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