1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Per Entity Load Tracking (PELT)
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 */
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
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
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
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 sched_balance_newidle() which calls
212 * sched_balance_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
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
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
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
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
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
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_HW_PRESSURE
388 /*
389 * hardware:
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 HW pressure is
397 * tracked through load_avg. This is because HW 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 HW event.
401 */
402
403 int update_hw_load_avg(u64 now, struct rq *rq, u64 capacity)
404 {
405 if (___update_load_sum(now, &rq->avg_hw,
406 capacity,
407 capacity,
408 capacity)) {
409 ___update_load_avg(&rq->avg_hw, 1);
410 trace_pelt_hw_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
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
471 /*
472 * Load avg and utiliztion metrics need to be updated periodically and before
473 * consumption. This function updates the metrics for all subsystems except for
474 * the fair class. @rq must be locked and have its clock updated.
475 */
476 bool update_other_load_avgs(struct rq *rq)
477 {
478 u64 now = rq_clock_pelt(rq);
479 const struct sched_class *curr_class = rq->curr->sched_class;
480 unsigned long hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
481
482 lockdep_assert_rq_held(rq);
483
484 /* hw_pressure doesn't care about invariance */
485 return update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
486 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
487 update_hw_load_avg(rq_clock_task(rq), rq, hw_pressure) |
488 update_irq_load_avg(rq, 0);
489 }
490
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.