 * unfortunately does during sched clock init when we swap over to TSC.
185 	 */
186 	if ((s64)delta < 0) {
187 		sa->last_update_time = now;
188 		return 0;
189 	}
190 
191 	/*
192 	 * Use 1024ns as the unit of measurement since it's a reasonable
193 	 * approximation of 1us and fast to compute.
194 	 */
195 	delta >>= 10;
196 	if (!delta)
197 		return 0;
198 
199 	sa->last_update_time += delta << 10;
200 
201 	/*
202 	 * running is a subset of runnable (weight) so running can't be set if
203 	 * runnable is clear. But there are some corner cases where the current
204 	 * se has been already dequeued but cfs_rq->curr still points to it.
205 	 * This means that weight will be 0 but not running for a sched_entity
206 	 * but also for a cfs_rq if the latter becomes idle. As an example,
207 	 * this happens during idle_balance() which calls
208 	 * update_blocked_averages()
209 	 */
210 	if (!load)
211 		runnable = running = 0;
212 
213 	/*
214 	 * Now we know we crossed measurement unit boundaries. The *_avg
215 	 * accrues by two steps:
216 	 *
217 	 * Step 1: accumulate *_sum since last_update_time. If we haven't
218 	 * crossed period boundaries, finish.
219 	 */
220 	if (!accumulate_sum(delta, sa, load, runnable, running))
221 		return 0;
222 
223 	return 1;
224 }
225 
226 static __always_inline void
227 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
228 {
229 	u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
230 
231 	/*
232 	 * Step 2: update *_avg.
233 	 */
234 	sa->load_avg = div_u64(load * sa->load_sum, divider);
235 	sa->runnable_load_avg =	div_u64(runnable * sa->runnable_load_sum, divider);
236 	WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
237 }
238 
239 /*
240  * sched_entity:
241  *
242  *   task:
243  *     se_runnable() == se_weight()
244  *
245  *   group: [ see update_cfs_group() ]
246  *     se_weight()   = tg->weight * grq->load_avg / tg->load_avg
247  *     se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
248  *
249  *   load_sum := runnable_sum
250  *   load_avg = se_weight(se) * runnable_avg
251  *
252  *   runnable_load_sum := runnable_sum
253  *   runnable_load_avg = se_runnable(se) * runnable_avg
254  *
255  * XXX collapse load_sum and runnable_load_sum
256  *
257  * cfq_rq:
258  *
259  *   load_sum = \Sum se_weight(se) * se->avg.load_sum
260  *   load_avg = \Sum se->avg.load_avg
261  *
262  *   runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
263  *   runnable_load_avg = \Sum se->avg.runable_load_avg
264  */
265 
266 int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
267 {
268 	if (___update_load_sum(now, &se->avg, 0, 0, 0)) {
269 		___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
270 		trace_pelt_se_tp(se);
271 		return 1;
272 	}
273 
274 	return 0;
275 }
276 
277 int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
278 {
279 	if (___update_load_sum(now, &se->avg, !!se->on_rq, !!se->on_rq,
280 				cfs_rq->curr == se)) {
281 
282 		___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
283 		cfs_se_util_change(&se->avg);
284 		trace_pelt_se_tp(se);
285 		return 1;
286 	}
287 
288 	return 0;
289 }
290 
291 int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
292 {
293 	if (___update_load_sum(now, &cfs_rq->avg,
294 				scale_load_down(cfs_rq->load.weight),
295 				scale_load_down(cfs_rq->runnable_weight),
296 				cfs_rq->curr != NULL)) {
297 
298 		___update_load_avg(&cfs_rq->avg, 1, 1);
299 		trace_pelt_cfs_tp(cfs_rq);
300 		return 1;
301 	}
302 
303 	return 0;
304 }
305 
306 /*
307  * rt_rq:
308  *
309  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
310  *   util_sum = cpu_scale * load_sum
311  *   runnable_load_sum = load_sum
312  *
313  *   load_avg and runnable_load_avg are not supported and meaningless.
314  *
315  */
316 
317 int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
318 {
319 	if (___update_load_sum(now, &rq->avg_rt,
320 				running,
321 				running,
322 				running)) {
323 
324 		___update_load_avg(&rq->avg_rt, 1, 1);
325 		trace_pelt_rt_tp(rq);
326 		return 1;
327 	}
328 
329 	return 0;
330 }
331 
332 /*
333  * dl_rq:
334  *
335  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
336  *   util_sum = cpu_scale * load_sum
337  *   runnable_load_sum = load_sum
338  *
339  */
340 
341 int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
342 {
343 	if (___update_load_sum(now, &rq->avg_dl,
344 				running,
345 				running,
346 				running)) {
347 
348 		___update_load_avg(&rq->avg_dl, 1, 1);
349 		trace_pelt_dl_tp(rq);
350 		return 1;
351 	}
352 
353 	return 0;
354 }
355 
356 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
357 /*
358  * irq:
359  *
360  *   util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
361  *   util_sum = cpu_scale * load_sum
362  *   runnable_load_sum = load_sum
363  *
364  */
365 
366 int update_irq_load_avg(struct rq *rq, u64 running)
367 {
368 	int ret = 0;
369 
370 	/*
371 	 * We can't use clock_pelt because irq time is not accounted in
372 	 * clock_task. Instead we directly scale the running time to
373 	 * reflect the real amount of computation
374 	 */
375 	running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
376 	running = cap_scale(running, arch_scale_cpu_capacity(cpu_of(rq)));
377 
378 	/*
379 	 * We know the time that has been used by interrupt since last update
380 	 * but we don't when. Let be pessimistic and assume that interrupt has
381 	 * happened just before the update. This is not so far from reality
382 	 * because interrupt will most probably wake up task and trig an update
383 	 * of rq clock during which the metric is updated.
384 	 * We start to decay with normal context time and then we add the
385 	 * interrupt context time.
386 	 * We can safely remove running from rq->clock because
387 	 * rq->clock += delta with delta >= running
388 	 */
389 	ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
390 				0,
391 				0,
392 				0);
393 	ret += ___update_load_sum(rq->clock, &rq->avg_irq,
394 				1,
395 				1,
396 				1);
397 
398 	if (ret) {
399 		___update_load_avg(&rq->avg_irq, 1, 1);
400 		trace_pelt_irq_tp(rq);
401 	}
402 
403 	return ret;
404 }
405 #endif
406