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 */
35static 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
61static 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#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
85
86/*
87 * Accumulate the three separate parts of the sum; d1 the remainder
88 * of the last (incomplete) period, d2 the span of full periods and d3
89 * the remainder of the (incomplete) current period.
90 *
91 * d1 d2 d3
92 * ^ ^ ^
93 * | | |
94 * |<->|<----------------->|<--->|
95 * ... |---x---|------| ... |------|-----x (now)
96 *
97 * p-1
98 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
99 * n=1
100 *
101 * = u y^p + (Step 1)
102 *
103 * p-1
104 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
105 * n=1
106 */
107static __always_inline u32
108accumulate_sum(u64 delta, struct sched_avg *sa,
109 unsigned long load, unsigned long runnable, int running)
110{
111 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
112 u64 periods;
113
114 delta += sa->period_contrib;
115 periods = delta / 1024; /* A period is 1024us (~1ms) */
116
117 /*
118 * Step 1: decay old *_sum if we crossed period boundaries.
119 */
120 if (periods) {
121 sa->load_sum = decay_load(sa->load_sum, periods);
122 sa->runnable_load_sum =
123 decay_load(sa->runnable_load_sum, periods);
124 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
125
126 /*
127 * Step 2
128 */
129 delta %= 1024;
130 contrib = __accumulate_pelt_segments(periods,
131 1024 - sa->period_contrib, delta);
132 }
133 sa->period_contrib = delta;
134
135 if (load)
136 sa->load_sum += load * contrib;
137 if (runnable)
138 sa->runnable_load_sum += runnable * contrib;
139 if (running)
140 sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
141
142 return periods;
143}
144
145/*
146 * We can represent the historical contribution to runnable average as the
147 * coefficients of a geometric series. To do this we sub-divide our runnable
148 * history into segments of approximately 1ms (1024us); label the segment that
149 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
150 *
151 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
152 * p0 p1 p2
153 * (now) (~1ms ago) (~2ms ago)
154 *
155 * Let u_i denote the fraction of p_i that the entity was runnable.
156 *
157 * We then designate the fractions u_i as our co-efficients, yielding the
158 * following representation of historical load:
159 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
160 *
161 * We choose y based on the with of a reasonably scheduling period, fixing:
162 * y^32 = 0.5
163 *
164 * This means that the contribution to load ~32ms ago (u_32) will be weighted
165 * approximately half as much as the contribution to load within the last ms
166 * (u_0).
167 *
168 * When a period "rolls over" and we have new u_0`, multiplying the previous
169 * sum again by y is sufficient to update:
170 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
171 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
172 */
173static __always_inline int
174___update_load_sum(u64 now, struct sched_avg *sa,
175 unsigned long load, unsigned long runnable, int running)
176{
177 u64 delta;
178
179 delta = now - sa->last_update_time;
180 /*
181 * This should only happen when time goes backwards, which it
182 * unfortunately does during sched clock init when we swap over to TSC.
183 */
184 if ((s64)delta < 0) {
185 sa->last_update_time = now;
186 return 0;
187 }
188
189 /*
190 * Use 1024ns as the unit of measurement since it's a reasonable
191 * approximation of 1us and fast to compute.
192 */
193 delta >>= 10;
194 if (!delta)
195 return 0;
196
197 sa->last_update_time += delta << 10;
198
199 /*
200 * running is a subset of runnable (weight) so running can't be set if
201 * runnable is clear. But there are some corner cases where the current
202 * se has been already dequeued but cfs_rq->curr still points to it.
203 * This means that weight will be 0 but not running for a sched_entity
204 * but also for a cfs_rq if the latter becomes idle. As an example,
205 * this happens during idle_balance() which calls
206 * update_blocked_averages()
207 */
208 if (!load)
209 runnable = running = 0;
210
211 /*
212 * Now we know we crossed measurement unit boundaries. The *_avg
213 * accrues by two steps:
214 *
215 * Step 1: accumulate *_sum since last_update_time. If we haven't
216 * crossed period boundaries, finish.
217 */
218 if (!accumulate_sum(delta, sa, load, runnable, running))
219 return 0;
220
221 return 1;
222}
223
224static __always_inline void
225___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
226{
227 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
228
229 /*
230 * Step 2: update *_avg.
231 */
232 sa->load_avg = div_u64(load * sa->load_sum, divider);
233 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
234 WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
235}
236
237/*
238 * sched_entity:
239 *
240 * task:
241 * se_runnable() == se_weight()
242 *
243 * group: [ see update_cfs_group() ]
244 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
245 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
246 *
247 * load_sum := runnable_sum
248 * load_avg = se_weight(se) * runnable_avg
249 *
250 * runnable_load_sum := runnable_sum
251 * runnable_load_avg = se_runnable(se) * runnable_avg
252 *
253 * XXX collapse load_sum and runnable_load_sum
254 *
255 * cfq_rq:
256 *
257 * load_sum = \Sum se_weight(se) * se->avg.load_sum
258 * load_avg = \Sum se->avg.load_avg
259 *
260 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
261 * runnable_load_avg = \Sum se->avg.runable_load_avg
262 */
263
264int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
265{
266 if (___update_load_sum(now, &se->avg, 0, 0, 0)) {
267 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
268 return 1;
269 }
270
271 return 0;
272}
273
274int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
275{
276 if (___update_load_sum(now, &se->avg, !!se->on_rq, !!se->on_rq,
277 cfs_rq->curr == se)) {
278
279 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
280 cfs_se_util_change(&se->avg);
281 return 1;
282 }
283
284 return 0;
285}
286
287int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
288{
289 if (___update_load_sum(now, &cfs_rq->avg,
290 scale_load_down(cfs_rq->load.weight),
291 scale_load_down(cfs_rq->runnable_weight),
292 cfs_rq->curr != NULL)) {
293
294 ___update_load_avg(&cfs_rq->avg, 1, 1);
295 return 1;
296 }
297
298 return 0;
299}
300
301/*
302 * rt_rq:
303 *
304 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
305 * util_sum = cpu_scale * load_sum
306 * runnable_load_sum = load_sum
307 *
308 * load_avg and runnable_load_avg are not supported and meaningless.
309 *
310 */
311
312int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
313{
314 if (___update_load_sum(now, &rq->avg_rt,
315 running,
316 running,
317 running)) {
318
319 ___update_load_avg(&rq->avg_rt, 1, 1);
320 return 1;
321 }
322
323 return 0;
324}
325
326/*
327 * dl_rq:
328 *
329 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
330 * util_sum = cpu_scale * load_sum
331 * runnable_load_sum = load_sum
332 *
333 */
334
335int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
336{
337 if (___update_load_sum(now, &rq->avg_dl,
338 running,
339 running,
340 running)) {
341
342 ___update_load_avg(&rq->avg_dl, 1, 1);
343 return 1;
344 }
345
346 return 0;
347}
348
349#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
350/*
351 * irq:
352 *
353 * util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
354 * util_sum = cpu_scale * load_sum
355 * runnable_load_sum = load_sum
356 *
357 */
358
359int update_irq_load_avg(struct rq *rq, u64 running)
360{
361 int ret = 0;
362
363 /*
364 * We can't use clock_pelt because irq time is not accounted in
365 * clock_task. Instead we directly scale the running time to
366 * reflect the real amount of computation
367 */
368 running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
369 running = cap_scale(running, arch_scale_cpu_capacity(NULL, cpu_of(rq)));
370
371 /*
372 * We know the time that has been used by interrupt since last update
373 * but we don't when. Let be pessimistic and assume that interrupt has
374 * happened just before the update. This is not so far from reality
375 * because interrupt will most probably wake up task and trig an update
376 * of rq clock during which the metric is updated.
377 * We start to decay with normal context time and then we add the
378 * interrupt context time.
379 * We can safely remove running from rq->clock because
380 * rq->clock += delta with delta >= running
381 */
382 ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
383 0,
384 0,
385 0);
386 ret += ___update_load_sum(rq->clock, &rq->avg_irq,
387 1,
388 1,
389 1);
390
391 if (ret)
392 ___update_load_avg(&rq->avg_irq, 1, 1);
393
394 return ret;
395}
396#endif
397