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 overlap 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. Prefer @prev_cpu's SMT sibling: 428 * - if @prev_cpu is busy and no fully idle core is available, try to 429 * place the task on an idle SMT sibling of @prev_cpu; keeping the 430 * task on the same core makes migration cheaper, preserves L1 cache 431 * locality and reduces wakeup latency. 432 * 433 * 4. Pick a CPU within the same LLC (Last-Level Cache): 434 * - if the above conditions aren't met, pick a CPU that shares the same 435 * LLC, if the LLC domain is a subset of @cpus_allowed, to maintain 436 * cache locality. 437 * 438 * 5. Pick a CPU within the same NUMA node, if enabled: 439 * - choose a CPU from the same NUMA node, if the node cpumask is a 440 * subset of @cpus_allowed, to reduce memory access latency. 441 * 442 * 6. Pick any idle CPU within the @cpus_allowed domain. 443 * 444 * Step 4 and 5 are performed only if the system has, respectively, 445 * multiple LLCs / multiple NUMA nodes (see scx_selcpu_topo_llc and 446 * scx_selcpu_topo_numa) and they don't contain the same subset of CPUs. 447 * 448 * If %SCX_OPS_BUILTIN_IDLE_PER_NODE is enabled, the search will always 449 * begin in @prev_cpu's node and proceed to other nodes in order of 450 * increasing distance. 451 * 452 * Return the picked CPU if idle, or a negative value otherwise. 453 * 454 * NOTE: tasks that can only run on 1 CPU are excluded by this logic, because 455 * we never call ops.select_cpu() for them, see select_task_rq(). 456 */ 457 s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags, 458 const struct cpumask *cpus_allowed, u64 flags) 459 { 460 const struct cpumask *llc_cpus = NULL, *numa_cpus = NULL; 461 const struct cpumask *allowed = cpus_allowed ?: p->cpus_ptr; 462 int node = scx_cpu_node_if_enabled(prev_cpu); 463 bool is_prev_allowed; 464 s32 cpu; 465 466 preempt_disable(); 467 468 /* 469 * Check whether @prev_cpu is still within the allowed set. If not, 470 * we can still try selecting a nearby CPU. 471 */ 472 is_prev_allowed = cpumask_test_cpu(prev_cpu, allowed); 473 474 /* 475 * Determine the subset of CPUs usable by @p within @cpus_allowed. 476 */ 477 if (allowed != p->cpus_ptr) { 478 struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_idle_cpumask); 479 480 if (task_affinity_all(p)) { 481 allowed = cpus_allowed; 482 } else if (cpumask_and(local_cpus, cpus_allowed, p->cpus_ptr)) { 483 allowed = local_cpus; 484 } else { 485 cpu = -EBUSY; 486 goto out_enable; 487 } 488 } 489 490 /* 491 * This is necessary to protect llc_cpus. 492 */ 493 rcu_read_lock(); 494 495 /* 496 * Determine the subset of CPUs that the task can use in its 497 * current LLC and node. 498 * 499 * If the task can run on all CPUs, use the node and LLC cpumasks 500 * directly. 501 */ 502 if (static_branch_maybe(CONFIG_NUMA, &scx_selcpu_topo_numa)) { 503 struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_numa_idle_cpumask); 504 const struct cpumask *cpus = numa_span(prev_cpu); 505 506 if (allowed == p->cpus_ptr && task_affinity_all(p)) 507 numa_cpus = cpus; 508 else if (cpus && cpumask_and(local_cpus, allowed, cpus)) 509 numa_cpus = local_cpus; 510 } 511 512 if (static_branch_maybe(CONFIG_SCHED_MC, &scx_selcpu_topo_llc)) { 513 struct cpumask *local_cpus = this_cpu_cpumask_var_ptr(local_llc_idle_cpumask); 514 const struct cpumask *cpus = llc_span(prev_cpu); 515 516 if (allowed == p->cpus_ptr && task_affinity_all(p)) 517 llc_cpus = cpus; 518 else if (cpus && cpumask_and(local_cpus, allowed, cpus)) 519 llc_cpus = local_cpus; 520 } 521 522 /* 523 * If WAKE_SYNC, try to migrate the wakee to the waker's CPU. 524 */ 525 if (wake_flags & SCX_WAKE_SYNC) { 526 int waker_node; 527 528 /* 529 * If the waker's CPU is cache affine and prev_cpu is idle, 530 * then avoid a migration. 531 */ 532 cpu = smp_processor_id(); 533 if (is_prev_allowed && cpus_share_cache(cpu, prev_cpu) && 534 scx_idle_test_and_clear_cpu(prev_cpu)) { 535 cpu = prev_cpu; 536 goto out_unlock; 537 } 538 539 /* 540 * If the waker's local DSQ is empty, and the system is under 541 * utilized, try to wake up @p to the local DSQ of the waker. 542 * 543 * Checking only for an empty local DSQ is insufficient as it 544 * could give the wakee an unfair advantage when the system is 545 * oversaturated. 546 * 547 * Checking only for the presence of idle CPUs is also 548 * insufficient as the local DSQ of the waker could have tasks 549 * piled up on it even if there is an idle core elsewhere on 550 * the system. 551 */ 552 waker_node = scx_cpu_node_if_enabled(cpu); 553 if (!(current->flags & PF_EXITING) && 554 cpu_rq(cpu)->scx.local_dsq.nr == 0 && 555 (!(flags & SCX_PICK_IDLE_IN_NODE) || (waker_node == node)) && 556 !cpumask_empty(idle_cpumask(waker_node)->cpu)) { 557 if (cpumask_test_cpu(cpu, allowed)) 558 goto out_unlock; 559 } 560 } 561 562 /* 563 * If CPU has SMT, any wholly idle CPU is likely a better pick than 564 * partially idle @prev_cpu. 565 */ 566 if (sched_smt_active()) { 567 /* 568 * Keep using @prev_cpu if it's part of a fully idle core. 569 */ 570 if (is_prev_allowed && 571 cpumask_test_cpu(prev_cpu, idle_cpumask(node)->smt) && 572 scx_idle_test_and_clear_cpu(prev_cpu)) { 573 cpu = prev_cpu; 574 goto out_unlock; 575 } 576 577 /* 578 * Search for any fully idle core in the same LLC domain. 579 */ 580 if (llc_cpus) { 581 cpu = pick_idle_cpu_in_node(llc_cpus, node, SCX_PICK_IDLE_CORE); 582 if (cpu >= 0) 583 goto out_unlock; 584 } 585 586 /* 587 * Search for any fully idle core in the same NUMA node. 588 */ 589 if (numa_cpus) { 590 cpu = pick_idle_cpu_in_node(numa_cpus, node, SCX_PICK_IDLE_CORE); 591 if (cpu >= 0) 592 goto out_unlock; 593 } 594 595 /* 596 * Search for any full-idle core usable by the task. 597 * 598 * If the node-aware idle CPU selection policy is enabled 599 * (%SCX_OPS_BUILTIN_IDLE_PER_NODE), the search will always 600 * begin in prev_cpu's node and proceed to other nodes in 601 * order of increasing distance. 602 */ 603 cpu = scx_pick_idle_cpu(allowed, node, flags | SCX_PICK_IDLE_CORE); 604 if (cpu >= 0) 605 goto out_unlock; 606 607 /* 608 * Give up if we're strictly looking for a full-idle SMT 609 * core. 610 */ 611 if (flags & SCX_PICK_IDLE_CORE) { 612 cpu = -EBUSY; 613 goto out_unlock; 614 } 615 } 616 617 /* 618 * Use @prev_cpu if it's idle. 619 */ 620 if (is_prev_allowed && scx_idle_test_and_clear_cpu(prev_cpu)) { 621 cpu = prev_cpu; 622 goto out_unlock; 623 } 624 625 #ifdef CONFIG_SCHED_SMT 626 /* 627 * Use @prev_cpu's sibling if it's idle. 628 */ 629 if (sched_smt_active()) { 630 for_each_cpu_and(cpu, cpu_smt_mask(prev_cpu), allowed) { 631 if (cpu == prev_cpu) 632 continue; 633 if (scx_idle_test_and_clear_cpu(cpu)) 634 goto out_unlock; 635 } 636 } 637 #endif 638 639 /* 640 * Search for any idle CPU in the same LLC domain. 641 */ 642 if (llc_cpus) { 643 cpu = pick_idle_cpu_in_node(llc_cpus, node, 0); 644 if (cpu >= 0) 645 goto out_unlock; 646 } 647 648 /* 649 * Search for any idle CPU in the same NUMA node. 650 */ 651 if (numa_cpus) { 652 cpu = pick_idle_cpu_in_node(numa_cpus, node, 0); 653 if (cpu >= 0) 654 goto out_unlock; 655 } 656 657 /* 658 * Search for any idle CPU usable by the task. 659 * 660 * If the node-aware idle CPU selection policy is enabled 661 * (%SCX_OPS_BUILTIN_IDLE_PER_NODE), the search will always begin 662 * in prev_cpu's node and proceed to other nodes in order of 663 * increasing distance. 664 */ 665 cpu = scx_pick_idle_cpu(allowed, node, flags); 666 667 out_unlock: 668 rcu_read_unlock(); 669 out_enable: 670 preempt_enable(); 671 672 return cpu; 673 } 674 675 /* 676 * Initialize global and per-node idle cpumasks. 677 */ 678 void scx_idle_init_masks(void) 679 { 680 int i; 681 682 /* Allocate global idle cpumasks */ 683 BUG_ON(!alloc_cpumask_var(&scx_idle_global_masks.cpu, GFP_KERNEL)); 684 BUG_ON(!alloc_cpumask_var(&scx_idle_global_masks.smt, GFP_KERNEL)); 685 686 /* Allocate per-node idle cpumasks (use nr_node_ids for non-contiguous NUMA nodes) */ 687 scx_idle_node_masks = kzalloc_objs(*scx_idle_node_masks, nr_node_ids); 688 BUG_ON(!scx_idle_node_masks); 689 690 for_each_node(i) { 691 scx_idle_node_masks[i] = kzalloc_node(sizeof(**scx_idle_node_masks), 692 GFP_KERNEL, i); 693 BUG_ON(!scx_idle_node_masks[i]); 694 695 BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->cpu, GFP_KERNEL, i)); 696 BUG_ON(!alloc_cpumask_var_node(&scx_idle_node_masks[i]->smt, GFP_KERNEL, i)); 697 } 698 699 /* Allocate local per-cpu idle cpumasks */ 700 for_each_possible_cpu(i) { 701 BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_idle_cpumask, i), 702 GFP_KERNEL, cpu_to_node(i))); 703 BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_llc_idle_cpumask, i), 704 GFP_KERNEL, cpu_to_node(i))); 705 BUG_ON(!alloc_cpumask_var_node(&per_cpu(local_numa_idle_cpumask, i), 706 GFP_KERNEL, cpu_to_node(i))); 707 } 708 } 709 710 static void update_builtin_idle(int cpu, bool idle) 711 { 712 int node = scx_cpu_node_if_enabled(cpu); 713 struct cpumask *idle_cpus = idle_cpumask(node)->cpu; 714 715 assign_cpu(cpu, idle_cpus, idle); 716 717 #ifdef CONFIG_SCHED_SMT 718 if (sched_smt_active()) { 719 const struct cpumask *smt = cpu_smt_mask(cpu); 720 struct cpumask *idle_smts = idle_cpumask(node)->smt; 721 722 if (idle) { 723 /* 724 * idle_smt handling is racy but that's fine as it's 725 * only for optimization and self-correcting. 726 */ 727 if (!cpumask_subset(smt, idle_cpus)) 728 return; 729 cpumask_or(idle_smts, idle_smts, smt); 730 } else { 731 cpumask_andnot(idle_smts, idle_smts, smt); 732 } 733 } 734 #endif 735 } 736 737 /* 738 * Update the idle state of a CPU to @idle. 739 * 740 * If @do_notify is true, ops.update_idle() is invoked to notify the scx 741 * scheduler of an actual idle state transition (idle to busy or vice 742 * versa). If @do_notify is false, only the idle state in the idle masks is 743 * refreshed without invoking ops.update_idle(). 744 * 745 * This distinction is necessary, because an idle CPU can be "reserved" and 746 * awakened via scx_bpf_pick_idle_cpu() + scx_bpf_kick_cpu(), marking it as 747 * busy even if no tasks are dispatched. In this case, the CPU may return 748 * to idle without a true state transition. Refreshing the idle masks 749 * without invoking ops.update_idle() ensures accurate idle state tracking 750 * while avoiding unnecessary updates and maintaining balanced state 751 * transitions. 752 */ 753 void __scx_update_idle(struct rq *rq, bool idle, bool do_notify) 754 { 755 struct scx_sched *sch = scx_root; 756 int cpu = cpu_of(rq); 757 758 lockdep_assert_rq_held(rq); 759 760 /* 761 * Update the idle masks: 762 * - for real idle transitions (do_notify == true) 763 * - for idle-to-idle transitions (indicated by the previous task 764 * being the idle thread, managed by pick_task_idle()) 765 * 766 * Skip updating idle masks if the previous task is not the idle 767 * thread, since set_next_task_idle() has already handled it when 768 * transitioning from a task to the idle thread (calling this 769 * function with do_notify == true). 770 * 771 * In this way we can avoid updating the idle masks twice, 772 * unnecessarily. 773 */ 774 if (static_branch_likely(&scx_builtin_idle_enabled)) 775 if (do_notify || is_idle_task(rq->curr)) 776 update_builtin_idle(cpu, idle); 777 778 /* 779 * Trigger ops.update_idle() only when transitioning from a task to 780 * the idle thread and vice versa. 781 * 782 * Idle transitions are indicated by do_notify being set to true, 783 * managed by put_prev_task_idle()/set_next_task_idle(). 784 * 785 * This must come after builtin idle update so that BPF schedulers can 786 * create interlocking between ops.update_idle() and ops.enqueue() - 787 * either enqueue() sees the idle bit or update_idle() sees the task 788 * that enqueue() queued. 789 */ 790 if (SCX_HAS_OP(sch, update_idle) && do_notify && 791 !scx_bypassing(sch, cpu_of(rq))) 792 SCX_CALL_OP(sch, update_idle, rq, cpu_of(rq), idle); 793 } 794 795 static void reset_idle_masks(struct sched_ext_ops *ops) 796 { 797 int node; 798 799 /* 800 * Consider all online cpus idle. Should converge to the actual state 801 * quickly. 802 */ 803 if (!(ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)) { 804 cpumask_copy(idle_cpumask(NUMA_NO_NODE)->cpu, cpu_online_mask); 805 cpumask_copy(idle_cpumask(NUMA_NO_NODE)->smt, cpu_online_mask); 806 return; 807 } 808 809 for_each_node(node) { 810 const struct cpumask *node_mask = cpumask_of_node(node); 811 812 cpumask_and(idle_cpumask(node)->cpu, cpu_online_mask, node_mask); 813 cpumask_and(idle_cpumask(node)->smt, cpu_online_mask, node_mask); 814 } 815 } 816 817 void scx_idle_enable(struct sched_ext_ops *ops) 818 { 819 if (!ops->update_idle || (ops->flags & SCX_OPS_KEEP_BUILTIN_IDLE)) 820 static_branch_enable_cpuslocked(&scx_builtin_idle_enabled); 821 else 822 static_branch_disable_cpuslocked(&scx_builtin_idle_enabled); 823 824 if (ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE) 825 static_branch_enable_cpuslocked(&scx_builtin_idle_per_node); 826 else 827 static_branch_disable_cpuslocked(&scx_builtin_idle_per_node); 828 829 reset_idle_masks(ops); 830 } 831 832 void scx_idle_disable(void) 833 { 834 static_branch_disable(&scx_builtin_idle_enabled); 835 static_branch_disable(&scx_builtin_idle_per_node); 836 } 837 838 /******************************************************************************** 839 * Helpers that can be called from the BPF scheduler. 840 */ 841 842 static int validate_node(struct scx_sched *sch, int node) 843 { 844 if (!static_branch_likely(&scx_builtin_idle_per_node)) { 845 scx_error(sch, "per-node idle tracking is disabled"); 846 return -EOPNOTSUPP; 847 } 848 849 /* Return no entry for NUMA_NO_NODE (not a critical scx error) */ 850 if (node == NUMA_NO_NODE) 851 return -ENOENT; 852 853 /* Make sure node is in a valid range */ 854 if (node < 0 || node >= nr_node_ids) { 855 scx_error(sch, "invalid node %d", node); 856 return -EINVAL; 857 } 858 859 /* Make sure the node is part of the set of possible nodes */ 860 if (!node_possible(node)) { 861 scx_error(sch, "unavailable node %d", node); 862 return -EINVAL; 863 } 864 865 return node; 866 } 867 868 __bpf_kfunc_start_defs(); 869 870 static bool check_builtin_idle_enabled(struct scx_sched *sch) 871 { 872 if (static_branch_likely(&scx_builtin_idle_enabled)) 873 return true; 874 875 scx_error(sch, "built-in idle tracking is disabled"); 876 return false; 877 } 878 879 /* 880 * Determine whether @p is a migration-disabled task in the context of BPF 881 * code. 882 * 883 * We can't simply check whether @p->migration_disabled is set in a 884 * sched_ext callback, because the BPF prolog (__bpf_prog_enter) may disable 885 * migration for the current task while running BPF code. 886 * 887 * Since the BPF prolog calls migrate_disable() only when CONFIG_PREEMPT_RCU 888 * is enabled (via rcu_read_lock_dont_migrate()), migration_disabled == 1 for 889 * the current task is ambiguous only in that case: it could be from the BPF 890 * prolog rather than a real migrate_disable() call. 891 * 892 * Without CONFIG_PREEMPT_RCU, the BPF prolog never calls migrate_disable(), 893 * so migration_disabled == 1 always means the task is truly 894 * migration-disabled. 895 * 896 * Therefore, when migration_disabled == 1 and CONFIG_PREEMPT_RCU is enabled, 897 * check whether @p is the current task or not: if it is, then migration was 898 * not disabled before entering the callback, otherwise migration was disabled. 899 * 900 * Returns true if @p is migration-disabled, false otherwise. 901 */ 902 static bool is_bpf_migration_disabled(const struct task_struct *p) 903 { 904 if (p->migration_disabled == 1) { 905 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) 906 return p != current; 907 return true; 908 } 909 return p->migration_disabled; 910 } 911 912 static s32 select_cpu_from_kfunc(struct scx_sched *sch, struct task_struct *p, 913 s32 prev_cpu, u64 wake_flags, 914 const struct cpumask *allowed, u64 flags) 915 { 916 unsigned long irq_flags; 917 bool we_locked = false; 918 s32 cpu; 919 920 if (!ops_cpu_valid(sch, prev_cpu, NULL)) 921 return -EINVAL; 922 923 if (!check_builtin_idle_enabled(sch)) 924 return -EBUSY; 925 926 /* 927 * Accessing p->cpus_ptr / p->nr_cpus_allowed needs either @p's rq 928 * lock or @p's pi_lock. Three cases: 929 * 930 * - inside ops.select_cpu(): try_to_wake_up() holds @p's pi_lock. 931 * - other rq-locked SCX op: scx_locked_rq() points at the held rq. 932 * - truly unlocked (UNLOCKED ops, SYSCALL, non-SCX struct_ops): 933 * nothing held, take pi_lock ourselves. 934 */ 935 if (this_rq()->scx.in_select_cpu) { 936 lockdep_assert_held(&p->pi_lock); 937 } else if (!scx_locked_rq()) { 938 raw_spin_lock_irqsave(&p->pi_lock, irq_flags); 939 we_locked = true; 940 } 941 942 /* 943 * This may also be called from ops.enqueue(), so we need to handle 944 * per-CPU tasks as well. For these tasks, we can skip all idle CPU 945 * selection optimizations and simply check whether the previously 946 * used CPU is idle and within the allowed cpumask. 947 */ 948 if (p->nr_cpus_allowed == 1 || is_bpf_migration_disabled(p)) { 949 if (cpumask_test_cpu(prev_cpu, allowed ?: p->cpus_ptr) && 950 scx_idle_test_and_clear_cpu(prev_cpu)) 951 cpu = prev_cpu; 952 else 953 cpu = -EBUSY; 954 } else { 955 cpu = scx_select_cpu_dfl(p, prev_cpu, wake_flags, 956 allowed ?: p->cpus_ptr, flags); 957 } 958 959 if (we_locked) 960 raw_spin_unlock_irqrestore(&p->pi_lock, irq_flags); 961 962 return cpu; 963 } 964 965 /** 966 * scx_bpf_cpu_node - Return the NUMA node the given @cpu belongs to, or 967 * trigger an error if @cpu is invalid 968 * @cpu: target CPU 969 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 970 */ 971 __bpf_kfunc s32 scx_bpf_cpu_node(s32 cpu, const struct bpf_prog_aux *aux) 972 { 973 struct scx_sched *sch; 974 975 guard(rcu)(); 976 977 sch = scx_prog_sched(aux); 978 if (unlikely(!sch) || !ops_cpu_valid(sch, cpu, NULL)) 979 return NUMA_NO_NODE; 980 return cpu_to_node(cpu); 981 } 982 983 /** 984 * scx_bpf_select_cpu_dfl - The default implementation of ops.select_cpu() 985 * @p: task_struct to select a CPU for 986 * @prev_cpu: CPU @p was on previously 987 * @wake_flags: %SCX_WAKE_* flags 988 * @is_idle: out parameter indicating whether the returned CPU is idle 989 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 990 * 991 * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked 992 * context such as a BPF test_run() call, as long as built-in CPU selection 993 * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE 994 * is set. 995 * 996 * Returns the picked CPU with *@is_idle indicating whether the picked CPU is 997 * currently idle and thus a good candidate for direct dispatching. 998 */ 999 __bpf_kfunc s32 scx_bpf_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, 1000 u64 wake_flags, bool *is_idle, 1001 const struct bpf_prog_aux *aux) 1002 { 1003 struct scx_sched *sch; 1004 s32 cpu; 1005 1006 guard(rcu)(); 1007 1008 sch = scx_prog_sched(aux); 1009 if (unlikely(!sch)) 1010 return -ENODEV; 1011 1012 cpu = select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags, NULL, 0); 1013 if (cpu >= 0) { 1014 *is_idle = true; 1015 return cpu; 1016 } 1017 *is_idle = false; 1018 return prev_cpu; 1019 } 1020 1021 struct scx_bpf_select_cpu_and_args { 1022 /* @p and @cpus_allowed can't be packed together as KF_RCU is not transitive */ 1023 s32 prev_cpu; 1024 u64 wake_flags; 1025 u64 flags; 1026 }; 1027 1028 /** 1029 * __scx_bpf_select_cpu_and - Arg-wrapped CPU selection with cpumask 1030 * @p: task_struct to select a CPU for 1031 * @cpus_allowed: cpumask of allowed CPUs 1032 * @args: struct containing the rest of the arguments 1033 * @args->prev_cpu: CPU @p was on previously 1034 * @args->wake_flags: %SCX_WAKE_* flags 1035 * @args->flags: %SCX_PICK_IDLE* flags 1036 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1037 * 1038 * Wrapper kfunc that takes arguments via struct to work around BPF's 5 argument 1039 * limit. BPF programs should use scx_bpf_select_cpu_and() which is provided 1040 * as an inline wrapper in common.bpf.h. 1041 * 1042 * Can be called from ops.select_cpu(), ops.enqueue(), or from an unlocked 1043 * context such as a BPF test_run() call, as long as built-in CPU selection 1044 * is enabled: ops.update_idle() is missing or %SCX_OPS_KEEP_BUILTIN_IDLE 1045 * is set. 1046 * 1047 * @p, @args->prev_cpu and @args->wake_flags match ops.select_cpu(). 1048 * 1049 * Returns the selected idle CPU, which will be automatically awakened upon 1050 * returning from ops.select_cpu() and can be used for direct dispatch, or 1051 * a negative value if no idle CPU is available. 1052 */ 1053 __bpf_kfunc s32 1054 __scx_bpf_select_cpu_and(struct task_struct *p, const struct cpumask *cpus_allowed, 1055 struct scx_bpf_select_cpu_and_args *args, 1056 const struct bpf_prog_aux *aux) 1057 { 1058 struct scx_sched *sch; 1059 1060 guard(rcu)(); 1061 1062 sch = scx_prog_sched(aux); 1063 if (unlikely(!sch)) 1064 return -ENODEV; 1065 1066 return select_cpu_from_kfunc(sch, p, args->prev_cpu, args->wake_flags, 1067 cpus_allowed, args->flags); 1068 } 1069 1070 /* 1071 * COMPAT: Will be removed in v6.22. 1072 */ 1073 __bpf_kfunc s32 scx_bpf_select_cpu_and(struct task_struct *p, s32 prev_cpu, u64 wake_flags, 1074 const struct cpumask *cpus_allowed, u64 flags) 1075 { 1076 struct scx_sched *sch; 1077 1078 guard(rcu)(); 1079 1080 sch = rcu_dereference(scx_root); 1081 if (unlikely(!sch)) 1082 return -ENODEV; 1083 1084 #ifdef CONFIG_EXT_SUB_SCHED 1085 /* 1086 * Disallow if any sub-scheds are attached. There is no way to tell 1087 * which scheduler called us, just error out @p's scheduler. 1088 */ 1089 if (unlikely(!list_empty(&sch->children))) { 1090 scx_error(scx_task_sched(p), "__scx_bpf_select_cpu_and() must be used"); 1091 return -EINVAL; 1092 } 1093 #endif 1094 1095 return select_cpu_from_kfunc(sch, p, prev_cpu, wake_flags, 1096 cpus_allowed, flags); 1097 } 1098 1099 /** 1100 * scx_bpf_get_idle_cpumask_node - Get a referenced kptr to the 1101 * idle-tracking per-CPU cpumask of a target NUMA node. 1102 * @node: target NUMA node 1103 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1104 * 1105 * Returns an empty cpumask if idle tracking is not enabled, if @node is 1106 * not valid, or running on a UP kernel. In this case the actual error will 1107 * be reported to the BPF scheduler via scx_error(). 1108 */ 1109 __bpf_kfunc const struct cpumask * 1110 scx_bpf_get_idle_cpumask_node(s32 node, const struct bpf_prog_aux *aux) 1111 { 1112 struct scx_sched *sch; 1113 1114 guard(rcu)(); 1115 1116 sch = scx_prog_sched(aux); 1117 if (unlikely(!sch)) 1118 return cpu_none_mask; 1119 1120 node = validate_node(sch, node); 1121 if (node < 0) 1122 return cpu_none_mask; 1123 1124 return idle_cpumask(node)->cpu; 1125 } 1126 1127 /** 1128 * scx_bpf_get_idle_cpumask - Get a referenced kptr to the idle-tracking 1129 * per-CPU cpumask. 1130 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1131 * 1132 * Returns an empty mask if idle tracking is not enabled, or running on a 1133 * UP kernel. 1134 */ 1135 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(const struct bpf_prog_aux *aux) 1136 { 1137 struct scx_sched *sch; 1138 1139 guard(rcu)(); 1140 1141 sch = scx_prog_sched(aux); 1142 if (unlikely(!sch)) 1143 return cpu_none_mask; 1144 1145 if (static_branch_unlikely(&scx_builtin_idle_per_node)) { 1146 scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled"); 1147 return cpu_none_mask; 1148 } 1149 1150 if (!check_builtin_idle_enabled(sch)) 1151 return cpu_none_mask; 1152 1153 return idle_cpumask(NUMA_NO_NODE)->cpu; 1154 } 1155 1156 /** 1157 * scx_bpf_get_idle_smtmask_node - Get a referenced kptr to the 1158 * idle-tracking, per-physical-core cpumask of a target NUMA node. Can be 1159 * used to determine if an entire physical core is free. 1160 * @node: target NUMA node 1161 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1162 * 1163 * Returns an empty cpumask if idle tracking is not enabled, if @node is 1164 * not valid, or running on a UP kernel. In this case the actual error will 1165 * be reported to the BPF scheduler via scx_error(). 1166 */ 1167 __bpf_kfunc const struct cpumask * 1168 scx_bpf_get_idle_smtmask_node(s32 node, const struct bpf_prog_aux *aux) 1169 { 1170 struct scx_sched *sch; 1171 1172 guard(rcu)(); 1173 1174 sch = scx_prog_sched(aux); 1175 if (unlikely(!sch)) 1176 return cpu_none_mask; 1177 1178 node = validate_node(sch, node); 1179 if (node < 0) 1180 return cpu_none_mask; 1181 1182 if (sched_smt_active()) 1183 return idle_cpumask(node)->smt; 1184 else 1185 return idle_cpumask(node)->cpu; 1186 } 1187 1188 /** 1189 * scx_bpf_get_idle_smtmask - Get a referenced kptr to the idle-tracking, 1190 * per-physical-core cpumask. Can be used to determine if an entire physical 1191 * core is free. 1192 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1193 * 1194 * Returns an empty mask if idle tracking is not enabled, or running on a 1195 * UP kernel. 1196 */ 1197 __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask(const struct bpf_prog_aux *aux) 1198 { 1199 struct scx_sched *sch; 1200 1201 guard(rcu)(); 1202 1203 sch = scx_prog_sched(aux); 1204 if (unlikely(!sch)) 1205 return cpu_none_mask; 1206 1207 if (static_branch_unlikely(&scx_builtin_idle_per_node)) { 1208 scx_error(sch, "SCX_OPS_BUILTIN_IDLE_PER_NODE enabled"); 1209 return cpu_none_mask; 1210 } 1211 1212 if (!check_builtin_idle_enabled(sch)) 1213 return cpu_none_mask; 1214 1215 if (sched_smt_active()) 1216 return idle_cpumask(NUMA_NO_NODE)->smt; 1217 else 1218 return idle_cpumask(NUMA_NO_NODE)->cpu; 1219 } 1220 1221 /** 1222 * scx_bpf_put_idle_cpumask - Release a previously acquired referenced kptr to 1223 * either the percpu, or SMT idle-tracking cpumask. 1224 * @idle_mask: &cpumask to use 1225 */ 1226 __bpf_kfunc void scx_bpf_put_idle_cpumask(const struct cpumask *idle_mask) 1227 { 1228 /* 1229 * Empty function body because we aren't actually acquiring or releasing 1230 * a reference to a global idle cpumask, which is read-only in the 1231 * caller and is never released. The acquire / release semantics here 1232 * are just used to make the cpumask a trusted pointer in the caller. 1233 */ 1234 } 1235 1236 /** 1237 * scx_bpf_test_and_clear_cpu_idle - Test and clear @cpu's idle state 1238 * @cpu: cpu to test and clear idle for 1239 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1240 * 1241 * Returns %true if @cpu was idle and its idle state was successfully cleared. 1242 * %false otherwise. 1243 * 1244 * Unavailable if ops.update_idle() is implemented and 1245 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set. 1246 */ 1247 __bpf_kfunc bool scx_bpf_test_and_clear_cpu_idle(s32 cpu, const struct bpf_prog_aux *aux) 1248 { 1249 struct scx_sched *sch; 1250 1251 guard(rcu)(); 1252 1253 sch = scx_prog_sched(aux); 1254 if (unlikely(!sch)) 1255 return false; 1256 1257 if (!check_builtin_idle_enabled(sch)) 1258 return false; 1259 1260 if (!ops_cpu_valid(sch, cpu, NULL)) 1261 return false; 1262 1263 return scx_idle_test_and_clear_cpu(cpu); 1264 } 1265 1266 /** 1267 * scx_bpf_pick_idle_cpu_node - Pick and claim an idle cpu from @node 1268 * @cpus_allowed: Allowed cpumask 1269 * @node: target NUMA node 1270 * @flags: %SCX_PICK_IDLE_* flags 1271 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1272 * 1273 * Pick and claim an idle cpu in @cpus_allowed from the NUMA node @node. 1274 * 1275 * Returns the picked idle cpu number on success, or -%EBUSY if no matching 1276 * cpu was found. 1277 * 1278 * The search starts from @node and proceeds to other online NUMA nodes in 1279 * order of increasing distance (unless SCX_PICK_IDLE_IN_NODE is specified, 1280 * in which case the search is limited to the target @node). 1281 * 1282 * Always returns an error if ops.update_idle() is implemented and 1283 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set, or if 1284 * %SCX_OPS_BUILTIN_IDLE_PER_NODE is not set. 1285 */ 1286 __bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(const struct cpumask *cpus_allowed, 1287 s32 node, u64 flags, 1288 const struct bpf_prog_aux *aux) 1289 { 1290 struct scx_sched *sch; 1291 1292 guard(rcu)(); 1293 1294 sch = scx_prog_sched(aux); 1295 if (unlikely(!sch)) 1296 return -ENODEV; 1297 1298 node = validate_node(sch, node); 1299 if (node < 0) 1300 return node; 1301 1302 return scx_pick_idle_cpu(cpus_allowed, node, flags); 1303 } 1304 1305 /** 1306 * scx_bpf_pick_idle_cpu - Pick and claim an idle cpu 1307 * @cpus_allowed: Allowed cpumask 1308 * @flags: %SCX_PICK_IDLE_CPU_* flags 1309 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1310 * 1311 * Pick and claim an idle cpu in @cpus_allowed. Returns the picked idle cpu 1312 * number on success. -%EBUSY if no matching cpu was found. 1313 * 1314 * Idle CPU tracking may race against CPU scheduling state transitions. For 1315 * example, this function may return -%EBUSY as CPUs are transitioning into the 1316 * idle state. If the caller then assumes that there will be dispatch events on 1317 * the CPUs as they were all busy, the scheduler may end up stalling with CPUs 1318 * idling while there are pending tasks. Use scx_bpf_pick_any_cpu() and 1319 * scx_bpf_kick_cpu() to guarantee that there will be at least one dispatch 1320 * event in the near future. 1321 * 1322 * Unavailable if ops.update_idle() is implemented and 1323 * %SCX_OPS_KEEP_BUILTIN_IDLE is not set. 1324 * 1325 * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use 1326 * scx_bpf_pick_idle_cpu_node() instead. 1327 */ 1328 __bpf_kfunc s32 scx_bpf_pick_idle_cpu(const struct cpumask *cpus_allowed, 1329 u64 flags, const struct bpf_prog_aux *aux) 1330 { 1331 struct scx_sched *sch; 1332 1333 guard(rcu)(); 1334 1335 sch = scx_prog_sched(aux); 1336 if (unlikely(!sch)) 1337 return -ENODEV; 1338 1339 if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) { 1340 scx_error(sch, "per-node idle tracking is enabled"); 1341 return -EBUSY; 1342 } 1343 1344 if (!check_builtin_idle_enabled(sch)) 1345 return -EBUSY; 1346 1347 return scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags); 1348 } 1349 1350 /** 1351 * scx_bpf_pick_any_cpu_node - Pick and claim an idle cpu if available 1352 * or pick any CPU from @node 1353 * @cpus_allowed: Allowed cpumask 1354 * @node: target NUMA node 1355 * @flags: %SCX_PICK_IDLE_CPU_* flags 1356 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1357 * 1358 * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any 1359 * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu 1360 * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is 1361 * empty. 1362 * 1363 * The search starts from @node and proceeds to other online NUMA nodes in 1364 * order of increasing distance (unless %SCX_PICK_IDLE_IN_NODE is specified, 1365 * in which case the search is limited to the target @node, regardless of 1366 * the CPU idle state). 1367 * 1368 * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not 1369 * set, this function can't tell which CPUs are idle and will always pick any 1370 * CPU. 1371 */ 1372 __bpf_kfunc s32 scx_bpf_pick_any_cpu_node(const struct cpumask *cpus_allowed, 1373 s32 node, u64 flags, 1374 const struct bpf_prog_aux *aux) 1375 { 1376 struct scx_sched *sch; 1377 s32 cpu; 1378 1379 guard(rcu)(); 1380 1381 sch = scx_prog_sched(aux); 1382 if (unlikely(!sch)) 1383 return -ENODEV; 1384 1385 node = validate_node(sch, node); 1386 if (node < 0) 1387 return node; 1388 1389 cpu = scx_pick_idle_cpu(cpus_allowed, node, flags); 1390 if (cpu >= 0) 1391 return cpu; 1392 1393 if (flags & SCX_PICK_IDLE_IN_NODE) 1394 cpu = cpumask_any_and_distribute(cpumask_of_node(node), cpus_allowed); 1395 else 1396 cpu = cpumask_any_distribute(cpus_allowed); 1397 if (cpu < nr_cpu_ids) 1398 return cpu; 1399 else 1400 return -EBUSY; 1401 } 1402 1403 /** 1404 * scx_bpf_pick_any_cpu - Pick and claim an idle cpu if available or pick any CPU 1405 * @cpus_allowed: Allowed cpumask 1406 * @flags: %SCX_PICK_IDLE_CPU_* flags 1407 * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs 1408 * 1409 * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any 1410 * CPU in @cpus_allowed. Guaranteed to succeed and returns the picked idle cpu 1411 * number if @cpus_allowed is not empty. -%EBUSY is returned if @cpus_allowed is 1412 * empty. 1413 * 1414 * If ops.update_idle() is implemented and %SCX_OPS_KEEP_BUILTIN_IDLE is not 1415 * set, this function can't tell which CPUs are idle and will always pick any 1416 * CPU. 1417 * 1418 * Always returns an error if %SCX_OPS_BUILTIN_IDLE_PER_NODE is set, use 1419 * scx_bpf_pick_any_cpu_node() instead. 1420 */ 1421 __bpf_kfunc s32 scx_bpf_pick_any_cpu(const struct cpumask *cpus_allowed, 1422 u64 flags, const struct bpf_prog_aux *aux) 1423 { 1424 struct scx_sched *sch; 1425 s32 cpu; 1426 1427 guard(rcu)(); 1428 1429 sch = scx_prog_sched(aux); 1430 if (unlikely(!sch)) 1431 return -ENODEV; 1432 1433 if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) { 1434 scx_error(sch, "per-node idle tracking is enabled"); 1435 return -EBUSY; 1436 } 1437 1438 if (static_branch_likely(&scx_builtin_idle_enabled)) { 1439 cpu = scx_pick_idle_cpu(cpus_allowed, NUMA_NO_NODE, flags); 1440 if (cpu >= 0) 1441 return cpu; 1442 } 1443 1444 cpu = cpumask_any_distribute(cpus_allowed); 1445 if (cpu < nr_cpu_ids) 1446 return cpu; 1447 else 1448 return -EBUSY; 1449 } 1450 1451 __bpf_kfunc_end_defs(); 1452 1453 BTF_KFUNCS_START(scx_kfunc_ids_idle) 1454 BTF_ID_FLAGS(func, scx_bpf_cpu_node, KF_IMPLICIT_ARGS) 1455 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask_node, KF_IMPLICIT_ARGS | KF_ACQUIRE) 1456 BTF_ID_FLAGS(func, scx_bpf_get_idle_cpumask, KF_IMPLICIT_ARGS | KF_ACQUIRE) 1457 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask_node, KF_IMPLICIT_ARGS | KF_ACQUIRE) 1458 BTF_ID_FLAGS(func, scx_bpf_get_idle_smtmask, KF_IMPLICIT_ARGS | KF_ACQUIRE) 1459 BTF_ID_FLAGS(func, scx_bpf_put_idle_cpumask, KF_RELEASE) 1460 BTF_ID_FLAGS(func, scx_bpf_test_and_clear_cpu_idle, KF_IMPLICIT_ARGS) 1461 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu_node, KF_IMPLICIT_ARGS | KF_RCU) 1462 BTF_ID_FLAGS(func, scx_bpf_pick_idle_cpu, KF_IMPLICIT_ARGS | KF_RCU) 1463 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu_node, KF_IMPLICIT_ARGS | KF_RCU) 1464 BTF_ID_FLAGS(func, scx_bpf_pick_any_cpu, KF_IMPLICIT_ARGS | KF_RCU) 1465 BTF_KFUNCS_END(scx_kfunc_ids_idle) 1466 1467 static const struct btf_kfunc_id_set scx_kfunc_set_idle = { 1468 .owner = THIS_MODULE, 1469 .set = &scx_kfunc_ids_idle, 1470 }; 1471 1472 /* 1473 * The select_cpu kfuncs internally call task_rq_lock() when invoked from an 1474 * rq-unlocked context, and thus cannot be safely called from arbitrary tracing 1475 * contexts where @p's pi_lock state is unknown. Keep them out of 1476 * BPF_PROG_TYPE_TRACING by registering them in their own set which is exposed 1477 * only to STRUCT_OPS and SYSCALL programs. 1478 * 1479 * These kfuncs are also members of scx_kfunc_ids_unlocked (see ext.c) because 1480 * they're callable from unlocked contexts in addition to ops.select_cpu() and 1481 * ops.enqueue(). 1482 */ 1483 BTF_KFUNCS_START(scx_kfunc_ids_select_cpu) 1484 BTF_ID_FLAGS(func, __scx_bpf_select_cpu_and, KF_IMPLICIT_ARGS | KF_RCU) 1485 BTF_ID_FLAGS(func, scx_bpf_select_cpu_and, KF_RCU) 1486 BTF_ID_FLAGS(func, scx_bpf_select_cpu_dfl, KF_IMPLICIT_ARGS | KF_RCU) 1487 BTF_KFUNCS_END(scx_kfunc_ids_select_cpu) 1488 1489 static const struct btf_kfunc_id_set scx_kfunc_set_select_cpu = { 1490 .owner = THIS_MODULE, 1491 .set = &scx_kfunc_ids_select_cpu, 1492 .filter = scx_kfunc_context_filter, 1493 }; 1494 1495 int scx_idle_init(void) 1496 { 1497 int ret; 1498 1499 ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_idle) || 1500 register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_idle) || 1501 register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_idle) || 1502 register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_select_cpu) || 1503 register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_select_cpu); 1504 1505 return ret; 1506 } 1507