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(struct scx_sched * sch,int node)822 static int validate_node(struct scx_sched *sch, int node)
823 {
824 if (!static_branch_likely(&scx_builtin_idle_per_node)) {
825 scx_error(sch, "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_error(sch, "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_error(sch, "unavailable node %d", node);
842 return -EINVAL;
843 }
844
845 return node;
846 }
847
848 __bpf_kfunc_start_defs();
849
check_builtin_idle_enabled(struct scx_sched * sch)850 static bool check_builtin_idle_enabled(struct scx_sched *sch)
851 {
852 if (static_branch_likely(&scx_builtin_idle_enabled))
853 return true;
854
855 scx_error(sch, "built-in idle tracking is disabled");
856 return false;
857 }
858
859 /*
860 * Determine whether @p is a migration-disabled task in the context of BPF
861 * code.
862 *
863 * We can't simply check whether @p->migration_disabled is set in a
864 * sched_ext callback, because migration is always disabled for the current
865 * task while running BPF code.
866 *
867 * The prolog (__bpf_prog_enter) and epilog (__bpf_prog_exit) respectively
868 * disable and re-enable migration. For this reason, the current task
869 * inside a sched_ext callback is always a migration-disabled task.
870 *
871 * Therefore, when @p->migration_disabled == 1, check whether @p is the
872 * current task or not: if it is, then migration was not disabled before
873 * entering the callback, otherwise migration was disabled.
874 *
875 * Returns true if @p is migration-disabled, false otherwise.
876 */
is_bpf_migration_disabled(const struct task_struct * p)877 static bool is_bpf_migration_disabled(const struct task_struct *p)
878 {
879 if (p->migration_disabled == 1)
880 return p != current;
881 else
882 return p->migration_disabled;
883 }
884
select_cpu_from_kfunc(struct scx_sched * sch,struct task_struct * p,s32 prev_cpu,u64 wake_flags,const struct cpumask * allowed,u64 flags)885 static s32 select_cpu_from_kfunc(struct scx_sched *sch, struct task_struct *p,
886 s32 prev_cpu, u64 wake_flags,
887 const struct cpumask *allowed, u64 flags)
888 {
889 struct rq *rq;
890 struct rq_flags rf;
891 s32 cpu;
892
893 if (!ops_cpu_valid(sch, prev_cpu, NULL))
894 return -EINVAL;
895
896 if (!check_builtin_idle_enabled(sch))
897 return -EBUSY;
898
899 /*
900 * If called from an unlocked context, acquire the task's rq lock,
901 * so that we can safely access p->cpus_ptr and p->nr_cpus_allowed.
902 *
903 * Otherwise, allow to use this kfunc only from ops.select_cpu()
904 * and ops.select_enqueue().
905 */
906 if (scx_kf_allowed_if_unlocked()) {
907 rq = task_rq_lock(p, &rf);
908 } else {
909 if (!scx_kf_allowed(sch, SCX_KF_SELECT_CPU | SCX_KF_ENQUEUE))
910 return -EPERM;
911 rq = scx_locked_rq();
912 }
913
914 /*
915 * Validate locking correctness to access p->cpus_ptr and
916 * p->nr_cpus_allowed: if we're holding an rq lock, we're safe;
917 * otherwise, assert that p->pi_lock is held.
918 */
919 if (!rq)
920 lockdep_assert_held(&p->pi_lock);
921
922 /*
923 * This may also be called from ops.enqueue(), so we need to handle
924 * per-CPU tasks as well. For these tasks, we can skip all idle CPU
925 * selection optimizations and simply check whether the previously
926 * used CPU is idle and within the allowed cpumask.
927 */
928 if (p->nr_cpus_allowed == 1 || is_bpf_migration_disabled(p)) {
929 if (cpumask_test_cpu(prev_cpu, allowed ?: p->cpus_ptr) &&
930 scx_idle_test_and_clear_cpu(prev_cpu))
931 cpu = prev_cpu;
932 else
933 cpu = -EBUSY;
934 } else {
935 cpu = scx_select_cpu_dfl(p, prev_cpu, wake_flags,
936 allowed ?: p->cpus_ptr, flags);
937 }
938
939 if (scx_kf_allowed_if_unlocked())
940 task_rq_unlock(rq, p, &rf);
941
942 return cpu;
943 }
944
945 /**
946 * scx_bpf_cpu_node - Return the NUMA node the given @cpu belongs to, or
947 * trigger an error if @cpu is invalid
948 * @cpu: target CPU
949 */
scx_bpf_cpu_node(s32 cpu)950 __bpf_kfunc int scx_bpf_cpu_node(s32 cpu)
951 {
952 struct scx_sched *sch;
953
954 guard(rcu)();
955
956 sch = rcu_dereference(scx_root);
957 if (unlikely(!sch) || !ops_cpu_valid(sch, cpu, NULL))
958 return NUMA_NO_NODE;
959 return cpu_to_node(cpu);
960 }
961
962 /**
963 * scx_bpf_select_cpu_dfl - The default implementation of ops.select_cpu()
964 * @p: task_struct to select a CPU for
965 * @prev_cpu: CPU @p was on previously
966 * @wake_flags: %SCX_WAKE_* flags
967 * @is_idle: out parameter indicating whether the returned CPU is idle
968 *
969 * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
970 * context such as a BPF test_run() call, as long as built-in CPU selection
971 * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
972 * is set.
973 *
974 * Returns the picked CPU with *@is_idle indicating whether the picked CPU is
975 * currently idle and thus a good candidate for direct dispatching.
976 */
scx_bpf_select_cpu_dfl(struct task_struct * p,s32 prev_cpu,u64 wake_flags,bool * is_idle)977 __bpf_kfunc s32 scx_bpf_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
978 u64 wake_flags, bool *is_idle)
979 {
980 struct scx_sched *sch;
981 s32 cpu;
982
983 guard(rcu)();
984
985 sch = rcu_dereference(scx_root);
986 if (unlikely(!sch))
987 return -ENODEV;
988
989 cpu = select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags, NULL, 0);
990 if (cpu >= 0) {
991 *is_idle = true;
992 return cpu;
993 }
994 *is_idle = false;
995 return prev_cpu;
996 }
997
998 /**
999 * scx_bpf_select_cpu_and - Pick an idle CPU usable by task @p,
1000 * prioritizing those in @cpus_allowed
1001 * @p: task_struct to select a CPU for
1002 * @prev_cpu: CPU @p was on previously
1003 * @wake_flags: %SCX_WAKE_* flags
1004 * @cpus_allowed: cpumask of allowed CPUs
1005 * @flags: %SCX_PICK_IDLE* flags
1006 *
1007 * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked
1008 * context such as a BPF test_run() call, as long as built-in CPU selection
1009 * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE
1010 * is set.
1011 *
1012 * @p, @prev_cpu and @wake_flags match ops.select_cpu().
1013 *
1014 * Returns the selected idle CPU, which will be automatically awakened upon
1015 * returning from ops.select_cpu() and can be used for direct dispatch, or
1016 * a negative value if no idle CPU is available.
1017 */
scx_bpf_select_cpu_and(struct task_struct * p,s32 prev_cpu,u64 wake_flags,const struct cpumask * cpus_allowed,u64 flags)1018 __bpf_kfunc s32 scx_bpf_select_cpu_and(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
1019 const struct cpumask *cpus_allowed, u64 flags)
1020 {
1021 struct scx_sched *sch;
1022
1023 guard(rcu)();
1024
1025 sch = rcu_dereference(scx_root);
1026 if (unlikely(!sch))
1027 return -ENODEV;
1028
1029 return select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags,
1030 cpus_allowed, flags);
1031 }
1032
1033 /**
1034 * scx_bpf_get_idle_cpumask_node - Get a referenced kptr to the
1035 * idle-tracking per-CPU cpumask of a target NUMA node.
1036 * @node: target NUMA node
1037 *
1038 * Returns an empty cpumask if idle tracking is not enabled, if @node is
1039 * not valid, or running on a UP kernel. In this case the actual error will
1040 * be reported to the BPF scheduler via scx_error().
1041 */
scx_bpf_get_idle_cpumask_node(int node)1042 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask_node(int node)
1043 {
1044 struct scx_sched *sch;
1045
1046 guard(rcu)();
1047
1048 sch = rcu_dereference(scx_root);
1049 if (unlikely(!sch))
1050 return cpu_none_mask;
1051
1052 node = validate_node(sch, node);
1053 if (node < 0)
1054 return cpu_none_mask;
1055
1056 return idle_cpumask(node)->cpu;
1057 }
1058
1059 /**
1060 * scx_bpf_get_idle_cpumask - Get a referenced kptr to the idle-tracking
1061 * per-CPU cpumask.
1062 *
1063 * Returns an empty mask if idle tracking is not enabled, or running on a
1064 * UP kernel.
1065 */
scx_bpf_get_idle_cpumask(void)1066 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(void)
1067 {
1068 struct scx_sched *sch;
1069
1070 guard(rcu)();
1071
1072 sch = rcu_dereference(scx_root);
1073 if (unlikely(!sch))
1074 return cpu_none_mask;
1075
1076 if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1077 scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1078 return cpu_none_mask;
1079 }
1080
1081 if (!check_builtin_idle_enabled(sch))
1082 return cpu_none_mask;
1083
1084 return idle_cpumask(NUMA_NO_NODE)->cpu;
1085 }
1086
1087 /**
1088 * scx_bpf_get_idle_smtmask_node - Get a referenced kptr to the
1089 * idle-tracking, per-physical-core cpumask of a target NUMA node. Can be
1090 * used to determine if an entire physical core is free.
1091 * @node: target NUMA node
1092 *
1093 * Returns an empty cpumask if idle tracking is not enabled, if @node is
1094 * not valid, or running on a UP kernel. In this case the actual error will
1095 * be reported to the BPF scheduler via scx_error().
1096 */
scx_bpf_get_idle_smtmask_node(int node)1097 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask_node(int node)
1098 {
1099 struct scx_sched *sch;
1100
1101 guard(rcu)();
1102
1103 sch = rcu_dereference(scx_root);
1104 if (unlikely(!sch))
1105 return cpu_none_mask;
1106
1107 node = validate_node(sch, node);
1108 if (node < 0)
1109 return cpu_none_mask;
1110
1111 if (sched_smt_active())
1112 return idle_cpumask(node)->smt;
1113 else
1114 return idle_cpumask(node)->cpu;
1115 }
1116
1117 /**
1118 * scx_bpf_get_idle_smtmask - Get a referenced kptr to the idle-tracking,
1119 * per-physical-core cpumask. Can be used to determine if an entire physical
1120 * core is free.
1121 *
1122 * Returns an empty mask if idle tracking is not enabled, or running on a
1123 * UP kernel.
1124 */
scx_bpf_get_idle_smtmask(void)1125 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask(void)
1126 {
1127 struct scx_sched *sch;
1128
1129 guard(rcu)();
1130
1131 sch = rcu_dereference(scx_root);
1132 if (unlikely(!sch))
1133 return cpu_none_mask;
1134
1135 if (static_branch_unlikely(&scx_builtin_idle_per_node)) {
1136 scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled");
1137 return cpu_none_mask;
1138 }
1139
1140 if (!check_builtin_idle_enabled(sch))
1141 return cpu_none_mask;
1142
1143 if (sched_smt_active())
1144 return idle_cpumask(NUMA_NO_NODE)->smt;
1145 else
1146 return idle_cpumask(NUMA_NO_NODE)->cpu;
1147 }
1148
1149 /**
1150 * scx_bpf_put_idle_cpumask - Release a previously acquired referenced kptr to
1151 * either the percpu, or SMT idle-tracking cpumask.
1152 * @idle_mask: &cpumask to use
1153 */
scx_bpf_put_idle_cpumask(const struct cpumask * idle_mask)1154 __bpf_kfunc void scx_bpf_put_idle_cpumask(const struct cpumask *idle_mask)
1155 {
1156 /*
1157 * Empty function body because we aren't actually acquiring or releasing
1158 * a reference to a global idle cpumask, which is read-only in the
1159 * caller and is never released. The acquire / release semantics here
1160 * are just used to make the cpumask a trusted pointer in the caller.
1161 */
1162 }
1163
1164 /**
1165 * scx_bpf_test_and_clear_cpu_idle - Test and clear @cpu's idle state
1166 * @cpu: cpu to test and clear idle for
1167 *
1168 * Returns %true if @cpu was idle and its idle state was successfully cleared.
1169 * %false otherwise.
1170 *
1171 * Unavailable if ops.update_idle() is implemented and
1172 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1173 */
scx_bpf_test_and_clear_cpu_idle(s32 cpu)1174 __bpf_kfunc bool scx_bpf_test_and_clear_cpu_idle(s32 cpu)
1175 {
1176 struct scx_sched *sch;
1177
1178 guard(rcu)();
1179
1180 sch = rcu_dereference(scx_root);
1181 if (unlikely(!sch))
1182 return false;
1183
1184 if (!check_builtin_idle_enabled(sch))
1185 return false;
1186
1187 if (!ops_cpu_valid(sch, cpu, NULL))
1188 return false;
1189
1190 return scx_idle_test_and_clear_cpu(cpu);
1191 }
1192
1193 /**
1194 * scx_bpf_pick_idle_cpu_node - Pick and claim an idle cpu from @node
1195 * @cpus_allowed: Allowed cpumask
1196 * @node: target NUMA node
1197 * @flags: %SCX_PICK_IDLE_* flags
1198 *
1199 * Pick and claim an idle cpu in @cpus_allowed from the NUMA node @node.
1200 *
1201 * Returns the picked idle cpu number on success, or -%EBUSY if no matching
1202 * cpu was found.
1203 *
1204 * The search starts from @node and proceeds to other online NUMA nodes in
1205 * order of increasing distance (unless SCX_PICK_IDLE_IN_NODE is specified,
1206 * in which case the search is limited to the target @node).
1207 *
1208 * Always returns an error if ops.update_idle() is implemented and
1209 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set, or if
1210 * %SCX_OPS_BUILTIN_IDLE_PER_NODE is not set.
1211 */
scx_bpf_pick_idle_cpu_node(const struct cpumask * cpus_allowed,int node,u64 flags)1212 __bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(const struct cpumask *cpus_allowed,
1213 int node, u64 flags)
1214 {
1215 struct scx_sched *sch;
1216
1217 guard(rcu)();
1218
1219 sch = rcu_dereference(scx_root);
1220 if (unlikely(!sch))
1221 return -ENODEV;
1222
1223 node = validate_node(sch, node);
1224 if (node < 0)
1225 return node;
1226
1227 return scx_pick_idle_cpu(cpus_allowed, node, flags);
1228 }
1229
1230 /**
1231 * scx_bpf_pick_idle_cpu - Pick and claim an idle cpu
1232 * @cpus_allowed: Allowed cpumask
1233 * @flags: %SCX_PICK_IDLE_CPU_* flags
1234 *
1235 * Pick and claim an idle cpu in @cpus_allowed. Returns the picked idle cpu
1236 * number on success. -%EBUSY if no matching cpu was found.
1237 *
1238 * Idle CPU tracking may race against CPU scheduling state transitions. For
1239 * example, this function may return -%EBUSY as CPUs are transitioning into the
1240 * idle state. If the caller then assumes that there will be dispatch events on
1241 * the CPUs as they were all busy, the scheduler may end up stalling with CPUs
1242 * idling while there are pending tasks. Use scx_bpf_pick_any_cpu() and
1243 * scx_bpf_kick_cpu() to guarantee that there will be at least one dispatch
1244 * event in the near future.
1245 *
1246 * Unavailable if ops.update_idle() is implemented and
1247 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set.
1248 *
1249 * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1250 * scx_bpf_pick_idle_cpu_node() instead.
1251 */
scx_bpf_pick_idle_cpu(const struct cpumask * cpus_allowed,u64 flags)1252 __bpf_kfunc s32 scx_bpf_pick_idle_cpu(const struct cpumask *cpus_allowed,
1253 u64 flags)
1254 {
1255 struct scx_sched *sch;
1256
1257 guard(rcu)();
1258
1259 sch = rcu_dereference(scx_root);
1260 if (unlikely(!sch))
1261 return -ENODEV;
1262
1263 if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1264 scx_error(sch, "per-node idle tracking is enabled");
1265 return -EBUSY;
1266 }
1267
1268 if (!check_builtin_idle_enabled(sch))
1269 return -EBUSY;
1270
1271 return scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1272 }
1273
1274 /**
1275 * scx_bpf_pick_any_cpu_node - Pick and claim an idle cpu if available
1276 * or pick any CPU from @node
1277 * @cpus_allowed: Allowed cpumask
1278 * @node: target NUMA node
1279 * @flags: %SCX_PICK_IDLE_CPU_* flags
1280 *
1281 * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1282 * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1283 * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1284 * empty.
1285 *
1286 * The search starts from @node and proceeds to other online NUMA nodes in
1287 * order of increasing distance (unless %SCX_PICK_IDLE_IN_NODE is specified,
1288 * in which case the search is limited to the target @node, regardless of
1289 * the CPU idle state).
1290 *
1291 * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1292 * set, this function can't tell which CPUs are idle and will always pick any
1293 * CPU.
1294 */
scx_bpf_pick_any_cpu_node(const struct cpumask * cpus_allowed,int node,u64 flags)1295 __bpf_kfunc s32 scx_bpf_pick_any_cpu_node(const struct cpumask *cpus_allowed,
1296 int node, u64 flags)
1297 {
1298 struct scx_sched *sch;
1299 s32 cpu;
1300
1301 guard(rcu)();
1302
1303 sch = rcu_dereference(scx_root);
1304 if (unlikely(!sch))
1305 return -ENODEV;
1306
1307 node = validate_node(sch, node);
1308 if (node < 0)
1309 return node;
1310
1311 cpu = scx_pick_idle_cpu(cpus_allowed, node, flags);
1312 if (cpu >= 0)
1313 return cpu;
1314
1315 if (flags & SCX_PICK_IDLE_IN_NODE)
1316 cpu = cpumask_any_and_distribute(cpumask_of_node(node), cpus_allowed);
1317 else
1318 cpu = cpumask_any_distribute(cpus_allowed);
1319 if (cpu < nr_cpu_ids)
1320 return cpu;
1321 else
1322 return -EBUSY;
1323 }
1324
1325 /**
1326 * scx_bpf_pick_any_cpu - Pick and claim an idle cpu if available or pick any CPU
1327 * @cpus_allowed: Allowed cpumask
1328 * @flags: %SCX_PICK_IDLE_CPU_* flags
1329 *
1330 * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
1331 * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu
1332 * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is
1333 * empty.
1334 *
1335 * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not
1336 * set, this function can't tell which CPUs are idle and will always pick any
1337 * CPU.
1338 *
1339 * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use
1340 * scx_bpf_pick_any_cpu_node() instead.
1341 */
scx_bpf_pick_any_cpu(const struct cpumask * cpus_allowed,u64 flags)1342 __bpf_kfunc s32 scx_bpf_pick_any_cpu(const struct cpumask *cpus_allowed,
1343 u64 flags)
1344 {
1345 struct scx_sched *sch;
1346 s32 cpu;
1347
1348 guard(rcu)();
1349
1350 sch = rcu_dereference(scx_root);
1351 if (unlikely(!sch))
1352 return -ENODEV;
1353
1354 if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
1355 scx_error(sch, "per-node idle tracking is enabled");
1356 return -EBUSY;
1357 }
1358
1359 if (static_branch_likely(&scx_builtin_idle_enabled)) {
1360 cpu = scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags);
1361 if (cpu >= 0)
1362 return cpu;
1363 }
1364
1365 cpu = cpumask_any_distribute(cpus_allowed);
1366 if (cpu < nr_cpu_ids)
1367 return cpu;
1368 else
1369 return -EBUSY;
1370 }
1371
1372 __bpf_kfunc_end_defs();
1373
1374 BTF_KFUNCS_START(scx_kfunc_ids_idle)
1375 BTF_ID_FLAGS(func, scx_bpf_cpu_node)
1376 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask_node, KF_ACQUIRE)
1377 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask, KF_ACQUIRE)
1378 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask_node, KF_ACQUIRE)
1379 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask, KF_ACQUIRE)
1380 BTF_ID_FLAGS(func, scx_bpf_put_idle_cpumask, KF_RELEASE)
1381 BTF_ID_FLAGS(func, scx_bpf_test_and_clear_cpu_idle)
1382 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu_node, KF_RCU)
1383 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu, KF_RCU)
1384 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu_node, KF_RCU)
1385 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu, KF_RCU)
1386 BTF_ID_FLAGS(func, scx_bpf_select_cpu_and, KF_RCU)
1387 BTF_ID_FLAGS(func, scx_bpf_select_cpu_dfl, KF_RCU)
1388 BTF_KFUNCS_END(scx_kfunc_ids_idle)
1389
1390 static const struct btf_kfunc_id_set scx_kfunc_set_idle = {
1391 .owner = THIS_MODULE,
1392 .set = &scx_kfunc_ids_idle,
1393 };
1394
scx_idle_init(void)1395 int scx_idle_init(void)
1396 {
1397 int ret;
1398
1399 ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_idle) ||
1400 register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_idle) ||
1401 register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_idle);
1402
1403 return ret;
1404 }
1405