xref: /linux/kernel/sched/ext_idle.c (revision e9ef810dfee7a2227da9d423aecb0ced35faddbe)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
4  *
5  * Built-in idle CPU tracking policy.
6  *
7  * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
8  * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
9  * Copyright (c) 2022 David Vernet <dvernet@meta.com>
10  * Copyright (c) 2024 Andrea Righi <arighi@nvidia.com>
11  */
12 #include "ext_idle.h"
13 
14 /* Enable/disable built-in idle CPU selection policy */
15 static DEFINE_STATIC_KEY_FALSE(scx_builtin_idle_enabled);
16 
17 /* Enable/disable per-node idle cpumasks */
18 static DEFINE_STATIC_KEY_FALSE(scx_builtin_idle_per_node);
19 
20 /* Enable/disable LLC aware optimizations */
21 static DEFINE_STATIC_KEY_FALSE(scx_selcpu_topo_llc);
22 
23 /* Enable/disable NUMA aware optimizations */
24 static DEFINE_STATIC_KEY_FALSE(scx_selcpu_topo_numa);
25 
26 /*
27  * cpumasks to track idle CPUs within each NUMA node.
28  *
29  * If SCX_OPS_BUILTIN_IDLE_PER_NODE is not enabled, a single global cpumask
30  * from is used to track all the idle CPUs in the system.
31  */
32 struct scx_idle_cpus {
33 	cpumask_var_t cpu;
34 	cpumask_var_t smt;
35 };
36 
37 /*
38  * Global host-wide idle cpumasks (used when SCX_OPS_BUILTIN_IDLE_PER_NODE
39  * is not enabled).
40  */
41 static struct scx_idle_cpus scx_idle_global_masks;
42 
43 /*
44  * Per-node idle cpumasks.
45  */
46 static struct scx_idle_cpus **scx_idle_node_masks;
47 
48 /*
49  * Local per-CPU cpumasks (used to generate temporary idle cpumasks).
50  */
51 static DEFINE_PER_CPU(cpumask_var_t, local_idle_cpumask);
52 static DEFINE_PER_CPU(cpumask_var_t, local_llc_idle_cpumask);
53 static DEFINE_PER_CPU(cpumask_var_t, local_numa_idle_cpumask);
54 
55 /*
56  * Return the idle masks associated to a target @node.
57  *
58  * NUMA_NO_NODE identifies the global idle cpumask.
59  */
idle_cpumask(int node)60 static struct scx_idle_cpus *idle_cpumask(int node)
61 {
62 	return node == NUMA_NO_NODE ? &scx_idle_global_masks : scx_idle_node_masks[node];
63 }
64 
65 /*
66  * Returns the NUMA node ID associated with a @cpu, or NUMA_NO_NODE if
67  * per-node idle cpumasks are disabled.
68  */
scx_cpu_node_if_enabled(int cpu)69 static int scx_cpu_node_if_enabled(int cpu)
70 {
71 	if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
72 		return NUMA_NO_NODE;
73 
74 	return cpu_to_node(cpu);
75 }
76 
scx_idle_test_and_clear_cpu(int cpu)77 static bool scx_idle_test_and_clear_cpu(int cpu)
78 {
79 	int node = scx_cpu_node_if_enabled(cpu);
80 	struct cpumask *idle_cpus = idle_cpumask(node)->cpu;
81 
82 #ifdef CONFIG_SCHED_SMT
83 	/*
84 	 * SMT mask should be cleared whether we can claim @cpu or not. The SMT
85 	 * cluster is not wholly idle either way. This also prevents
86 	 * scx_pick_idle_cpu() from getting caught in an infinite loop.
87 	 */
88 	if (sched_smt_active()) {
89 		const struct cpumask *smt = cpu_smt_mask(cpu);
90 		struct cpumask *idle_smts = idle_cpumask(node)->smt;
91 
92 		/*
93 		 * If offline, @cpu is not its own sibling and
94 		 * scx_pick_idle_cpu() can get caught in an infinite loop as
95 		 * @cpu is never cleared from the idle SMT mask. Ensure that
96 		 * @cpu is eventually cleared.
97 		 *
98 		 * NOTE: Use cpumask_intersects() and cpumask_test_cpu() to
99 		 * reduce memory writes, which may help alleviate cache
100 		 * coherence pressure.
101 		 */
102 		if (cpumask_intersects(smt, idle_smts))
103 			cpumask_andnot(idle_smts, idle_smts, smt);
104 		else if (cpumask_test_cpu(cpu, idle_smts))
105 			__cpumask_clear_cpu(cpu, idle_smts);
106 	}
107 #endif
108 
109 	return cpumask_test_and_clear_cpu(cpu, idle_cpus);
110 }
111 
112 /*
113  * Pick an idle CPU in a specific NUMA node.
114  */
pick_idle_cpu_in_node(const struct cpumask * cpus_allowed,int node,u64 flags)115 static s32 pick_idle_cpu_in_node(const struct cpumask *cpus_allowed, int node, u64 flags)
116 {
117 	int cpu;
118 
119 retry:
120 	if (sched_smt_active()) {
121 		cpu = cpumask_any_and_distribute(idle_cpumask(node)->smt, cpus_allowed);
122 		if (cpu < nr_cpu_ids)
123 			goto found;
124 
125 		if (flags & SCX_PICK_IDLE_CORE)
126 			return -EBUSY;
127 	}
128 
129 	cpu = cpumask_any_and_distribute(idle_cpumask(node)->cpu, cpus_allowed);
130 	if (cpu >= nr_cpu_ids)
131 		return -EBUSY;
132 
133 found:
134 	if (scx_idle_test_and_clear_cpu(cpu))
135 		return cpu;
136 	else
137 		goto retry;
138 }
139 
140 #ifdef CONFIG_NUMA
141 /*
142  * Tracks nodes that have not yet been visited when searching for an idle
143  * CPU across all available nodes.
144  */
145 static DEFINE_PER_CPU(nodemask_t, per_cpu_unvisited);
146 
147 /*
148  * Search for an idle CPU across all nodes, excluding @node.
149  */
pick_idle_cpu_from_online_nodes(const struct cpumask * cpus_allowed,int node,u64 flags)150 static s32 pick_idle_cpu_from_online_nodes(const struct cpumask *cpus_allowed, int node, u64 flags)
151 {
152 	nodemask_t *unvisited;
153 	s32 cpu = -EBUSY;
154 
155 	preempt_disable();
156 	unvisited = this_cpu_ptr(&per_cpu_unvisited);
157 
158 	/*
159 	 * Restrict the search to the online nodes (excluding the current
160 	 * node that has been visited already).
161 	 */
162 	nodes_copy(*unvisited, node_states[N_ONLINE]);
163 	node_clear(node, *unvisited);
164 
165 	/*
166 	 * Traverse all nodes in order of increasing distance, starting
167 	 * from @node.
168 	 *
169 	 * This loop is O(N^2), with N being the amount of NUMA nodes,
170 	 * which might be quite expensive in large NUMA systems. However,
171 	 * this complexity comes into play only when a scheduler enables
172 	 * SCX_OPS_BUILTIN_IDLE_PER_NODE and it's requesting an idle CPU
173 	 * without specifying a target NUMA node, so it shouldn't be a
174 	 * bottleneck is most cases.
175 	 *
176 	 * As a future optimization we may want to cache the list of nodes
177 	 * in a per-node array, instead of actually traversing them every
178 	 * time.
179 	 */
180 	for_each_node_numadist(node, *unvisited) {
181 		cpu = pick_idle_cpu_in_node(cpus_allowed, node, flags);
182 		if (cpu >= 0)
183 			break;
184 	}
185 	preempt_enable();
186 
187 	return cpu;
188 }
189 #else
190 static inline s32
pick_idle_cpu_from_online_nodes(const struct cpumask * cpus_allowed,int node,u64 flags)191 pick_idle_cpu_from_online_nodes(const struct cpumask *cpus_allowed, int node, u64 flags)
192 {
193 	return -EBUSY;
194 }
195 #endif
196 
197 /*
198  * Find an idle CPU in the system, starting from @node.
199  */
scx_pick_idle_cpu(const struct cpumask * cpus_allowed,int node,u64 flags)200 static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, int node, u64 flags)
201 {
202 	s32 cpu;
203 
204 	/*
205 	 * Always search in the starting node first (this is an
206 	 * optimization that can save some cycles even when the search is
207 	 * not limited to a single node).
208 	 */
209 	cpu = pick_idle_cpu_in_node(cpus_allowed, node, flags);
210 	if (cpu >= 0)
211 		return cpu;
212 
213 	/*
214 	 * Stop the search if we are using only a single global cpumask
215 	 * (NUMA_NO_NODE) or if the search is restricted to the first node
216 	 * only.
217 	 */
218 	if (node == NUMA_NO_NODE || flags & SCX_PICK_IDLE_IN_NODE)
219 		return -EBUSY;
220 
221 	/*
222 	 * Extend the search to the other online nodes.
223 	 */
224 	return pick_idle_cpu_from_online_nodes(cpus_allowed, node, flags);
225 }
226 
227 /*
228  * Return the amount of CPUs in the same LLC domain of @cpu (or zero if the LLC
229  * domain is not defined).
230  */
llc_weight(s32 cpu)231 static unsigned int llc_weight(s32 cpu)
232 {
233 	struct sched_domain *sd;
234 
235 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
236 	if (!sd)
237 		return 0;
238 
239 	return sd->span_weight;
240 }
241 
242 /*
243  * Return the cpumask representing the LLC domain of @cpu (or NULL if the LLC
244  * domain is not defined).
245  */
llc_span(s32 cpu)246 static struct cpumask *llc_span(s32 cpu)
247 {
248 	struct sched_domain *sd;
249 
250 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
251 	if (!sd)
252 		return NULL;
253 
254 	return sched_domain_span(sd);
255 }
256 
257 /*
258  * Return the amount of CPUs in the same NUMA domain of @cpu (or zero if the
259  * NUMA domain is not defined).
260  */
numa_weight(s32 cpu)261 static unsigned int numa_weight(s32 cpu)
262 {
263 	struct sched_domain *sd;
264 	struct sched_group *sg;
265 
266 	sd = rcu_dereference(per_cpu(sd_numa, cpu));
267 	if (!sd)
268 		return 0;
269 	sg = sd->groups;
270 	if (!sg)
271 		return 0;
272 
273 	return sg->group_weight;
274 }
275 
276 /*
277  * Return the cpumask representing the NUMA domain of @cpu (or NULL if the NUMA
278  * domain is not defined).
279  */
numa_span(s32 cpu)280 static struct cpumask *numa_span(s32 cpu)
281 {
282 	struct sched_domain *sd;
283 	struct sched_group *sg;
284 
285 	sd = rcu_dereference(per_cpu(sd_numa, cpu));
286 	if (!sd)
287 		return NULL;
288 	sg = sd->groups;
289 	if (!sg)
290 		return NULL;
291 
292 	return sched_group_span(sg);
293 }
294 
295 /*
296  * Return true if the LLC domains do not perfectly overlap with the NUMA
297  * domains, false otherwise.
298  */
llc_numa_mismatch(void)299 static bool llc_numa_mismatch(void)
300 {
301 	int cpu;
302 
303 	/*
304 	 * We need to scan all online CPUs to verify whether their scheduling
305 	 * domains overlap.
306 	 *
307 	 * While it is rare to encounter architectures with asymmetric NUMA
308 	 * topologies, CPU hotplugging or virtualized environments can result
309 	 * in asymmetric configurations.
310 	 *
311 	 * For example:
312 	 *
313 	 *  NUMA 0:
314 	 *    - LLC 0: cpu0..cpu7
315 	 *    - LLC 1: cpu8..cpu15 [offline]
316 	 *
317 	 *  NUMA 1:
318 	 *    - LLC 0: cpu16..cpu23
319 	 *    - LLC 1: cpu24..cpu31
320 	 *
321 	 * In this case, if we only check the first online CPU (cpu0), we might
322 	 * incorrectly assume that the LLC and NUMA domains are fully
323 	 * overlapping, which is incorrect (as NUMA 1 has two distinct LLC
324 	 * domains).
325 	 */
326 	for_each_online_cpu(cpu)
327 		if (llc_weight(cpu) != numa_weight(cpu))
328 			return true;
329 
330 	return false;
331 }
332 
333 /*
334  * Initialize topology-aware scheduling.
335  *
336  * Detect if the system has multiple LLC or multiple NUMA domains and enable
337  * cache-aware / NUMA-aware scheduling optimizations in the default CPU idle
338  * selection policy.
339  *
340  * Assumption: the kernel's internal topology representation assumes that each
341  * CPU belongs to a single LLC domain, and that each LLC domain is entirely
342  * contained within a single NUMA node.
343  */
scx_idle_update_selcpu_topology(struct sched_ext_ops * ops)344 void scx_idle_update_selcpu_topology(struct sched_ext_ops *ops)
345 {
346 	bool enable_llc = false, enable_numa = false;
347 	unsigned int nr_cpus;
348 	s32 cpu = cpumask_first(cpu_online_mask);
349 
350 	/*
351 	 * Enable LLC domain optimization only when there are multiple LLC
352 	 * domains among the online CPUs. If all online CPUs are part of a
353 	 * single LLC domain, the idle CPU selection logic can choose any
354 	 * online CPU without bias.
355 	 *
356 	 * Note that it is sufficient to check the LLC domain of the first
357 	 * online CPU to determine whether a single LLC domain includes all
358 	 * CPUs.
359 	 */
360 	rcu_read_lock();
361 	nr_cpus = llc_weight(cpu);
362 	if (nr_cpus > 0) {
363 		if (nr_cpus < num_online_cpus())
364 			enable_llc = true;
365 		pr_debug("sched_ext: LLC=%*pb weight=%u\n",
366 			 cpumask_pr_args(llc_span(cpu)), llc_weight(cpu));
367 	}
368 
369 	/*
370 	 * Enable NUMA optimization only when there are multiple NUMA domains
371 	 * among the online CPUs and the NUMA domains don't perfectly overlaps
372 	 * with the LLC domains.
373 	 *
374 	 * If all CPUs belong to the same NUMA node and the same LLC domain,
375 	 * enabling both NUMA and LLC optimizations is unnecessary, as checking
376 	 * for an idle CPU in the same domain twice is redundant.
377 	 *
378 	 * If SCX_OPS_BUILTIN_IDLE_PER_NODE is enabled ignore the NUMA
379 	 * optimization, as we would naturally select idle CPUs within
380 	 * specific NUMA nodes querying the corresponding per-node cpumask.
381 	 */
382 	if (!(ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)) {
383 		nr_cpus = numa_weight(cpu);
384 		if (nr_cpus > 0) {
385 			if (nr_cpus < num_online_cpus() && llc_numa_mismatch())
386 				enable_numa = true;
387 			pr_debug("sched_ext: NUMA=%*pb weight=%u\n",
388 				 cpumask_pr_args(numa_span(cpu)), nr_cpus);
389 		}
390 	}
391 	rcu_read_unlock();
392 
393 	pr_debug("sched_ext: LLC idle selection %s\n",
394 		 str_enabled_disabled(enable_llc));
395 	pr_debug("sched_ext: NUMA idle selection %s\n",
396 		 str_enabled_disabled(enable_numa));
397 
398 	if (enable_llc)
399 		static_branch_enable_cpuslocked(&scx_selcpu_topo_llc);
400 	else
401 		static_branch_disable_cpuslocked(&scx_selcpu_topo_llc);
402 	if (enable_numa)
403 		static_branch_enable_cpuslocked(&scx_selcpu_topo_numa);
404 	else
405 		static_branch_disable_cpuslocked(&scx_selcpu_topo_numa);
406 }
407 
408 /*
409  * Return true if @p can run on all possible CPUs, false otherwise.
410  */
task_affinity_all(const struct task_struct * p)411 static inline bool task_affinity_all(const struct task_struct *p)
412 {
413 	return p->nr_cpus_allowed >= num_possible_cpus();
414 }
415 
416 /*
417  * Built-in CPU idle selection policy:
418  *
419  * 1. Prioritize full-idle cores:
420  *   - always prioritize CPUs from fully idle cores (both logical CPUs are
421  *     idle) to avoid interference caused by SMT.
422  *
423  * 2. Reuse the same CPU:
424  *   - prefer the last used CPU to take advantage of cached data (L1, L2) and
425  *     branch prediction optimizations.
426  *
427  * 3. Pick a CPU within the same LLC (Last-Level Cache):
428  *   - if the above conditions aren't met, pick a CPU that shares the same
429  *     LLC, if the LLC domain is a subset of @cpus_allowed, to maintain
430  *     cache locality.
431  *
432  * 4. Pick a CPU within the same NUMA node, if enabled:
433  *   - choose a CPU from the same NUMA node, if the node cpumask is a
434  *     subset of @cpus_allowed, to reduce memory access latency.
435  *
436  * 5. Pick any idle CPU within the @cpus_allowed domain.
437  *
438  * Step 3 and 4 are performed only if the system has, respectively,
439  * multiple LLCs / multiple NUMA nodes (see scx_selcpu_topo_llc and
440  * scx_selcpu_topo_numa) and they don't contain the same subset of CPUs.
441  *
442  * If %SCX_OPS_BUILTIN_IDLE_PER_NODE is enabled, the search will always
443  * begin in @prev_cpu's node and proceed to other nodes in order of
444  * increasing distance.
445  *
446  * Return the picked CPU if idle, or a negative value otherwise.
447  *
448  * NOTE: tasks that can only run on 1 CPU are excluded by this logic, because
449  * we never call ops.select_cpu() for them, see select_task_rq().
450  */
scx_select_cpu_dfl(struct task_struct * p,s32 prev_cpu,u64 wake_flags,const struct cpumask * cpus_allowed,u64 flags)451 s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
452 		       const struct cpumask *cpus_allowed, u64 flags)
453 {
454 	const struct cpumask *llc_cpus = NULL, *numa_cpus = NULL;
455 	const struct cpumask *allowed = cpus_allowed ?: p->cpus_ptr;
456 	int node = scx_cpu_node_if_enabled(prev_cpu);
457 	bool is_prev_allowed;
458 	s32 cpu;
459 
460 	preempt_disable();
461 
462 	/*
463 	 * Check whether @prev_cpu is still within the allowed set. If not,
464 	 * we can still try selecting a nearby CPU.
465 	 */
466 	is_prev_allowed = cpumask_test_cpu(prev_cpu, allowed);
467 
468 	/*
469 	 * Determine the subset of CPUs usable by @p within @cpus_allowed.
470 	 */
471 	if (allowed != p->cpus_ptr) {
472 		struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_idle_cpumask);
473 
474 		if (task_affinity_all(p)) {
475 			allowed = cpus_allowed;
476 		} else if (cpumask_and(local_cpus, cpus_allowed, p->cpus_ptr)) {
477 			allowed = local_cpus;
478 		} else {
479 			cpu = -EBUSY;
480 			goto out_enable;
481 		}
482 	}
483 
484 	/*
485 	 * This is necessary to protect llc_cpus.
486 	 */
487 	rcu_read_lock();
488 
489 	/*
490 	 * Determine the subset of CPUs that the task can use in its
491 	 * current LLC and node.
492 	 *
493 	 * If the task can run on all CPUs, use the node and LLC cpumasks
494 	 * directly.
495 	 */
496 	if (static_branch_maybe(CONFIG_NUMA, &scx_selcpu_topo_numa)) {
497 		struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_numa_idle_cpumask);
498 		const struct cpumask *cpus = numa_span(prev_cpu);
499 
500 		if (allowed == p->cpus_ptr && task_affinity_all(p))
501 			numa_cpus = cpus;
502 		else if (cpus && cpumask_and(local_cpus, allowed, cpus))
503 			numa_cpus = local_cpus;
504 	}
505 
506 	if (static_branch_maybe(CONFIG_SCHED_MC, &scx_selcpu_topo_llc)) {
507 		struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_llc_idle_cpumask);
508 		const struct cpumask *cpus = llc_span(prev_cpu);
509 
510 		if (allowed == p->cpus_ptr && task_affinity_all(p))
511 			llc_cpus = cpus;
512 		else if (cpus && cpumask_and(local_cpus, allowed, cpus))
513 			llc_cpus = local_cpus;
514 	}
515 
516 	/*
517 	 * If WAKE_SYNC, try to migrate the wakee to the waker's CPU.
518 	 */
519 	if (wake_flags & SCX_WAKE_SYNC) {
520 		int waker_node;
521 
522 		/*
523 		 * If the waker's CPU is cache affine and prev_cpu is idle,
524 		 * then avoid a migration.
525 		 */
526 		cpu = smp_processor_id();
527 		if (is_prev_allowed && cpus_share_cache(cpu, prev_cpu) &&
528 		    scx_idle_test_and_clear_cpu(prev_cpu)) {
529 			cpu = prev_cpu;
530 			goto out_unlock;
531 		}
532 
533 		/*
534 		 * If the waker's local DSQ is empty, and the system is under
535 		 * utilized, try to wake up @p to the local DSQ of the waker.
536 		 *
537 		 * Checking only for an empty local DSQ is insufficient as it
538 		 * could give the wakee an unfair advantage when the system is
539 		 * oversaturated.
540 		 *
541 		 * Checking only for the presence of idle CPUs is also
542 		 * insufficient as the local DSQ of the waker could have tasks
543 		 * piled up on it even if there is an idle core elsewhere on
544 		 * the system.
545 		 */
546 		waker_node = cpu_to_node(cpu);
547 		if (!(current->flags & PF_EXITING) &&
548 		    cpu_rq(cpu)->scx.local_dsq.nr == 0 &&
549 		    (!(flags & SCX_PICK_IDLE_IN_NODE) || (waker_node == node)) &&
550 		    !cpumask_empty(idle_cpumask(waker_node)->cpu)) {
551 			if (cpumask_test_cpu(cpu, allowed))
552 				goto out_unlock;
553 		}
554 	}
555 
556 	/*
557 	 * If CPU has SMT, any wholly idle CPU is likely a better pick than
558 	 * partially idle @prev_cpu.
559 	 */
560 	if (sched_smt_active()) {
561 		/*
562 		 * Keep using @prev_cpu if it's part of a fully idle core.
563 		 */
564 		if (is_prev_allowed &&
565 		    cpumask_test_cpu(prev_cpu, idle_cpumask(node)->smt) &&
566 		    scx_idle_test_and_clear_cpu(prev_cpu)) {
567 			cpu = prev_cpu;
568 			goto out_unlock;
569 		}
570 
571 		/*
572 		 * Search for any fully idle core in the same LLC domain.
573 		 */
574 		if (llc_cpus) {
575 			cpu = pick_idle_cpu_in_node(llc_cpus, node, SCX_PICK_IDLE_CORE);
576 			if (cpu >= 0)
577 				goto out_unlock;
578 		}
579 
580 		/*
581 		 * Search for any fully idle core in the same NUMA node.
582 		 */
583 		if (numa_cpus) {
584 			cpu = pick_idle_cpu_in_node(numa_cpus, node, SCX_PICK_IDLE_CORE);
585 			if (cpu >= 0)
586 				goto out_unlock;
587 		}
588 
589 		/*
590 		 * Search for any full-idle core usable by the task.
591 		 *
592 		 * If the node-aware idle CPU selection policy is enabled
593 		 * (%SCX_OPS_BUILTIN_IDLE_PER_NODE), the search will always
594 		 * begin in prev_cpu's node and proceed to other nodes in
595 		 * order of increasing distance.
596 		 */
597 		cpu = scx_pick_idle_cpu(allowed, node, flags | SCX_PICK_IDLE_CORE);
598 		if (cpu >= 0)
599 			goto out_unlock;
600 
601 		/*
602 		 * Give up if we're strictly looking for a full-idle SMT
603 		 * core.
604 		 */
605 		if (flags & SCX_PICK_IDLE_CORE) {
606 			cpu = -EBUSY;
607 			goto out_unlock;
608 		}
609 	}
610 
611 	/*
612 	 * Use @prev_cpu if it's idle.
613 	 */
614 	if (is_prev_allowed && scx_idle_test_and_clear_cpu(prev_cpu)) {
615 		cpu = prev_cpu;
616 		goto out_unlock;
617 	}
618 
619 	/*
620 	 * Search for any idle CPU in the same LLC domain.
621 	 */
622 	if (llc_cpus) {
623 		cpu = pick_idle_cpu_in_node(llc_cpus, node, 0);
624 		if (cpu >= 0)
625 			goto out_unlock;
626 	}
627 
628 	/*
629 	 * Search for any idle CPU in the same NUMA node.
630 	 */
631 	if (numa_cpus) {
632 		cpu = pick_idle_cpu_in_node(numa_cpus, node, 0);
633 		if (cpu >= 0)
634 			goto out_unlock;
635 	}
636 
637 	/*
638 	 * Search for any idle CPU usable by the task.
639 	 *
640 	 * If the node-aware idle CPU selection policy is enabled
641 	 * (%SCX_OPS_BUILTIN_IDLE_PER_NODE), the search will always begin
642 	 * in prev_cpu's node and proceed to other nodes in order of
643 	 * increasing distance.
644 	 */
645 	cpu = scx_pick_idle_cpu(allowed, node, flags);
646 
647 out_unlock:
648 	rcu_read_unlock();
649 out_enable:
650 	preempt_enable();
651 
652 	return cpu;
653 }
654 
655 /*
656  * Initialize global and per-node idle cpumasks.
657  */
scx_idle_init_masks(void)658 void scx_idle_init_masks(void)
659 {
660 	int i;
661 
662 	/* Allocate global idle cpumasks */
663 	BUG_ON(!alloc_cpumask_var(&scx_idle_global_masks.cpu, GFP_KERNEL));
664 	BUG_ON(!alloc_cpumask_var(&scx_idle_global_masks.smt, GFP_KERNEL));
665 
666 	/* Allocate per-node idle cpumasks */
667 	scx_idle_node_masks = kcalloc(num_possible_nodes(),
668 				      sizeof(*scx_idle_node_masks), GFP_KERNEL);
669 	BUG_ON(!scx_idle_node_masks);
670 
671 	for_each_node(i) {
672 		scx_idle_node_masks[i] = kzalloc_node(sizeof(**scx_idle_node_masks),
673 							 GFP_KERNEL, i);
674 		BUG_ON(!scx_idle_node_masks[i]);
675 
676 		BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->cpu, GFP_KERNEL, i));
677 		BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->smt, GFP_KERNEL, i));
678 	}
679 
680 	/* Allocate local per-cpu idle cpumasks */
681 	for_each_possible_cpu(i) {
682 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_idle_cpumask, i),
683 					       GFP_KERNEL, cpu_to_node(i)));
684 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_llc_idle_cpumask, i),
685 					       GFP_KERNEL, cpu_to_node(i)));
686 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_numa_idle_cpumask, i),
687 					       GFP_KERNEL, cpu_to_node(i)));
688 	}
689 }
690 
update_builtin_idle(int cpu,bool idle)691 static void update_builtin_idle(int cpu, bool idle)
692 {
693 	int node = scx_cpu_node_if_enabled(cpu);
694 	struct cpumask *idle_cpus = idle_cpumask(node)->cpu;
695 
696 	assign_cpu(cpu, idle_cpus, idle);
697 
698 #ifdef CONFIG_SCHED_SMT
699 	if (sched_smt_active()) {
700 		const struct cpumask *smt = cpu_smt_mask(cpu);
701 		struct cpumask *idle_smts = idle_cpumask(node)->smt;
702 
703 		if (idle) {
704 			/*
705 			 * idle_smt handling is racy but that's fine as it's
706 			 * only for optimization and self-correcting.
707 			 */
708 			if (!cpumask_subset(smt, idle_cpus))
709 				return;
710 			cpumask_or(idle_smts, idle_smts, smt);
711 		} else {
712 			cpumask_andnot(idle_smts, idle_smts, smt);
713 		}
714 	}
715 #endif
716 }
717 
718 /*
719  * Update the idle state of a CPU to @idle.
720  *
721  * If @do_notify is true, ops.update_idle() is invoked to notify the scx
722  * scheduler of an actual idle state transition (idle to busy or vice
723  * versa). If @do_notify is false, only the idle state in the idle masks is
724  * refreshed without invoking ops.update_idle().
725  *
726  * This distinction is necessary, because an idle CPU can be "reserved" and
727  * awakened via scx_bpf_pick_idle_cpu() + scx_bpf_kick_cpu(), marking it as
728  * busy even if no tasks are dispatched. In this case, the CPU may return
729  * to idle without a true state transition. Refreshing the idle masks
730  * without invoking ops.update_idle() ensures accurate idle state tracking
731  * while avoiding unnecessary updates and maintaining balanced state
732  * transitions.
733  */
__scx_update_idle(struct rq * rq,bool idle,bool do_notify)734 void __scx_update_idle(struct rq *rq, bool idle, bool do_notify)
735 {
736 	struct scx_sched *sch = scx_root;
737 	int cpu = cpu_of(rq);
738 
739 	lockdep_assert_rq_held(rq);
740 
741 	/*
742 	 * Update the idle masks:
743 	 * - for real idle transitions (do_notify == true)
744 	 * - for idle-to-idle transitions (indicated by the previous task
745 	 *   being the idle thread, managed by pick_task_idle())
746 	 *
747 	 * Skip updating idle masks if the previous task is not the idle
748 	 * thread, since set_next_task_idle() has already handled it when
749 	 * transitioning from a task to the idle thread (calling this
750 	 * function with do_notify == true).
751 	 *
752 	 * In this way we can avoid updating the idle masks twice,
753 	 * unnecessarily.
754 	 */
755 	if (static_branch_likely(&scx_builtin_idle_enabled))
756 		if (do_notify || is_idle_task(rq->curr))
757 			update_builtin_idle(cpu, idle);
758 
759 	/*
760 	 * Trigger ops.update_idle() only when transitioning from a task to
761 	 * the idle thread and vice versa.
762 	 *
763 	 * Idle transitions are indicated by do_notify being set to true,
764 	 * managed by put_prev_task_idle()/set_next_task_idle().
765 	 *
766 	 * This must come after builtin idle update so that BPF schedulers can
767 	 * create interlocking between ops.update_idle() and ops.enqueue() -
768 	 * either enqueue() sees the idle bit or update_idle() sees the task
769 	 * that enqueue() queued.
770 	 */
771 	if (SCX_HAS_OP(sch, update_idle) && do_notify && !scx_rq_bypassing(rq))
772 		SCX_CALL_OP(sch, SCX_KF_REST, update_idle, rq, cpu_of(rq), idle);
773 }
774 
reset_idle_masks(struct sched_ext_ops * ops)775 static void reset_idle_masks(struct sched_ext_ops *ops)
776 {
777 	int node;
778 
779 	/*
780 	 * Consider all online cpus idle. Should converge to the actual state
781 	 * quickly.
782 	 */
783 	if (!(ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)) {
784 		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->cpu, cpu_online_mask);
785 		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->smt, cpu_online_mask);
786 		return;
787 	}
788 
789 	for_each_node(node) {
790 		const struct cpumask *node_mask = cpumask_of_node(node);
791 
792 		cpumask_and(idle_cpumask(node)->cpu, cpu_online_mask, node_mask);
793 		cpumask_and(idle_cpumask(node)->smt, cpu_online_mask, node_mask);
794 	}
795 }
796 
scx_idle_enable(struct sched_ext_ops * ops)797 void scx_idle_enable(struct sched_ext_ops *ops)
798 {
799 	if (!ops->update_idle || (ops->flags & SCX_OPS_KEEP_BUILTIN_IDLE))
800 		static_branch_enable_cpuslocked(&scx_builtin_idle_enabled);
801 	else
802 		static_branch_disable_cpuslocked(&scx_builtin_idle_enabled);
803 
804 	if (ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)
805 		static_branch_enable_cpuslocked(&scx_builtin_idle_per_node);
806 	else
807 		static_branch_disable_cpuslocked(&scx_builtin_idle_per_node);
808 
809 	reset_idle_masks(ops);
810 }
811 
scx_idle_disable(void)812 void scx_idle_disable(void)
813 {
814 	static_branch_disable(&scx_builtin_idle_enabled);
815 	static_branch_disable(&scx_builtin_idle_per_node);
816 }
817 
818 /********************************************************************************
819  * Helpers that can be called from the BPF scheduler.
820  */
821 
validate_node(int node)822 static int validate_node(int node)
823 {
824 	if (!static_branch_likely(&scx_builtin_idle_per_node)) {
825 		scx_kf_error("per-node idle tracking is disabled");
826 		return -EOPNOTSUPP;
827 	}
828 
829 	/* Return no entry for NUMA_NO_NODE (not a critical scx error) */
830 	if (node == NUMA_NO_NODE)
831 		return -ENOENT;
832 
833 	/* Make sure node is in a valid range */
834 	if (node < 0 || node >= nr_node_ids) {
835 		scx_kf_error("invalid node %d", node);
836 		return -EINVAL;
837 	}
838 
839 	/* Make sure the node is part of the set of possible nodes */
840 	if (!node_possible(node)) {
841 		scx_kf_error("unavailable node %d", node);
842 		return -EINVAL;
843 	}
844 
845 	return node;
846 }
847 
848 __bpf_kfunc_start_defs();
849 
check_builtin_idle_enabled(void)850 static bool check_builtin_idle_enabled(void)
851 {
852 	if (static_branch_likely(&scx_builtin_idle_enabled))
853 		return true;
854 
855 	scx_kf_error("built-in idle tracking is disabled");
856 	return false;
857 }
858 
select_cpu_from_kfunc(struct task_struct * p,s32 prev_cpu,u64 wake_flags,const struct cpumask * allowed,u64 flags)859 static s32 select_cpu_from_kfunc(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
860 				 const struct cpumask *allowed, u64 flags)
861 {
862 	struct rq *rq;
863 	struct rq_flags rf;
864 	s32 cpu;
865 
866 	if (!kf_cpu_valid(prev_cpu, NULL))
867 		return -EINVAL;
868 
869 	if (!check_builtin_idle_enabled())
870 		return -EBUSY;
871 
872 	/*
873 	 * If called from an unlocked context, acquire the task's rq lock,
874 	 * so that we can safely access p->cpus_ptr and p->nr_cpus_allowed.
875 	 *
876 	 * Otherwise, allow to use this kfunc only from ops.select_cpu()
877 	 * and ops.select_enqueue().
878 	 */
879 	if (scx_kf_allowed_if_unlocked()) {
880 		rq = task_rq_lock(p, &rf);
881 	} else {
882 		if (!scx_kf_allowed(SCX_KF_SELECT_CPU | SCX_KF_ENQUEUE))
883 			return -EPERM;
884 		rq = scx_locked_rq();
885 	}
886 
887 	/*
888 	 * Validate locking correctness to access p->cpus_ptr and
889 	 * p->nr_cpus_allowed: if we're holding an rq lock, we're safe;
890 	 * otherwise, assert that p->pi_lock is held.
891 	 */
892 	if (!rq)
893 		lockdep_assert_held(&p->pi_lock);
894 
895 	/*
896 	 * This may also be called from ops.enqueue(), so we need to handle
897 	 * per-CPU tasks as well. For these tasks, we can skip all idle CPU
898 	 * selection optimizations and simply check whether the previously
899 	 * used CPU is idle and within the allowed cpumask.
900 	 */
901 	if (p->nr_cpus_allowed == 1 || is_migration_disabled(p)) {
902 		if (cpumask_test_cpu(prev_cpu, allowed ?: p->cpus_ptr) &&
903 		    scx_idle_test_and_clear_cpu(prev_cpu))
904 			cpu = prev_cpu;
905 		else
906 			cpu = -EBUSY;
907 	} else {
908 		cpu = scx_select_cpu_dfl(p, prev_cpu, wake_flags,
909 					 allowed ?: p->cpus_ptr, flags);
910 	}
911 
912 	if (scx_kf_allowed_if_unlocked())
913 		task_rq_unlock(rq, p, &rf);
914 
915 	return cpu;
916 }
917 
918 /**
919  * scx_bpf_cpu_node - Return the NUMA node the given @cpu belongs to, or
920  *		      trigger an error if @cpu is invalid
921  * @cpu: target CPU
922  */
scx_bpf_cpu_node(s32 cpu)923 __bpf_kfunc int scx_bpf_cpu_node(s32 cpu)
924 {
925 	if (!kf_cpu_valid(cpu, NULL))
926 		return NUMA_NO_NODE;
927 
928 	return cpu_to_node(cpu);
929 }
930 
931 /**
932  * scx_bpf_select_cpu_dfl - The default implementation of ops.select_cpu()
933  * @p: task_struct to select a CPU for
934  * @prev_cpu: CPU @p was on previously
935  * @wake_flags: %SCX_WAKE_* flags
936  * @is_idle: out parameter indicating whether the returned CPU is idle
937  *
938  * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
939  * context such as a BPF test_run() call, as long as built-in CPU selection
940  * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
941  * is set.
942  *
943  * Returns the picked CPU with *@is_idle indicating whether the picked CPU is
944  * currently idle and thus a good candidate for direct dispatching.
945  */
scx_bpf_select_cpu_dfl(struct task_struct * p,s32 prev_cpu,u64 wake_flags,bool * is_idle)946 __bpf_kfunc s32 scx_bpf_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
947 				       u64 wake_flags, bool *is_idle)
948 {
949 	s32 cpu;
950 
951 	cpu = select_cpu_from_kfunc(p, prev_cpu, wake_flags, NULL, 0);
952 	if (cpu >= 0) {
953 		*is_idle = true;
954 		return cpu;
955 	}
956 	*is_idle = false;
957 
958 	return prev_cpu;
959 }
960 
961 /**
962  * scx_bpf_select_cpu_and - Pick an idle CPU usable by task @p,
963  *			    prioritizing those in @cpus_allowed
964  * @p: task_struct to select a CPU for
965  * @prev_cpu: CPU @p was on previously
966  * @wake_flags: %SCX_WAKE_* flags
967  * @cpus_allowed: cpumask of allowed CPUs
968  * @flags: %SCX_PICK_IDLE* flags
969  *
970  * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
971  * context such as a BPF test_run() call, as long as built-in CPU selection
972  * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
973  * is set.
974  *
975  * @p, @prev_cpu and @wake_flags match ops.select_cpu().
976  *
977  * Returns the selected idle CPU, which will be automatically awakened upon
978  * returning from ops.select_cpu() and can be used for direct dispatch, or
979  * a negative value if no idle CPU is available.
980  */
scx_bpf_select_cpu_and(struct task_struct * p,s32 prev_cpu,u64 wake_flags,const struct cpumask * cpus_allowed,u64 flags)981 __bpf_kfunc s32 scx_bpf_select_cpu_and(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
982 				       const struct cpumask *cpus_allowed, u64 flags)
983 {
984 	return select_cpu_from_kfunc(p, prev_cpu, wake_flags, cpus_allowed, flags);
985 }
986 
987 /**
988  * scx_bpf_get_idle_cpumask_node - Get a referenced kptr to the
989  * idle-tracking per-CPU cpumask of a target NUMA node.
990  * @node: target NUMA node
991  *
992  * Returns an empty cpumask if idle tracking is not enabled, if @node is
993  * not valid, or running on a UP kernel. In this case the actual error will
994  * be reported to the BPF scheduler via scx_error().
995  */
scx_bpf_get_idle_cpumask_node(int node)996 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask_node(int node)
997 {
998 	node = validate_node(node);
999 	if (node < 0)
1000 		return cpu_none_mask;
1001 
1002 	return idle_cpumask(node)->cpu;
1003 }
1004 
1005 /**
1006  * scx_bpf_get_idle_cpumask - Get a referenced kptr to the idle-tracking
1007  * per-CPU cpumask.
1008  *
1009  * Returns an empty mask if idle tracking is not enabled, or running on a
1010  * UP kernel.
1011  */
scx_bpf_get_idle_cpumask(void)1012 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(void)
1013 {
1014 	if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1015 		scx_kf_error("SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1016 		return cpu_none_mask;
1017 	}
1018 
1019 	if (!check_builtin_idle_enabled())
1020 		return cpu_none_mask;
1021 
1022 	return idle_cpumask(NUMA_NO_NODE)->cpu;
1023 }
1024 
1025 /**
1026  * scx_bpf_get_idle_smtmask_node - Get a referenced kptr to the
1027  * idle-tracking, per-physical-core cpumask of a target NUMA node. Can be
1028  * used to determine if an entire physical core is free.
1029  * @node: target NUMA node
1030  *
1031  * Returns an empty cpumask if idle tracking is not enabled, if @node is
1032  * not valid, or running on a UP kernel. In this case the actual error will
1033  * be reported to the BPF scheduler via scx_error().
1034  */
scx_bpf_get_idle_smtmask_node(int node)1035 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask_node(int node)
1036 {
1037 	node = validate_node(node);
1038 	if (node < 0)
1039 		return cpu_none_mask;
1040 
1041 	if (sched_smt_active())
1042 		return idle_cpumask(node)->smt;
1043 	else
1044 		return idle_cpumask(node)->cpu;
1045 }
1046 
1047 /**
1048  * scx_bpf_get_idle_smtmask - Get a referenced kptr to the idle-tracking,
1049  * per-physical-core cpumask. Can be used to determine if an entire physical
1050  * core is free.
1051  *
1052  * Returns an empty mask if idle tracking is not enabled, or running on a
1053  * UP kernel.
1054  */
scx_bpf_get_idle_smtmask(void)1055 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask(void)
1056 {
1057 	if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1058 		scx_kf_error("SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1059 		return cpu_none_mask;
1060 	}
1061 
1062 	if (!check_builtin_idle_enabled())
1063 		return cpu_none_mask;
1064 
1065 	if (sched_smt_active())
1066 		return idle_cpumask(NUMA_NO_NODE)->smt;
1067 	else
1068 		return idle_cpumask(NUMA_NO_NODE)->cpu;
1069 }
1070 
1071 /**
1072  * scx_bpf_put_idle_cpumask - Release a previously acquired referenced kptr to
1073  * either the percpu, or SMT idle-tracking cpumask.
1074  * @idle_mask: &cpumask to use
1075  */
scx_bpf_put_idle_cpumask(const struct cpumask * idle_mask)1076 __bpf_kfunc void scx_bpf_put_idle_cpumask(const struct cpumask *idle_mask)
1077 {
1078 	/*
1079 	 * Empty function body because we aren't actually acquiring or releasing
1080 	 * a reference to a global idle cpumask, which is read-only in the
1081 	 * caller and is never released. The acquire / release semantics here
1082 	 * are just used to make the cpumask a trusted pointer in the caller.
1083 	 */
1084 }
1085 
1086 /**
1087  * scx_bpf_test_and_clear_cpu_idle - Test and clear @cpu's idle state
1088  * @cpu: cpu to test and clear idle for
1089  *
1090  * Returns %true if @cpu was idle and its idle state was successfully cleared.
1091  * %false otherwise.
1092  *
1093  * Unavailable if ops.update_idle() is implemented and
1094  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1095  */
scx_bpf_test_and_clear_cpu_idle(s32 cpu)1096 __bpf_kfunc bool scx_bpf_test_and_clear_cpu_idle(s32 cpu)
1097 {
1098 	if (!check_builtin_idle_enabled())
1099 		return false;
1100 
1101 	if (!kf_cpu_valid(cpu, NULL))
1102 		return false;
1103 
1104 	return scx_idle_test_and_clear_cpu(cpu);
1105 }
1106 
1107 /**
1108  * scx_bpf_pick_idle_cpu_node - Pick and claim an idle cpu from @node
1109  * @cpus_allowed: Allowed cpumask
1110  * @node: target NUMA node
1111  * @flags: %SCX_PICK_IDLE_* flags
1112  *
1113  * Pick and claim an idle cpu in @cpus_allowed from the NUMA node @node.
1114  *
1115  * Returns the picked idle cpu number on success, or -%EBUSY if no matching
1116  * cpu was found.
1117  *
1118  * The search starts from @node and proceeds to other online NUMA nodes in
1119  * order of increasing distance (unless SCX_PICK_IDLE_IN_NODE is specified,
1120  * in which case the search is limited to the target @node).
1121  *
1122  * Always returns an error if ops.update_idle() is implemented and
1123  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set, or if
1124  * %SCX_OPS_BUILTIN_IDLE_PER_NODE is not set.
1125  */
scx_bpf_pick_idle_cpu_node(const struct cpumask * cpus_allowed,int node,u64 flags)1126 __bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(const struct cpumask *cpus_allowed,
1127 					   int node, u64 flags)
1128 {
1129 	node = validate_node(node);
1130 	if (node < 0)
1131 		return node;
1132 
1133 	return scx_pick_idle_cpu(cpus_allowed, node, flags);
1134 }
1135 
1136 /**
1137  * scx_bpf_pick_idle_cpu - Pick and claim an idle cpu
1138  * @cpus_allowed: Allowed cpumask
1139  * @flags: %SCX_PICK_IDLE_CPU_* flags
1140  *
1141  * Pick and claim an idle cpu in @cpus_allowed. Returns the picked idle cpu
1142  * number on success. -%EBUSY if no matching cpu was found.
1143  *
1144  * Idle CPU tracking may race against CPU scheduling state transitions. For
1145  * example, this function may return -%EBUSY as CPUs are transitioning into the
1146  * idle state. If the caller then assumes that there will be dispatch events on
1147  * the CPUs as they were all busy, the scheduler may end up stalling with CPUs
1148  * idling while there are pending tasks. Use scx_bpf_pick_any_cpu() and
1149  * scx_bpf_kick_cpu() to guarantee that there will be at least one dispatch
1150  * event in the near future.
1151  *
1152  * Unavailable if ops.update_idle() is implemented and
1153  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1154  *
1155  * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1156  * scx_bpf_pick_idle_cpu_node() instead.
1157  */
scx_bpf_pick_idle_cpu(const struct cpumask * cpus_allowed,u64 flags)1158 __bpf_kfunc s32 scx_bpf_pick_idle_cpu(const struct cpumask *cpus_allowed,
1159 				      u64 flags)
1160 {
1161 	if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1162 		scx_kf_error("per-node idle tracking is enabled");
1163 		return -EBUSY;
1164 	}
1165 
1166 	if (!check_builtin_idle_enabled())
1167 		return -EBUSY;
1168 
1169 	return scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1170 }
1171 
1172 /**
1173  * scx_bpf_pick_any_cpu_node - Pick and claim an idle cpu if available
1174  *			       or pick any CPU from @node
1175  * @cpus_allowed: Allowed cpumask
1176  * @node: target NUMA node
1177  * @flags: %SCX_PICK_IDLE_CPU_* flags
1178  *
1179  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1180  * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1181  * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1182  * empty.
1183  *
1184  * The search starts from @node and proceeds to other online NUMA nodes in
1185  * order of increasing distance (unless %SCX_PICK_IDLE_IN_NODE is specified,
1186  * in which case the search is limited to the target @node, regardless of
1187  * the CPU idle state).
1188  *
1189  * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1190  * set, this function can't tell which CPUs are idle and will always pick any
1191  * CPU.
1192  */
scx_bpf_pick_any_cpu_node(const struct cpumask * cpus_allowed,int node,u64 flags)1193 __bpf_kfunc s32 scx_bpf_pick_any_cpu_node(const struct cpumask *cpus_allowed,
1194 					  int node, u64 flags)
1195 {
1196 	s32 cpu;
1197 
1198 	node = validate_node(node);
1199 	if (node < 0)
1200 		return node;
1201 
1202 	cpu = scx_pick_idle_cpu(cpus_allowed, node, flags);
1203 	if (cpu >= 0)
1204 		return cpu;
1205 
1206 	if (flags & SCX_PICK_IDLE_IN_NODE)
1207 		cpu = cpumask_any_and_distribute(cpumask_of_node(node), cpus_allowed);
1208 	else
1209 		cpu = cpumask_any_distribute(cpus_allowed);
1210 	if (cpu < nr_cpu_ids)
1211 		return cpu;
1212 	else
1213 		return -EBUSY;
1214 }
1215 
1216 /**
1217  * scx_bpf_pick_any_cpu - Pick and claim an idle cpu if available or pick any CPU
1218  * @cpus_allowed: Allowed cpumask
1219  * @flags: %SCX_PICK_IDLE_CPU_* flags
1220  *
1221  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1222  * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1223  * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1224  * empty.
1225  *
1226  * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1227  * set, this function can't tell which CPUs are idle and will always pick any
1228  * CPU.
1229  *
1230  * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1231  * scx_bpf_pick_any_cpu_node() instead.
1232  */
scx_bpf_pick_any_cpu(const struct cpumask * cpus_allowed,u64 flags)1233 __bpf_kfunc s32 scx_bpf_pick_any_cpu(const struct cpumask *cpus_allowed,
1234 				     u64 flags)
1235 {
1236 	s32 cpu;
1237 
1238 	if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1239 		scx_kf_error("per-node idle tracking is enabled");
1240 		return -EBUSY;
1241 	}
1242 
1243 	if (static_branch_likely(&scx_builtin_idle_enabled)) {
1244 		cpu = scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1245 		if (cpu >= 0)
1246 			return cpu;
1247 	}
1248 
1249 	cpu = cpumask_any_distribute(cpus_allowed);
1250 	if (cpu < nr_cpu_ids)
1251 		return cpu;
1252 	else
1253 		return -EBUSY;
1254 }
1255 
1256 __bpf_kfunc_end_defs();
1257 
1258 BTF_KFUNCS_START(scx_kfunc_ids_idle)
1259 BTF_ID_FLAGS(func, scx_bpf_cpu_node)
1260 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask_node, KF_ACQUIRE)
1261 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask, KF_ACQUIRE)
1262 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask_node, KF_ACQUIRE)
1263 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask, KF_ACQUIRE)
1264 BTF_ID_FLAGS(func, scx_bpf_put_idle_cpumask, KF_RELEASE)
1265 BTF_ID_FLAGS(func, scx_bpf_test_and_clear_cpu_idle)
1266 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu_node, KF_RCU)
1267 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu, KF_RCU)
1268 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu_node, KF_RCU)
1269 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu, KF_RCU)
1270 BTF_ID_FLAGS(func, scx_bpf_select_cpu_and, KF_RCU)
1271 BTF_ID_FLAGS(func, scx_bpf_select_cpu_dfl, KF_RCU)
1272 BTF_KFUNCS_END(scx_kfunc_ids_idle)
1273 
1274 static const struct btf_kfunc_id_set scx_kfunc_set_idle = {
1275 	.owner			= THIS_MODULE,
1276 	.set			= &scx_kfunc_ids_idle,
1277 };
1278 
scx_idle_init(void)1279 int scx_idle_init(void)
1280 {
1281 	int ret;
1282 
1283 	ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_idle) ||
1284 	      register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_idle) ||
1285 	      register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_idle);
1286 
1287 	return ret;
1288 }
1289