xref: /linux/kernel/sched/ext_idle.c (revision 631919fb12fe7b1f0453fe1035e62ce704bc3923)
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  */
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  */
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 
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  */
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  */
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
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  */
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  */
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  */
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  */
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  */
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  */
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  */
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  */
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  */
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 = scx_cpu_node_if_enabled(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  */
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 (use nr_node_ids for non-contiguous NUMA nodes) */
667 	scx_idle_node_masks = kzalloc_objs(*scx_idle_node_masks, nr_node_ids);
668 	BUG_ON(!scx_idle_node_masks);
669 
670 	for_each_node(i) {
671 		scx_idle_node_masks[i] = kzalloc_node(sizeof(**scx_idle_node_masks),
672 							 GFP_KERNEL, i);
673 		BUG_ON(!scx_idle_node_masks[i]);
674 
675 		BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->cpu, GFP_KERNEL, i));
676 		BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->smt, GFP_KERNEL, i));
677 	}
678 
679 	/* Allocate local per-cpu idle cpumasks */
680 	for_each_possible_cpu(i) {
681 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_idle_cpumask, i),
682 					       GFP_KERNEL, cpu_to_node(i)));
683 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_llc_idle_cpumask, i),
684 					       GFP_KERNEL, cpu_to_node(i)));
685 		BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_numa_idle_cpumask, i),
686 					       GFP_KERNEL, cpu_to_node(i)));
687 	}
688 }
689 
690 static void update_builtin_idle(int cpu, bool idle)
691 {
692 	int node = scx_cpu_node_if_enabled(cpu);
693 	struct cpumask *idle_cpus = idle_cpumask(node)->cpu;
694 
695 	assign_cpu(cpu, idle_cpus, idle);
696 
697 #ifdef CONFIG_SCHED_SMT
698 	if (sched_smt_active()) {
699 		const struct cpumask *smt = cpu_smt_mask(cpu);
700 		struct cpumask *idle_smts = idle_cpumask(node)->smt;
701 
702 		if (idle) {
703 			/*
704 			 * idle_smt handling is racy but that's fine as it's
705 			 * only for optimization and self-correcting.
706 			 */
707 			if (!cpumask_subset(smt, idle_cpus))
708 				return;
709 			cpumask_or(idle_smts, idle_smts, smt);
710 		} else {
711 			cpumask_andnot(idle_smts, idle_smts, smt);
712 		}
713 	}
714 #endif
715 }
716 
717 /*
718  * Update the idle state of a CPU to @idle.
719  *
720  * If @do_notify is true, ops.update_idle() is invoked to notify the scx
721  * scheduler of an actual idle state transition (idle to busy or vice
722  * versa). If @do_notify is false, only the idle state in the idle masks is
723  * refreshed without invoking ops.update_idle().
724  *
725  * This distinction is necessary, because an idle CPU can be "reserved" and
726  * awakened via scx_bpf_pick_idle_cpu() + scx_bpf_kick_cpu(), marking it as
727  * busy even if no tasks are dispatched. In this case, the CPU may return
728  * to idle without a true state transition. Refreshing the idle masks
729  * without invoking ops.update_idle() ensures accurate idle state tracking
730  * while avoiding unnecessary updates and maintaining balanced state
731  * transitions.
732  */
733 void __scx_update_idle(struct rq *rq, bool idle, bool do_notify)
734 {
735 	struct scx_sched *sch = scx_root;
736 	int cpu = cpu_of(rq);
737 
738 	lockdep_assert_rq_held(rq);
739 
740 	/*
741 	 * Update the idle masks:
742 	 * - for real idle transitions (do_notify == true)
743 	 * - for idle-to-idle transitions (indicated by the previous task
744 	 *   being the idle thread, managed by pick_task_idle())
745 	 *
746 	 * Skip updating idle masks if the previous task is not the idle
747 	 * thread, since set_next_task_idle() has already handled it when
748 	 * transitioning from a task to the idle thread (calling this
749 	 * function with do_notify == true).
750 	 *
751 	 * In this way we can avoid updating the idle masks twice,
752 	 * unnecessarily.
753 	 */
754 	if (static_branch_likely(&scx_builtin_idle_enabled))
755 		if (do_notify || is_idle_task(rq->curr))
756 			update_builtin_idle(cpu, idle);
757 
758 	/*
759 	 * Trigger ops.update_idle() only when transitioning from a task to
760 	 * the idle thread and vice versa.
761 	 *
762 	 * Idle transitions are indicated by do_notify being set to true,
763 	 * managed by put_prev_task_idle()/set_next_task_idle().
764 	 *
765 	 * This must come after builtin idle update so that BPF schedulers can
766 	 * create interlocking between ops.update_idle() and ops.enqueue() -
767 	 * either enqueue() sees the idle bit or update_idle() sees the task
768 	 * that enqueue() queued.
769 	 */
770 	if (SCX_HAS_OP(sch, update_idle) && do_notify && !scx_rq_bypassing(rq))
771 		SCX_CALL_OP(sch, SCX_KF_REST, update_idle, rq, cpu_of(rq), idle);
772 }
773 
774 static void reset_idle_masks(struct sched_ext_ops *ops)
775 {
776 	int node;
777 
778 	/*
779 	 * Consider all online cpus idle. Should converge to the actual state
780 	 * quickly.
781 	 */
782 	if (!(ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)) {
783 		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->cpu, cpu_online_mask);
784 		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->smt, cpu_online_mask);
785 		return;
786 	}
787 
788 	for_each_node(node) {
789 		const struct cpumask *node_mask = cpumask_of_node(node);
790 
791 		cpumask_and(idle_cpumask(node)->cpu, cpu_online_mask, node_mask);
792 		cpumask_and(idle_cpumask(node)->smt, cpu_online_mask, node_mask);
793 	}
794 }
795 
796 void scx_idle_enable(struct sched_ext_ops *ops)
797 {
798 	if (!ops->update_idle || (ops->flags & SCX_OPS_KEEP_BUILTIN_IDLE))
799 		static_branch_enable_cpuslocked(&scx_builtin_idle_enabled);
800 	else
801 		static_branch_disable_cpuslocked(&scx_builtin_idle_enabled);
802 
803 	if (ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)
804 		static_branch_enable_cpuslocked(&scx_builtin_idle_per_node);
805 	else
806 		static_branch_disable_cpuslocked(&scx_builtin_idle_per_node);
807 
808 	reset_idle_masks(ops);
809 }
810 
811 void scx_idle_disable(void)
812 {
813 	static_branch_disable(&scx_builtin_idle_enabled);
814 	static_branch_disable(&scx_builtin_idle_per_node);
815 }
816 
817 /********************************************************************************
818  * Helpers that can be called from the BPF scheduler.
819  */
820 
821 static int validate_node(struct scx_sched *sch, int node)
822 {
823 	if (!static_branch_likely(&scx_builtin_idle_per_node)) {
824 		scx_error(sch, "per-node idle tracking is disabled");
825 		return -EOPNOTSUPP;
826 	}
827 
828 	/* Return no entry for NUMA_NO_NODE (not a critical scx error) */
829 	if (node == NUMA_NO_NODE)
830 		return -ENOENT;
831 
832 	/* Make sure node is in a valid range */
833 	if (node < 0 || node >= nr_node_ids) {
834 		scx_error(sch, "invalid node %d", node);
835 		return -EINVAL;
836 	}
837 
838 	/* Make sure the node is part of the set of possible nodes */
839 	if (!node_possible(node)) {
840 		scx_error(sch, "unavailable node %d", node);
841 		return -EINVAL;
842 	}
843 
844 	return node;
845 }
846 
847 __bpf_kfunc_start_defs();
848 
849 static bool check_builtin_idle_enabled(struct scx_sched *sch)
850 {
851 	if (static_branch_likely(&scx_builtin_idle_enabled))
852 		return true;
853 
854 	scx_error(sch, "built-in idle tracking is disabled");
855 	return false;
856 }
857 
858 /*
859  * Determine whether @p is a migration-disabled task in the context of BPF
860  * code.
861  *
862  * We can't simply check whether @p->migration_disabled is set in a
863  * sched_ext callback, because the BPF prolog (__bpf_prog_enter) may disable
864  * migration for the current task while running BPF code.
865  *
866  * Since the BPF prolog calls migrate_disable() only when CONFIG_PREEMPT_RCU
867  * is enabled (via rcu_read_lock_dont_migrate()), migration_disabled == 1 for
868  * the current task is ambiguous only in that case: it could be from the BPF
869  * prolog rather than a real migrate_disable() call.
870  *
871  * Without CONFIG_PREEMPT_RCU, the BPF prolog never calls migrate_disable(),
872  * so migration_disabled == 1 always means the task is truly
873  * migration-disabled.
874  *
875  * Therefore, when migration_disabled == 1 and CONFIG_PREEMPT_RCU is enabled,
876  * check whether @p is the current task or not: if it is, then migration was
877  * not disabled before entering the callback, otherwise migration was disabled.
878  *
879  * Returns true if @p is migration-disabled, false otherwise.
880  */
881 static bool is_bpf_migration_disabled(const struct task_struct *p)
882 {
883 	if (p->migration_disabled == 1) {
884 		if (IS_ENABLED(CONFIG_PREEMPT_RCU))
885 			return p != current;
886 		return true;
887 	}
888 	return p->migration_disabled;
889 }
890 
891 static s32 select_cpu_from_kfunc(struct scx_sched *sch, struct task_struct *p,
892 				 s32 prev_cpu, u64 wake_flags,
893 				 const struct cpumask *allowed, u64 flags)
894 {
895 	struct rq *rq;
896 	struct rq_flags rf;
897 	s32 cpu;
898 
899 	if (!ops_cpu_valid(sch, prev_cpu, NULL))
900 		return -EINVAL;
901 
902 	if (!check_builtin_idle_enabled(sch))
903 		return -EBUSY;
904 
905 	/*
906 	 * If called from an unlocked context, acquire the task's rq lock,
907 	 * so that we can safely access p->cpus_ptr and p->nr_cpus_allowed.
908 	 *
909 	 * Otherwise, allow to use this kfunc only from ops.select_cpu()
910 	 * and ops.select_enqueue().
911 	 */
912 	if (scx_kf_allowed_if_unlocked()) {
913 		rq = task_rq_lock(p, &rf);
914 	} else {
915 		if (!scx_kf_allowed(sch, SCX_KF_SELECT_CPU | SCX_KF_ENQUEUE))
916 			return -EPERM;
917 		rq = scx_locked_rq();
918 	}
919 
920 	/*
921 	 * Validate locking correctness to access p->cpus_ptr and
922 	 * p->nr_cpus_allowed: if we're holding an rq lock, we're safe;
923 	 * otherwise, assert that p->pi_lock is held.
924 	 */
925 	if (!rq)
926 		lockdep_assert_held(&p->pi_lock);
927 
928 	/*
929 	 * This may also be called from ops.enqueue(), so we need to handle
930 	 * per-CPU tasks as well. For these tasks, we can skip all idle CPU
931 	 * selection optimizations and simply check whether the previously
932 	 * used CPU is idle and within the allowed cpumask.
933 	 */
934 	if (p->nr_cpus_allowed == 1 || is_bpf_migration_disabled(p)) {
935 		if (cpumask_test_cpu(prev_cpu, allowed ?: p->cpus_ptr) &&
936 		    scx_idle_test_and_clear_cpu(prev_cpu))
937 			cpu = prev_cpu;
938 		else
939 			cpu = -EBUSY;
940 	} else {
941 		cpu = scx_select_cpu_dfl(p, prev_cpu, wake_flags,
942 					 allowed ?: p->cpus_ptr, flags);
943 	}
944 
945 	if (scx_kf_allowed_if_unlocked())
946 		task_rq_unlock(rq, p, &rf);
947 
948 	return cpu;
949 }
950 
951 /**
952  * scx_bpf_cpu_node - Return the NUMA node the given @cpu belongs to, or
953  *		      trigger an error if @cpu is invalid
954  * @cpu: target CPU
955  */
956 __bpf_kfunc int scx_bpf_cpu_node(s32 cpu)
957 {
958 	struct scx_sched *sch;
959 
960 	guard(rcu)();
961 
962 	sch = rcu_dereference(scx_root);
963 	if (unlikely(!sch) || !ops_cpu_valid(sch, cpu, NULL))
964 		return NUMA_NO_NODE;
965 	return cpu_to_node(cpu);
966 }
967 
968 /**
969  * scx_bpf_select_cpu_dfl - The default implementation of ops.select_cpu()
970  * @p: task_struct to select a CPU for
971  * @prev_cpu: CPU @p was on previously
972  * @wake_flags: %SCX_WAKE_* flags
973  * @is_idle: out parameter indicating whether the returned CPU is idle
974  *
975  * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
976  * context such as a BPF test_run() call, as long as built-in CPU selection
977  * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
978  * is set.
979  *
980  * Returns the picked CPU with *@is_idle indicating whether the picked CPU is
981  * currently idle and thus a good candidate for direct dispatching.
982  */
983 __bpf_kfunc s32 scx_bpf_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
984 				       u64 wake_flags, bool *is_idle)
985 {
986 	struct scx_sched *sch;
987 	s32 cpu;
988 
989 	guard(rcu)();
990 
991 	sch = rcu_dereference(scx_root);
992 	if (unlikely(!sch))
993 		return -ENODEV;
994 
995 	cpu = select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags, NULL, 0);
996 	if (cpu >= 0) {
997 		*is_idle = true;
998 		return cpu;
999 	}
1000 	*is_idle = false;
1001 	return prev_cpu;
1002 }
1003 
1004 struct scx_bpf_select_cpu_and_args {
1005 	/* @p and @cpus_allowed can't be packed together as KF_RCU is not transitive */
1006 	s32			prev_cpu;
1007 	u64			wake_flags;
1008 	u64			flags;
1009 };
1010 
1011 /**
1012  * __scx_bpf_select_cpu_and - Arg-wrapped CPU selection with cpumask
1013  * @p: task_struct to select a CPU for
1014  * @cpus_allowed: cpumask of allowed CPUs
1015  * @args: struct containing the rest of the arguments
1016  *       @args->prev_cpu: CPU @p was on previously
1017  *       @args->wake_flags: %SCX_WAKE_* flags
1018  *       @args->flags: %SCX_PICK_IDLE* flags
1019  *
1020  * Wrapper kfunc that takes arguments via struct to work around BPF's 5 argument
1021  * limit. BPF programs should use scx_bpf_select_cpu_and() which is provided
1022  * as an inline wrapper in common.bpf.h.
1023  *
1024  * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
1025  * context such as a BPF test_run() call, as long as built-in CPU selection
1026  * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
1027  * is set.
1028  *
1029  * @p, @args->prev_cpu and @args->wake_flags match ops.select_cpu().
1030  *
1031  * Returns the selected idle CPU, which will be automatically awakened upon
1032  * returning from ops.select_cpu() and can be used for direct dispatch, or
1033  * a negative value if no idle CPU is available.
1034  */
1035 __bpf_kfunc s32
1036 __scx_bpf_select_cpu_and(struct task_struct *p, const struct cpumask *cpus_allowed,
1037 			 struct scx_bpf_select_cpu_and_args *args)
1038 {
1039 	struct scx_sched *sch;
1040 
1041 	guard(rcu)();
1042 
1043 	sch = rcu_dereference(scx_root);
1044 	if (unlikely(!sch))
1045 		return -ENODEV;
1046 
1047 	return select_cpu_from_kfunc(sch, p, args->prev_cpu, args->wake_flags,
1048 				     cpus_allowed, args->flags);
1049 }
1050 
1051 /*
1052  * COMPAT: Will be removed in v6.22.
1053  */
1054 __bpf_kfunc s32 scx_bpf_select_cpu_and(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
1055 				       const struct cpumask *cpus_allowed, u64 flags)
1056 {
1057 	struct scx_sched *sch;
1058 
1059 	guard(rcu)();
1060 
1061 	sch = rcu_dereference(scx_root);
1062 	if (unlikely(!sch))
1063 		return -ENODEV;
1064 
1065 	return select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags,
1066 				     cpus_allowed, flags);
1067 }
1068 
1069 /**
1070  * scx_bpf_get_idle_cpumask_node - Get a referenced kptr to the
1071  * idle-tracking per-CPU cpumask of a target NUMA node.
1072  * @node: target NUMA node
1073  *
1074  * Returns an empty cpumask if idle tracking is not enabled, if @node is
1075  * not valid, or running on a UP kernel. In this case the actual error will
1076  * be reported to the BPF scheduler via scx_error().
1077  */
1078 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask_node(int node)
1079 {
1080 	struct scx_sched *sch;
1081 
1082 	guard(rcu)();
1083 
1084 	sch = rcu_dereference(scx_root);
1085 	if (unlikely(!sch))
1086 		return cpu_none_mask;
1087 
1088 	node = validate_node(sch, node);
1089 	if (node < 0)
1090 		return cpu_none_mask;
1091 
1092 	return idle_cpumask(node)->cpu;
1093 }
1094 
1095 /**
1096  * scx_bpf_get_idle_cpumask - Get a referenced kptr to the idle-tracking
1097  * per-CPU cpumask.
1098  *
1099  * Returns an empty mask if idle tracking is not enabled, or running on a
1100  * UP kernel.
1101  */
1102 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(void)
1103 {
1104 	struct scx_sched *sch;
1105 
1106 	guard(rcu)();
1107 
1108 	sch = rcu_dereference(scx_root);
1109 	if (unlikely(!sch))
1110 		return cpu_none_mask;
1111 
1112 	if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1113 		scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1114 		return cpu_none_mask;
1115 	}
1116 
1117 	if (!check_builtin_idle_enabled(sch))
1118 		return cpu_none_mask;
1119 
1120 	return idle_cpumask(NUMA_NO_NODE)->cpu;
1121 }
1122 
1123 /**
1124  * scx_bpf_get_idle_smtmask_node - Get a referenced kptr to the
1125  * idle-tracking, per-physical-core cpumask of a target NUMA node. Can be
1126  * used to determine if an entire physical core is free.
1127  * @node: target NUMA node
1128  *
1129  * Returns an empty cpumask if idle tracking is not enabled, if @node is
1130  * not valid, or running on a UP kernel. In this case the actual error will
1131  * be reported to the BPF scheduler via scx_error().
1132  */
1133 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask_node(int node)
1134 {
1135 	struct scx_sched *sch;
1136 
1137 	guard(rcu)();
1138 
1139 	sch = rcu_dereference(scx_root);
1140 	if (unlikely(!sch))
1141 		return cpu_none_mask;
1142 
1143 	node = validate_node(sch, node);
1144 	if (node < 0)
1145 		return cpu_none_mask;
1146 
1147 	if (sched_smt_active())
1148 		return idle_cpumask(node)->smt;
1149 	else
1150 		return idle_cpumask(node)->cpu;
1151 }
1152 
1153 /**
1154  * scx_bpf_get_idle_smtmask - Get a referenced kptr to the idle-tracking,
1155  * per-physical-core cpumask. Can be used to determine if an entire physical
1156  * core is free.
1157  *
1158  * Returns an empty mask if idle tracking is not enabled, or running on a
1159  * UP kernel.
1160  */
1161 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask(void)
1162 {
1163 	struct scx_sched *sch;
1164 
1165 	guard(rcu)();
1166 
1167 	sch = rcu_dereference(scx_root);
1168 	if (unlikely(!sch))
1169 		return cpu_none_mask;
1170 
1171 	if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1172 		scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1173 		return cpu_none_mask;
1174 	}
1175 
1176 	if (!check_builtin_idle_enabled(sch))
1177 		return cpu_none_mask;
1178 
1179 	if (sched_smt_active())
1180 		return idle_cpumask(NUMA_NO_NODE)->smt;
1181 	else
1182 		return idle_cpumask(NUMA_NO_NODE)->cpu;
1183 }
1184 
1185 /**
1186  * scx_bpf_put_idle_cpumask - Release a previously acquired referenced kptr to
1187  * either the percpu, or SMT idle-tracking cpumask.
1188  * @idle_mask: &cpumask to use
1189  */
1190 __bpf_kfunc void scx_bpf_put_idle_cpumask(const struct cpumask *idle_mask)
1191 {
1192 	/*
1193 	 * Empty function body because we aren't actually acquiring or releasing
1194 	 * a reference to a global idle cpumask, which is read-only in the
1195 	 * caller and is never released. The acquire / release semantics here
1196 	 * are just used to make the cpumask a trusted pointer in the caller.
1197 	 */
1198 }
1199 
1200 /**
1201  * scx_bpf_test_and_clear_cpu_idle - Test and clear @cpu's idle state
1202  * @cpu: cpu to test and clear idle for
1203  *
1204  * Returns %true if @cpu was idle and its idle state was successfully cleared.
1205  * %false otherwise.
1206  *
1207  * Unavailable if ops.update_idle() is implemented and
1208  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1209  */
1210 __bpf_kfunc bool scx_bpf_test_and_clear_cpu_idle(s32 cpu)
1211 {
1212 	struct scx_sched *sch;
1213 
1214 	guard(rcu)();
1215 
1216 	sch = rcu_dereference(scx_root);
1217 	if (unlikely(!sch))
1218 		return false;
1219 
1220 	if (!check_builtin_idle_enabled(sch))
1221 		return false;
1222 
1223 	if (!ops_cpu_valid(sch, cpu, NULL))
1224 		return false;
1225 
1226 	return scx_idle_test_and_clear_cpu(cpu);
1227 }
1228 
1229 /**
1230  * scx_bpf_pick_idle_cpu_node - Pick and claim an idle cpu from @node
1231  * @cpus_allowed: Allowed cpumask
1232  * @node: target NUMA node
1233  * @flags: %SCX_PICK_IDLE_* flags
1234  *
1235  * Pick and claim an idle cpu in @cpus_allowed from the NUMA node @node.
1236  *
1237  * Returns the picked idle cpu number on success, or -%EBUSY if no matching
1238  * cpu was found.
1239  *
1240  * The search starts from @node and proceeds to other online NUMA nodes in
1241  * order of increasing distance (unless SCX_PICK_IDLE_IN_NODE is specified,
1242  * in which case the search is limited to the target @node).
1243  *
1244  * Always returns an error if ops.update_idle() is implemented and
1245  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set, or if
1246  * %SCX_OPS_BUILTIN_IDLE_PER_NODE is not set.
1247  */
1248 __bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(const struct cpumask *cpus_allowed,
1249 					   int node, u64 flags)
1250 {
1251 	struct scx_sched *sch;
1252 
1253 	guard(rcu)();
1254 
1255 	sch = rcu_dereference(scx_root);
1256 	if (unlikely(!sch))
1257 		return -ENODEV;
1258 
1259 	node = validate_node(sch, node);
1260 	if (node < 0)
1261 		return node;
1262 
1263 	return scx_pick_idle_cpu(cpus_allowed, node, flags);
1264 }
1265 
1266 /**
1267  * scx_bpf_pick_idle_cpu - Pick and claim an idle cpu
1268  * @cpus_allowed: Allowed cpumask
1269  * @flags: %SCX_PICK_IDLE_CPU_* flags
1270  *
1271  * Pick and claim an idle cpu in @cpus_allowed. Returns the picked idle cpu
1272  * number on success. -%EBUSY if no matching cpu was found.
1273  *
1274  * Idle CPU tracking may race against CPU scheduling state transitions. For
1275  * example, this function may return -%EBUSY as CPUs are transitioning into the
1276  * idle state. If the caller then assumes that there will be dispatch events on
1277  * the CPUs as they were all busy, the scheduler may end up stalling with CPUs
1278  * idling while there are pending tasks. Use scx_bpf_pick_any_cpu() and
1279  * scx_bpf_kick_cpu() to guarantee that there will be at least one dispatch
1280  * event in the near future.
1281  *
1282  * Unavailable if ops.update_idle() is implemented and
1283  * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1284  *
1285  * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1286  * scx_bpf_pick_idle_cpu_node() instead.
1287  */
1288 __bpf_kfunc s32 scx_bpf_pick_idle_cpu(const struct cpumask *cpus_allowed,
1289 				      u64 flags)
1290 {
1291 	struct scx_sched *sch;
1292 
1293 	guard(rcu)();
1294 
1295 	sch = rcu_dereference(scx_root);
1296 	if (unlikely(!sch))
1297 		return -ENODEV;
1298 
1299 	if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1300 		scx_error(sch, "per-node idle tracking is enabled");
1301 		return -EBUSY;
1302 	}
1303 
1304 	if (!check_builtin_idle_enabled(sch))
1305 		return -EBUSY;
1306 
1307 	return scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1308 }
1309 
1310 /**
1311  * scx_bpf_pick_any_cpu_node - Pick and claim an idle cpu if available
1312  *			       or pick any CPU from @node
1313  * @cpus_allowed: Allowed cpumask
1314  * @node: target NUMA node
1315  * @flags: %SCX_PICK_IDLE_CPU_* flags
1316  *
1317  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1318  * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1319  * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1320  * empty.
1321  *
1322  * The search starts from @node and proceeds to other online NUMA nodes in
1323  * order of increasing distance (unless %SCX_PICK_IDLE_IN_NODE is specified,
1324  * in which case the search is limited to the target @node, regardless of
1325  * the CPU idle state).
1326  *
1327  * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1328  * set, this function can't tell which CPUs are idle and will always pick any
1329  * CPU.
1330  */
1331 __bpf_kfunc s32 scx_bpf_pick_any_cpu_node(const struct cpumask *cpus_allowed,
1332 					  int node, u64 flags)
1333 {
1334 	struct scx_sched *sch;
1335 	s32 cpu;
1336 
1337 	guard(rcu)();
1338 
1339 	sch = rcu_dereference(scx_root);
1340 	if (unlikely(!sch))
1341 		return -ENODEV;
1342 
1343 	node = validate_node(sch, node);
1344 	if (node < 0)
1345 		return node;
1346 
1347 	cpu = scx_pick_idle_cpu(cpus_allowed, node, flags);
1348 	if (cpu >= 0)
1349 		return cpu;
1350 
1351 	if (flags & SCX_PICK_IDLE_IN_NODE)
1352 		cpu = cpumask_any_and_distribute(cpumask_of_node(node), cpus_allowed);
1353 	else
1354 		cpu = cpumask_any_distribute(cpus_allowed);
1355 	if (cpu < nr_cpu_ids)
1356 		return cpu;
1357 	else
1358 		return -EBUSY;
1359 }
1360 
1361 /**
1362  * scx_bpf_pick_any_cpu - Pick and claim an idle cpu if available or pick any CPU
1363  * @cpus_allowed: Allowed cpumask
1364  * @flags: %SCX_PICK_IDLE_CPU_* flags
1365  *
1366  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1367  * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1368  * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1369  * empty.
1370  *
1371  * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1372  * set, this function can't tell which CPUs are idle and will always pick any
1373  * CPU.
1374  *
1375  * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1376  * scx_bpf_pick_any_cpu_node() instead.
1377  */
1378 __bpf_kfunc s32 scx_bpf_pick_any_cpu(const struct cpumask *cpus_allowed,
1379 				     u64 flags)
1380 {
1381 	struct scx_sched *sch;
1382 	s32 cpu;
1383 
1384 	guard(rcu)();
1385 
1386 	sch = rcu_dereference(scx_root);
1387 	if (unlikely(!sch))
1388 		return -ENODEV;
1389 
1390 	if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1391 		scx_error(sch, "per-node idle tracking is enabled");
1392 		return -EBUSY;
1393 	}
1394 
1395 	if (static_branch_likely(&scx_builtin_idle_enabled)) {
1396 		cpu = scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1397 		if (cpu >= 0)
1398 			return cpu;
1399 	}
1400 
1401 	cpu = cpumask_any_distribute(cpus_allowed);
1402 	if (cpu < nr_cpu_ids)
1403 		return cpu;
1404 	else
1405 		return -EBUSY;
1406 }
1407 
1408 __bpf_kfunc_end_defs();
1409 
1410 BTF_KFUNCS_START(scx_kfunc_ids_idle)
1411 BTF_ID_FLAGS(func, scx_bpf_cpu_node)
1412 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask_node, KF_ACQUIRE)
1413 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask, KF_ACQUIRE)
1414 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask_node, KF_ACQUIRE)
1415 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask, KF_ACQUIRE)
1416 BTF_ID_FLAGS(func, scx_bpf_put_idle_cpumask, KF_RELEASE)
1417 BTF_ID_FLAGS(func, scx_bpf_test_and_clear_cpu_idle)
1418 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu_node, KF_RCU)
1419 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu, KF_RCU)
1420 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu_node, KF_RCU)
1421 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu, KF_RCU)
1422 BTF_ID_FLAGS(func, __scx_bpf_select_cpu_and, KF_RCU)
1423 BTF_ID_FLAGS(func, scx_bpf_select_cpu_and, KF_RCU)
1424 BTF_ID_FLAGS(func, scx_bpf_select_cpu_dfl, KF_RCU)
1425 BTF_KFUNCS_END(scx_kfunc_ids_idle)
1426 
1427 static const struct btf_kfunc_id_set scx_kfunc_set_idle = {
1428 	.owner			= THIS_MODULE,
1429 	.set			= &scx_kfunc_ids_idle,
1430 };
1431 
1432 int scx_idle_init(void)
1433 {
1434 	int ret;
1435 
1436 	ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_idle) ||
1437 	      register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_idle) ||
1438 	      register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_idle);
1439 
1440 	return ret;
1441 }
1442