| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * kernel/sched/core.c |
| 4 | * |
| 5 | * Core kernel CPU scheduler code |
| 6 | * |
| 7 | * Copyright (C) 1991-2002 Linus Torvalds |
| 8 | * Copyright (C) 1998-2024 Ingo Molnar, Red Hat |
| 9 | */ |
| 10 | #define INSTANTIATE_EXPORTED_MIGRATE_DISABLE |
| 11 | #include <linux/sched.h> |
| 12 | #include <linux/highmem.h> |
| 13 | #include <linux/hrtimer_api.h> |
| 14 | #include <linux/ktime_api.h> |
| 15 | #include <linux/sched/signal.h> |
| 16 | #include <linux/syscalls_api.h> |
| 17 | #include <linux/debug_locks.h> |
| 18 | #include <linux/prefetch.h> |
| 19 | #include <linux/capability.h> |
| 20 | #include <linux/pgtable_api.h> |
| 21 | #include <linux/wait_bit.h> |
| 22 | #include <linux/jiffies.h> |
| 23 | #include <linux/spinlock_api.h> |
| 24 | #include <linux/cpumask_api.h> |
| 25 | #include <linux/lockdep_api.h> |
| 26 | #include <linux/hardirq.h> |
| 27 | #include <linux/softirq.h> |
| 28 | #include <linux/refcount_api.h> |
| 29 | #include <linux/topology.h> |
| 30 | #include <linux/sched/clock.h> |
| 31 | #include <linux/sched/cond_resched.h> |
| 32 | #include <linux/sched/cputime.h> |
| 33 | #include <linux/sched/debug.h> |
| 34 | #include <linux/sched/hotplug.h> |
| 35 | #include <linux/sched/init.h> |
| 36 | #include <linux/sched/isolation.h> |
| 37 | #include <linux/sched/loadavg.h> |
| 38 | #include <linux/sched/mm.h> |
| 39 | #include <linux/sched/nohz.h> |
| 40 | #include <linux/sched/rseq_api.h> |
| 41 | #include <linux/sched/rt.h> |
| 42 | |
| 43 | #include <linux/context_tracking.h> |
| 44 | #include <linux/cpuset.h> |
| 45 | #include <linux/delayacct.h> |
| 46 | #include <linux/init_task.h> |
| 47 | #include <linux/interrupt.h> |
| 48 | #include <linux/ioprio.h> |
| 49 | #include <linux/kallsyms.h> |
| 50 | #include <linux/kcov.h> |
| 51 | #include <linux/kprobes.h> |
| 52 | #include <linux/llist_api.h> |
| 53 | #include <linux/mmu_context.h> |
| 54 | #include <linux/mmzone.h> |
| 55 | #include <linux/mutex_api.h> |
| 56 | #include <linux/nmi.h> |
| 57 | #include <linux/nospec.h> |
| 58 | #include <linux/perf_event_api.h> |
| 59 | #include <linux/profile.h> |
| 60 | #include <linux/psi.h> |
| 61 | #include <linux/rcuwait_api.h> |
| 62 | #include <linux/rseq.h> |
| 63 | #include <linux/sched/wake_q.h> |
| 64 | #include <linux/scs.h> |
| 65 | #include <linux/slab.h> |
| 66 | #include <linux/syscalls.h> |
| 67 | #include <linux/vtime.h> |
| 68 | #include <linux/wait_api.h> |
| 69 | #include <linux/workqueue_api.h> |
| 70 | #include <linux/livepatch_sched.h> |
| 71 | |
| 72 | #ifdef CONFIG_PREEMPT_DYNAMIC |
| 73 | # ifdef CONFIG_GENERIC_IRQ_ENTRY |
| 74 | # include <linux/irq-entry-common.h> |
| 75 | # endif |
| 76 | #endif |
| 77 | |
| 78 | #include <uapi/linux/sched/types.h> |
| 79 | |
| 80 | #include <asm/irq_regs.h> |
| 81 | #include <asm/switch_to.h> |
| 82 | #include <asm/tlb.h> |
| 83 | |
| 84 | #define CREATE_TRACE_POINTS |
| 85 | #include <linux/sched/rseq_api.h> |
| 86 | #include <trace/events/sched.h> |
| 87 | #include <trace/events/ipi.h> |
| 88 | #undef CREATE_TRACE_POINTS |
| 89 | |
| 90 | #include "sched.h" |
| 91 | #include "stats.h" |
| 92 | |
| 93 | #include "autogroup.h" |
| 94 | #include "pelt.h" |
| 95 | #include "smp.h" |
| 96 | |
| 97 | #include "../workqueue_internal.h" |
| 98 | #include "../../io_uring/io-wq.h" |
| 99 | #include "../smpboot.h" |
| 100 | #include "../locking/mutex.h" |
| 101 | |
| 102 | EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu); |
| 103 | EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask); |
| 104 | |
| 105 | /* |
| 106 | * Export tracepoints that act as a bare tracehook (ie: have no trace event |
| 107 | * associated with them) to allow external modules to probe them. |
| 108 | */ |
| 109 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp); |
| 110 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp); |
| 111 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp); |
| 112 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp); |
| 113 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp); |
| 114 | EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_hw_tp); |
| 115 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp); |
| 116 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp); |
| 117 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp); |
| 118 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp); |
| 119 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp); |
| 120 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp); |
| 121 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_entry_tp); |
| 122 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_exit_tp); |
| 123 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_set_need_resched_tp); |
| 124 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_dl_throttle_tp); |
| 125 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_dl_replenish_tp); |
| 126 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_dl_update_tp); |
| 127 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_dl_server_start_tp); |
| 128 | EXPORT_TRACEPOINT_SYMBOL_GPL(sched_dl_server_stop_tp); |
| 129 | |
| 130 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
| 131 | DEFINE_PER_CPU(struct rnd_state, sched_rnd_state); |
| 132 | |
| 133 | #ifdef CONFIG_SCHED_PROXY_EXEC |
| 134 | DEFINE_STATIC_KEY_TRUE(__sched_proxy_exec); |
| 135 | static int __init setup_proxy_exec(char *str) |
| 136 | { |
| 137 | bool proxy_enable = true; |
| 138 | |
| 139 | if (*str && kstrtobool(str + 1, &proxy_enable)) { |
| 140 | pr_warn("Unable to parse sched_proxy_exec=\n"); |
| 141 | return 0; |
| 142 | } |
| 143 | |
| 144 | if (proxy_enable) { |
| 145 | pr_info("sched_proxy_exec enabled via boot arg\n"); |
| 146 | static_branch_enable(&__sched_proxy_exec); |
| 147 | } else { |
| 148 | pr_info("sched_proxy_exec disabled via boot arg\n"); |
| 149 | static_branch_disable(&__sched_proxy_exec); |
| 150 | } |
| 151 | return 1; |
| 152 | } |
| 153 | #else |
| 154 | static int __init setup_proxy_exec(char *str) |
| 155 | { |
| 156 | pr_warn("CONFIG_SCHED_PROXY_EXEC=n, so it cannot be enabled or disabled at boot time\n"); |
| 157 | return 0; |
| 158 | } |
| 159 | #endif |
| 160 | __setup("sched_proxy_exec", setup_proxy_exec); |
| 161 | |
| 162 | /* |
| 163 | * Debugging: various feature bits |
| 164 | * |
| 165 | * If SCHED_DEBUG is disabled, each compilation unit has its own copy of |
| 166 | * sysctl_sched_features, defined in sched.h, to allow constants propagation |
| 167 | * at compile time and compiler optimization based on features default. |
| 168 | */ |
| 169 | #define SCHED_FEAT(name, enabled) \ |
| 170 | (1UL << __SCHED_FEAT_##name) * enabled | |
| 171 | __read_mostly unsigned int sysctl_sched_features = |
| 172 | #include "features.h" |
| 173 | 0; |
| 174 | #undef SCHED_FEAT |
| 175 | |
| 176 | /* |
| 177 | * Print a warning if need_resched is set for the given duration (if |
| 178 | * LATENCY_WARN is enabled). |
| 179 | * |
| 180 | * If sysctl_resched_latency_warn_once is set, only one warning will be shown |
| 181 | * per boot. |
| 182 | */ |
| 183 | __read_mostly int sysctl_resched_latency_warn_ms = 100; |
| 184 | __read_mostly int sysctl_resched_latency_warn_once = 1; |
| 185 | |
| 186 | /* |
| 187 | * Number of tasks to iterate in a single balance run. |
| 188 | * Limited because this is done with IRQs disabled. |
| 189 | */ |
| 190 | __read_mostly unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK; |
| 191 | |
| 192 | __read_mostly int scheduler_running; |
| 193 | |
| 194 | #ifdef CONFIG_SCHED_CORE |
| 195 | |
| 196 | DEFINE_STATIC_KEY_FALSE(__sched_core_enabled); |
| 197 | |
| 198 | /* kernel prio, less is more */ |
| 199 | static inline int __task_prio(const struct task_struct *p) |
| 200 | { |
| 201 | if (p->sched_class == &stop_sched_class) /* trumps deadline */ |
| 202 | return -2; |
| 203 | |
| 204 | if (p->dl_server) |
| 205 | return -1; /* deadline */ |
| 206 | |
| 207 | if (rt_or_dl_prio(p->prio)) |
| 208 | return p->prio; /* [-1, 99] */ |
| 209 | |
| 210 | if (p->sched_class == &idle_sched_class) |
| 211 | return MAX_RT_PRIO + NICE_WIDTH; /* 140 */ |
| 212 | |
| 213 | if (task_on_scx(p)) |
| 214 | return MAX_RT_PRIO + MAX_NICE + 1; /* 120, squash ext */ |
| 215 | |
| 216 | return MAX_RT_PRIO + MAX_NICE; /* 119, squash fair */ |
| 217 | } |
| 218 | |
| 219 | /* |
| 220 | * l(a,b) |
| 221 | * le(a,b) := !l(b,a) |
| 222 | * g(a,b) := l(b,a) |
| 223 | * ge(a,b) := !l(a,b) |
| 224 | */ |
| 225 | |
| 226 | /* real prio, less is less */ |
| 227 | static inline bool prio_less(const struct task_struct *a, |
| 228 | const struct task_struct *b, bool in_fi) |
| 229 | { |
| 230 | |
| 231 | int pa = __task_prio(a), pb = __task_prio(b); |
| 232 | |
| 233 | if (-pa < -pb) |
| 234 | return true; |
| 235 | |
| 236 | if (-pb < -pa) |
| 237 | return false; |
| 238 | |
| 239 | if (pa == -1) { /* dl_prio() doesn't work because of stop_class above */ |
| 240 | const struct sched_dl_entity *a_dl, *b_dl; |
| 241 | |
| 242 | a_dl = &a->dl; |
| 243 | /* |
| 244 | * Since,'a' and 'b' can be CFS tasks served by DL server, |
| 245 | * __task_prio() can return -1 (for DL) even for those. In that |
| 246 | * case, get to the dl_server's DL entity. |
| 247 | */ |
| 248 | if (a->dl_server) |
| 249 | a_dl = a->dl_server; |
| 250 | |
| 251 | b_dl = &b->dl; |
| 252 | if (b->dl_server) |
| 253 | b_dl = b->dl_server; |
| 254 | |
| 255 | return !dl_time_before(a_dl->deadline, b_dl->deadline); |
| 256 | } |
| 257 | |
| 258 | if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */ |
| 259 | return cfs_prio_less(a, b, in_fi); |
| 260 | |
| 261 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 262 | if (pa == MAX_RT_PRIO + MAX_NICE + 1) /* ext */ |
| 263 | return scx_prio_less(a, b, in_fi); |
| 264 | #endif |
| 265 | |
| 266 | return false; |
| 267 | } |
| 268 | |
| 269 | static inline bool __sched_core_less(const struct task_struct *a, |
| 270 | const struct task_struct *b) |
| 271 | { |
| 272 | if (a->core_cookie < b->core_cookie) |
| 273 | return true; |
| 274 | |
| 275 | if (a->core_cookie > b->core_cookie) |
| 276 | return false; |
| 277 | |
| 278 | /* flip prio, so high prio is leftmost */ |
| 279 | if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count)) |
| 280 | return true; |
| 281 | |
| 282 | return false; |
| 283 | } |
| 284 | |
| 285 | #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node) |
| 286 | |
| 287 | static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b) |
| 288 | { |
| 289 | return __sched_core_less(__node_2_sc(a), __node_2_sc(b)); |
| 290 | } |
| 291 | |
| 292 | static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node) |
| 293 | { |
| 294 | const struct task_struct *p = __node_2_sc(node); |
| 295 | unsigned long cookie = (unsigned long)key; |
| 296 | |
| 297 | if (cookie < p->core_cookie) |
| 298 | return -1; |
| 299 | |
| 300 | if (cookie > p->core_cookie) |
| 301 | return 1; |
| 302 | |
| 303 | return 0; |
| 304 | } |
| 305 | |
| 306 | void sched_core_enqueue(struct rq *rq, struct task_struct *p) |
| 307 | { |
| 308 | if (p->se.sched_delayed) |
| 309 | return; |
| 310 | |
| 311 | rq->core->core_task_seq++; |
| 312 | |
| 313 | if (!p->core_cookie) |
| 314 | return; |
| 315 | |
| 316 | rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less); |
| 317 | } |
| 318 | |
| 319 | void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) |
| 320 | { |
| 321 | if (p->se.sched_delayed) |
| 322 | return; |
| 323 | |
| 324 | rq->core->core_task_seq++; |
| 325 | |
| 326 | if (sched_core_enqueued(p)) { |
| 327 | rb_erase(&p->core_node, &rq->core_tree); |
| 328 | RB_CLEAR_NODE(&p->core_node); |
| 329 | } |
| 330 | |
| 331 | /* |
| 332 | * Migrating the last task off the cpu, with the cpu in forced idle |
| 333 | * state. Reschedule to create an accounting edge for forced idle, |
| 334 | * and re-examine whether the core is still in forced idle state. |
| 335 | */ |
| 336 | if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 && |
| 337 | rq->core->core_forceidle_count && rq->curr == rq->idle) |
| 338 | resched_curr(rq); |
| 339 | } |
| 340 | |
| 341 | static int sched_task_is_throttled(struct task_struct *p, int cpu) |
| 342 | { |
| 343 | if (p->sched_class->task_is_throttled) |
| 344 | return p->sched_class->task_is_throttled(p, cpu); |
| 345 | |
| 346 | return 0; |
| 347 | } |
| 348 | |
| 349 | static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie) |
| 350 | { |
| 351 | struct rb_node *node = &p->core_node; |
| 352 | int cpu = task_cpu(p); |
| 353 | |
| 354 | do { |
| 355 | node = rb_next(node); |
| 356 | if (!node) |
| 357 | return NULL; |
| 358 | |
| 359 | p = __node_2_sc(node); |
| 360 | if (p->core_cookie != cookie) |
| 361 | return NULL; |
| 362 | |
| 363 | } while (sched_task_is_throttled(p, cpu)); |
| 364 | |
| 365 | return p; |
| 366 | } |
| 367 | |
| 368 | /* |
| 369 | * Find left-most (aka, highest priority) and unthrottled task matching @cookie. |
| 370 | * If no suitable task is found, NULL will be returned. |
| 371 | */ |
| 372 | static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie) |
| 373 | { |
| 374 | struct task_struct *p; |
| 375 | struct rb_node *node; |
| 376 | |
| 377 | node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp); |
| 378 | if (!node) |
| 379 | return NULL; |
| 380 | |
| 381 | p = __node_2_sc(node); |
| 382 | if (!sched_task_is_throttled(p, rq->cpu)) |
| 383 | return p; |
| 384 | |
| 385 | return sched_core_next(p, cookie); |
| 386 | } |
| 387 | |
| 388 | /* |
| 389 | * Magic required such that: |
| 390 | * |
| 391 | * raw_spin_rq_lock(rq); |
| 392 | * ... |
| 393 | * raw_spin_rq_unlock(rq); |
| 394 | * |
| 395 | * ends up locking and unlocking the _same_ lock, and all CPUs |
| 396 | * always agree on what rq has what lock. |
| 397 | * |
| 398 | * XXX entirely possible to selectively enable cores, don't bother for now. |
| 399 | */ |
| 400 | |
| 401 | static DEFINE_MUTEX(sched_core_mutex); |
| 402 | static atomic_t sched_core_count; |
| 403 | static struct cpumask sched_core_mask; |
| 404 | |
| 405 | static void sched_core_lock(int cpu, unsigned long *flags) |
| 406 | __context_unsafe(/* acquires multiple */) |
| 407 | __acquires(&runqueues.__lock) /* overapproximation */ |
| 408 | { |
| 409 | const struct cpumask *smt_mask = cpu_smt_mask(cpu); |
| 410 | int t, i = 0; |
| 411 | |
| 412 | local_irq_save(*flags); |
| 413 | for_each_cpu(t, smt_mask) |
| 414 | raw_spin_lock_nested(&cpu_rq(t)->__lock, i++); |
| 415 | } |
| 416 | |
| 417 | static void sched_core_unlock(int cpu, unsigned long *flags) |
| 418 | __context_unsafe(/* releases multiple */) |
| 419 | __releases(&runqueues.__lock) /* overapproximation */ |
| 420 | { |
| 421 | const struct cpumask *smt_mask = cpu_smt_mask(cpu); |
| 422 | int t; |
| 423 | |
| 424 | for_each_cpu(t, smt_mask) |
| 425 | raw_spin_unlock(&cpu_rq(t)->__lock); |
| 426 | local_irq_restore(*flags); |
| 427 | } |
| 428 | |
| 429 | static void __sched_core_flip(bool enabled) |
| 430 | { |
| 431 | unsigned long flags; |
| 432 | int cpu, t; |
| 433 | |
| 434 | cpus_read_lock(); |
| 435 | |
| 436 | /* |
| 437 | * Toggle the online cores, one by one. |
| 438 | */ |
| 439 | cpumask_copy(&sched_core_mask, cpu_online_mask); |
| 440 | for_each_cpu(cpu, &sched_core_mask) { |
| 441 | const struct cpumask *smt_mask = cpu_smt_mask(cpu); |
| 442 | |
| 443 | sched_core_lock(cpu, &flags); |
| 444 | |
| 445 | /* |
| 446 | * A core-wide selection may have the shared rq lock temporarily |
| 447 | * released by a lock-dropping ->pick_task(). Flipping would |
| 448 | * rebind rq_lockp() under it. Wait it out. |
| 449 | */ |
| 450 | while (cpu_rq(cpu)->core->core_pick_in_flight) { |
| 451 | sched_core_unlock(cpu, &flags); |
| 452 | cpu_relax(); |
| 453 | sched_core_lock(cpu, &flags); |
| 454 | } |
| 455 | |
| 456 | for_each_cpu(t, smt_mask) |
| 457 | cpu_rq(t)->core_enabled = enabled; |
| 458 | |
| 459 | cpu_rq(cpu)->core->core_forceidle_start = 0; |
| 460 | |
| 461 | sched_core_unlock(cpu, &flags); |
| 462 | |
| 463 | cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask); |
| 464 | } |
| 465 | |
| 466 | /* |
| 467 | * Toggle the offline CPUs. |
| 468 | */ |
| 469 | for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask) |
| 470 | cpu_rq(cpu)->core_enabled = enabled; |
| 471 | |
| 472 | cpus_read_unlock(); |
| 473 | } |
| 474 | |
| 475 | static void sched_core_assert_empty(void) |
| 476 | { |
| 477 | int cpu; |
| 478 | |
| 479 | for_each_possible_cpu(cpu) |
| 480 | WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree)); |
| 481 | } |
| 482 | |
| 483 | static void __sched_core_enable(void) |
| 484 | { |
| 485 | static_branch_enable(&__sched_core_enabled); |
| 486 | /* |
| 487 | * Ensure all previous instances of raw_spin_rq_*lock() have finished |
| 488 | * and future ones will observe !sched_core_disabled(). |
| 489 | */ |
| 490 | synchronize_rcu(); |
| 491 | __sched_core_flip(true); |
| 492 | sched_core_assert_empty(); |
| 493 | } |
| 494 | |
| 495 | static void __sched_core_disable(void) |
| 496 | { |
| 497 | sched_core_assert_empty(); |
| 498 | __sched_core_flip(false); |
| 499 | static_branch_disable(&__sched_core_enabled); |
| 500 | } |
| 501 | |
| 502 | void sched_core_get(void) |
| 503 | { |
| 504 | if (atomic_inc_not_zero(&sched_core_count)) |
| 505 | return; |
| 506 | |
| 507 | mutex_lock(&sched_core_mutex); |
| 508 | if (!atomic_read(&sched_core_count)) |
| 509 | __sched_core_enable(); |
| 510 | |
| 511 | smp_mb__before_atomic(); |
| 512 | atomic_inc(&sched_core_count); |
| 513 | mutex_unlock(&sched_core_mutex); |
| 514 | } |
| 515 | |
| 516 | static void __sched_core_put(struct work_struct *work) |
| 517 | { |
| 518 | if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) { |
| 519 | __sched_core_disable(); |
| 520 | mutex_unlock(&sched_core_mutex); |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | void sched_core_put(void) |
| 525 | { |
| 526 | static DECLARE_WORK(_work, __sched_core_put); |
| 527 | |
| 528 | /* |
| 529 | * "There can be only one" |
| 530 | * |
| 531 | * Either this is the last one, or we don't actually need to do any |
| 532 | * 'work'. If it is the last *again*, we rely on |
| 533 | * WORK_STRUCT_PENDING_BIT. |
| 534 | */ |
| 535 | if (!atomic_add_unless(&sched_core_count, -1, 1)) |
| 536 | schedule_work(&_work); |
| 537 | } |
| 538 | |
| 539 | #else /* !CONFIG_SCHED_CORE: */ |
| 540 | |
| 541 | static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { } |
| 542 | static inline void |
| 543 | sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { } |
| 544 | |
| 545 | #endif /* !CONFIG_SCHED_CORE */ |
| 546 | |
| 547 | /* need a wrapper since we may need to trace from modules */ |
| 548 | EXPORT_TRACEPOINT_SYMBOL(sched_set_state_tp); |
| 549 | |
| 550 | /* |
| 551 | * Call via the helper macro trace_set_current_state. |
| 552 | * Calls to this function MUST be guarded by a |
| 553 | * tracepoint_enabled(sched_set_state_tp) |
| 554 | */ |
| 555 | void __trace_set_current_state(int state_value) |
| 556 | { |
| 557 | trace_call__sched_set_state_tp(current, state_value); |
| 558 | } |
| 559 | EXPORT_SYMBOL(__trace_set_current_state); |
| 560 | |
| 561 | int task_llc(const struct task_struct *p) |
| 562 | { |
| 563 | return per_cpu(sd_llc_id, task_cpu(p)); |
| 564 | } |
| 565 | |
| 566 | /* |
| 567 | * Serialization rules: |
| 568 | * |
| 569 | * Lock order: |
| 570 | * |
| 571 | * p->pi_lock |
| 572 | * rq->lock |
| 573 | * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls) |
| 574 | * |
| 575 | * rq1->lock |
| 576 | * rq2->lock where: rq1 < rq2 |
| 577 | * |
| 578 | * Regular state: |
| 579 | * |
| 580 | * Normal scheduling state is serialized by rq->lock. __schedule() takes the |
| 581 | * local CPU's rq->lock, it optionally removes the task from the runqueue and |
| 582 | * always looks at the local rq data structures to find the most eligible task |
| 583 | * to run next. |
| 584 | * |
| 585 | * Task enqueue is also under rq->lock, possibly taken from another CPU. |
| 586 | * Wakeups from another LLC domain might use an IPI to transfer the enqueue to |
| 587 | * the local CPU to avoid bouncing the runqueue state around [ see |
| 588 | * ttwu_queue_wakelist() ] |
| 589 | * |
| 590 | * Task wakeup, specifically wakeups that involve migration, are horribly |
| 591 | * complicated to avoid having to take two rq->locks. |
| 592 | * |
| 593 | * Special state: |
| 594 | * |
| 595 | * System-calls and anything external will use task_rq_lock() which acquires |
| 596 | * both p->pi_lock and rq->lock. As a consequence the state they change is |
| 597 | * stable while holding either lock: |
| 598 | * |
| 599 | * - sched_setaffinity()/ |
| 600 | * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed |
| 601 | * - set_user_nice(): p->se.load, p->*prio |
| 602 | * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio, |
| 603 | * p->se.load, p->rt_priority, |
| 604 | * p->dl.dl_{runtime, deadline, period, flags, bw, density} |
| 605 | * - sched_setnuma(): p->numa_preferred_nid |
| 606 | * - sched_move_task(): p->sched_task_group |
| 607 | * - uclamp_update_active() p->uclamp* |
| 608 | * |
| 609 | * p->state <- TASK_*: |
| 610 | * |
| 611 | * is changed locklessly using set_current_state(), __set_current_state() or |
| 612 | * set_special_state(), see their respective comments, or by |
| 613 | * try_to_wake_up(). This latter uses p->pi_lock to serialize against |
| 614 | * concurrent self. |
| 615 | * |
| 616 | * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }: |
| 617 | * |
| 618 | * is set by activate_task() and cleared by deactivate_task()/block_task(), |
| 619 | * under rq->lock. Non-zero indicates the task is runnable, the special |
| 620 | * ON_RQ_MIGRATING state is used for migration without holding both |
| 621 | * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock(). |
| 622 | * |
| 623 | * Additionally it is possible to be ->on_rq but still be considered not |
| 624 | * runnable when p->se.sched_delayed is true. These tasks are on the runqueue |
| 625 | * but will be dequeued as soon as they get picked again. See the |
| 626 | * task_is_runnable() helper. |
| 627 | * |
| 628 | * p->on_cpu <- { 0, 1 }: |
| 629 | * |
| 630 | * is set by prepare_task() and cleared by finish_task() such that it will be |
| 631 | * set before p is scheduled-in and cleared after p is scheduled-out, both |
| 632 | * under rq->lock. Non-zero indicates the task is running on its CPU. |
| 633 | * |
| 634 | * [ The astute reader will observe that it is possible for two tasks on one |
| 635 | * CPU to have ->on_cpu = 1 at the same time. ] |
| 636 | * |
| 637 | * p->is_blocked <- { 0, 1 }: |
| 638 | * |
| 639 | * is set by try_to_block_task() and cleared by ttwu_do_wakeup() and tracks |
| 640 | * if the task is blocked. Traditionally this would mirror p->on_rq, however |
| 641 | * due things like DELAY_DEQUEUE and PROXY_EXEC, this can diverge. |
| 642 | * |
| 643 | * task_cpu(p): is changed by set_task_cpu(), the rules are: |
| 644 | * |
| 645 | * - Don't call set_task_cpu() on a blocked task: |
| 646 | * |
| 647 | * We don't care what CPU we're not running on, this simplifies hotplug, |
| 648 | * the CPU assignment of blocked tasks isn't required to be valid. |
| 649 | * |
| 650 | * - for try_to_wake_up(), called under p->pi_lock: |
| 651 | * |
| 652 | * This allows try_to_wake_up() to only take one rq->lock, see its comment. |
| 653 | * |
| 654 | * - for migration called under rq->lock: |
| 655 | * [ see task_on_rq_migrating() in task_rq_lock() ] |
| 656 | * |
| 657 | * o move_queued_task() |
| 658 | * o detach_task() |
| 659 | * |
| 660 | * - for migration called under double_rq_lock(): |
| 661 | * |
| 662 | * o __migrate_swap_task() |
| 663 | * o push_rt_task() / pull_rt_task() |
| 664 | * o push_dl_task() / pull_dl_task() |
| 665 | * o dl_task_offline_migration() |
| 666 | * |
| 667 | */ |
| 668 | |
| 669 | void raw_spin_rq_lock_nested(struct rq *rq, int subclass) |
| 670 | __context_unsafe() |
| 671 | { |
| 672 | raw_spinlock_t *lock; |
| 673 | |
| 674 | /* Matches synchronize_rcu() in __sched_core_enable() */ |
| 675 | preempt_disable(); |
| 676 | if (sched_core_disabled()) { |
| 677 | raw_spin_lock_nested(&rq->__lock, subclass); |
| 678 | /* preempt_count *MUST* be > 1 */ |
| 679 | preempt_enable_no_resched(); |
| 680 | return; |
| 681 | } |
| 682 | |
| 683 | for (;;) { |
| 684 | lock = __rq_lockp(rq); |
| 685 | raw_spin_lock_nested(lock, subclass); |
| 686 | if (likely(lock == __rq_lockp(rq))) { |
| 687 | /* preempt_count *MUST* be > 1 */ |
| 688 | preempt_enable_no_resched(); |
| 689 | return; |
| 690 | } |
| 691 | raw_spin_unlock(lock); |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | bool raw_spin_rq_trylock(struct rq *rq) |
| 696 | __context_unsafe() |
| 697 | { |
| 698 | raw_spinlock_t *lock; |
| 699 | bool ret; |
| 700 | |
| 701 | /* Matches synchronize_rcu() in __sched_core_enable() */ |
| 702 | preempt_disable(); |
| 703 | if (sched_core_disabled()) { |
| 704 | ret = raw_spin_trylock(&rq->__lock); |
| 705 | preempt_enable(); |
| 706 | return ret; |
| 707 | } |
| 708 | |
| 709 | for (;;) { |
| 710 | lock = __rq_lockp(rq); |
| 711 | ret = raw_spin_trylock(lock); |
| 712 | if (!ret || (likely(lock == __rq_lockp(rq)))) { |
| 713 | preempt_enable(); |
| 714 | return ret; |
| 715 | } |
| 716 | raw_spin_unlock(lock); |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | /* |
| 721 | * double_rq_lock - safely lock two runqueues |
| 722 | */ |
| 723 | void double_rq_lock(struct rq *rq1, struct rq *rq2) |
| 724 | { |
| 725 | lockdep_assert_irqs_disabled(); |
| 726 | |
| 727 | if (rq_order_less(rq2, rq1)) |
| 728 | swap(rq1, rq2); |
| 729 | |
| 730 | raw_spin_rq_lock(rq1); |
| 731 | if (__rq_lockp(rq1) != __rq_lockp(rq2)) |
| 732 | raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING); |
| 733 | else |
| 734 | __acquire_ctx_lock(__rq_lockp(rq2)); /* fake acquire */ |
| 735 | |
| 736 | double_rq_clock_clear_update(rq1, rq2); |
| 737 | } |
| 738 | |
| 739 | /* |
| 740 | * ___task_rq_lock - lock the rq @p resides on. |
| 741 | */ |
| 742 | struct rq *___task_rq_lock(struct task_struct *p, struct rq_flags *rf) |
| 743 | { |
| 744 | struct rq *rq; |
| 745 | |
| 746 | lockdep_assert_held(&p->pi_lock); |
| 747 | |
| 748 | for (;;) { |
| 749 | rq = task_rq(p); |
| 750 | raw_spin_rq_lock(rq); |
| 751 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { |
| 752 | rq_pin_lock(rq, rf); |
| 753 | return rq; |
| 754 | } |
| 755 | raw_spin_rq_unlock(rq); |
| 756 | |
| 757 | while (unlikely(task_on_rq_migrating(p))) |
| 758 | cpu_relax(); |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | /* |
| 763 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. |
| 764 | */ |
| 765 | struct rq *_task_rq_lock(struct task_struct *p, struct rq_flags *rf) |
| 766 | { |
| 767 | struct rq *rq; |
| 768 | |
| 769 | for (;;) { |
| 770 | raw_spin_lock_irqsave(&p->pi_lock, rf->flags); |
| 771 | rq = task_rq(p); |
| 772 | raw_spin_rq_lock(rq); |
| 773 | /* |
| 774 | * move_queued_task() task_rq_lock() |
| 775 | * |
| 776 | * ACQUIRE (rq->lock) |
| 777 | * [S] ->on_rq = MIGRATING [L] rq = task_rq() |
| 778 | * WMB (__set_task_cpu()) ACQUIRE (rq->lock); |
| 779 | * [S] ->cpu = new_cpu [L] task_rq() |
| 780 | * [L] ->on_rq |
| 781 | * RELEASE (rq->lock) |
| 782 | * |
| 783 | * If we observe the old CPU in task_rq_lock(), the acquire of |
| 784 | * the old rq->lock will fully serialize against the stores. |
| 785 | * |
| 786 | * If we observe the new CPU in task_rq_lock(), the address |
| 787 | * dependency headed by '[L] rq = task_rq()' and the acquire |
| 788 | * will pair with the WMB to ensure we then also see migrating. |
| 789 | */ |
| 790 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { |
| 791 | rq_pin_lock(rq, rf); |
| 792 | return rq; |
| 793 | } |
| 794 | raw_spin_rq_unlock(rq); |
| 795 | raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); |
| 796 | |
| 797 | while (unlikely(task_on_rq_migrating(p))) |
| 798 | cpu_relax(); |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | /* |
| 803 | * RQ-clock updating methods: |
| 804 | */ |
| 805 | |
| 806 | /* Use CONFIG_PARAVIRT as this will avoid more #ifdef in arch code. */ |
| 807 | #ifdef CONFIG_PARAVIRT |
| 808 | DEFINE_STATIC_KEY_FALSE(paravirt_steal_rq_enabled); |
| 809 | #endif |
| 810 | |
| 811 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
| 812 | { |
| 813 | /* |
| 814 | * In theory, the compile should just see 0 here, and optimize out the call |
| 815 | * to sched_rt_avg_update. But I don't trust it... |
| 816 | */ |
| 817 | s64 __maybe_unused steal = 0, irq_delta = 0; |
| 818 | |
| 819 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
| 820 | if (irqtime_enabled()) { |
| 821 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; |
| 822 | |
| 823 | /* |
| 824 | * Since irq_time is only updated on {soft,}irq_exit, we might run into |
| 825 | * this case when a previous update_rq_clock() happened inside a |
| 826 | * {soft,}IRQ region. |
| 827 | * |
| 828 | * When this happens, we stop ->clock_task and only update the |
| 829 | * prev_irq_time stamp to account for the part that fit, so that a next |
| 830 | * update will consume the rest. This ensures ->clock_task is |
| 831 | * monotonic. |
| 832 | * |
| 833 | * It does however cause some slight miss-attribution of {soft,}IRQ |
| 834 | * time, a more accurate solution would be to update the irq_time using |
| 835 | * the current rq->clock timestamp, except that would require using |
| 836 | * atomic ops. |
| 837 | */ |
| 838 | if (irq_delta > delta) |
| 839 | irq_delta = delta; |
| 840 | |
| 841 | rq->prev_irq_time += irq_delta; |
| 842 | delta -= irq_delta; |
| 843 | delayacct_irq(rq->curr, irq_delta); |
| 844 | } |
| 845 | #endif |
| 846 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING |
| 847 | if (static_branch_unlikely(¶virt_steal_rq_enabled)) { |
| 848 | u64 prev_steal; |
| 849 | |
| 850 | steal = prev_steal = paravirt_steal_clock(cpu_of(rq)); |
| 851 | steal -= rq->prev_steal_time_rq; |
| 852 | |
| 853 | if (unlikely(steal > delta)) |
| 854 | steal = delta; |
| 855 | |
| 856 | rq->prev_steal_time_rq = prev_steal; |
| 857 | delta -= steal; |
| 858 | } |
| 859 | #endif |
| 860 | |
| 861 | rq->clock_task += delta; |
| 862 | |
| 863 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
| 864 | if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) |
| 865 | update_irq_load_avg(rq, irq_delta + steal); |
| 866 | #endif |
| 867 | update_rq_clock_pelt(rq, delta); |
| 868 | } |
| 869 | |
| 870 | void update_rq_clock(struct rq *rq) |
| 871 | { |
| 872 | s64 delta; |
| 873 | u64 clock; |
| 874 | |
| 875 | lockdep_assert_rq_held(rq); |
| 876 | |
| 877 | if (rq->clock_update_flags & RQCF_ACT_SKIP) |
| 878 | return; |
| 879 | |
| 880 | if (sched_feat(WARN_DOUBLE_CLOCK)) |
| 881 | WARN_ON_ONCE(rq->clock_update_flags & RQCF_UPDATED); |
| 882 | rq->clock_update_flags |= RQCF_UPDATED; |
| 883 | |
| 884 | clock = sched_clock_cpu(cpu_of(rq)); |
| 885 | scx_rq_clock_update(rq, clock); |
| 886 | |
| 887 | delta = clock - rq->clock; |
| 888 | if (delta < 0) |
| 889 | return; |
| 890 | rq->clock += delta; |
| 891 | |
| 892 | update_rq_clock_task(rq, delta); |
| 893 | } |
| 894 | |
| 895 | #ifdef CONFIG_SCHED_HRTICK |
| 896 | /* |
| 897 | * Use HR-timers to deliver accurate preemption points. |
| 898 | */ |
| 899 | |
| 900 | enum { |
| 901 | HRTICK_SCHED_NONE = 0, |
| 902 | HRTICK_SCHED_DEFER = BIT(1), |
| 903 | HRTICK_SCHED_START = BIT(2), |
| 904 | HRTICK_SCHED_REARM_HRTIMER = BIT(3) |
| 905 | }; |
| 906 | |
| 907 | static void __used hrtick_clear(struct rq *rq) |
| 908 | { |
| 909 | if (hrtimer_active(&rq->hrtick_timer)) |
| 910 | hrtimer_cancel(&rq->hrtick_timer); |
| 911 | } |
| 912 | |
| 913 | /* |
| 914 | * High-resolution timer tick. |
| 915 | * Runs from hardirq context with interrupts disabled. |
| 916 | */ |
| 917 | static enum hrtimer_restart hrtick(struct hrtimer *timer) |
| 918 | { |
| 919 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); |
| 920 | struct rq_flags rf; |
| 921 | |
| 922 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
| 923 | |
| 924 | rq_lock(rq, &rf); |
| 925 | update_rq_clock(rq); |
| 926 | rq->donor->sched_class->task_tick(rq, rq->donor, 1); |
| 927 | rq_unlock(rq, &rf); |
| 928 | |
| 929 | return HRTIMER_NORESTART; |
| 930 | } |
| 931 | |
| 932 | static inline bool hrtick_needs_rearm(struct hrtimer *timer, ktime_t expires) |
| 933 | { |
| 934 | /* |
| 935 | * Queued is false when the timer is not started or currently |
| 936 | * running the callback. In both cases, restart. If queued check |
| 937 | * whether the expiry time actually changes substantially. |
| 938 | */ |
| 939 | return !hrtimer_is_queued(timer) || |
| 940 | abs(expires - hrtimer_get_expires(timer)) > 5000; |
| 941 | } |
| 942 | |
| 943 | static void hrtick_cond_restart(struct rq *rq) |
| 944 | { |
| 945 | struct hrtimer *timer = &rq->hrtick_timer; |
| 946 | ktime_t time = rq->hrtick_time; |
| 947 | |
| 948 | if (hrtick_needs_rearm(timer, time)) |
| 949 | hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD); |
| 950 | } |
| 951 | |
| 952 | /* |
| 953 | * called from hardirq (IPI) context |
| 954 | */ |
| 955 | static void __hrtick_start(void *arg) |
| 956 | { |
| 957 | struct rq *rq = arg; |
| 958 | struct rq_flags rf; |
| 959 | |
| 960 | rq_lock(rq, &rf); |
| 961 | hrtick_cond_restart(rq); |
| 962 | rq_unlock(rq, &rf); |
| 963 | } |
| 964 | |
| 965 | /* |
| 966 | * Called to set the hrtick timer state. |
| 967 | * |
| 968 | * called with rq->lock held and IRQs disabled |
| 969 | */ |
| 970 | void hrtick_start(struct rq *rq, u64 delay) |
| 971 | { |
| 972 | s64 delta; |
| 973 | |
| 974 | /* |
| 975 | * Don't schedule slices shorter than 10000ns, that just |
| 976 | * doesn't make sense and can cause timer DoS. |
| 977 | */ |
| 978 | delta = max_t(s64, delay, 10000LL); |
| 979 | |
| 980 | /* |
| 981 | * If this is in the middle of schedule() only note the delay |
| 982 | * and let hrtick_schedule_exit() deal with it. |
| 983 | */ |
| 984 | if (rq->hrtick_sched) { |
| 985 | rq->hrtick_sched |= HRTICK_SCHED_START; |
| 986 | rq->hrtick_delay = delta; |
| 987 | return; |
| 988 | } |
| 989 | |
| 990 | rq->hrtick_time = ktime_add_ns(ktime_get(), delta); |
| 991 | if (!hrtick_needs_rearm(&rq->hrtick_timer, rq->hrtick_time)) |
| 992 | return; |
| 993 | |
| 994 | if (rq == this_rq()) |
| 995 | hrtimer_start(&rq->hrtick_timer, rq->hrtick_time, HRTIMER_MODE_ABS_PINNED_HARD); |
| 996 | else |
| 997 | smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); |
| 998 | } |
| 999 | |
| 1000 | static inline void hrtick_schedule_enter(struct rq *rq) |
| 1001 | { |
| 1002 | rq->hrtick_sched = HRTICK_SCHED_DEFER; |
| 1003 | if (hrtimer_test_and_clear_rearm_deferred()) |
| 1004 | rq->hrtick_sched |= HRTICK_SCHED_REARM_HRTIMER; |
| 1005 | } |
| 1006 | |
| 1007 | static inline void hrtick_schedule_exit(struct rq *rq) |
| 1008 | { |
| 1009 | if (rq->hrtick_sched & HRTICK_SCHED_START) { |
| 1010 | rq->hrtick_time = ktime_add_ns(ktime_get(), rq->hrtick_delay); |
| 1011 | hrtick_cond_restart(rq); |
| 1012 | } else if (idle_rq(rq)) { |
| 1013 | /* |
| 1014 | * No need for using hrtimer_is_active(). The timer is CPU local |
| 1015 | * and interrupts are disabled, so the callback cannot be |
| 1016 | * running and the queued state is valid. |
| 1017 | */ |
| 1018 | if (hrtimer_is_queued(&rq->hrtick_timer)) |
| 1019 | hrtimer_cancel(&rq->hrtick_timer); |
| 1020 | } |
| 1021 | |
| 1022 | if (rq->hrtick_sched & HRTICK_SCHED_REARM_HRTIMER) |
| 1023 | __hrtimer_rearm_deferred(); |
| 1024 | |
| 1025 | rq->hrtick_sched = HRTICK_SCHED_NONE; |
| 1026 | } |
| 1027 | |
| 1028 | static void hrtick_rq_init(struct rq *rq) |
| 1029 | { |
| 1030 | INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq); |
| 1031 | rq->hrtick_sched = HRTICK_SCHED_NONE; |
| 1032 | hrtimer_setup(&rq->hrtick_timer, hrtick, CLOCK_MONOTONIC, |
| 1033 | HRTIMER_MODE_REL_HARD | HRTIMER_MODE_LAZY_REARM); |
| 1034 | } |
| 1035 | #else /* !CONFIG_SCHED_HRTICK: */ |
| 1036 | static inline void hrtick_clear(struct rq *rq) { } |
| 1037 | static inline void hrtick_rq_init(struct rq *rq) { } |
| 1038 | static inline void hrtick_schedule_enter(struct rq *rq) { } |
| 1039 | static inline void hrtick_schedule_exit(struct rq *rq) { } |
| 1040 | #endif /* !CONFIG_SCHED_HRTICK */ |
| 1041 | |
| 1042 | /* |
| 1043 | * try_cmpxchg based fetch_or() macro so it works for different integer types: |
| 1044 | */ |
| 1045 | #define fetch_or(ptr, mask) \ |
| 1046 | ({ \ |
| 1047 | typeof(ptr) _ptr = (ptr); \ |
| 1048 | typeof(mask) _mask = (mask); \ |
| 1049 | typeof(*_ptr) _val = *_ptr; \ |
| 1050 | \ |
| 1051 | do { \ |
| 1052 | } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \ |
| 1053 | _val; \ |
| 1054 | }) |
| 1055 | |
| 1056 | #ifdef TIF_POLLING_NRFLAG |
| 1057 | /* |
| 1058 | * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, |
| 1059 | * this avoids any races wrt polling state changes and thereby avoids |
| 1060 | * spurious IPIs. |
| 1061 | */ |
| 1062 | static inline bool set_nr_and_not_polling(struct thread_info *ti, int tif) |
| 1063 | { |
| 1064 | return !(fetch_or(&ti->flags, 1 << tif) & _TIF_POLLING_NRFLAG); |
| 1065 | } |
| 1066 | |
| 1067 | /* |
| 1068 | * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. |
| 1069 | * |
| 1070 | * If this returns true, then the idle task promises to call |
| 1071 | * sched_ttwu_pending() and reschedule soon. |
| 1072 | */ |
| 1073 | static bool set_nr_if_polling(struct task_struct *p) |
| 1074 | { |
| 1075 | struct thread_info *ti = task_thread_info(p); |
| 1076 | typeof(ti->flags) val = READ_ONCE(ti->flags); |
| 1077 | |
| 1078 | do { |
| 1079 | if (!(val & _TIF_POLLING_NRFLAG)) |
| 1080 | return false; |
| 1081 | if (val & _TIF_NEED_RESCHED) |
| 1082 | return true; |
| 1083 | } while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED)); |
| 1084 | |
| 1085 | return true; |
| 1086 | } |
| 1087 | |
| 1088 | #else |
| 1089 | static inline bool set_nr_and_not_polling(struct thread_info *ti, int tif) |
| 1090 | { |
| 1091 | set_ti_thread_flag(ti, tif); |
| 1092 | return true; |
| 1093 | } |
| 1094 | |
| 1095 | static inline bool set_nr_if_polling(struct task_struct *p) |
| 1096 | { |
| 1097 | return false; |
| 1098 | } |
| 1099 | #endif |
| 1100 | |
| 1101 | static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task) |
| 1102 | { |
| 1103 | struct wake_q_node *node = &task->wake_q; |
| 1104 | |
| 1105 | /* |
| 1106 | * Atomically grab the task, if ->wake_q is !nil already it means |
| 1107 | * it's already queued (either by us or someone else) and will get the |
| 1108 | * wakeup due to that. |
| 1109 | * |
| 1110 | * In order to ensure that a pending wakeup will observe our pending |
| 1111 | * state, even in the failed case, an explicit smp_mb() must be used. |
| 1112 | */ |
| 1113 | smp_mb__before_atomic(); |
| 1114 | if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))) |
| 1115 | return false; |
| 1116 | |
| 1117 | /* |
| 1118 | * The head is context local, there can be no concurrency. |
| 1119 | */ |
| 1120 | *head->lastp = node; |
| 1121 | head->lastp = &node->next; |
| 1122 | return true; |
| 1123 | } |
| 1124 | |
| 1125 | /** |
| 1126 | * wake_q_add() - queue a wakeup for 'later' waking. |
| 1127 | * @head: the wake_q_head to add @task to |
| 1128 | * @task: the task to queue for 'later' wakeup |
| 1129 | * |
| 1130 | * Queue a task for later wakeup, most likely by the wake_up_q() call in the |
| 1131 | * same context, _HOWEVER_ this is not guaranteed, the wakeup can come |
| 1132 | * instantly. |
| 1133 | * |
| 1134 | * This function must be used as-if it were wake_up_process(); IOW the task |
| 1135 | * must be ready to be woken at this location. |
| 1136 | */ |
| 1137 | void wake_q_add(struct wake_q_head *head, struct task_struct *task) |
| 1138 | { |
| 1139 | if (__wake_q_add(head, task)) |
| 1140 | get_task_struct(task); |
| 1141 | } |
| 1142 | |
| 1143 | /** |
| 1144 | * wake_q_add_safe() - safely queue a wakeup for 'later' waking. |
| 1145 | * @head: the wake_q_head to add @task to |
| 1146 | * @task: the task to queue for 'later' wakeup |
| 1147 | * |
| 1148 | * Queue a task for later wakeup, most likely by the wake_up_q() call in the |
| 1149 | * same context, _HOWEVER_ this is not guaranteed, the wakeup can come |
| 1150 | * instantly. |
| 1151 | * |
| 1152 | * This function must be used as-if it were wake_up_process(); IOW the task |
| 1153 | * must be ready to be woken at this location. |
| 1154 | * |
| 1155 | * This function is essentially a task-safe equivalent to wake_q_add(). Callers |
| 1156 | * that already hold reference to @task can call the 'safe' version and trust |
| 1157 | * wake_q to do the right thing depending whether or not the @task is already |
| 1158 | * queued for wakeup. |
| 1159 | */ |
| 1160 | void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task) |
| 1161 | { |
| 1162 | if (!__wake_q_add(head, task)) |
| 1163 | put_task_struct(task); |
| 1164 | } |
| 1165 | |
| 1166 | void wake_up_q(struct wake_q_head *head) |
| 1167 | { |
| 1168 | struct wake_q_node *node = head->first; |
| 1169 | |
| 1170 | while (node != WAKE_Q_TAIL) { |
| 1171 | struct task_struct *task; |
| 1172 | |
| 1173 | task = container_of(node, struct task_struct, wake_q); |
| 1174 | node = node->next; |
| 1175 | /* pairs with cmpxchg_relaxed() in __wake_q_add() */ |
| 1176 | WRITE_ONCE(task->wake_q.next, NULL); |
| 1177 | /* Task can safely be re-inserted now. */ |
| 1178 | |
| 1179 | /* |
| 1180 | * wake_up_process() executes a full barrier, which pairs with |
| 1181 | * the queueing in wake_q_add() so as not to miss wakeups. |
| 1182 | */ |
| 1183 | wake_up_process(task); |
| 1184 | put_task_struct(task); |
| 1185 | } |
| 1186 | } |
| 1187 | |
| 1188 | /* |
| 1189 | * resched_curr - mark rq's current task 'to be rescheduled now'. |
| 1190 | * |
| 1191 | * On UP this means the setting of the need_resched flag, on SMP it |
| 1192 | * might also involve a cross-CPU call to trigger the scheduler on |
| 1193 | * the target CPU. |
| 1194 | */ |
| 1195 | static void __resched_curr(struct rq *rq, int tif) |
| 1196 | { |
| 1197 | struct task_struct *curr = rq->curr; |
| 1198 | struct thread_info *cti = task_thread_info(curr); |
| 1199 | bool need_ipi; |
| 1200 | int cpu; |
| 1201 | |
| 1202 | lockdep_assert_rq_held(rq); |
| 1203 | |
| 1204 | /* |
| 1205 | * Always immediately preempt the idle task; no point in delaying doing |
| 1206 | * actual work. |
| 1207 | */ |
| 1208 | if (is_idle_task(curr) && tif == TIF_NEED_RESCHED_LAZY) |
| 1209 | tif = TIF_NEED_RESCHED; |
| 1210 | |
| 1211 | if (cti->flags & ((1 << tif) | _TIF_NEED_RESCHED)) |
| 1212 | return; |
| 1213 | |
| 1214 | cpu = cpu_of(rq); |
| 1215 | |
| 1216 | if (cpu == smp_processor_id()) { |
| 1217 | set_ti_thread_flag(cti, tif); |
| 1218 | if (tif == TIF_NEED_RESCHED) |
| 1219 | set_preempt_need_resched(); |
| 1220 | trace_sched_set_need_resched_tp(curr, cpu, tif); |
| 1221 | return; |
| 1222 | } |
| 1223 | |
| 1224 | need_ipi = set_nr_and_not_polling(cti, tif); |
| 1225 | trace_sched_set_need_resched_tp(curr, cpu, tif); |
| 1226 | if (need_ipi) { |
| 1227 | if (tif == TIF_NEED_RESCHED) |
| 1228 | smp_send_reschedule(cpu); |
| 1229 | } else { |
| 1230 | trace_sched_wake_idle_without_ipi(cpu); |
| 1231 | } |
| 1232 | } |
| 1233 | |
| 1234 | /* |
| 1235 | * Calls to this function MUST be guarded by a |
| 1236 | * tracepoint_enabled(sched_set_need_resched_tp) |
| 1237 | */ |
| 1238 | void __trace_set_need_resched(struct task_struct *curr, int tif) |
| 1239 | { |
| 1240 | trace_call__sched_set_need_resched_tp(curr, smp_processor_id(), tif); |
| 1241 | } |
| 1242 | EXPORT_SYMBOL_GPL(__trace_set_need_resched); |
| 1243 | |
| 1244 | void resched_curr(struct rq *rq) |
| 1245 | { |
| 1246 | __resched_curr(rq, TIF_NEED_RESCHED); |
| 1247 | } |
| 1248 | |
| 1249 | #ifdef CONFIG_PREEMPT_DYNAMIC |
| 1250 | static DEFINE_STATIC_KEY_FALSE(sk_dynamic_preempt_lazy); |
| 1251 | static __always_inline bool dynamic_preempt_lazy(void) |
| 1252 | { |
| 1253 | return static_branch_unlikely(&sk_dynamic_preempt_lazy); |
| 1254 | } |
| 1255 | #else |
| 1256 | static __always_inline bool dynamic_preempt_lazy(void) |
| 1257 | { |
| 1258 | return IS_ENABLED(CONFIG_PREEMPT_LAZY); |
| 1259 | } |
| 1260 | #endif |
| 1261 | |
| 1262 | static __always_inline int get_lazy_tif_bit(void) |
| 1263 | { |
| 1264 | if (dynamic_preempt_lazy()) |
| 1265 | return TIF_NEED_RESCHED_LAZY; |
| 1266 | |
| 1267 | return TIF_NEED_RESCHED; |
| 1268 | } |
| 1269 | |
| 1270 | void resched_curr_lazy(struct rq *rq) |
| 1271 | { |
| 1272 | __resched_curr(rq, get_lazy_tif_bit()); |
| 1273 | } |
| 1274 | |
| 1275 | void resched_cpu(int cpu) |
| 1276 | { |
| 1277 | struct rq *rq = cpu_rq(cpu); |
| 1278 | unsigned long flags; |
| 1279 | |
| 1280 | raw_spin_rq_lock_irqsave(rq, flags); |
| 1281 | if (cpu_online(cpu) || cpu == smp_processor_id()) |
| 1282 | resched_curr(rq); |
| 1283 | raw_spin_rq_unlock_irqrestore(rq, flags); |
| 1284 | } |
| 1285 | |
| 1286 | #ifdef CONFIG_NO_HZ_COMMON |
| 1287 | /* |
| 1288 | * In the semi idle case, use the nearest busy CPU for migrating timers |
| 1289 | * from an idle CPU. This is good for power-savings. |
| 1290 | * |
| 1291 | * We don't do similar optimization for completely idle system, as |
| 1292 | * selecting an idle CPU will add more delays to the timers than intended |
| 1293 | * (as that CPU's timer base may not be up to date wrt jiffies etc). |
| 1294 | */ |
| 1295 | int get_nohz_timer_target(void) |
| 1296 | { |
| 1297 | int i, cpu = smp_processor_id(), default_cpu = -1; |
| 1298 | struct sched_domain *sd; |
| 1299 | const struct cpumask *hk_mask; |
| 1300 | |
| 1301 | if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) { |
| 1302 | if (!idle_cpu(cpu)) |
| 1303 | return cpu; |
| 1304 | default_cpu = cpu; |
| 1305 | } |
| 1306 | |
| 1307 | hk_mask = housekeeping_cpumask(HK_TYPE_KERNEL_NOISE); |
| 1308 | |
| 1309 | guard(rcu)(); |
| 1310 | |
| 1311 | for_each_domain(cpu, sd) { |
| 1312 | for_each_cpu_and(i, sched_domain_span(sd), hk_mask) { |
| 1313 | if (cpu == i) |
| 1314 | continue; |
| 1315 | |
| 1316 | if (!idle_cpu(i)) |
| 1317 | return i; |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | if (default_cpu == -1) |
| 1322 | default_cpu = housekeeping_any_cpu(HK_TYPE_KERNEL_NOISE); |
| 1323 | |
| 1324 | return default_cpu; |
| 1325 | } |
| 1326 | |
| 1327 | /* |
| 1328 | * When add_timer_on() enqueues a timer into the timer wheel of an |
| 1329 | * idle CPU then this timer might expire before the next timer event |
| 1330 | * which is scheduled to wake up that CPU. In case of a completely |
| 1331 | * idle system the next event might even be infinite time into the |
| 1332 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and |
| 1333 | * leaves the inner idle loop so the newly added timer is taken into |
| 1334 | * account when the CPU goes back to idle and evaluates the timer |
| 1335 | * wheel for the next timer event. |
| 1336 | */ |
| 1337 | static void wake_up_idle_cpu(int cpu) |
| 1338 | { |
| 1339 | struct rq *rq = cpu_rq(cpu); |
| 1340 | |
| 1341 | if (cpu == smp_processor_id()) |
| 1342 | return; |
| 1343 | |
| 1344 | /* |
| 1345 | * Set TIF_NEED_RESCHED and send an IPI if in the non-polling |
| 1346 | * part of the idle loop. This forces an exit from the idle loop |
| 1347 | * and a round trip to schedule(). Now this could be optimized |
| 1348 | * because a simple new idle loop iteration is enough to |
| 1349 | * re-evaluate the next tick. Provided some re-ordering of tick |
| 1350 | * nohz functions that would need to follow TIF_NR_POLLING |
| 1351 | * clearing: |
| 1352 | * |
| 1353 | * - On most architectures, a simple fetch_or on ti::flags with a |
| 1354 | * "0" value would be enough to know if an IPI needs to be sent. |
| 1355 | * |
| 1356 | * - x86 needs to perform a last need_resched() check between |
| 1357 | * monitor and mwait which doesn't take timers into account. |
| 1358 | * There a dedicated TIF_TIMER flag would be required to |
| 1359 | * fetch_or here and be checked along with TIF_NEED_RESCHED |
| 1360 | * before mwait(). |
| 1361 | * |
| 1362 | * However, remote timer enqueue is not such a frequent event |
| 1363 | * and testing of the above solutions didn't appear to report |
| 1364 | * much benefits. |
| 1365 | */ |
| 1366 | if (set_nr_and_not_polling(task_thread_info(rq->idle), TIF_NEED_RESCHED)) |
| 1367 | smp_send_reschedule(cpu); |
| 1368 | else |
| 1369 | trace_sched_wake_idle_without_ipi(cpu); |
| 1370 | } |
| 1371 | |
| 1372 | static bool wake_up_full_nohz_cpu(int cpu) |
| 1373 | { |
| 1374 | /* |
| 1375 | * We just need the target to call irq_exit() and re-evaluate |
| 1376 | * the next tick. The nohz full kick at least implies that. |
| 1377 | * If needed we can still optimize that later with an |
| 1378 | * empty IRQ. |
| 1379 | */ |
| 1380 | if (cpu_is_offline(cpu)) |
| 1381 | return true; /* Don't try to wake offline CPUs. */ |
| 1382 | if (tick_nohz_full_cpu(cpu)) { |
| 1383 | if (cpu != smp_processor_id() || |
| 1384 | tick_nohz_tick_stopped()) |
| 1385 | tick_nohz_full_kick_cpu(cpu); |
| 1386 | return true; |
| 1387 | } |
| 1388 | |
| 1389 | return false; |
| 1390 | } |
| 1391 | |
| 1392 | /* |
| 1393 | * Wake up the specified CPU. If the CPU is going offline, it is the |
| 1394 | * caller's responsibility to deal with the lost wakeup, for example, |
| 1395 | * by hooking into the CPU_DEAD notifier like timers and hrtimers do. |
| 1396 | */ |
| 1397 | void wake_up_nohz_cpu(int cpu) |
| 1398 | { |
| 1399 | if (!wake_up_full_nohz_cpu(cpu)) |
| 1400 | wake_up_idle_cpu(cpu); |
| 1401 | } |
| 1402 | |
| 1403 | static void nohz_csd_func(void *info) |
| 1404 | { |
| 1405 | struct rq *rq = info; |
| 1406 | int cpu = cpu_of(rq); |
| 1407 | unsigned int flags; |
| 1408 | |
| 1409 | /* |
| 1410 | * Release the rq::nohz_csd. |
| 1411 | */ |
| 1412 | flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu)); |
| 1413 | WARN_ON(!(flags & NOHZ_KICK_MASK)); |
| 1414 | |
| 1415 | rq->idle_balance = idle_cpu(cpu); |
| 1416 | if (rq->idle_balance) { |
| 1417 | rq->nohz_idle_balance = flags; |
| 1418 | __raise_softirq_irqoff(SCHED_SOFTIRQ); |
| 1419 | } |
| 1420 | } |
| 1421 | |
| 1422 | #endif /* CONFIG_NO_HZ_COMMON */ |
| 1423 | |
| 1424 | #ifdef CONFIG_NO_HZ_FULL |
| 1425 | static inline bool __need_bw_check(struct rq *rq, struct task_struct *p) |
| 1426 | { |
| 1427 | if (rq->nr_running != 1) |
| 1428 | return false; |
| 1429 | |
| 1430 | if (p->sched_class != &fair_sched_class) |
| 1431 | return false; |
| 1432 | |
| 1433 | if (!task_on_rq_queued(p)) |
| 1434 | return false; |
| 1435 | |
| 1436 | return true; |
| 1437 | } |
| 1438 | |
| 1439 | bool sched_can_stop_tick(struct rq *rq) |
| 1440 | { |
| 1441 | int fifo_nr_running; |
| 1442 | |
| 1443 | /* Deadline tasks, even if single, need the tick */ |
| 1444 | if (rq->dl.dl_nr_running) |
| 1445 | return false; |
| 1446 | |
| 1447 | /* |
| 1448 | * If there are more than one RR tasks, we need the tick to affect the |
| 1449 | * actual RR behaviour. |
| 1450 | */ |
| 1451 | if (rq->rt.rr_nr_running) { |
| 1452 | if (rq->rt.rr_nr_running == 1) |
| 1453 | return true; |
| 1454 | else |
| 1455 | return false; |
| 1456 | } |
| 1457 | |
| 1458 | /* |
| 1459 | * If there's no RR tasks, but FIFO tasks, we can skip the tick, no |
| 1460 | * forced preemption between FIFO tasks. |
| 1461 | */ |
| 1462 | fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running; |
| 1463 | if (fifo_nr_running) |
| 1464 | return true; |
| 1465 | |
| 1466 | /* |
| 1467 | * If there are no DL,RR/FIFO tasks, there must only be CFS or SCX tasks |
| 1468 | * left. For CFS, if there's more than one we need the tick for |
| 1469 | * involuntary preemption. For SCX, ask. |
| 1470 | */ |
| 1471 | if (scx_enabled() && !scx_can_stop_tick(rq)) |
| 1472 | return false; |
| 1473 | |
| 1474 | if (rq->cfs.h_nr_queued > 1) |
| 1475 | return false; |
| 1476 | |
| 1477 | /* |
| 1478 | * If there is one task and it has CFS runtime bandwidth constraints |
| 1479 | * and it's on the cpu now we don't want to stop the tick. |
| 1480 | * This check prevents clearing the bit if a newly enqueued task here is |
| 1481 | * dequeued by migrating while the constrained task continues to run. |
| 1482 | * E.g. going from 2->1 without going through pick_next_task(). |
| 1483 | */ |
| 1484 | if (__need_bw_check(rq, rq->curr)) { |
| 1485 | if (cfs_task_bw_constrained(rq->curr)) |
| 1486 | return false; |
| 1487 | } |
| 1488 | |
| 1489 | return true; |
| 1490 | } |
| 1491 | #endif /* CONFIG_NO_HZ_FULL */ |
| 1492 | |
| 1493 | #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_FAIR_GROUP_SCHED) |
| 1494 | /* |
| 1495 | * Iterate task_group tree rooted at *from, calling @down when first entering a |
| 1496 | * node and @up when leaving it for the final time. |
| 1497 | * |
| 1498 | * Caller must hold rcu_lock or sufficient equivalent. |
| 1499 | */ |
| 1500 | int walk_tg_tree_from(struct task_group *from, |
| 1501 | tg_visitor down, tg_visitor up, void *data) |
| 1502 | { |
| 1503 | struct task_group *parent, *child; |
| 1504 | int ret; |
| 1505 | |
| 1506 | parent = from; |
| 1507 | |
| 1508 | down: |
| 1509 | ret = (*down)(parent, data); |
| 1510 | if (ret) |
| 1511 | goto out; |
| 1512 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
| 1513 | parent = child; |
| 1514 | goto down; |
| 1515 | |
| 1516 | up: |
| 1517 | continue; |
| 1518 | } |
| 1519 | ret = (*up)(parent, data); |
| 1520 | if (ret || parent == from) |
| 1521 | goto out; |
| 1522 | |
| 1523 | child = parent; |
| 1524 | parent = parent->parent; |
| 1525 | if (parent) |
| 1526 | goto up; |
| 1527 | out: |
| 1528 | return ret; |
| 1529 | } |
| 1530 | |
| 1531 | int tg_nop(struct task_group *tg, void *data) |
| 1532 | { |
| 1533 | return 0; |
| 1534 | } |
| 1535 | #endif |
| 1536 | |
| 1537 | void set_load_weight(struct task_struct *p, bool update_load) |
| 1538 | { |
| 1539 | int prio = p->static_prio - MAX_RT_PRIO; |
| 1540 | struct load_weight lw; |
| 1541 | |
| 1542 | if (task_has_idle_policy(p)) { |
| 1543 | lw.weight = scale_load(WEIGHT_IDLEPRIO); |
| 1544 | lw.inv_weight = WMULT_IDLEPRIO; |
| 1545 | } else { |
| 1546 | lw.weight = scale_load(sched_prio_to_weight[prio]); |
| 1547 | lw.inv_weight = sched_prio_to_wmult[prio]; |
| 1548 | } |
| 1549 | |
| 1550 | /* |
| 1551 | * SCHED_OTHER tasks have to update their load when changing their |
| 1552 | * weight |
| 1553 | */ |
| 1554 | if (update_load && p->sched_class->reweight_task) |
| 1555 | p->sched_class->reweight_task(task_rq(p), p, &lw); |
| 1556 | else |
| 1557 | p->se.load = lw; |
| 1558 | } |
| 1559 | |
| 1560 | #ifdef CONFIG_UCLAMP_TASK |
| 1561 | /* |
| 1562 | * Serializes updates of utilization clamp values |
| 1563 | * |
| 1564 | * The (slow-path) user-space triggers utilization clamp value updates which |
| 1565 | * can require updates on (fast-path) scheduler's data structures used to |
| 1566 | * support enqueue/dequeue operations. |
| 1567 | * While the per-CPU rq lock protects fast-path update operations, user-space |
| 1568 | * requests are serialized using a mutex to reduce the risk of conflicting |
| 1569 | * updates or API abuses. |
| 1570 | */ |
| 1571 | static __maybe_unused DEFINE_MUTEX(uclamp_mutex); |
| 1572 | |
| 1573 | /* Max allowed minimum utilization */ |
| 1574 | static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE; |
| 1575 | |
| 1576 | /* Max allowed maximum utilization */ |
| 1577 | static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE; |
| 1578 | |
| 1579 | /* |
| 1580 | * By default RT tasks run at the maximum performance point/capacity of the |
| 1581 | * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to |
| 1582 | * SCHED_CAPACITY_SCALE. |
| 1583 | * |
| 1584 | * This knob allows admins to change the default behavior when uclamp is being |
| 1585 | * used. In battery powered devices, particularly, running at the maximum |
| 1586 | * capacity and frequency will increase energy consumption and shorten the |
| 1587 | * battery life. |
| 1588 | * |
| 1589 | * This knob only affects RT tasks that their uclamp_se->user_defined == false. |
| 1590 | * |
| 1591 | * This knob will not override the system default sched_util_clamp_min defined |
| 1592 | * above. |
| 1593 | */ |
| 1594 | unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE; |
| 1595 | |
| 1596 | /* All clamps are required to be less or equal than these values */ |
| 1597 | static struct uclamp_se uclamp_default[UCLAMP_CNT]; |
| 1598 | |
| 1599 | /* |
| 1600 | * This static key is used to reduce the uclamp overhead in the fast path. It |
| 1601 | * primarily disables the call to uclamp_rq_{inc, dec}() in |
| 1602 | * enqueue/dequeue_task(). |
| 1603 | * |
| 1604 | * This allows users to continue to enable uclamp in their kernel config with |
| 1605 | * minimum uclamp overhead in the fast path. |
| 1606 | * |
| 1607 | * As soon as userspace modifies any of the uclamp knobs, the static key is |
| 1608 | * enabled, since we have an actual users that make use of uclamp |
| 1609 | * functionality. |
| 1610 | * |
| 1611 | * The knobs that would enable this static key are: |
| 1612 | * |
| 1613 | * * A task modifying its uclamp value with sched_setattr(). |
| 1614 | * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs. |
| 1615 | * * An admin modifying the cgroup cpu.uclamp.{min, max} |
| 1616 | */ |
| 1617 | DEFINE_STATIC_KEY_FALSE(sched_uclamp_used); |
| 1618 | |
| 1619 | static inline unsigned int |
| 1620 | uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id, |
| 1621 | unsigned int clamp_value) |
| 1622 | { |
| 1623 | /* |
| 1624 | * Avoid blocked utilization pushing up the frequency when we go |
| 1625 | * idle (which drops the max-clamp) by retaining the last known |
| 1626 | * max-clamp. |
| 1627 | */ |
| 1628 | if (clamp_id == UCLAMP_MAX) { |
| 1629 | rq->uclamp_flags |= UCLAMP_FLAG_IDLE; |
| 1630 | return clamp_value; |
| 1631 | } |
| 1632 | |
| 1633 | return uclamp_none(UCLAMP_MIN); |
| 1634 | } |
| 1635 | |
| 1636 | static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id, |
| 1637 | unsigned int clamp_value) |
| 1638 | { |
| 1639 | /* Reset max-clamp retention only on idle exit */ |
| 1640 | if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE)) |
| 1641 | return; |
| 1642 | |
| 1643 | uclamp_rq_set(rq, clamp_id, clamp_value); |
| 1644 | } |
| 1645 | |
| 1646 | static inline |
| 1647 | unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id, |
| 1648 | unsigned int clamp_value) |
| 1649 | { |
| 1650 | struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket; |
| 1651 | int bucket_id = UCLAMP_BUCKETS - 1; |
| 1652 | |
| 1653 | /* |
| 1654 | * Since both min and max clamps are max aggregated, find the |
| 1655 | * top most bucket with tasks in. |
| 1656 | */ |
| 1657 | for ( ; bucket_id >= 0; bucket_id--) { |
| 1658 | if (!bucket[bucket_id].tasks) |
| 1659 | continue; |
| 1660 | return bucket[bucket_id].value; |
| 1661 | } |
| 1662 | |
| 1663 | /* No tasks -- default clamp values */ |
| 1664 | return uclamp_idle_value(rq, clamp_id, clamp_value); |
| 1665 | } |
| 1666 | |
| 1667 | static void __uclamp_update_util_min_rt_default(struct task_struct *p) |
| 1668 | { |
| 1669 | unsigned int default_util_min; |
| 1670 | struct uclamp_se *uc_se; |
| 1671 | |
| 1672 | lockdep_assert_held(&p->pi_lock); |
| 1673 | |
| 1674 | uc_se = &p->uclamp_req[UCLAMP_MIN]; |
| 1675 | |
| 1676 | /* Only sync if user didn't override the default */ |
| 1677 | if (uc_se->user_defined) |
| 1678 | return; |
| 1679 | |
| 1680 | default_util_min = sysctl_sched_uclamp_util_min_rt_default; |
| 1681 | uclamp_se_set(uc_se, default_util_min, false); |
| 1682 | } |
| 1683 | |
| 1684 | static void uclamp_update_util_min_rt_default(struct task_struct *p) |
| 1685 | { |
| 1686 | if (!rt_task(p)) |
| 1687 | return; |
| 1688 | |
| 1689 | /* Protect updates to p->uclamp_* */ |
| 1690 | guard(task_rq_lock)(p); |
| 1691 | __uclamp_update_util_min_rt_default(p); |
| 1692 | } |
| 1693 | |
| 1694 | static inline struct uclamp_se |
| 1695 | uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id) |
| 1696 | { |
| 1697 | /* Copy by value as we could modify it */ |
| 1698 | struct uclamp_se uc_req = p->uclamp_req[clamp_id]; |
| 1699 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 1700 | unsigned int tg_min, tg_max, value; |
| 1701 | |
| 1702 | /* |
| 1703 | * Tasks in autogroups or root task group will be |
| 1704 | * restricted by system defaults. |
| 1705 | */ |
| 1706 | if (task_group_is_autogroup(task_group(p))) |
| 1707 | return uc_req; |
| 1708 | if (task_group(p) == &root_task_group) |
| 1709 | return uc_req; |
| 1710 | |
| 1711 | tg_min = task_group(p)->uclamp[UCLAMP_MIN].value; |
| 1712 | tg_max = task_group(p)->uclamp[UCLAMP_MAX].value; |
| 1713 | value = uc_req.value; |
| 1714 | value = clamp(value, tg_min, tg_max); |
| 1715 | uclamp_se_set(&uc_req, value, false); |
| 1716 | #endif |
| 1717 | |
| 1718 | return uc_req; |
| 1719 | } |
| 1720 | |
| 1721 | /* |
| 1722 | * The effective clamp bucket index of a task depends on, by increasing |
| 1723 | * priority: |
| 1724 | * - the task specific clamp value, when explicitly requested from userspace |
| 1725 | * - the task group effective clamp value, for tasks not either in the root |
| 1726 | * group or in an autogroup |
| 1727 | * - the system default clamp value, defined by the sysadmin |
| 1728 | */ |
| 1729 | static inline struct uclamp_se |
| 1730 | uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id) |
| 1731 | { |
| 1732 | struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id); |
| 1733 | struct uclamp_se uc_max = uclamp_default[clamp_id]; |
| 1734 | |
| 1735 | /* System default restrictions always apply */ |
| 1736 | if (unlikely(uc_req.value > uc_max.value)) |
| 1737 | return uc_max; |
| 1738 | |
| 1739 | return uc_req; |
| 1740 | } |
| 1741 | |
| 1742 | unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id) |
| 1743 | { |
| 1744 | struct uclamp_se uc_eff; |
| 1745 | |
| 1746 | /* Task currently refcounted: use back-annotated (effective) value */ |
| 1747 | if (p->uclamp[clamp_id].active) |
| 1748 | return (unsigned long)p->uclamp[clamp_id].value; |
| 1749 | |
| 1750 | uc_eff = uclamp_eff_get(p, clamp_id); |
| 1751 | |
| 1752 | return (unsigned long)uc_eff.value; |
| 1753 | } |
| 1754 | |
| 1755 | /* |
| 1756 | * When a task is enqueued on a rq, the clamp bucket currently defined by the |
| 1757 | * task's uclamp::bucket_id is refcounted on that rq. This also immediately |
| 1758 | * updates the rq's clamp value if required. |
| 1759 | * |
| 1760 | * Tasks can have a task-specific value requested from user-space, track |
| 1761 | * within each bucket the maximum value for tasks refcounted in it. |
| 1762 | * This "local max aggregation" allows to track the exact "requested" value |
| 1763 | * for each bucket when all its RUNNABLE tasks require the same clamp. |
| 1764 | */ |
| 1765 | static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p, |
| 1766 | enum uclamp_id clamp_id) |
| 1767 | { |
| 1768 | struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; |
| 1769 | struct uclamp_se *uc_se = &p->uclamp[clamp_id]; |
| 1770 | struct uclamp_bucket *bucket; |
| 1771 | |
| 1772 | lockdep_assert_rq_held(rq); |
| 1773 | |
| 1774 | /* Update task effective clamp */ |
| 1775 | p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id); |
| 1776 | |
| 1777 | bucket = &uc_rq->bucket[uc_se->bucket_id]; |
| 1778 | bucket->tasks++; |
| 1779 | uc_se->active = true; |
| 1780 | |
| 1781 | uclamp_idle_reset(rq, clamp_id, uc_se->value); |
| 1782 | |
| 1783 | /* |
| 1784 | * Local max aggregation: rq buckets always track the max |
| 1785 | * "requested" clamp value of its RUNNABLE tasks. |
| 1786 | */ |
| 1787 | if (bucket->tasks == 1 || uc_se->value > bucket->value) |
| 1788 | bucket->value = uc_se->value; |
| 1789 | |
| 1790 | if (uc_se->value > uclamp_rq_get(rq, clamp_id)) |
| 1791 | uclamp_rq_set(rq, clamp_id, uc_se->value); |
| 1792 | } |
| 1793 | |
| 1794 | /* |
| 1795 | * When a task is dequeued from a rq, the clamp bucket refcounted by the task |
| 1796 | * is released. If this is the last task reference counting the rq's max |
| 1797 | * active clamp value, then the rq's clamp value is updated. |
| 1798 | * |
| 1799 | * Both refcounted tasks and rq's cached clamp values are expected to be |
| 1800 | * always valid. If it's detected they are not, as defensive programming, |
| 1801 | * enforce the expected state and warn. |
| 1802 | */ |
| 1803 | static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p, |
| 1804 | enum uclamp_id clamp_id) |
| 1805 | { |
| 1806 | struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; |
| 1807 | struct uclamp_se *uc_se = &p->uclamp[clamp_id]; |
| 1808 | struct uclamp_bucket *bucket; |
| 1809 | unsigned int bkt_clamp; |
| 1810 | unsigned int rq_clamp; |
| 1811 | |
| 1812 | lockdep_assert_rq_held(rq); |
| 1813 | |
| 1814 | /* |
| 1815 | * If sched_uclamp_used was enabled after task @p was enqueued, |
| 1816 | * we could end up with unbalanced call to uclamp_rq_dec_id(). |
| 1817 | * |
| 1818 | * In this case the uc_se->active flag should be false since no uclamp |
| 1819 | * accounting was performed at enqueue time and we can just return |
| 1820 | * here. |
| 1821 | * |
| 1822 | * Need to be careful of the following enqueue/dequeue ordering |
| 1823 | * problem too |
| 1824 | * |
| 1825 | * enqueue(taskA) |
| 1826 | * // sched_uclamp_used gets enabled |
| 1827 | * enqueue(taskB) |
| 1828 | * dequeue(taskA) |
| 1829 | * // Must not decrement bucket->tasks here |
| 1830 | * dequeue(taskB) |
| 1831 | * |
| 1832 | * where we could end up with stale data in uc_se and |
| 1833 | * bucket[uc_se->bucket_id]. |
| 1834 | * |
| 1835 | * The following check here eliminates the possibility of such race. |
| 1836 | */ |
| 1837 | if (unlikely(!uc_se->active)) |
| 1838 | return; |
| 1839 | |
| 1840 | bucket = &uc_rq->bucket[uc_se->bucket_id]; |
| 1841 | |
| 1842 | WARN_ON_ONCE(!bucket->tasks); |
| 1843 | if (likely(bucket->tasks)) |
| 1844 | bucket->tasks--; |
| 1845 | |
| 1846 | uc_se->active = false; |
| 1847 | |
| 1848 | /* |
| 1849 | * Keep "local max aggregation" simple and accept to (possibly) |
| 1850 | * overboost some RUNNABLE tasks in the same bucket. |
| 1851 | * The rq clamp bucket value is reset to its base value whenever |
| 1852 | * there are no more RUNNABLE tasks refcounting it. |
| 1853 | */ |
| 1854 | if (likely(bucket->tasks)) |
| 1855 | return; |
| 1856 | |
| 1857 | rq_clamp = uclamp_rq_get(rq, clamp_id); |
| 1858 | /* |
| 1859 | * Defensive programming: this should never happen. If it happens, |
| 1860 | * e.g. due to future modification, warn and fix up the expected value. |
| 1861 | */ |
| 1862 | WARN_ON_ONCE(bucket->value > rq_clamp); |
| 1863 | if (bucket->value >= rq_clamp) { |
| 1864 | bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value); |
| 1865 | uclamp_rq_set(rq, clamp_id, bkt_clamp); |
| 1866 | } |
| 1867 | } |
| 1868 | |
| 1869 | static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p, int flags) |
| 1870 | { |
| 1871 | enum uclamp_id clamp_id; |
| 1872 | |
| 1873 | /* |
| 1874 | * Avoid any overhead until uclamp is actually used by the userspace. |
| 1875 | * |
| 1876 | * The condition is constructed such that a NOP is generated when |
| 1877 | * sched_uclamp_used is disabled. |
| 1878 | */ |
| 1879 | if (!uclamp_is_used()) |
| 1880 | return; |
| 1881 | |
| 1882 | if (unlikely(!p->sched_class->uclamp_enabled)) |
| 1883 | return; |
| 1884 | |
| 1885 | /* Only inc the delayed task which being woken up. */ |
| 1886 | if (p->se.sched_delayed && !(flags & ENQUEUE_DELAYED)) |
| 1887 | return; |
| 1888 | |
| 1889 | for_each_clamp_id(clamp_id) |
| 1890 | uclamp_rq_inc_id(rq, p, clamp_id); |
| 1891 | |
| 1892 | /* Reset clamp idle holding when there is one RUNNABLE task */ |
| 1893 | if (rq->uclamp_flags & UCLAMP_FLAG_IDLE) |
| 1894 | rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE; |
| 1895 | } |
| 1896 | |
| 1897 | static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) |
| 1898 | { |
| 1899 | enum uclamp_id clamp_id; |
| 1900 | |
| 1901 | /* |
| 1902 | * Avoid any overhead until uclamp is actually used by the userspace. |
| 1903 | * |
| 1904 | * The condition is constructed such that a NOP is generated when |
| 1905 | * sched_uclamp_used is disabled. |
| 1906 | */ |
| 1907 | if (!uclamp_is_used()) |
| 1908 | return; |
| 1909 | |
| 1910 | if (unlikely(!p->sched_class->uclamp_enabled)) |
| 1911 | return; |
| 1912 | |
| 1913 | if (p->se.sched_delayed) |
| 1914 | return; |
| 1915 | |
| 1916 | for_each_clamp_id(clamp_id) |
| 1917 | uclamp_rq_dec_id(rq, p, clamp_id); |
| 1918 | } |
| 1919 | |
| 1920 | static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p, |
| 1921 | enum uclamp_id clamp_id) |
| 1922 | { |
| 1923 | if (!p->uclamp[clamp_id].active) |
| 1924 | return; |
| 1925 | |
| 1926 | uclamp_rq_dec_id(rq, p, clamp_id); |
| 1927 | uclamp_rq_inc_id(rq, p, clamp_id); |
| 1928 | |
| 1929 | /* |
| 1930 | * Make sure to clear the idle flag if we've transiently reached 0 |
| 1931 | * active tasks on rq. |
| 1932 | */ |
| 1933 | if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE)) |
| 1934 | rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE; |
| 1935 | } |
| 1936 | |
| 1937 | static inline void |
| 1938 | uclamp_update_active(struct task_struct *p) |
| 1939 | { |
| 1940 | enum uclamp_id clamp_id; |
| 1941 | struct rq_flags rf; |
| 1942 | struct rq *rq; |
| 1943 | |
| 1944 | /* |
| 1945 | * Lock the task and the rq where the task is (or was) queued. |
| 1946 | * |
| 1947 | * We might lock the (previous) rq of a !RUNNABLE task, but that's the |
| 1948 | * price to pay to safely serialize util_{min,max} updates with |
| 1949 | * enqueues, dequeues and migration operations. |
| 1950 | * This is the same locking schema used by __set_cpus_allowed_ptr(). |
| 1951 | */ |
| 1952 | rq = task_rq_lock(p, &rf); |
| 1953 | |
| 1954 | /* |
| 1955 | * Setting the clamp bucket is serialized by task_rq_lock(). |
| 1956 | * If the task is not yet RUNNABLE and its task_struct is not |
| 1957 | * affecting a valid clamp bucket, the next time it's enqueued, |
| 1958 | * it will already see the updated clamp bucket value. |
| 1959 | */ |
| 1960 | for_each_clamp_id(clamp_id) |
| 1961 | uclamp_rq_reinc_id(rq, p, clamp_id); |
| 1962 | |
| 1963 | task_rq_unlock(rq, p, &rf); |
| 1964 | } |
| 1965 | |
| 1966 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 1967 | static inline void |
| 1968 | uclamp_update_active_tasks(struct cgroup_subsys_state *css) |
| 1969 | { |
| 1970 | struct css_task_iter it; |
| 1971 | struct task_struct *p; |
| 1972 | |
| 1973 | css_task_iter_start(css, 0, &it); |
| 1974 | while ((p = css_task_iter_next(&it))) |
| 1975 | uclamp_update_active(p); |
| 1976 | css_task_iter_end(&it); |
| 1977 | } |
| 1978 | |
| 1979 | static void cpu_util_update_eff(struct cgroup_subsys_state *css); |
| 1980 | #endif |
| 1981 | |
| 1982 | #ifdef CONFIG_SYSCTL |
| 1983 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 1984 | static void uclamp_update_root_tg(void) |
| 1985 | { |
| 1986 | struct task_group *tg = &root_task_group; |
| 1987 | |
| 1988 | uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN], |
| 1989 | sysctl_sched_uclamp_util_min, false); |
| 1990 | uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX], |
| 1991 | sysctl_sched_uclamp_util_max, false); |
| 1992 | |
| 1993 | guard(rcu)(); |
| 1994 | cpu_util_update_eff(&root_task_group.css); |
| 1995 | } |
| 1996 | #else |
| 1997 | static void uclamp_update_root_tg(void) { } |
| 1998 | #endif |
| 1999 | |
| 2000 | static void uclamp_sync_util_min_rt_default(void) |
| 2001 | { |
| 2002 | struct task_struct *g, *p; |
| 2003 | |
| 2004 | /* |
| 2005 | * copy_process() sysctl_uclamp |
| 2006 | * uclamp_min_rt = X; |
| 2007 | * write_lock(&tasklist_lock) read_lock(&tasklist_lock) |
| 2008 | * // link thread smp_mb__after_spinlock() |
| 2009 | * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock); |
| 2010 | * sched_post_fork() for_each_process_thread() |
| 2011 | * __uclamp_sync_rt() __uclamp_sync_rt() |
| 2012 | * |
| 2013 | * Ensures that either sched_post_fork() will observe the new |
| 2014 | * uclamp_min_rt or for_each_process_thread() will observe the new |
| 2015 | * task. |
| 2016 | */ |
| 2017 | read_lock(&tasklist_lock); |
| 2018 | smp_mb__after_spinlock(); |
| 2019 | read_unlock(&tasklist_lock); |
| 2020 | |
| 2021 | guard(rcu)(); |
| 2022 | for_each_process_thread(g, p) |
| 2023 | uclamp_update_util_min_rt_default(p); |
| 2024 | } |
| 2025 | |
| 2026 | static int sysctl_sched_uclamp_handler(const struct ctl_table *table, int write, |
| 2027 | void *buffer, size_t *lenp, loff_t *ppos) |
| 2028 | { |
| 2029 | bool update_root_tg = false; |
| 2030 | int old_min, old_max, old_min_rt; |
| 2031 | int result; |
| 2032 | |
| 2033 | guard(mutex)(&uclamp_mutex); |
| 2034 | |
| 2035 | old_min = sysctl_sched_uclamp_util_min; |
| 2036 | old_max = sysctl_sched_uclamp_util_max; |
| 2037 | old_min_rt = sysctl_sched_uclamp_util_min_rt_default; |
| 2038 | |
| 2039 | result = proc_dointvec(table, write, buffer, lenp, ppos); |
| 2040 | if (result) |
| 2041 | goto undo; |
| 2042 | if (!write) |
| 2043 | return 0; |
| 2044 | |
| 2045 | if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max || |
| 2046 | sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE || |
| 2047 | sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) { |
| 2048 | |
| 2049 | result = -EINVAL; |
| 2050 | goto undo; |
| 2051 | } |
| 2052 | |
| 2053 | if (old_min != sysctl_sched_uclamp_util_min) { |
| 2054 | uclamp_se_set(&uclamp_default[UCLAMP_MIN], |
| 2055 | sysctl_sched_uclamp_util_min, false); |
| 2056 | update_root_tg = true; |
| 2057 | } |
| 2058 | if (old_max != sysctl_sched_uclamp_util_max) { |
| 2059 | uclamp_se_set(&uclamp_default[UCLAMP_MAX], |
| 2060 | sysctl_sched_uclamp_util_max, false); |
| 2061 | update_root_tg = true; |
| 2062 | } |
| 2063 | |
| 2064 | if (update_root_tg) { |
| 2065 | sched_uclamp_enable(); |
| 2066 | uclamp_update_root_tg(); |
| 2067 | } |
| 2068 | |
| 2069 | if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) { |
| 2070 | sched_uclamp_enable(); |
| 2071 | uclamp_sync_util_min_rt_default(); |
| 2072 | } |
| 2073 | |
| 2074 | /* |
| 2075 | * We update all RUNNABLE tasks only when task groups are in use. |
| 2076 | * Otherwise, keep it simple and do just a lazy update at each next |
| 2077 | * task enqueue time. |
| 2078 | */ |
| 2079 | return 0; |
| 2080 | |
| 2081 | undo: |
| 2082 | sysctl_sched_uclamp_util_min = old_min; |
| 2083 | sysctl_sched_uclamp_util_max = old_max; |
| 2084 | sysctl_sched_uclamp_util_min_rt_default = old_min_rt; |
| 2085 | return result; |
| 2086 | } |
| 2087 | #endif /* CONFIG_SYSCTL */ |
| 2088 | |
| 2089 | static void uclamp_fork(struct task_struct *p) |
| 2090 | { |
| 2091 | enum uclamp_id clamp_id; |
| 2092 | |
| 2093 | /* |
| 2094 | * We don't need to hold task_rq_lock() when updating p->uclamp_* here |
| 2095 | * as the task is still at its early fork stages. |
| 2096 | */ |
| 2097 | for_each_clamp_id(clamp_id) |
| 2098 | p->uclamp[clamp_id].active = false; |
| 2099 | |
| 2100 | if (likely(!p->sched_reset_on_fork)) |
| 2101 | return; |
| 2102 | |
| 2103 | for_each_clamp_id(clamp_id) { |
| 2104 | uclamp_se_set(&p->uclamp_req[clamp_id], |
| 2105 | uclamp_none(clamp_id), false); |
| 2106 | } |
| 2107 | } |
| 2108 | |
| 2109 | static void uclamp_post_fork(struct task_struct *p) |
| 2110 | { |
| 2111 | uclamp_update_util_min_rt_default(p); |
| 2112 | } |
| 2113 | |
| 2114 | static void __init init_uclamp_rq(struct rq *rq) |
| 2115 | { |
| 2116 | enum uclamp_id clamp_id; |
| 2117 | struct uclamp_rq *uc_rq = rq->uclamp; |
| 2118 | |
| 2119 | for_each_clamp_id(clamp_id) { |
| 2120 | uc_rq[clamp_id] = (struct uclamp_rq) { |
| 2121 | .value = uclamp_none(clamp_id) |
| 2122 | }; |
| 2123 | } |
| 2124 | |
| 2125 | rq->uclamp_flags = UCLAMP_FLAG_IDLE; |
| 2126 | } |
| 2127 | |
| 2128 | static void __init init_uclamp(void) |
| 2129 | { |
| 2130 | struct uclamp_se uc_max = {}; |
| 2131 | enum uclamp_id clamp_id; |
| 2132 | int cpu; |
| 2133 | |
| 2134 | for_each_possible_cpu(cpu) |
| 2135 | init_uclamp_rq(cpu_rq(cpu)); |
| 2136 | |
| 2137 | for_each_clamp_id(clamp_id) { |
| 2138 | uclamp_se_set(&init_task.uclamp_req[clamp_id], |
| 2139 | uclamp_none(clamp_id), false); |
| 2140 | } |
| 2141 | |
| 2142 | /* System defaults allow max clamp values for both indexes */ |
| 2143 | uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false); |
| 2144 | for_each_clamp_id(clamp_id) { |
| 2145 | uclamp_default[clamp_id] = uc_max; |
| 2146 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 2147 | root_task_group.uclamp_req[clamp_id] = uc_max; |
| 2148 | root_task_group.uclamp[clamp_id] = uc_max; |
| 2149 | #endif |
| 2150 | } |
| 2151 | } |
| 2152 | |
| 2153 | #else /* !CONFIG_UCLAMP_TASK: */ |
| 2154 | static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p, int flags) { } |
| 2155 | static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { } |
| 2156 | static inline void uclamp_fork(struct task_struct *p) { } |
| 2157 | static inline void uclamp_post_fork(struct task_struct *p) { } |
| 2158 | static inline void init_uclamp(void) { } |
| 2159 | #endif /* !CONFIG_UCLAMP_TASK */ |
| 2160 | |
| 2161 | bool sched_task_on_rq(struct task_struct *p) |
| 2162 | { |
| 2163 | return task_on_rq_queued(p); |
| 2164 | } |
| 2165 | |
| 2166 | unsigned long get_wchan(struct task_struct *p) |
| 2167 | { |
| 2168 | unsigned long ip = 0; |
| 2169 | unsigned int state; |
| 2170 | |
| 2171 | if (!p || p == current) |
| 2172 | return 0; |
| 2173 | |
| 2174 | /* Only get wchan if task is blocked and we can keep it that way. */ |
| 2175 | raw_spin_lock_irq(&p->pi_lock); |
| 2176 | state = READ_ONCE(p->__state); |
| 2177 | smp_rmb(); /* see try_to_wake_up() */ |
| 2178 | if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq) |
| 2179 | ip = __get_wchan(p); |
| 2180 | raw_spin_unlock_irq(&p->pi_lock); |
| 2181 | |
| 2182 | return ip; |
| 2183 | } |
| 2184 | |
| 2185 | void enqueue_task(struct rq *rq, struct task_struct *p, int flags) |
| 2186 | { |
| 2187 | if (!(flags & ENQUEUE_NOCLOCK)) |
| 2188 | update_rq_clock(rq); |
| 2189 | |
| 2190 | /* |
| 2191 | * Can be before ->enqueue_task() because uclamp considers the |
| 2192 | * ENQUEUE_DELAYED task before its ->sched_delayed gets cleared |
| 2193 | * in ->enqueue_task(). |
| 2194 | */ |
| 2195 | uclamp_rq_inc(rq, p, flags); |
| 2196 | |
| 2197 | p->sched_class->enqueue_task(rq, p, flags); |
| 2198 | |
| 2199 | psi_enqueue(p, flags); |
| 2200 | |
| 2201 | if (!(flags & ENQUEUE_RESTORE)) |
| 2202 | sched_info_enqueue(rq, p); |
| 2203 | |
| 2204 | if (sched_core_enabled(rq)) |
| 2205 | sched_core_enqueue(rq, p); |
| 2206 | } |
| 2207 | |
| 2208 | /* |
| 2209 | * Must only return false when DEQUEUE_SLEEP. |
| 2210 | */ |
| 2211 | inline bool dequeue_task(struct rq *rq, struct task_struct *p, int flags) |
| 2212 | { |
| 2213 | if (sched_core_enabled(rq)) |
| 2214 | sched_core_dequeue(rq, p, flags); |
| 2215 | |
| 2216 | if (!(flags & DEQUEUE_NOCLOCK)) |
| 2217 | update_rq_clock(rq); |
| 2218 | |
| 2219 | if (!(flags & DEQUEUE_SAVE)) |
| 2220 | sched_info_dequeue(rq, p); |
| 2221 | |
| 2222 | psi_dequeue(p, flags); |
| 2223 | |
| 2224 | /* |
| 2225 | * Must be before ->dequeue_task() because ->dequeue_task() can 'fail' |
| 2226 | * and mark the task ->sched_delayed. |
| 2227 | */ |
| 2228 | uclamp_rq_dec(rq, p); |
| 2229 | return p->sched_class->dequeue_task(rq, p, flags); |
| 2230 | } |
| 2231 | |
| 2232 | void activate_task(struct rq *rq, struct task_struct *p, int flags) |
| 2233 | { |
| 2234 | if (task_on_rq_migrating(p)) |
| 2235 | flags |= ENQUEUE_MIGRATED; |
| 2236 | |
| 2237 | enqueue_task(rq, p, flags); |
| 2238 | |
| 2239 | WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED); |
| 2240 | ASSERT_EXCLUSIVE_WRITER(p->on_rq); |
| 2241 | } |
| 2242 | |
| 2243 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) |
| 2244 | { |
| 2245 | WARN_ON_ONCE(flags & DEQUEUE_SLEEP); |
| 2246 | |
| 2247 | WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING); |
| 2248 | ASSERT_EXCLUSIVE_WRITER(p->on_rq); |
| 2249 | |
| 2250 | /* |
| 2251 | * Code explicitly relies on TASK_ON_RQ_MIGRATING begin set *before* |
| 2252 | * dequeue_task() and cleared *after* enqueue_task(). |
| 2253 | */ |
| 2254 | |
| 2255 | dequeue_task(rq, p, flags); |
| 2256 | } |
| 2257 | |
| 2258 | static bool dequeue_block_task(struct rq *rq, struct task_struct *p, |
| 2259 | unsigned long task_state) |
| 2260 | { |
| 2261 | int flags = DEQUEUE_NOCLOCK; |
| 2262 | |
| 2263 | p->sched_contributes_to_load = |
| 2264 | (task_state & TASK_UNINTERRUPTIBLE) && |
| 2265 | !(task_state & TASK_NOLOAD) && |
| 2266 | !(task_state & TASK_FROZEN); |
| 2267 | |
| 2268 | if (unlikely(is_special_task_state(task_state))) |
| 2269 | flags |= DEQUEUE_SPECIAL; |
| 2270 | |
| 2271 | /* |
| 2272 | * __schedule() ttwu() |
| 2273 | * prev_state = prev->state; if (p->on_rq && ...) |
| 2274 | * if (prev_state) goto out; |
| 2275 | * p->on_rq = 0; smp_acquire__after_ctrl_dep(); |
| 2276 | * p->state = TASK_WAKING |
| 2277 | * |
| 2278 | * Where __schedule() and ttwu() have matching control dependencies. |
| 2279 | * |
| 2280 | * Once the caller invokes __block_task(), schedule() must not care about |
| 2281 | * p->state any more. |
| 2282 | */ |
| 2283 | return dequeue_task(rq, p, DEQUEUE_SLEEP | flags); |
| 2284 | } |
| 2285 | |
| 2286 | static void block_task(struct rq *rq, struct task_struct *p, unsigned long task_state) |
| 2287 | { |
| 2288 | if (dequeue_block_task(rq, p, task_state)) |
| 2289 | __block_task(rq, p); |
| 2290 | } |
| 2291 | |
| 2292 | /** |
| 2293 | * task_curr - is this task currently executing on a CPU? |
| 2294 | * @p: the task in question. |
| 2295 | * |
| 2296 | * Return: 1 if the task is currently executing. 0 otherwise. |
| 2297 | */ |
| 2298 | inline int task_curr(const struct task_struct *p) |
| 2299 | { |
| 2300 | return cpu_curr(task_cpu(p)) == p; |
| 2301 | } |
| 2302 | |
| 2303 | void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags) |
| 2304 | { |
| 2305 | struct task_struct *donor = rq->donor; |
| 2306 | |
| 2307 | if (p->sched_class == rq->next_class) { |
| 2308 | rq->next_class->wakeup_preempt(rq, p, flags); |
| 2309 | |
| 2310 | } else if (sched_class_above(p->sched_class, rq->next_class)) { |
| 2311 | rq->next_class->wakeup_preempt(rq, p, flags); |
| 2312 | resched_curr(rq); |
| 2313 | rq->next_class = p->sched_class; |
| 2314 | } |
| 2315 | |
| 2316 | /* |
| 2317 | * A queue event has occurred, and we're going to schedule. In |
| 2318 | * this case, we can save a useless back to back clock update. |
| 2319 | */ |
| 2320 | if (task_on_rq_queued(donor) && test_tsk_need_resched(rq->curr)) |
| 2321 | rq_clock_skip_update(rq); |
| 2322 | } |
| 2323 | |
| 2324 | static __always_inline |
| 2325 | int __task_state_match(struct task_struct *p, unsigned int state) |
| 2326 | { |
| 2327 | if (READ_ONCE(p->__state) & state) |
| 2328 | return 1; |
| 2329 | |
| 2330 | if (READ_ONCE(p->saved_state) & state) |
| 2331 | return -1; |
| 2332 | |
| 2333 | return 0; |
| 2334 | } |
| 2335 | |
| 2336 | static __always_inline |
| 2337 | int task_state_match(struct task_struct *p, unsigned int state) |
| 2338 | { |
| 2339 | /* |
| 2340 | * Serialize against current_save_and_set_rtlock_wait_state(), |
| 2341 | * current_restore_rtlock_saved_state(), and __refrigerator(). |
| 2342 | */ |
| 2343 | guard(raw_spinlock_irq)(&p->pi_lock); |
| 2344 | return __task_state_match(p, state); |
| 2345 | } |
| 2346 | |
| 2347 | /* |
| 2348 | * wait_task_inactive - wait for a thread to unschedule. |
| 2349 | * |
| 2350 | * Wait for the thread to block in any of the states set in @match_state. |
| 2351 | * If it changes, i.e. @p might have woken up, then return zero. When we |
| 2352 | * succeed in waiting for @p to be off its CPU, we return a positive number |
| 2353 | * (its total switch count). If a second call a short while later returns the |
| 2354 | * same number, the caller can be sure that @p has remained unscheduled the |
| 2355 | * whole time. |
| 2356 | * |
| 2357 | * The caller must ensure that the task *will* unschedule sometime soon, |
| 2358 | * else this function might spin for a *long* time. This function can't |
| 2359 | * be called with interrupts off, or it may introduce deadlock with |
| 2360 | * smp_call_function() if an IPI is sent by the same process we are |
| 2361 | * waiting to become inactive. |
| 2362 | */ |
| 2363 | unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state) |
| 2364 | { |
| 2365 | int running, queued, match; |
| 2366 | struct rq_flags rf; |
| 2367 | unsigned long ncsw; |
| 2368 | struct rq *rq; |
| 2369 | |
| 2370 | for (;;) { |
| 2371 | /* |
| 2372 | * We do the initial early heuristics without holding |
| 2373 | * any task-queue locks at all. We'll only try to get |
| 2374 | * the runqueue lock when things look like they will |
| 2375 | * work out! |
| 2376 | */ |
| 2377 | rq = task_rq(p); |
| 2378 | |
| 2379 | /* |
| 2380 | * If the task is actively running on another CPU |
| 2381 | * still, just relax and busy-wait without holding |
| 2382 | * any locks. |
| 2383 | * |
| 2384 | * NOTE! Since we don't hold any locks, it's not |
| 2385 | * even sure that "rq" stays as the right runqueue! |
| 2386 | * But we don't care, since "task_on_cpu()" will |
| 2387 | * return false if the runqueue has changed and p |
| 2388 | * is actually now running somewhere else! |
| 2389 | */ |
| 2390 | while (task_on_cpu(rq, p)) { |
| 2391 | if (!task_state_match(p, match_state)) |
| 2392 | return 0; |
| 2393 | cpu_relax(); |
| 2394 | } |
| 2395 | |
| 2396 | /* |
| 2397 | * Ok, time to look more closely! We need the rq |
| 2398 | * lock now, to be *sure*. If we're wrong, we'll |
| 2399 | * just go back and repeat. |
| 2400 | */ |
| 2401 | rq = task_rq_lock(p, &rf); |
| 2402 | /* |
| 2403 | * If task is sched_delayed, force dequeue it, to avoid always |
| 2404 | * hitting the tick timeout in the queued case |
| 2405 | */ |
| 2406 | if (p->se.sched_delayed) |
| 2407 | dequeue_task(rq, p, DEQUEUE_SLEEP | DEQUEUE_DELAYED); |
| 2408 | trace_sched_wait_task(p); |
| 2409 | running = task_on_cpu(rq, p); |
| 2410 | queued = task_on_rq_queued(p); |
| 2411 | ncsw = 0; |
| 2412 | if ((match = __task_state_match(p, match_state))) { |
| 2413 | /* |
| 2414 | * When matching on p->saved_state, consider this task |
| 2415 | * still queued so it will wait. |
| 2416 | */ |
| 2417 | if (match < 0) |
| 2418 | queued = 1; |
| 2419 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
| 2420 | } |
| 2421 | task_rq_unlock(rq, p, &rf); |
| 2422 | |
| 2423 | /* |
| 2424 | * If it changed from the expected state, bail out now. |
| 2425 | */ |
| 2426 | if (unlikely(!ncsw)) |
| 2427 | break; |
| 2428 | |
| 2429 | /* |
| 2430 | * Was it really running after all now that we |
| 2431 | * checked with the proper locks actually held? |
| 2432 | * |
| 2433 | * Oops. Go back and try again.. |
| 2434 | */ |
| 2435 | if (unlikely(running)) { |
| 2436 | cpu_relax(); |
| 2437 | continue; |
| 2438 | } |
| 2439 | |
| 2440 | /* |
| 2441 | * It's not enough that it's not actively running, |
| 2442 | * it must be off the runqueue _entirely_, and not |
| 2443 | * preempted! |
| 2444 | * |
| 2445 | * So if it was still runnable (but just not actively |
| 2446 | * running right now), it's preempted, and we should |
| 2447 | * yield - it could be a while. |
| 2448 | */ |
| 2449 | if (unlikely(queued)) { |
| 2450 | ktime_t to = NSEC_PER_SEC / HZ; |
| 2451 | |
| 2452 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 2453 | schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD); |
| 2454 | continue; |
| 2455 | } |
| 2456 | |
| 2457 | /* |
| 2458 | * Ahh, all good. It wasn't running, and it wasn't |
| 2459 | * runnable, which means that it will never become |
| 2460 | * running in the future either. We're all done! |
| 2461 | */ |
| 2462 | break; |
| 2463 | } |
| 2464 | |
| 2465 | return ncsw; |
| 2466 | } |
| 2467 | |
| 2468 | static void |
| 2469 | do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx); |
| 2470 | |
| 2471 | static void migrate_disable_switch(struct rq *rq, struct task_struct *p) |
| 2472 | { |
| 2473 | struct affinity_context ac = { |
| 2474 | .new_mask = cpumask_of(rq->cpu), |
| 2475 | .flags = SCA_MIGRATE_DISABLE, |
| 2476 | }; |
| 2477 | |
| 2478 | if (likely(!p->migration_disabled)) |
| 2479 | return; |
| 2480 | |
| 2481 | if (p->cpus_ptr != &p->cpus_mask) |
| 2482 | return; |
| 2483 | |
| 2484 | scoped_guard (task_rq_lock, p) |
| 2485 | do_set_cpus_allowed(p, &ac); |
| 2486 | } |
| 2487 | |
| 2488 | void ___migrate_enable(void) |
| 2489 | { |
| 2490 | struct task_struct *p = current; |
| 2491 | struct affinity_context ac = { |
| 2492 | .new_mask = &p->cpus_mask, |
| 2493 | .flags = SCA_MIGRATE_ENABLE, |
| 2494 | }; |
| 2495 | |
| 2496 | __set_cpus_allowed_ptr(p, &ac); |
| 2497 | } |
| 2498 | EXPORT_SYMBOL_GPL(___migrate_enable); |
| 2499 | |
| 2500 | void migrate_disable(void) |
| 2501 | { |
| 2502 | __migrate_disable(); |
| 2503 | } |
| 2504 | EXPORT_SYMBOL_GPL(migrate_disable); |
| 2505 | |
| 2506 | void migrate_enable(void) |
| 2507 | { |
| 2508 | __migrate_enable(); |
| 2509 | } |
| 2510 | EXPORT_SYMBOL_GPL(migrate_enable); |
| 2511 | |
| 2512 | static inline bool rq_has_pinned_tasks(struct rq *rq) |
| 2513 | { |
| 2514 | return rq->nr_pinned; |
| 2515 | } |
| 2516 | |
| 2517 | static inline bool task_can_migrate_to_preferred(struct task_struct *p, int cpu) |
| 2518 | { |
| 2519 | /* No need to migrate from a preferred CPU */ |
| 2520 | if (cpu_preferred(cpu)) |
| 2521 | return false; |
| 2522 | |
| 2523 | /* Only FAIR tasks honor preferred CPU state */ |
| 2524 | if (unlikely(p->sched_class != &fair_sched_class)) |
| 2525 | return false; |
| 2526 | |
| 2527 | /* Ignore preferred state if task affinity is changing */ |
| 2528 | if (unlikely(!cpumask_test_cpu(task_cpu(p), p->cpus_ptr))) |
| 2529 | return false; |
| 2530 | |
| 2531 | return cpumask_intersects_and(p->cpus_ptr, cpu_preferred_mask, |
| 2532 | task_cpu_possible_mask(p)); |
| 2533 | } |
| 2534 | |
| 2535 | /* |
| 2536 | * Per-CPU kthreads are allowed to run on !active && online CPUs, see |
| 2537 | * __set_cpus_allowed_ptr() and select_fallback_rq(). |
| 2538 | */ |
| 2539 | static inline bool is_cpu_allowed(struct task_struct *p, int cpu) |
| 2540 | { |
| 2541 | /* When not in the task's cpumask, no point in looking further. */ |
| 2542 | if (!task_allowed_on_cpu(p, cpu)) |
| 2543 | return false; |
| 2544 | |
| 2545 | /* migrate_disabled() must be allowed to finish. */ |
| 2546 | if (is_migration_disabled(p)) |
| 2547 | return cpu_online(cpu); |
| 2548 | |
| 2549 | /* Non kernel threads are not allowed during either online or offline. */ |
| 2550 | if (!(p->flags & PF_KTHREAD)) { |
| 2551 | /* Try to use preferred CPU if task's affinity allows */ |
| 2552 | if (task_can_migrate_to_preferred(p, cpu)) |
| 2553 | return false; |
| 2554 | return cpu_active(cpu); |
| 2555 | } |
| 2556 | |
| 2557 | /* KTHREAD_IS_PER_CPU is always allowed. */ |
| 2558 | if (kthread_is_per_cpu(p)) |
| 2559 | return cpu_online(cpu); |
| 2560 | |
| 2561 | /* Regular kernel threads don't get to stay during offline. */ |
| 2562 | if (cpu_dying(cpu)) |
| 2563 | return false; |
| 2564 | |
| 2565 | /* Try to keep unbound kthreads on a preferred CPU if possible. */ |
| 2566 | if (task_can_migrate_to_preferred(p, cpu)) |
| 2567 | return false; |
| 2568 | |
| 2569 | /* Otherwise, they are allowed to run on online CPU. */ |
| 2570 | return cpu_online(cpu); |
| 2571 | } |
| 2572 | |
| 2573 | /* |
| 2574 | * This is how migration works: |
| 2575 | * |
| 2576 | * 1) we invoke migration_cpu_stop() on the target CPU using |
| 2577 | * stop_one_cpu(). |
| 2578 | * 2) stopper starts to run (implicitly forcing the migrated thread |
| 2579 | * off the CPU) |
| 2580 | * 3) it checks whether the migrated task is still in the wrong runqueue. |
| 2581 | * 4) if it's in the wrong runqueue then the migration thread removes |
| 2582 | * it and puts it into the right queue. |
| 2583 | * 5) stopper completes and stop_one_cpu() returns and the migration |
| 2584 | * is done. |
| 2585 | */ |
| 2586 | |
| 2587 | /* |
| 2588 | * move_queued_task - move a queued task to new rq. |
| 2589 | * |
| 2590 | * Returns (locked) new rq. Old rq's lock is released. |
| 2591 | */ |
| 2592 | static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf, |
| 2593 | struct task_struct *p, int new_cpu) |
| 2594 | __must_hold(__rq_lockp(rq)) |
| 2595 | { |
| 2596 | lockdep_assert_rq_held(rq); |
| 2597 | |
| 2598 | deactivate_task(rq, p, DEQUEUE_NOCLOCK); |
| 2599 | set_task_cpu(p, new_cpu); |
| 2600 | rq_unlock(rq, rf); |
| 2601 | |
| 2602 | rq = cpu_rq(new_cpu); |
| 2603 | |
| 2604 | rq_lock(rq, rf); |
| 2605 | WARN_ON_ONCE(task_cpu(p) != new_cpu); |
| 2606 | activate_task(rq, p, 0); |
| 2607 | wakeup_preempt(rq, p, 0); |
| 2608 | |
| 2609 | return rq; |
| 2610 | } |
| 2611 | |
| 2612 | struct migration_arg { |
| 2613 | struct task_struct *task; |
| 2614 | int dest_cpu; |
| 2615 | struct set_affinity_pending *pending; |
| 2616 | }; |
| 2617 | |
| 2618 | /* |
| 2619 | * @refs: number of wait_for_completion() |
| 2620 | * @stop_pending: is @stop_work in use |
| 2621 | */ |
| 2622 | struct set_affinity_pending { |
| 2623 | refcount_t refs; |
| 2624 | unsigned int stop_pending; |
| 2625 | struct completion done; |
| 2626 | struct cpu_stop_work stop_work; |
| 2627 | struct migration_arg arg; |
| 2628 | }; |
| 2629 | |
| 2630 | /* |
| 2631 | * Move (not current) task off this CPU, onto the destination CPU. We're doing |
| 2632 | * this because either it can't run here any more (set_cpus_allowed() |
| 2633 | * away from this CPU, or CPU going down), or because we're |
| 2634 | * attempting to rebalance this task on exec (sched_exec). |
| 2635 | * |
| 2636 | * So we race with normal scheduler movements, but that's OK, as long |
| 2637 | * as the task is no longer on this CPU. |
| 2638 | */ |
| 2639 | static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf, |
| 2640 | struct task_struct *p, int dest_cpu) |
| 2641 | __must_hold(__rq_lockp(rq)) |
| 2642 | { |
| 2643 | /* Affinity changed (again). */ |
| 2644 | if (!is_cpu_allowed(p, dest_cpu)) |
| 2645 | return rq; |
| 2646 | |
| 2647 | rq = move_queued_task(rq, rf, p, dest_cpu); |
| 2648 | |
| 2649 | return rq; |
| 2650 | } |
| 2651 | |
| 2652 | /* |
| 2653 | * migration_cpu_stop - this will be executed by a high-prio stopper thread |
| 2654 | * and performs thread migration by bumping thread off CPU then |
| 2655 | * 'pushing' onto another runqueue. |
| 2656 | */ |
| 2657 | static int migration_cpu_stop(void *data) |
| 2658 | { |
| 2659 | struct migration_arg *arg = data; |
| 2660 | struct set_affinity_pending *pending = arg->pending; |
| 2661 | struct task_struct *p = arg->task; |
| 2662 | struct rq *rq = this_rq(); |
| 2663 | bool complete = false; |
| 2664 | struct rq_flags rf; |
| 2665 | |
| 2666 | /* |
| 2667 | * The original target CPU might have gone down and we might |
| 2668 | * be on another CPU but it doesn't matter. |
| 2669 | */ |
| 2670 | local_irq_save(rf.flags); |
| 2671 | /* |
| 2672 | * We need to explicitly wake pending tasks before running |
| 2673 | * __migrate_task() such that we will not miss enforcing cpus_ptr |
| 2674 | * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. |
| 2675 | */ |
| 2676 | flush_smp_call_function_queue(); |
| 2677 | |
| 2678 | /* |
| 2679 | * We may change the underlying rq, but the locks held will |
| 2680 | * appropriately be "transferred" when switching. |
| 2681 | */ |
| 2682 | context_unsafe_alias(rq); |
| 2683 | |
| 2684 | raw_spin_lock(&p->pi_lock); |
| 2685 | rq_lock(rq, &rf); |
| 2686 | |
| 2687 | /* |
| 2688 | * If we were passed a pending, then ->stop_pending was set, thus |
| 2689 | * p->migration_pending must have remained stable. |
| 2690 | */ |
| 2691 | WARN_ON_ONCE(pending && pending != p->migration_pending); |
| 2692 | |
| 2693 | /* |
| 2694 | * If task_rq(p) != rq, it cannot be migrated here, because we're |
| 2695 | * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because |
| 2696 | * we're holding p->pi_lock. |
| 2697 | */ |
| 2698 | if (task_rq(p) == rq) { |
| 2699 | if (is_migration_disabled(p)) |
| 2700 | goto out; |
| 2701 | |
| 2702 | if (pending) { |
| 2703 | p->migration_pending = NULL; |
| 2704 | complete = true; |
| 2705 | |
| 2706 | if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) |
| 2707 | goto out; |
| 2708 | } |
| 2709 | |
| 2710 | if (task_on_rq_queued(p)) { |
| 2711 | update_rq_clock(rq); |
| 2712 | rq = __migrate_task(rq, &rf, p, arg->dest_cpu); |
| 2713 | } else { |
| 2714 | p->wake_cpu = arg->dest_cpu; |
| 2715 | } |
| 2716 | |
| 2717 | /* |
| 2718 | * XXX __migrate_task() can fail, at which point we might end |
| 2719 | * up running on a dodgy CPU, AFAICT this can only happen |
| 2720 | * during CPU hotplug, at which point we'll get pushed out |
| 2721 | * anyway, so it's probably not a big deal. |
| 2722 | */ |
| 2723 | |
| 2724 | } else if (pending) { |
| 2725 | /* |
| 2726 | * This happens when we get migrated between migrate_enable()'s |
| 2727 | * preempt_enable() and scheduling the stopper task. At that |
| 2728 | * point we're a regular task again and not current anymore. |
| 2729 | * |
| 2730 | * A !PREEMPT kernel has a giant hole here, which makes it far |
| 2731 | * more likely. |
| 2732 | */ |
| 2733 | |
| 2734 | /* |
| 2735 | * The task moved before the stopper got to run. We're holding |
| 2736 | * ->pi_lock, so the allowed mask is stable - if it got |
| 2737 | * somewhere allowed, we're done. |
| 2738 | */ |
| 2739 | if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) { |
| 2740 | p->migration_pending = NULL; |
| 2741 | complete = true; |
| 2742 | goto out; |
| 2743 | } |
| 2744 | |
| 2745 | /* |
| 2746 | * When migrate_enable() hits a rq mis-match we can't reliably |
| 2747 | * determine is_migration_disabled() and so have to chase after |
| 2748 | * it. |
| 2749 | */ |
| 2750 | WARN_ON_ONCE(!pending->stop_pending); |
| 2751 | preempt_disable(); |
| 2752 | rq_unlock(rq, &rf); |
| 2753 | raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags); |
| 2754 | stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop, |
| 2755 | &pending->arg, &pending->stop_work); |
| 2756 | preempt_enable(); |
| 2757 | return 0; |
| 2758 | } |
| 2759 | out: |
| 2760 | if (pending) |
| 2761 | pending->stop_pending = false; |
| 2762 | rq_unlock(rq, &rf); |
| 2763 | raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags); |
| 2764 | |
| 2765 | if (complete) |
| 2766 | complete_all(&pending->done); |
| 2767 | |
| 2768 | return 0; |
| 2769 | } |
| 2770 | |
| 2771 | int push_cpu_stop(void *arg) |
| 2772 | { |
| 2773 | struct rq *lowest_rq = NULL, *rq = this_rq(); |
| 2774 | struct task_struct *p = arg; |
| 2775 | |
| 2776 | raw_spin_lock_irq(&p->pi_lock); |
| 2777 | raw_spin_rq_lock(rq); |
| 2778 | |
| 2779 | if (task_rq(p) != rq) |
| 2780 | goto out_unlock; |
| 2781 | |
| 2782 | if (is_migration_disabled(p)) { |
| 2783 | p->migration_flags |= MDF_PUSH; |
| 2784 | goto out_unlock; |
| 2785 | } |
| 2786 | |
| 2787 | p->migration_flags &= ~MDF_PUSH; |
| 2788 | |
| 2789 | if (p->sched_class->find_lock_rq) |
| 2790 | lowest_rq = p->sched_class->find_lock_rq(p, rq); |
| 2791 | |
| 2792 | if (!lowest_rq) |
| 2793 | goto out_unlock; |
| 2794 | |
| 2795 | lockdep_assert_rq_held(lowest_rq); |
| 2796 | |
| 2797 | // XXX validate p is still the highest prio task |
| 2798 | if (task_rq(p) == rq) { |
| 2799 | move_queued_task_locked(rq, lowest_rq, p); |
| 2800 | resched_curr(lowest_rq); |
| 2801 | } |
| 2802 | |
| 2803 | double_unlock_balance(rq, lowest_rq); |
| 2804 | |
| 2805 | out_unlock: |
| 2806 | rq->push_busy = false; |
| 2807 | raw_spin_rq_unlock(rq); |
| 2808 | raw_spin_unlock_irq(&p->pi_lock); |
| 2809 | |
| 2810 | put_task_struct(p); |
| 2811 | return 0; |
| 2812 | } |
| 2813 | |
| 2814 | static inline void mm_update_cpus_allowed(struct mm_struct *mm, const cpumask_t *affmask); |
| 2815 | |
| 2816 | /* |
| 2817 | * sched_class::set_cpus_allowed must do the below, but is not required to |
| 2818 | * actually call this function. |
| 2819 | */ |
| 2820 | void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx) |
| 2821 | { |
| 2822 | if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) { |
| 2823 | p->cpus_ptr = ctx->new_mask; |
| 2824 | return; |
| 2825 | } |
| 2826 | |
| 2827 | cpumask_copy(&p->cpus_mask, ctx->new_mask); |
| 2828 | p->nr_cpus_allowed = cpumask_weight(ctx->new_mask); |
| 2829 | mm_update_cpus_allowed(p->mm, ctx->new_mask); |
| 2830 | |
| 2831 | /* |
| 2832 | * Swap in a new user_cpus_ptr if SCA_USER flag set |
| 2833 | */ |
| 2834 | if (ctx->flags & SCA_USER) |
| 2835 | swap(p->user_cpus_ptr, ctx->user_mask); |
| 2836 | } |
| 2837 | |
| 2838 | static void |
| 2839 | do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx) |
| 2840 | { |
| 2841 | scoped_guard (sched_change, p, DEQUEUE_SAVE) |
| 2842 | p->sched_class->set_cpus_allowed(p, ctx); |
| 2843 | } |
| 2844 | |
| 2845 | /* |
| 2846 | * Used for kthread_bind() and select_fallback_rq(), in both cases the user |
| 2847 | * affinity (if any) should be destroyed too. |
| 2848 | */ |
| 2849 | void set_cpus_allowed_force(struct task_struct *p, const struct cpumask *new_mask) |
| 2850 | { |
| 2851 | struct affinity_context ac = { |
| 2852 | .new_mask = new_mask, |
| 2853 | .user_mask = NULL, |
| 2854 | .flags = SCA_USER, /* clear the user requested mask */ |
| 2855 | }; |
| 2856 | union cpumask_rcuhead { |
| 2857 | cpumask_t cpumask; |
| 2858 | struct rcu_head rcu; |
| 2859 | }; |
| 2860 | |
| 2861 | scoped_guard (__task_rq_lock, p) |
| 2862 | do_set_cpus_allowed(p, &ac); |
| 2863 | |
| 2864 | /* |
| 2865 | * Because this is called with p->pi_lock held, it is not possible |
| 2866 | * to use kfree() here (when PREEMPT_RT=y), therefore punt to using |
| 2867 | * kfree_rcu(). |
| 2868 | */ |
| 2869 | kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu); |
| 2870 | } |
| 2871 | |
| 2872 | int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src, |
| 2873 | int node) |
| 2874 | { |
| 2875 | cpumask_t *user_mask; |
| 2876 | unsigned long flags; |
| 2877 | |
| 2878 | /* |
| 2879 | * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's |
| 2880 | * may differ by now due to racing. |
| 2881 | */ |
| 2882 | dst->user_cpus_ptr = NULL; |
| 2883 | |
| 2884 | /* |
| 2885 | * This check is racy and losing the race is a valid situation. |
| 2886 | * It is not worth the extra overhead of taking the pi_lock on |
| 2887 | * every fork/clone. |
| 2888 | */ |
| 2889 | if (data_race(!src->user_cpus_ptr)) |
| 2890 | return 0; |
| 2891 | |
| 2892 | user_mask = alloc_user_cpus_ptr(node); |
| 2893 | if (!user_mask) |
| 2894 | return -ENOMEM; |
| 2895 | |
| 2896 | /* |
| 2897 | * Use pi_lock to protect content of user_cpus_ptr |
| 2898 | * |
| 2899 | * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent |
| 2900 | * set_cpus_allowed_force(). |
| 2901 | */ |
| 2902 | raw_spin_lock_irqsave(&src->pi_lock, flags); |
| 2903 | if (src->user_cpus_ptr) { |
| 2904 | swap(dst->user_cpus_ptr, user_mask); |
| 2905 | cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr); |
| 2906 | } |
| 2907 | raw_spin_unlock_irqrestore(&src->pi_lock, flags); |
| 2908 | |
| 2909 | if (unlikely(user_mask)) |
| 2910 | kfree(user_mask); |
| 2911 | |
| 2912 | return 0; |
| 2913 | } |
| 2914 | |
| 2915 | static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p) |
| 2916 | { |
| 2917 | struct cpumask *user_mask = NULL; |
| 2918 | |
| 2919 | swap(p->user_cpus_ptr, user_mask); |
| 2920 | |
| 2921 | return user_mask; |
| 2922 | } |
| 2923 | |
| 2924 | void release_user_cpus_ptr(struct task_struct *p) |
| 2925 | { |
| 2926 | kfree(clear_user_cpus_ptr(p)); |
| 2927 | } |
| 2928 | |
| 2929 | /* |
| 2930 | * This function is wildly self concurrent; here be dragons. |
| 2931 | * |
| 2932 | * |
| 2933 | * When given a valid mask, __set_cpus_allowed_ptr() must block until the |
| 2934 | * designated task is enqueued on an allowed CPU. If that task is currently |
| 2935 | * running, we have to kick it out using the CPU stopper. |
| 2936 | * |
| 2937 | * Migrate-Disable comes along and tramples all over our nice sandcastle. |
| 2938 | * Consider: |
| 2939 | * |
| 2940 | * Initial conditions: P0->cpus_mask = [0, 1] |
| 2941 | * |
| 2942 | * P0@CPU0 P1 |
| 2943 | * |
| 2944 | * migrate_disable(); |
| 2945 | * <preempted> |
| 2946 | * set_cpus_allowed_ptr(P0, [1]); |
| 2947 | * |
| 2948 | * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes |
| 2949 | * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region). |
| 2950 | * This means we need the following scheme: |
| 2951 | * |
| 2952 | * P0@CPU0 P1 |
| 2953 | * |
| 2954 | * migrate_disable(); |
| 2955 | * <preempted> |
| 2956 | * set_cpus_allowed_ptr(P0, [1]); |
| 2957 | * <blocks> |
| 2958 | * <resumes> |
| 2959 | * migrate_enable(); |
| 2960 | * __set_cpus_allowed_ptr(); |
| 2961 | * <wakes local stopper> |
| 2962 | * `--> <woken on migration completion> |
| 2963 | * |
| 2964 | * Now the fun stuff: there may be several P1-like tasks, i.e. multiple |
| 2965 | * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any |
| 2966 | * task p are serialized by p->pi_lock, which we can leverage: the one that |
| 2967 | * should come into effect at the end of the Migrate-Disable region is the last |
| 2968 | * one. This means we only need to track a single cpumask (i.e. p->cpus_mask), |
| 2969 | * but we still need to properly signal those waiting tasks at the appropriate |
| 2970 | * moment. |
| 2971 | * |
| 2972 | * This is implemented using struct set_affinity_pending. The first |
| 2973 | * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will |
| 2974 | * setup an instance of that struct and install it on the targeted task_struct. |
| 2975 | * Any and all further callers will reuse that instance. Those then wait for |
| 2976 | * a completion signaled at the tail of the CPU stopper callback (1), triggered |
| 2977 | * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()). |
| 2978 | * |
| 2979 | * |
| 2980 | * (1) In the cases covered above. There is one more where the completion is |
| 2981 | * signaled within affine_move_task() itself: when a subsequent affinity request |
| 2982 | * occurs after the stopper bailed out due to the targeted task still being |
| 2983 | * Migrate-Disable. Consider: |
| 2984 | * |
| 2985 | * Initial conditions: P0->cpus_mask = [0, 1] |
| 2986 | * |
| 2987 | * CPU0 P1 P2 |
| 2988 | * <P0> |
| 2989 | * migrate_disable(); |
| 2990 | * <preempted> |
| 2991 | * set_cpus_allowed_ptr(P0, [1]); |
| 2992 | * <blocks> |
| 2993 | * <migration/0> |
| 2994 | * migration_cpu_stop() |
| 2995 | * is_migration_disabled() |
| 2996 | * <bails> |
| 2997 | * set_cpus_allowed_ptr(P0, [0, 1]); |
| 2998 | * <signal completion> |
| 2999 | * <awakes> |
| 3000 | * |
| 3001 | * Note that the above is safe vs a concurrent migrate_enable(), as any |
| 3002 | * pending affinity completion is preceded by an uninstallation of |
| 3003 | * p->migration_pending done with p->pi_lock held. |
| 3004 | */ |
| 3005 | static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf, |
| 3006 | int dest_cpu, unsigned int flags) |
| 3007 | __releases(__rq_lockp(rq), &p->pi_lock) |
| 3008 | { |
| 3009 | struct set_affinity_pending my_pending = { }, *pending = NULL; |
| 3010 | bool stop_pending, complete = false; |
| 3011 | |
| 3012 | /* |
| 3013 | * Can the task run on the task's current CPU? If so, we're done |
| 3014 | * |
| 3015 | * We are also done if the task is the current donor, boosting a lock- |
| 3016 | * holding proxy, (and potentially has been migrated outside its |
| 3017 | * current or previous affinity mask) |
| 3018 | */ |
| 3019 | if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask) || |
| 3020 | (task_current_donor(rq, p) && !task_current(rq, p))) { |
| 3021 | struct task_struct *push_task = NULL; |
| 3022 | |
| 3023 | if ((flags & SCA_MIGRATE_ENABLE) && |
| 3024 | (p->migration_flags & MDF_PUSH) && !rq->push_busy) { |
| 3025 | rq->push_busy = true; |
| 3026 | push_task = get_task_struct(p); |
| 3027 | } |
| 3028 | |
| 3029 | /* |
| 3030 | * If there are pending waiters, but no pending stop_work, |
| 3031 | * then complete now. |
| 3032 | */ |
| 3033 | pending = p->migration_pending; |
| 3034 | if (pending && !pending->stop_pending) { |
| 3035 | p->migration_pending = NULL; |
| 3036 | complete = true; |
| 3037 | } |
| 3038 | |
| 3039 | preempt_disable(); |
| 3040 | task_rq_unlock(rq, p, rf); |
| 3041 | if (push_task) { |
| 3042 | stop_one_cpu_nowait(rq->cpu, push_cpu_stop, |
| 3043 | p, &rq->push_work); |
| 3044 | } |
| 3045 | preempt_enable(); |
| 3046 | |
| 3047 | if (complete) |
| 3048 | complete_all(&pending->done); |
| 3049 | |
| 3050 | return 0; |
| 3051 | } |
| 3052 | |
| 3053 | if (!(flags & SCA_MIGRATE_ENABLE)) { |
| 3054 | /* serialized by p->pi_lock */ |
| 3055 | if (!p->migration_pending) { |
| 3056 | /* Install the request */ |
| 3057 | refcount_set(&my_pending.refs, 1); |
| 3058 | init_completion(&my_pending.done); |
| 3059 | my_pending.arg = (struct migration_arg) { |
| 3060 | .task = p, |
| 3061 | .dest_cpu = dest_cpu, |
| 3062 | .pending = &my_pending, |
| 3063 | }; |
| 3064 | |
| 3065 | p->migration_pending = &my_pending; |
| 3066 | } else { |
| 3067 | pending = p->migration_pending; |
| 3068 | refcount_inc(&pending->refs); |
| 3069 | /* |
| 3070 | * Affinity has changed, but we've already installed a |
| 3071 | * pending. migration_cpu_stop() *must* see this, else |
| 3072 | * we risk a completion of the pending despite having a |
| 3073 | * task on a disallowed CPU. |
| 3074 | * |
| 3075 | * Serialized by p->pi_lock, so this is safe. |
| 3076 | */ |
| 3077 | pending->arg.dest_cpu = dest_cpu; |
| 3078 | } |
| 3079 | } |
| 3080 | pending = p->migration_pending; |
| 3081 | /* |
| 3082 | * - !MIGRATE_ENABLE: |
| 3083 | * we'll have installed a pending if there wasn't one already. |
| 3084 | * |
| 3085 | * - MIGRATE_ENABLE: |
| 3086 | * we're here because the current CPU isn't matching anymore, |
| 3087 | * the only way that can happen is because of a concurrent |
| 3088 | * set_cpus_allowed_ptr() call, which should then still be |
| 3089 | * pending completion. |
| 3090 | * |
| 3091 | * Either way, we really should have a @pending here. |
| 3092 | */ |
| 3093 | if (WARN_ON_ONCE(!pending)) { |
| 3094 | task_rq_unlock(rq, p, rf); |
| 3095 | return -EINVAL; |
| 3096 | } |
| 3097 | |
| 3098 | if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) { |
| 3099 | /* |
| 3100 | * MIGRATE_ENABLE gets here because 'p == current', but for |
| 3101 | * anything else we cannot do is_migration_disabled(), punt |
| 3102 | * and have the stopper function handle it all race-free. |
| 3103 | */ |
| 3104 | stop_pending = pending->stop_pending; |
| 3105 | if (!stop_pending) |
| 3106 | pending->stop_pending = true; |
| 3107 | |
| 3108 | if (flags & SCA_MIGRATE_ENABLE) |
| 3109 | p->migration_flags &= ~MDF_PUSH; |
| 3110 | |
| 3111 | preempt_disable(); |
| 3112 | task_rq_unlock(rq, p, rf); |
| 3113 | if (!stop_pending) { |
| 3114 | stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop, |
| 3115 | &pending->arg, &pending->stop_work); |
| 3116 | } |
| 3117 | preempt_enable(); |
| 3118 | |
| 3119 | if (flags & SCA_MIGRATE_ENABLE) |
| 3120 | return 0; |
| 3121 | } else { |
| 3122 | |
| 3123 | if (!is_migration_disabled(p)) { |
| 3124 | if (task_on_rq_queued(p)) |
| 3125 | rq = move_queued_task(rq, rf, p, dest_cpu); |
| 3126 | |
| 3127 | if (!pending->stop_pending) { |
| 3128 | p->migration_pending = NULL; |
| 3129 | complete = true; |
| 3130 | } |
| 3131 | } |
| 3132 | task_rq_unlock(rq, p, rf); |
| 3133 | |
| 3134 | if (complete) |
| 3135 | complete_all(&pending->done); |
| 3136 | } |
| 3137 | |
| 3138 | wait_for_completion(&pending->done); |
| 3139 | |
| 3140 | if (refcount_dec_and_test(&pending->refs)) |
| 3141 | wake_up_var(&pending->refs); /* No UaF, just an address */ |
| 3142 | |
| 3143 | /* |
| 3144 | * Block the original owner of &pending until all subsequent callers |
| 3145 | * have seen the completion and decremented the refcount |
| 3146 | */ |
| 3147 | wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs)); |
| 3148 | |
| 3149 | /* ARGH */ |
| 3150 | WARN_ON_ONCE(my_pending.stop_pending); |
| 3151 | |
| 3152 | return 0; |
| 3153 | } |
| 3154 | |
| 3155 | /* |
| 3156 | * Called with both p->pi_lock and rq->lock held; drops both before returning. |
| 3157 | */ |
| 3158 | static int __set_cpus_allowed_ptr_locked(struct task_struct *p, |
| 3159 | struct affinity_context *ctx, |
| 3160 | struct rq *rq, |
| 3161 | struct rq_flags *rf) |
| 3162 | __releases(__rq_lockp(rq), &p->pi_lock) |
| 3163 | { |
| 3164 | const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p); |
| 3165 | const struct cpumask *cpu_valid_mask = cpu_active_mask; |
| 3166 | bool kthread = p->flags & PF_KTHREAD; |
| 3167 | unsigned int dest_cpu; |
| 3168 | int ret = 0; |
| 3169 | |
| 3170 | if (kthread || is_migration_disabled(p)) { |
| 3171 | /* |
| 3172 | * Kernel threads are allowed on online && !active CPUs, |
| 3173 | * however, during cpu-hot-unplug, even these might get pushed |
| 3174 | * away if not KTHREAD_IS_PER_CPU. |
| 3175 | * |
| 3176 | * Specifically, migration_disabled() tasks must not fail the |
| 3177 | * cpumask_any_and_distribute() pick below, esp. so on |
| 3178 | * SCA_MIGRATE_ENABLE, otherwise we'll not call |
| 3179 | * set_cpus_allowed_common() and actually reset p->cpus_ptr. |
| 3180 | */ |
| 3181 | cpu_valid_mask = cpu_online_mask; |
| 3182 | } |
| 3183 | |
| 3184 | if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) { |
| 3185 | ret = -EINVAL; |
| 3186 | goto out; |
| 3187 | } |
| 3188 | |
| 3189 | /* |
| 3190 | * Must re-check here, to close a race against __kthread_bind(), |
| 3191 | * sched_setaffinity() is not guaranteed to observe the flag. |
| 3192 | */ |
| 3193 | if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) { |
| 3194 | ret = -EINVAL; |
| 3195 | goto out; |
| 3196 | } |
| 3197 | |
| 3198 | if (!(ctx->flags & SCA_MIGRATE_ENABLE)) { |
| 3199 | if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) { |
| 3200 | if (ctx->flags & SCA_USER) |
| 3201 | swap(p->user_cpus_ptr, ctx->user_mask); |
| 3202 | goto out; |
| 3203 | } |
| 3204 | |
| 3205 | if (WARN_ON_ONCE(p == current && |
| 3206 | is_migration_disabled(p) && |
| 3207 | !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) { |
| 3208 | ret = -EBUSY; |
| 3209 | goto out; |
| 3210 | } |
| 3211 | } |
| 3212 | |
| 3213 | /* |
| 3214 | * Picking a ~random cpu helps in cases where we are changing affinity |
| 3215 | * for groups of tasks (ie. cpuset), so that load balancing is not |
| 3216 | * immediately required to distribute the tasks within their new mask. |
| 3217 | */ |
| 3218 | dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask); |
| 3219 | if (dest_cpu >= nr_cpu_ids) { |
| 3220 | ret = -EINVAL; |
| 3221 | goto out; |
| 3222 | } |
| 3223 | |
| 3224 | do_set_cpus_allowed(p, ctx); |
| 3225 | |
| 3226 | return affine_move_task(rq, p, rf, dest_cpu, ctx->flags); |
| 3227 | |
| 3228 | out: |
| 3229 | task_rq_unlock(rq, p, rf); |
| 3230 | |
| 3231 | return ret; |
| 3232 | } |
| 3233 | |
| 3234 | /* |
| 3235 | * Change a given task's CPU affinity. Migrate the thread to a |
| 3236 | * proper CPU and schedule it away if the CPU it's executing on |
| 3237 | * is removed from the allowed bitmask. |
| 3238 | * |
| 3239 | * NOTE: the caller must have a valid reference to the task, the |
| 3240 | * task must not exit() & deallocate itself prematurely. The |
| 3241 | * call is not atomic; no spinlocks may be held. |
| 3242 | */ |
| 3243 | int __set_cpus_allowed_ptr(struct task_struct *p, struct affinity_context *ctx) |
| 3244 | { |
| 3245 | struct rq_flags rf; |
| 3246 | struct rq *rq; |
| 3247 | |
| 3248 | rq = task_rq_lock(p, &rf); |
| 3249 | /* |
| 3250 | * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_* |
| 3251 | * flags are set. |
| 3252 | */ |
| 3253 | if (p->user_cpus_ptr && |
| 3254 | !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) && |
| 3255 | cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr)) |
| 3256 | ctx->new_mask = rq->scratch_mask; |
| 3257 | |
| 3258 | return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf); |
| 3259 | } |
| 3260 | |
| 3261 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) |
| 3262 | { |
| 3263 | struct affinity_context ac = { |
| 3264 | .new_mask = new_mask, |
| 3265 | .flags = 0, |
| 3266 | }; |
| 3267 | |
| 3268 | return __set_cpus_allowed_ptr(p, &ac); |
| 3269 | } |
| 3270 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
| 3271 | |
| 3272 | /* |
| 3273 | * Change a given task's CPU affinity to the intersection of its current |
| 3274 | * affinity mask and @subset_mask, writing the resulting mask to @new_mask. |
| 3275 | * If user_cpus_ptr is defined, use it as the basis for restricting CPU |
| 3276 | * affinity or use cpu_online_mask instead. |
| 3277 | * |
| 3278 | * If the resulting mask is empty, leave the affinity unchanged and return |
| 3279 | * -EINVAL. |
| 3280 | */ |
| 3281 | static int restrict_cpus_allowed_ptr(struct task_struct *p, |
| 3282 | struct cpumask *new_mask, |
| 3283 | const struct cpumask *subset_mask) |
| 3284 | { |
| 3285 | struct affinity_context ac = { |
| 3286 | .new_mask = new_mask, |
| 3287 | .flags = 0, |
| 3288 | }; |
| 3289 | struct rq_flags rf; |
| 3290 | struct rq *rq; |
| 3291 | int err; |
| 3292 | |
| 3293 | rq = task_rq_lock(p, &rf); |
| 3294 | |
| 3295 | /* |
| 3296 | * Forcefully restricting the affinity of a deadline task is |
| 3297 | * likely to cause problems, so fail and noisily override the |
| 3298 | * mask entirely. |
| 3299 | */ |
| 3300 | if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { |
| 3301 | err = -EPERM; |
| 3302 | goto err_unlock; |
| 3303 | } |
| 3304 | |
| 3305 | if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) { |
| 3306 | err = -EINVAL; |
| 3307 | goto err_unlock; |
| 3308 | } |
| 3309 | |
| 3310 | return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf); |
| 3311 | |
| 3312 | err_unlock: |
| 3313 | task_rq_unlock(rq, p, &rf); |
| 3314 | return err; |
| 3315 | } |
| 3316 | |
| 3317 | /* |
| 3318 | * Restrict the CPU affinity of task @p so that it is a subset of |
| 3319 | * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the |
| 3320 | * old affinity mask. If the resulting mask is empty, we warn and walk |
| 3321 | * up the cpuset hierarchy until we find a suitable mask. |
| 3322 | */ |
| 3323 | void force_compatible_cpus_allowed_ptr(struct task_struct *p) |
| 3324 | { |
| 3325 | cpumask_var_t new_mask; |
| 3326 | const struct cpumask *override_mask = task_cpu_possible_mask(p); |
| 3327 | |
| 3328 | alloc_cpumask_var(&new_mask, GFP_KERNEL); |
| 3329 | |
| 3330 | /* |
| 3331 | * __migrate_task() can fail silently in the face of concurrent |
| 3332 | * offlining of the chosen destination CPU, so take the hotplug |
| 3333 | * lock to ensure that the migration succeeds. |
| 3334 | */ |
| 3335 | cpus_read_lock(); |
| 3336 | if (!cpumask_available(new_mask)) |
| 3337 | goto out_set_mask; |
| 3338 | |
| 3339 | if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask)) |
| 3340 | goto out_free_mask; |
| 3341 | |
| 3342 | /* |
| 3343 | * We failed to find a valid subset of the affinity mask for the |
| 3344 | * task, so override it based on its cpuset hierarchy. |
| 3345 | */ |
| 3346 | cpuset_cpus_allowed(p, new_mask); |
| 3347 | override_mask = new_mask; |
| 3348 | |
| 3349 | out_set_mask: |
| 3350 | if (printk_ratelimit()) { |
| 3351 | printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n", |
| 3352 | task_pid_nr(p), p->comm, |
| 3353 | cpumask_pr_args(override_mask)); |
| 3354 | } |
| 3355 | |
| 3356 | WARN_ON(set_cpus_allowed_ptr(p, override_mask)); |
| 3357 | out_free_mask: |
| 3358 | cpus_read_unlock(); |
| 3359 | free_cpumask_var(new_mask); |
| 3360 | } |
| 3361 | |
| 3362 | /* |
| 3363 | * Restore the affinity of a task @p which was previously restricted by a |
| 3364 | * call to force_compatible_cpus_allowed_ptr(). |
| 3365 | * |
| 3366 | * It is the caller's responsibility to serialise this with any calls to |
| 3367 | * force_compatible_cpus_allowed_ptr(@p). |
| 3368 | */ |
| 3369 | void relax_compatible_cpus_allowed_ptr(struct task_struct *p) |
| 3370 | { |
| 3371 | struct affinity_context ac = { |
| 3372 | .new_mask = task_user_cpus(p), |
| 3373 | .flags = 0, |
| 3374 | }; |
| 3375 | int ret; |
| 3376 | |
| 3377 | /* |
| 3378 | * Try to restore the old affinity mask with __sched_setaffinity(). |
| 3379 | * Cpuset masking will be done there too. |
| 3380 | */ |
| 3381 | ret = __sched_setaffinity(p, &ac); |
| 3382 | WARN_ON_ONCE(ret); |
| 3383 | } |
| 3384 | |
| 3385 | #ifdef CONFIG_SMP |
| 3386 | |
| 3387 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
| 3388 | { |
| 3389 | unsigned int state = READ_ONCE(p->__state); |
| 3390 | bool proxy_migrated = sched_proxy_exec() && p->is_blocked && |
| 3391 | task_cpu(p) != p->wake_cpu; |
| 3392 | |
| 3393 | /* |
| 3394 | * We should never call set_task_cpu() on a blocked task, |
| 3395 | * ttwu() will sort out the placement. |
| 3396 | */ |
| 3397 | WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq); |
| 3398 | |
| 3399 | /* |
| 3400 | * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING, |
| 3401 | * because schedstat_wait_{start,end} rebase migrating task's wait_start |
| 3402 | * time relying on p->on_rq. |
| 3403 | */ |
| 3404 | WARN_ON_ONCE(state == TASK_RUNNING && |
| 3405 | p->sched_class == &fair_sched_class && |
| 3406 | (p->on_rq && !task_on_rq_migrating(p))); |
| 3407 | |
| 3408 | #ifdef CONFIG_LOCKDEP |
| 3409 | /* |
| 3410 | * The caller should hold either p->pi_lock or rq->lock, when changing |
| 3411 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. |
| 3412 | * |
| 3413 | * sched_move_task() holds both and thus holding either pins the cgroup, |
| 3414 | * see task_group(). |
| 3415 | * |
| 3416 | * Furthermore, all task_rq users should acquire both locks, see |
| 3417 | * task_rq_lock(). |
| 3418 | */ |
| 3419 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || |
| 3420 | lockdep_is_held(__rq_lockp(task_rq(p))))); |
| 3421 | #endif |
| 3422 | /* |
| 3423 | * Clearly, migrating tasks to offline CPUs is a fairly daft thing. |
| 3424 | */ |
| 3425 | WARN_ON_ONCE(!cpu_online(new_cpu)); |
| 3426 | |
| 3427 | /* |
| 3428 | * Proxy execution can move a blocked task's scheduling context to any |
| 3429 | * CPU without moving its migration-disabled execution context. The |
| 3430 | * wakeup path will return the task to a CPU where it can execute. |
| 3431 | */ |
| 3432 | WARN_ON_ONCE(is_migration_disabled(p) && !proxy_migrated); |
| 3433 | |
| 3434 | trace_sched_migrate_task(p, new_cpu); |
| 3435 | |
| 3436 | if (task_cpu(p) != new_cpu) { |
| 3437 | if (p->sched_class->migrate_task_rq) |
| 3438 | p->sched_class->migrate_task_rq(p, new_cpu); |
| 3439 | p->se.nr_migrations++; |
| 3440 | perf_event_task_migrate(p); |
| 3441 | } |
| 3442 | |
| 3443 | __set_task_cpu(p, new_cpu); |
| 3444 | } |
| 3445 | #endif /* CONFIG_SMP */ |
| 3446 | |
| 3447 | #ifdef CONFIG_NUMA_BALANCING |
| 3448 | static void __migrate_swap_task(struct task_struct *p, int cpu) |
| 3449 | { |
| 3450 | if (task_on_rq_queued(p)) { |
| 3451 | struct rq *src_rq, *dst_rq; |
| 3452 | struct rq_flags srf, drf; |
| 3453 | |
| 3454 | src_rq = task_rq(p); |
| 3455 | dst_rq = cpu_rq(cpu); |
| 3456 | |
| 3457 | rq_pin_lock(src_rq, &srf); |
| 3458 | rq_pin_lock(dst_rq, &drf); |
| 3459 | |
| 3460 | move_queued_task_locked(src_rq, dst_rq, p); |
| 3461 | wakeup_preempt(dst_rq, p, 0); |
| 3462 | |
| 3463 | rq_unpin_lock(dst_rq, &drf); |
| 3464 | rq_unpin_lock(src_rq, &srf); |
| 3465 | |
| 3466 | } else { |
| 3467 | /* |
| 3468 | * Task isn't running anymore; make it appear like we migrated |
| 3469 | * it before it went to sleep. This means on wakeup we make the |
| 3470 | * previous CPU our target instead of where it really is. |
| 3471 | */ |
| 3472 | p->wake_cpu = cpu; |
| 3473 | } |
| 3474 | } |
| 3475 | |
| 3476 | struct migration_swap_arg { |
| 3477 | struct task_struct *src_task, *dst_task; |
| 3478 | int src_cpu, dst_cpu; |
| 3479 | }; |
| 3480 | |
| 3481 | static int migrate_swap_stop(void *data) |
| 3482 | { |
| 3483 | struct migration_swap_arg *arg = data; |
| 3484 | struct rq *src_rq, *dst_rq; |
| 3485 | |
| 3486 | if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu)) |
| 3487 | return -EAGAIN; |
| 3488 | |
| 3489 | src_rq = cpu_rq(arg->src_cpu); |
| 3490 | dst_rq = cpu_rq(arg->dst_cpu); |
| 3491 | |
| 3492 | guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock); |
| 3493 | guard(double_rq_lock)(src_rq, dst_rq); |
| 3494 | |
| 3495 | if (task_cpu(arg->dst_task) != arg->dst_cpu) |
| 3496 | return -EAGAIN; |
| 3497 | |
| 3498 | if (task_cpu(arg->src_task) != arg->src_cpu) |
| 3499 | return -EAGAIN; |
| 3500 | |
| 3501 | if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr)) |
| 3502 | return -EAGAIN; |
| 3503 | |
| 3504 | if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr)) |
| 3505 | return -EAGAIN; |
| 3506 | |
| 3507 | __migrate_swap_task(arg->src_task, arg->dst_cpu); |
| 3508 | __migrate_swap_task(arg->dst_task, arg->src_cpu); |
| 3509 | |
| 3510 | return 0; |
| 3511 | } |
| 3512 | |
| 3513 | /* |
| 3514 | * Cross migrate two tasks |
| 3515 | */ |
| 3516 | int migrate_swap(struct task_struct *cur, struct task_struct *p, |
| 3517 | int target_cpu, int curr_cpu) |
| 3518 | { |
| 3519 | struct migration_swap_arg arg; |
| 3520 | int ret = -EINVAL; |
| 3521 | |
| 3522 | arg = (struct migration_swap_arg){ |
| 3523 | .src_task = cur, |
| 3524 | .src_cpu = curr_cpu, |
| 3525 | .dst_task = p, |
| 3526 | .dst_cpu = target_cpu, |
| 3527 | }; |
| 3528 | |
| 3529 | if (arg.src_cpu == arg.dst_cpu) |
| 3530 | goto out; |
| 3531 | |
| 3532 | /* |
| 3533 | * These three tests are all lockless; this is OK since all of them |
| 3534 | * will be re-checked with proper locks held further down the line. |
| 3535 | */ |
| 3536 | if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) |
| 3537 | goto out; |
| 3538 | |
| 3539 | if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr)) |
| 3540 | goto out; |
| 3541 | |
| 3542 | if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr)) |
| 3543 | goto out; |
| 3544 | |
| 3545 | trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); |
| 3546 | ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); |
| 3547 | |
| 3548 | out: |
| 3549 | return ret; |
| 3550 | } |
| 3551 | #endif /* CONFIG_NUMA_BALANCING */ |
| 3552 | |
| 3553 | /*** |
| 3554 | * kick_process - kick a running thread to enter/exit the kernel |
| 3555 | * @p: the to-be-kicked thread |
| 3556 | * |
| 3557 | * Cause a process which is running on another CPU to enter |
| 3558 | * kernel-mode, without any delay. (to get signals handled.) |
| 3559 | * |
| 3560 | * NOTE: this function doesn't have to take the runqueue lock, |
| 3561 | * because all it wants to ensure is that the remote task enters |
| 3562 | * the kernel. If the IPI races and the task has been migrated |
| 3563 | * to another CPU then no harm is done and the purpose has been |
| 3564 | * achieved as well. |
| 3565 | */ |
| 3566 | void kick_process(struct task_struct *p) |
| 3567 | { |
| 3568 | guard(preempt)(); |
| 3569 | int cpu = task_cpu(p); |
| 3570 | |
| 3571 | if ((cpu != smp_processor_id()) && task_curr(p)) |
| 3572 | smp_send_reschedule(cpu); |
| 3573 | } |
| 3574 | EXPORT_SYMBOL_GPL(kick_process); |
| 3575 | |
| 3576 | /* |
| 3577 | * ->cpus_ptr is protected by both rq->lock and p->pi_lock |
| 3578 | * |
| 3579 | * A few notes on cpu_active vs cpu_online: |
| 3580 | * |
| 3581 | * - cpu_active must be a subset of cpu_online |
| 3582 | * |
| 3583 | * - on CPU-up we allow per-CPU kthreads on the online && !active CPU, |
| 3584 | * see __set_cpus_allowed_ptr(). At this point the newly online |
| 3585 | * CPU isn't yet part of the sched domains, and balancing will not |
| 3586 | * see it. |
| 3587 | * |
| 3588 | * - on CPU-down we clear cpu_active() to mask the sched domains and |
| 3589 | * avoid the load balancer to place new tasks on the to be removed |
| 3590 | * CPU. Existing tasks will remain running there and will be taken |
| 3591 | * off. |
| 3592 | * |
| 3593 | * This means that fallback selection must not select !active CPUs. |
| 3594 | * And can assume that any active CPU must be online. Conversely |
| 3595 | * select_task_rq() below may allow selection of !active CPUs in order |
| 3596 | * to satisfy the above rules. |
| 3597 | */ |
| 3598 | static int select_fallback_rq(int cpu, struct task_struct *p) |
| 3599 | { |
| 3600 | int nid = cpu_to_node(cpu); |
| 3601 | const struct cpumask *nodemask = NULL; |
| 3602 | enum { cpuset, possible, fail } state = cpuset; |
| 3603 | int dest_cpu; |
| 3604 | |
| 3605 | /* |
| 3606 | * If the node that the CPU is on has been offlined, cpu_to_node() |
| 3607 | * will return -1. There is no CPU on the node, and we should |
| 3608 | * select the CPU on the other node. |
| 3609 | */ |
| 3610 | if (nid != -1) { |
| 3611 | nodemask = cpumask_of_node(nid); |
| 3612 | |
| 3613 | /* Look for allowed, online CPU in same node. */ |
| 3614 | for_each_cpu(dest_cpu, nodemask) { |
| 3615 | if (is_cpu_allowed(p, dest_cpu)) |
| 3616 | return dest_cpu; |
| 3617 | } |
| 3618 | } |
| 3619 | |
| 3620 | for (;;) { |
| 3621 | /* Any allowed, online CPU? */ |
| 3622 | for_each_cpu(dest_cpu, p->cpus_ptr) { |
| 3623 | if (!is_cpu_allowed(p, dest_cpu)) |
| 3624 | continue; |
| 3625 | |
| 3626 | goto out; |
| 3627 | } |
| 3628 | |
| 3629 | /* No more Mr. Nice Guy. */ |
| 3630 | switch (state) { |
| 3631 | case cpuset: |
| 3632 | if (cpuset_cpus_allowed_fallback(p)) { |
| 3633 | state = possible; |
| 3634 | break; |
| 3635 | } |
| 3636 | fallthrough; |
| 3637 | case possible: |
| 3638 | set_cpus_allowed_force(p, task_cpu_fallback_mask(p)); |
| 3639 | state = fail; |
| 3640 | break; |
| 3641 | case fail: |
| 3642 | BUG(); |
| 3643 | break; |
| 3644 | } |
| 3645 | } |
| 3646 | |
| 3647 | out: |
| 3648 | if (state != cpuset) { |
| 3649 | /* |
| 3650 | * Don't tell them about moving exiting tasks or |
| 3651 | * kernel threads (both mm NULL), since they never |
| 3652 | * leave kernel. |
| 3653 | */ |
| 3654 | if (p->mm && printk_ratelimit()) { |
| 3655 | printk_deferred("process %d (%s) no longer affine to cpu%d\n", |
| 3656 | task_pid_nr(p), p->comm, cpu); |
| 3657 | } |
| 3658 | } |
| 3659 | |
| 3660 | return dest_cpu; |
| 3661 | } |
| 3662 | |
| 3663 | /* |
| 3664 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable. |
| 3665 | */ |
| 3666 | static inline |
| 3667 | int select_task_rq(struct task_struct *p, int cpu, int *wake_flags) |
| 3668 | { |
| 3669 | lockdep_assert_held(&p->pi_lock); |
| 3670 | |
| 3671 | if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p)) { |
| 3672 | cpu = p->sched_class->select_task_rq(p, cpu, *wake_flags); |
| 3673 | *wake_flags |= WF_RQ_SELECTED; |
| 3674 | } else { |
| 3675 | cpu = cpumask_any(p->cpus_ptr); |
| 3676 | } |
| 3677 | |
| 3678 | /* |
| 3679 | * In order not to call set_task_cpu() on a blocking task we need |
| 3680 | * to rely on ttwu() to place the task on a valid ->cpus_ptr |
| 3681 | * CPU. |
| 3682 | * |
| 3683 | * Since this is common to all placement strategies, this lives here. |
| 3684 | * |
| 3685 | * [ this allows ->select_task() to simply return task_cpu(p) and |
| 3686 | * not worry about this generic constraint ] |
| 3687 | */ |
| 3688 | if (unlikely(!is_cpu_allowed(p, cpu))) |
| 3689 | cpu = select_fallback_rq(task_cpu(p), p); |
| 3690 | |
| 3691 | return cpu; |
| 3692 | } |
| 3693 | |
| 3694 | void sched_set_stop_task(int cpu, struct task_struct *stop) |
| 3695 | { |
| 3696 | static struct lock_class_key stop_pi_lock; |
| 3697 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; |
| 3698 | struct task_struct *old_stop = cpu_rq(cpu)->stop; |
| 3699 | |
| 3700 | if (stop) { |
| 3701 | /* |
| 3702 | * Make it appear like a SCHED_FIFO task, its something |
| 3703 | * userspace knows about and won't get confused about. |
| 3704 | * |
| 3705 | * Also, it will make PI more or less work without too |
| 3706 | * much confusion -- but then, stop work should not |
| 3707 | * rely on PI working anyway. |
| 3708 | */ |
| 3709 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); |
| 3710 | |
| 3711 | stop->sched_class = &stop_sched_class; |
| 3712 | |
| 3713 | /* |
| 3714 | * The PI code calls rt_mutex_setprio() with ->pi_lock held to |
| 3715 | * adjust the effective priority of a task. As a result, |
| 3716 | * rt_mutex_setprio() can trigger (RT) balancing operations, |
| 3717 | * which can then trigger wakeups of the stop thread to push |
| 3718 | * around the current task. |
| 3719 | * |
| 3720 | * The stop task itself will never be part of the PI-chain, it |
| 3721 | * never blocks, therefore that ->pi_lock recursion is safe. |
| 3722 | * Tell lockdep about this by placing the stop->pi_lock in its |
| 3723 | * own class. |
| 3724 | */ |
| 3725 | lockdep_set_class(&stop->pi_lock, &stop_pi_lock); |
| 3726 | } |
| 3727 | |
| 3728 | cpu_rq(cpu)->stop = stop; |
| 3729 | |
| 3730 | if (old_stop) { |
| 3731 | /* |
| 3732 | * Reset it back to a normal scheduling class so that |
| 3733 | * it can die in pieces. |
| 3734 | */ |
| 3735 | old_stop->sched_class = &rt_sched_class; |
| 3736 | } |
| 3737 | } |
| 3738 | |
| 3739 | static void |
| 3740 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) |
| 3741 | { |
| 3742 | struct rq *rq; |
| 3743 | |
| 3744 | if (!schedstat_enabled()) |
| 3745 | return; |
| 3746 | |
| 3747 | rq = this_rq(); |
| 3748 | |
| 3749 | if (cpu == rq->cpu) { |
| 3750 | __schedstat_inc(rq->ttwu_local); |
| 3751 | __schedstat_inc(p->stats.nr_wakeups_local); |
| 3752 | } else { |
| 3753 | struct sched_domain *sd; |
| 3754 | |
| 3755 | __schedstat_inc(p->stats.nr_wakeups_remote); |
| 3756 | |
| 3757 | guard(rcu)(); |
| 3758 | for_each_domain(rq->cpu, sd) { |
| 3759 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
| 3760 | __schedstat_inc(sd->ttwu_wake_remote); |
| 3761 | break; |
| 3762 | } |
| 3763 | } |
| 3764 | } |
| 3765 | |
| 3766 | if (wake_flags & WF_MIGRATED) |
| 3767 | __schedstat_inc(p->stats.nr_wakeups_migrate); |
| 3768 | |
| 3769 | __schedstat_inc(rq->ttwu_count); |
| 3770 | __schedstat_inc(p->stats.nr_wakeups); |
| 3771 | |
| 3772 | if (wake_flags & WF_SYNC) |
| 3773 | __schedstat_inc(p->stats.nr_wakeups_sync); |
| 3774 | } |
| 3775 | |
| 3776 | /* |
| 3777 | * Mark the task runnable. |
| 3778 | */ |
| 3779 | static inline void ttwu_do_wakeup(struct task_struct *p) |
| 3780 | { |
| 3781 | p->is_blocked = 0; |
| 3782 | WRITE_ONCE(p->__state, TASK_RUNNING); |
| 3783 | trace_sched_wakeup(p); |
| 3784 | } |
| 3785 | |
| 3786 | void update_rq_avg_idle(struct rq *rq) |
| 3787 | { |
| 3788 | u64 idle_stamp = rq->idle_stamp; |
| 3789 | u64 delta, max; |
| 3790 | |
| 3791 | if (!idle_stamp) |
| 3792 | return; |
| 3793 | |
| 3794 | delta = rq_clock(rq) - idle_stamp; |
| 3795 | |
| 3796 | update_avg(&rq->avg_idle, delta); |
| 3797 | |
| 3798 | max = 2 * rq->max_idle_balance_cost; |
| 3799 | if (rq->avg_idle > max) |
| 3800 | rq->avg_idle = max; |
| 3801 | rq->idle_stamp = 0; |
| 3802 | } |
| 3803 | |
| 3804 | #ifdef CONFIG_SCHED_PROXY_EXEC |
| 3805 | static void zap_balance_callbacks(struct rq *rq); |
| 3806 | |
| 3807 | static inline void proxy_reset_donor(struct rq *rq) |
| 3808 | { |
| 3809 | WARN_ON_ONCE(rq->donor == rq->curr); |
| 3810 | |
| 3811 | put_prev_set_next_task(rq, rq->donor, rq->curr); |
| 3812 | rq->next_class = rq->curr->sched_class; |
| 3813 | rq_set_donor(rq, rq->curr); |
| 3814 | zap_balance_callbacks(rq); |
| 3815 | resched_curr(rq); |
| 3816 | } |
| 3817 | |
| 3818 | /* |
| 3819 | * Checks to see if task p has been proxy-migrated to another rq |
| 3820 | * and needs to be returned. If so, we deactivate the task here |
| 3821 | * so that it can be properly woken up on the p->wake_cpu |
| 3822 | * (or whichever cpu select_task_rq() picks at the bottom of |
| 3823 | * try_to_wake_up() |
| 3824 | */ |
| 3825 | static inline bool proxy_needs_return(struct rq *rq, struct task_struct *p) |
| 3826 | { |
| 3827 | bool dequeued; |
| 3828 | |
| 3829 | /* |
| 3830 | * Typically per __set_task_cpu(), task_cpu(p) == p->wake_cpu. |
| 3831 | * |
| 3832 | * However, proxy_set_task_cpu() is such that it preserves the |
| 3833 | * original cpu in p->wake_cpu while migrating p for proxy reasons |
| 3834 | * (possibly outside of the allowed p->cpus_ptr). |
| 3835 | * |
| 3836 | * Furthermore, migration_cpu_stop() / __migrate_swap_task(), will |
| 3837 | * only set p->wake_cpu when !p->on_rq, and since here p->on_rq, this |
| 3838 | * will not apply. But if it did, this check is the safe way around |
| 3839 | * and would migrate. |
| 3840 | */ |
| 3841 | if (task_cpu(p) == p->wake_cpu) |
| 3842 | return false; |
| 3843 | |
| 3844 | scoped_guard(raw_spinlock, &p->blocked_lock) { |
| 3845 | /* Task is waking up; clear any blocked_on relationship */ |
| 3846 | __clear_task_blocked_on(p, NULL); |
| 3847 | |
| 3848 | /* If already current, don't need to return migrate */ |
| 3849 | if (task_current(rq, p)) |
| 3850 | return false; |
| 3851 | } |
| 3852 | |
| 3853 | dequeued = dequeue_block_task(rq, p, TASK_WAKING); |
| 3854 | |
| 3855 | /* |
| 3856 | * Dequeue @p from its scheduling class before resetting rq->donor. |
| 3857 | * In particular, sched_ext needs to end the donor's running session |
| 3858 | * and clear SCX_TASK_QUEUED before put_prev_task_scx() is called by |
| 3859 | * proxy_reset_donor(); otherwise it would reenqueue the blocked donor. |
| 3860 | * |
| 3861 | * Keep on_rq set until all donor references have been replaced. |
| 3862 | */ |
| 3863 | if (task_current_donor(rq, p)) |
| 3864 | proxy_reset_donor(rq); |
| 3865 | |
| 3866 | if (dequeued) |
| 3867 | __block_task(rq, p); |
| 3868 | return true; |
| 3869 | } |
| 3870 | #else /* !CONFIG_SCHED_PROXY_EXEC */ |
| 3871 | static inline bool proxy_needs_return(struct rq *rq, struct task_struct *p) |
| 3872 | { |
| 3873 | return false; |
| 3874 | } |
| 3875 | #endif /* CONFIG_SCHED_PROXY_EXEC */ |
| 3876 | |
| 3877 | static void |
| 3878 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags, |
| 3879 | struct rq_flags *rf) |
| 3880 | { |
| 3881 | int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK; |
| 3882 | |
| 3883 | lockdep_assert_rq_held(rq); |
| 3884 | |
| 3885 | if (p->sched_contributes_to_load) |
| 3886 | rq->nr_uninterruptible--; |
| 3887 | |
| 3888 | if (wake_flags & WF_RQ_SELECTED) |
| 3889 | en_flags |= ENQUEUE_RQ_SELECTED; |
| 3890 | if (wake_flags & WF_MIGRATED) |
| 3891 | en_flags |= ENQUEUE_MIGRATED; |
| 3892 | else if (p->in_iowait) { |
| 3893 | delayacct_blkio_end(p); |
| 3894 | atomic_dec(&task_rq(p)->nr_iowait); |
| 3895 | } |
| 3896 | |
| 3897 | activate_task(rq, p, en_flags); |
| 3898 | wakeup_preempt(rq, p, wake_flags); |
| 3899 | |
| 3900 | ttwu_do_wakeup(p); |
| 3901 | |
| 3902 | if (p->sched_class->task_woken) { |
| 3903 | /* |
| 3904 | * Our task @p is fully woken up and running; so it's safe to |
| 3905 | * drop the rq->lock, hereafter rq is only used for statistics. |
| 3906 | */ |
| 3907 | rq_unpin_lock(rq, rf); |
| 3908 | p->sched_class->task_woken(rq, p); |
| 3909 | rq_repin_lock(rq, rf); |
| 3910 | } |
| 3911 | } |
| 3912 | |
| 3913 | /* |
| 3914 | * Consider @p being inside a wait loop: |
| 3915 | * |
| 3916 | * for (;;) { |
| 3917 | * set_current_state(TASK_UNINTERRUPTIBLE); |
| 3918 | * |
| 3919 | * if (CONDITION) |
| 3920 | * break; |
| 3921 | * |
| 3922 | * schedule(); |
| 3923 | * } |
| 3924 | * __set_current_state(TASK_RUNNING); |
| 3925 | * |
| 3926 | * between set_current_state() and schedule(). In this case @p is still |
| 3927 | * runnable, so all that needs doing is change p->state back to TASK_RUNNING in |
| 3928 | * an atomic manner. |
| 3929 | * |
| 3930 | * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq |
| 3931 | * then schedule() must still happen and p->state can be changed to |
| 3932 | * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we |
| 3933 | * need to do a full wakeup with enqueue. |
| 3934 | * |
| 3935 | * Returns: %true when the wakeup is done, |
| 3936 | * %false otherwise. |
| 3937 | */ |
| 3938 | static int ttwu_runnable(struct task_struct *p, int wake_flags) |
| 3939 | { |
| 3940 | ACQUIRE(__task_rq_lock, guard)(p); |
| 3941 | struct rq *rq = guard.rq; |
| 3942 | |
| 3943 | if (!task_on_rq_queued(p)) |
| 3944 | return 0; |
| 3945 | |
| 3946 | update_rq_clock(rq); |
| 3947 | if (p->is_blocked) { |
| 3948 | if (p->se.sched_delayed) |
| 3949 | enqueue_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_DELAYED); |
| 3950 | if (proxy_needs_return(rq, p)) |
| 3951 | return 0; |
| 3952 | } |
| 3953 | if (!task_on_cpu(rq, p)) { |
| 3954 | /* |
| 3955 | * When on_rq && !on_cpu the task is preempted, see if |
| 3956 | * it should preempt the task that is current now. |
| 3957 | */ |
| 3958 | wakeup_preempt(rq, p, wake_flags | WF_TTWU_RQ); |
| 3959 | } |
| 3960 | ttwu_do_wakeup(p); |
| 3961 | return 1; |
| 3962 | } |
| 3963 | |
| 3964 | void sched_ttwu_pending(void *arg) |
| 3965 | { |
| 3966 | struct llist_node *llist = arg; |
| 3967 | struct rq *rq = this_rq(); |
| 3968 | struct task_struct *p, *t; |
| 3969 | struct rq_flags rf; |
| 3970 | |
| 3971 | if (!llist) |
| 3972 | return; |
| 3973 | |
| 3974 | rq_lock_irqsave(rq, &rf); |
| 3975 | update_rq_clock(rq); |
| 3976 | |
| 3977 | llist_for_each_entry_safe(p, t, llist, wake_entry.llist) { |
| 3978 | if (WARN_ON_ONCE(p->on_cpu)) |
| 3979 | smp_cond_load_acquire(&p->on_cpu, !VAL); |
| 3980 | |
| 3981 | if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq))) |
| 3982 | set_task_cpu(p, cpu_of(rq)); |
| 3983 | |
| 3984 | ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf); |
| 3985 | } |
| 3986 | |
| 3987 | /* |
| 3988 | * Must be after enqueueing at least once task such that |
| 3989 | * idle_cpu() does not observe a false-negative -- if it does, |
| 3990 | * it is possible for select_idle_siblings() to stack a number |
| 3991 | * of tasks on this CPU during that window. |
| 3992 | * |
| 3993 | * It is OK to clear ttwu_pending when another task pending. |
| 3994 | * We will receive IPI after local IRQ enabled and then enqueue it. |
| 3995 | * Since now nr_running > 0, idle_cpu() will always get correct result. |
| 3996 | */ |
| 3997 | WRITE_ONCE(rq->ttwu_pending, 0); |
| 3998 | rq_unlock_irqrestore(rq, &rf); |
| 3999 | } |
| 4000 | |
| 4001 | /* |
| 4002 | * Prepare the scene for sending an IPI for a remote smp_call |
| 4003 | * |
| 4004 | * Returns true if the caller can proceed with sending the IPI. |
| 4005 | * Returns false otherwise. |
| 4006 | */ |
| 4007 | bool call_function_single_prep_ipi(int cpu) |
| 4008 | { |
| 4009 | if (set_nr_if_polling(cpu_rq(cpu)->idle)) { |
| 4010 | trace_sched_wake_idle_without_ipi(cpu); |
| 4011 | return false; |
| 4012 | } |
| 4013 | |
| 4014 | return true; |
| 4015 | } |
| 4016 | |
| 4017 | /* |
| 4018 | * Queue a task on the target CPUs wake_list and wake the CPU via IPI if |
| 4019 | * necessary. The wakee CPU on receipt of the IPI will queue the task |
| 4020 | * via sched_ttwu_wakeup() for activation so the wakee incurs the cost |
| 4021 | * of the wakeup instead of the waker. |
| 4022 | */ |
| 4023 | static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) |
| 4024 | { |
| 4025 | struct rq *rq = cpu_rq(cpu); |
| 4026 | |
| 4027 | p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED); |
| 4028 | |
| 4029 | WRITE_ONCE(rq->ttwu_pending, 1); |
| 4030 | #ifdef CONFIG_SMP |
| 4031 | __smp_call_single_queue(cpu, &p->wake_entry.llist); |
| 4032 | #endif |
| 4033 | } |
| 4034 | |
| 4035 | void wake_up_if_idle(int cpu) |
| 4036 | { |
| 4037 | struct rq *rq = cpu_rq(cpu); |
| 4038 | |
| 4039 | guard(rcu)(); |
| 4040 | if (is_idle_task(rcu_dereference(rq->curr))) { |
| 4041 | guard(rq_lock_irqsave)(rq); |
| 4042 | if (is_idle_task(rq->curr)) |
| 4043 | resched_curr(rq); |
| 4044 | } |
| 4045 | } |
| 4046 | |
| 4047 | bool cpus_equal_capacity(int this_cpu, int that_cpu) |
| 4048 | { |
| 4049 | if (!sched_asym_cpucap_active()) |
| 4050 | return true; |
| 4051 | |
| 4052 | if (this_cpu == that_cpu) |
| 4053 | return true; |
| 4054 | |
| 4055 | return arch_scale_cpu_capacity(this_cpu) == arch_scale_cpu_capacity(that_cpu); |
| 4056 | } |
| 4057 | |
| 4058 | bool cpus_share_cache(int this_cpu, int that_cpu) |
| 4059 | { |
| 4060 | if (this_cpu == that_cpu) |
| 4061 | return true; |
| 4062 | |
| 4063 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); |
| 4064 | } |
| 4065 | |
| 4066 | /* |
| 4067 | * Whether CPUs are share cache resources, which means LLC on non-cluster |
| 4068 | * machines and LLC tag or L2 on machines with clusters. |
| 4069 | */ |
| 4070 | bool cpus_share_resources(int this_cpu, int that_cpu) |
| 4071 | { |
| 4072 | if (this_cpu == that_cpu) |
| 4073 | return true; |
| 4074 | |
| 4075 | return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu); |
| 4076 | } |
| 4077 | |
| 4078 | static inline bool ttwu_queue_cond(struct task_struct *p, int cpu) |
| 4079 | { |
| 4080 | int this_cpu = smp_processor_id(); |
| 4081 | |
| 4082 | /* See SCX_OPS_ALLOW_QUEUED_WAKEUP. */ |
| 4083 | if (!scx_allow_ttwu_queue(p)) |
| 4084 | return false; |
| 4085 | |
| 4086 | #ifdef CONFIG_SMP |
| 4087 | if (p->sched_class == &stop_sched_class) |
| 4088 | return false; |
| 4089 | #endif |
| 4090 | |
| 4091 | /* |
| 4092 | * Do not complicate things with the async wake_list while the CPU is |
| 4093 | * in hotplug state. |
| 4094 | */ |
| 4095 | if (!cpu_active(cpu)) |
| 4096 | return false; |
| 4097 | |
| 4098 | /* Ensure the task will still be allowed to run on the CPU. */ |
| 4099 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
| 4100 | return false; |
| 4101 | |
| 4102 | /* |
| 4103 | * If the CPU does not share cache, then queue the task on the |
| 4104 | * remote rqs wakelist to avoid accessing remote data. |
| 4105 | */ |
| 4106 | if (!cpus_share_cache(this_cpu, cpu)) |
| 4107 | return true; |
| 4108 | |
| 4109 | if (cpu == this_cpu) |
| 4110 | return false; |
| 4111 | |
| 4112 | /* |
| 4113 | * If the wakee cpu is idle, or the task is descheduling and the |
| 4114 | * only running task on the CPU, then use the wakelist to offload |
| 4115 | * the task activation to the idle (or soon-to-be-idle) CPU as |
| 4116 | * the current CPU is likely busy. nr_running is checked to |
| 4117 | * avoid unnecessary task stacking. |
| 4118 | * |
| 4119 | * Note that we can only get here with (wakee) p->on_rq=0, |
| 4120 | * p->on_cpu can be whatever, we've done the dequeue, so |
| 4121 | * the wakee has been accounted out of ->nr_running. |
| 4122 | */ |
| 4123 | if (!cpu_rq(cpu)->nr_running) |
| 4124 | return true; |
| 4125 | |
| 4126 | return false; |
| 4127 | } |
| 4128 | |
| 4129 | static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) |
| 4130 | { |
| 4131 | if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) { |
| 4132 | sched_clock_cpu(cpu); /* Sync clocks across CPUs */ |
| 4133 | __ttwu_queue_wakelist(p, cpu, wake_flags); |
| 4134 | return true; |
| 4135 | } |
| 4136 | |
| 4137 | return false; |
| 4138 | } |
| 4139 | |
| 4140 | static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags) |
| 4141 | { |
| 4142 | struct rq *rq = cpu_rq(cpu); |
| 4143 | struct rq_flags rf; |
| 4144 | |
| 4145 | if (ttwu_queue_wakelist(p, cpu, wake_flags)) |
| 4146 | return; |
| 4147 | |
| 4148 | rq_lock(rq, &rf); |
| 4149 | update_rq_clock(rq); |
| 4150 | ttwu_do_activate(rq, p, wake_flags, &rf); |
| 4151 | rq_unlock(rq, &rf); |
| 4152 | } |
| 4153 | |
| 4154 | /* |
| 4155 | * Invoked from try_to_wake_up() to check whether the task can be woken up. |
| 4156 | * |
| 4157 | * The caller holds p::pi_lock if p != current or has preemption |
| 4158 | * disabled when p == current. |
| 4159 | * |
| 4160 | * The rules of saved_state: |
| 4161 | * |
| 4162 | * The related locking code always holds p::pi_lock when updating |
| 4163 | * p::saved_state, which means the code is fully serialized in both cases. |
| 4164 | * |
| 4165 | * For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. |
| 4166 | * No other bits set. This allows to distinguish all wakeup scenarios. |
| 4167 | * |
| 4168 | * For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This |
| 4169 | * allows us to prevent early wakeup of tasks before they can be run on |
| 4170 | * asymmetric ISA architectures (eg ARMv9). |
| 4171 | */ |
| 4172 | static __always_inline |
| 4173 | bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success) |
| 4174 | { |
| 4175 | int match; |
| 4176 | |
| 4177 | if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) { |
| 4178 | WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) && |
| 4179 | state != TASK_RTLOCK_WAIT); |
| 4180 | } |
| 4181 | |
| 4182 | *success = !!(match = __task_state_match(p, state)); |
| 4183 | |
| 4184 | /* |
| 4185 | * Saved state preserves the task state across blocking on |
| 4186 | * an RT lock or TASK_FREEZABLE tasks. If the state matches, |
| 4187 | * set p::saved_state to TASK_RUNNING, but do not wake the task |
| 4188 | * because it waits for a lock wakeup or __thaw_task(). Also |
| 4189 | * indicate success because from the regular waker's point of |
| 4190 | * view this has succeeded. |
| 4191 | * |
| 4192 | * After acquiring the lock the task will restore p::__state |
| 4193 | * from p::saved_state which ensures that the regular |
| 4194 | * wakeup is not lost. The restore will also set |
| 4195 | * p::saved_state to TASK_RUNNING so any further tests will |
| 4196 | * not result in false positives vs. @success |
| 4197 | */ |
| 4198 | if (match < 0) |
| 4199 | p->saved_state = TASK_RUNNING; |
| 4200 | |
| 4201 | return match > 0; |
| 4202 | } |
| 4203 | |
| 4204 | /* |
| 4205 | * Notes on Program-Order guarantees on SMP systems. |
| 4206 | * |
| 4207 | * MIGRATION |
| 4208 | * |
| 4209 | * The basic program-order guarantee on SMP systems is that when a task [t] |
| 4210 | * migrates, all its activity on its old CPU [c0] happens-before any subsequent |
| 4211 | * execution on its new CPU [c1]. |
| 4212 | * |
| 4213 | * For migration (of runnable tasks) this is provided by the following means: |
| 4214 | * |
| 4215 | * A) UNLOCK of the rq(c0)->lock scheduling out task t |
| 4216 | * B) migration for t is required to synchronize *both* rq(c0)->lock and |
| 4217 | * rq(c1)->lock (if not at the same time, then in that order). |
| 4218 | * C) LOCK of the rq(c1)->lock scheduling in task |
| 4219 | * |
| 4220 | * Release/acquire chaining guarantees that B happens after A and C after B. |
| 4221 | * Note: the CPU doing B need not be c0 or c1 |
| 4222 | * |
| 4223 | * Example: |
| 4224 | * |
| 4225 | * CPU0 CPU1 CPU2 |
| 4226 | * |
| 4227 | * LOCK rq(0)->lock |
| 4228 | * sched-out X |
| 4229 | * sched-in Y |
| 4230 | * UNLOCK rq(0)->lock |
| 4231 | * |
| 4232 | * LOCK rq(0)->lock // orders against CPU0 |
| 4233 | * dequeue X |
| 4234 | * UNLOCK rq(0)->lock |
| 4235 | * |
| 4236 | * LOCK rq(1)->lock |
| 4237 | * enqueue X |
| 4238 | * UNLOCK rq(1)->lock |
| 4239 | * |
| 4240 | * LOCK rq(1)->lock // orders against CPU2 |
| 4241 | * sched-out Z |
| 4242 | * sched-in X |
| 4243 | * UNLOCK rq(1)->lock |
| 4244 | * |
| 4245 | * |
| 4246 | * BLOCKING -- aka. SLEEP + WAKEUP |
| 4247 | * |
| 4248 | * For blocking we (obviously) need to provide the same guarantee as for |
| 4249 | * migration. However the means are completely different as there is no lock |
| 4250 | * chain to provide order. Instead we do: |
| 4251 | * |
| 4252 | * 1) smp_store_release(X->on_cpu, 0) -- finish_task() |
| 4253 | * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up() |
| 4254 | * |
| 4255 | * Example: |
| 4256 | * |
| 4257 | * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule) |
| 4258 | * |
| 4259 | * LOCK rq(0)->lock LOCK X->pi_lock |
| 4260 | * dequeue X |
| 4261 | * sched-out X |
| 4262 | * smp_store_release(X->on_cpu, 0); |
| 4263 | * |
| 4264 | * smp_cond_load_acquire(&X->on_cpu, !VAL); |
| 4265 | * X->state = WAKING |
| 4266 | * set_task_cpu(X,2) |
| 4267 | * |
| 4268 | * LOCK rq(2)->lock |
| 4269 | * enqueue X |
| 4270 | * X->state = RUNNING |
| 4271 | * UNLOCK rq(2)->lock |
| 4272 | * |
| 4273 | * LOCK rq(2)->lock // orders against CPU1 |
| 4274 | * sched-out Z |
| 4275 | * sched-in X |
| 4276 | * UNLOCK rq(2)->lock |
| 4277 | * |
| 4278 | * UNLOCK X->pi_lock |
| 4279 | * UNLOCK rq(0)->lock |
| 4280 | * |
| 4281 | * |
| 4282 | * However, for wakeups there is a second guarantee we must provide, namely we |
| 4283 | * must ensure that CONDITION=1 done by the caller can not be reordered with |
| 4284 | * accesses to the task state; see try_to_wake_up() and set_current_state(). |
| 4285 | */ |
| 4286 | |
| 4287 | /** |
| 4288 | * try_to_wake_up - wake up a thread |
| 4289 | * @p: the thread to be awakened |
| 4290 | * @state: the mask of task states that can be woken |
| 4291 | * @wake_flags: wake modifier flags (WF_*) |
| 4292 | * |
| 4293 | * Conceptually does: |
| 4294 | * |
| 4295 | * If (@state & @p->state) @p->state = TASK_RUNNING. |
| 4296 | * |
| 4297 | * If the task was not queued/runnable, also place it back on a runqueue. |
| 4298 | * |
| 4299 | * This function is atomic against schedule() which would dequeue the task. |
| 4300 | * |
| 4301 | * It issues a full memory barrier before accessing @p->state, see the comment |
| 4302 | * with set_current_state(). |
| 4303 | * |
| 4304 | * Uses p->pi_lock to serialize against concurrent wake-ups. |
| 4305 | * |
| 4306 | * Relies on p->pi_lock stabilizing: |
| 4307 | * - p->sched_class |
| 4308 | * - p->cpus_ptr |
| 4309 | * - p->sched_task_group |
| 4310 | * in order to do migration, see its use of select_task_rq()/set_task_cpu(). |
| 4311 | * |
| 4312 | * Tries really hard to only take one task_rq(p)->lock for performance. |
| 4313 | * Takes rq->lock in: |
| 4314 | * - ttwu_runnable() -- old rq, unavoidable, see comment there; |
| 4315 | * - ttwu_queue() -- new rq, for enqueue of the task; |
| 4316 | * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us. |
| 4317 | * |
| 4318 | * As a consequence we race really badly with just about everything. See the |
| 4319 | * many memory barriers and their comments for details. |
| 4320 | * |
| 4321 | * Return: %true if @p->state changes (an actual wakeup was done), |
| 4322 | * %false otherwise. |
| 4323 | */ |
| 4324 | int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) |
| 4325 | { |
| 4326 | guard(preempt)(); |
| 4327 | int cpu, success = 0; |
| 4328 | |
| 4329 | wake_flags |= WF_TTWU; |
| 4330 | |
| 4331 | if (p == current) { |
| 4332 | /* |
| 4333 | * We're waking current, this means 'p->on_rq' and 'task_cpu(p) |
| 4334 | * == smp_processor_id()'. Together this means we can special |
| 4335 | * case the whole 'p->on_rq && ttwu_runnable()' case below |
| 4336 | * without taking any locks. |
| 4337 | * |
| 4338 | * Specifically, given current runs ttwu() we must be before |
| 4339 | * schedule()'s block_task(), as such this must not observe |
| 4340 | * sched_delayed. |
| 4341 | * |
| 4342 | * In particular: |
| 4343 | * - we rely on Program-Order guarantees for all the ordering, |
| 4344 | * - we're serialized against set_special_state() by virtue of |
| 4345 | * it disabling IRQs (this allows not taking ->pi_lock). |
| 4346 | */ |
| 4347 | WARN_ON_ONCE(p->se.sched_delayed); |
| 4348 | WARN_ON_ONCE(p->is_blocked); |
| 4349 | /* If p is current, we know we can run here, so clear blocked_on */ |
| 4350 | clear_task_blocked_on(p, NULL); |
| 4351 | if (!ttwu_state_match(p, state, &success)) |
| 4352 | goto out; |
| 4353 | |
| 4354 | trace_sched_waking(p); |
| 4355 | ttwu_do_wakeup(p); |
| 4356 | goto out; |
| 4357 | } |
| 4358 | |
| 4359 | /* |
| 4360 | * If we are going to wake up a thread waiting for CONDITION we |
| 4361 | * need to ensure that CONDITION=1 done by the caller can not be |
| 4362 | * reordered with p->state check below. This pairs with smp_store_mb() |
| 4363 | * in set_current_state() that the waiting thread does. |
| 4364 | */ |
| 4365 | scoped_guard (raw_spinlock_irqsave, &p->pi_lock) { |
| 4366 | smp_mb__after_spinlock(); |
| 4367 | |
| 4368 | if (!ttwu_state_match(p, state, &success)) |
| 4369 | break; |
| 4370 | |
| 4371 | trace_sched_waking(p); |
| 4372 | |
| 4373 | /* |
| 4374 | * Ensure we load p->on_rq _after_ p->state, otherwise it would |
| 4375 | * be possible to, falsely, observe p->on_rq == 0 and get stuck |
| 4376 | * in smp_cond_load_acquire() below. |
| 4377 | * |
| 4378 | * sched_ttwu_pending() try_to_wake_up() |
| 4379 | * STORE p->on_rq = 1 LOAD p->state |
| 4380 | * UNLOCK rq->lock |
| 4381 | * |
| 4382 | * __schedule() (switch to task 'p') |
| 4383 | * LOCK rq->lock smp_rmb(); |
| 4384 | * smp_mb__after_spinlock(); |
| 4385 | * UNLOCK rq->lock |
| 4386 | * |
| 4387 | * [task p] |
| 4388 | * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq |
| 4389 | * |
| 4390 | * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in |
| 4391 | * __schedule(). See the comment for smp_mb__after_spinlock(). |
| 4392 | * |
| 4393 | * A similar smp_rmb() lives in __task_needs_rq_lock(). |
| 4394 | */ |
| 4395 | smp_rmb(); |
| 4396 | if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags)) |
| 4397 | break; |
| 4398 | |
| 4399 | /* |
| 4400 | * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be |
| 4401 | * possible to, falsely, observe p->on_cpu == 0. |
| 4402 | * |
| 4403 | * One must be running (->on_cpu == 1) in order to remove oneself |
| 4404 | * from the runqueue. |
| 4405 | * |
| 4406 | * __schedule() (switch to task 'p') try_to_wake_up() |
| 4407 | * STORE p->on_cpu = 1 LOAD p->on_rq |
| 4408 | * UNLOCK rq->lock |
| 4409 | * |
| 4410 | * __schedule() (put 'p' to sleep) |
| 4411 | * LOCK rq->lock smp_rmb(); |
| 4412 | * smp_mb__after_spinlock(); |
| 4413 | * STORE p->on_rq = 0 LOAD p->on_cpu |
| 4414 | * |
| 4415 | * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in |
| 4416 | * __schedule(). See the comment for smp_mb__after_spinlock(). |
| 4417 | * |
| 4418 | * Form a control-dep-acquire with p->on_rq == 0 above, to ensure |
| 4419 | * schedule()'s block_task() has 'happened' and p will no longer |
| 4420 | * care about it's own p->state. See the comment in __schedule(). |
| 4421 | */ |
| 4422 | smp_acquire__after_ctrl_dep(); |
| 4423 | |
| 4424 | /* |
| 4425 | * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq |
| 4426 | * == 0), which means we need to do an enqueue, change p->state to |
| 4427 | * TASK_WAKING such that we can unlock p->pi_lock before doing the |
| 4428 | * enqueue, such as ttwu_queue_wakelist(). |
| 4429 | */ |
| 4430 | WRITE_ONCE(p->__state, TASK_WAKING); |
| 4431 | |
| 4432 | /* |
| 4433 | * If the owning (remote) CPU is still in the middle of schedule() with |
| 4434 | * this task as prev, considering queueing p on the remote CPUs wake_list |
| 4435 | * which potentially sends an IPI instead of spinning on p->on_cpu to |
| 4436 | * let the waker make forward progress. This is safe because IRQs are |
| 4437 | * disabled and the IPI will deliver after on_cpu is cleared. |
| 4438 | * |
| 4439 | * Ensure we load task_cpu(p) after p->on_cpu: |
| 4440 | * |
| 4441 | * set_task_cpu(p, cpu); |
| 4442 | * STORE p->cpu = @cpu |
| 4443 | * __schedule() (switch to task 'p') |
| 4444 | * LOCK rq->lock |
| 4445 | * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu) |
| 4446 | * STORE p->on_cpu = 1 LOAD p->cpu |
| 4447 | * |
| 4448 | * to ensure we observe the correct CPU on which the task is currently |
| 4449 | * scheduling. |
| 4450 | */ |
| 4451 | if (smp_load_acquire(&p->on_cpu) && |
| 4452 | ttwu_queue_wakelist(p, task_cpu(p), wake_flags)) |
| 4453 | break; |
| 4454 | |
| 4455 | /* |
| 4456 | * If the owning (remote) CPU is still in the middle of schedule() with |
| 4457 | * this task as prev, wait until it's done referencing the task. |
| 4458 | * |
| 4459 | * Pairs with the smp_store_release() in finish_task(). |
| 4460 | * |
| 4461 | * This ensures that tasks getting woken will be fully ordered against |
| 4462 | * their previous state and preserve Program Order. |
| 4463 | */ |
| 4464 | smp_cond_load_acquire(&p->on_cpu, !VAL); |
| 4465 | |
| 4466 | cpu = select_task_rq(p, p->wake_cpu, &wake_flags); |
| 4467 | if (task_cpu(p) != cpu) { |
| 4468 | if (p->in_iowait) { |
| 4469 | delayacct_blkio_end(p); |
| 4470 | atomic_dec(&task_rq(p)->nr_iowait); |
| 4471 | } |
| 4472 | |
| 4473 | wake_flags |= WF_MIGRATED; |
| 4474 | psi_ttwu_dequeue(p); |
| 4475 | set_task_cpu(p, cpu); |
| 4476 | } else if (cpu != p->wake_cpu) { |
| 4477 | /* |
| 4478 | * If we were proxy-migrated to cpu, then |
| 4479 | * select_task_rq() picks cpu instead of wake_cpu |
| 4480 | * to return to, we won't call set_task_cpu(), |
| 4481 | * leaving a stale wake_cpu pointing to where we |
| 4482 | * proxy-migrated from. So just fixup wake_cpu here |
| 4483 | * if its not correct |
| 4484 | */ |
| 4485 | p->wake_cpu = cpu; |
| 4486 | } |
| 4487 | |
| 4488 | ttwu_queue(p, cpu, wake_flags); |
| 4489 | } |
| 4490 | out: |
| 4491 | if (success) |
| 4492 | ttwu_stat(p, task_cpu(p), wake_flags); |
| 4493 | |
| 4494 | return success; |
| 4495 | } |
| 4496 | |
| 4497 | static bool __task_needs_rq_lock(struct task_struct *p) |
| 4498 | { |
| 4499 | unsigned int state = READ_ONCE(p->__state); |
| 4500 | |
| 4501 | /* |
| 4502 | * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when |
| 4503 | * the task is blocked. Make sure to check @state since ttwu() can drop |
| 4504 | * locks at the end, see ttwu_queue_wakelist(). |
| 4505 | */ |
| 4506 | if (state == TASK_RUNNING || state == TASK_WAKING) |
| 4507 | return true; |
| 4508 | |
| 4509 | /* |
| 4510 | * Ensure we load p->on_rq after p->__state, otherwise it would be |
| 4511 | * possible to, falsely, observe p->on_rq == 0. |
| 4512 | * |
| 4513 | * See try_to_wake_up() for a longer comment. |
| 4514 | */ |
| 4515 | smp_rmb(); |
| 4516 | if (p->on_rq) |
| 4517 | return true; |
| 4518 | |
| 4519 | /* |
| 4520 | * Ensure the task has finished __schedule() and will not be referenced |
| 4521 | * anymore. Again, see try_to_wake_up() for a longer comment. |
| 4522 | */ |
| 4523 | smp_rmb(); |
| 4524 | smp_cond_load_acquire(&p->on_cpu, !VAL); |
| 4525 | |
| 4526 | return false; |
| 4527 | } |
| 4528 | |
| 4529 | /** |
| 4530 | * task_call_func - Invoke a function on task in fixed state |
| 4531 | * @p: Process for which the function is to be invoked, can be @current. |
| 4532 | * @func: Function to invoke. |
| 4533 | * @arg: Argument to function. |
| 4534 | * |
| 4535 | * Fix the task in it's current state by avoiding wakeups and or rq operations |
| 4536 | * and call @func(@arg) on it. This function can use task_is_runnable() and |
| 4537 | * task_curr() to work out what the state is, if required. Given that @func |
| 4538 | * can be invoked with a runqueue lock held, it had better be quite |
| 4539 | * lightweight. |
| 4540 | * |
| 4541 | * Returns: |
| 4542 | * Whatever @func returns |
| 4543 | */ |
| 4544 | int task_call_func(struct task_struct *p, task_call_f func, void *arg) |
| 4545 | { |
| 4546 | struct rq_flags rf; |
| 4547 | int ret; |
| 4548 | |
| 4549 | raw_spin_lock_irqsave(&p->pi_lock, rf.flags); |
| 4550 | |
| 4551 | if (__task_needs_rq_lock(p)) { |
| 4552 | struct rq *rq = __task_rq_lock(p, &rf); |
| 4553 | |
| 4554 | /* |
| 4555 | * At this point the task is pinned; either: |
| 4556 | * - blocked and we're holding off wakeups (pi->lock) |
| 4557 | * - woken, and we're holding off enqueue (rq->lock) |
| 4558 | * - queued, and we're holding off schedule (rq->lock) |
| 4559 | * - running, and we're holding off de-schedule (rq->lock) |
| 4560 | * |
| 4561 | * The called function (@func) can use: task_curr(), p->on_rq and |
| 4562 | * p->__state to differentiate between these states. |
| 4563 | */ |
| 4564 | ret = func(p, arg); |
| 4565 | |
| 4566 | __task_rq_unlock(rq, p, &rf); |
| 4567 | } else { |
| 4568 | ret = func(p, arg); |
| 4569 | } |
| 4570 | |
| 4571 | raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags); |
| 4572 | return ret; |
| 4573 | } |
| 4574 | |
| 4575 | /** |
| 4576 | * cpu_curr_snapshot - Return a snapshot of the currently running task |
| 4577 | * @cpu: The CPU on which to snapshot the task. |
| 4578 | * |
| 4579 | * Returns the task_struct pointer of the task "currently" running on |
| 4580 | * the specified CPU. |
| 4581 | * |
| 4582 | * If the specified CPU was offline, the return value is whatever it |
| 4583 | * is, perhaps a pointer to the task_struct structure of that CPU's idle |
| 4584 | * task, but there is no guarantee. Callers wishing a useful return |
| 4585 | * value must take some action to ensure that the specified CPU remains |
| 4586 | * online throughout. |
| 4587 | * |
| 4588 | * This function executes full memory barriers before and after fetching |
| 4589 | * the pointer, which permits the caller to confine this function's fetch |
| 4590 | * with respect to the caller's accesses to other shared variables. |
| 4591 | */ |
| 4592 | struct task_struct *cpu_curr_snapshot(int cpu) |
| 4593 | { |
| 4594 | struct rq *rq = cpu_rq(cpu); |
| 4595 | struct task_struct *t; |
| 4596 | struct rq_flags rf; |
| 4597 | |
| 4598 | rq_lock_irqsave(rq, &rf); |
| 4599 | smp_mb__after_spinlock(); /* Pairing determined by caller's synchronization design. */ |
| 4600 | t = rcu_dereference(cpu_curr(cpu)); |
| 4601 | rq_unlock_irqrestore(rq, &rf); |
| 4602 | smp_mb(); /* Pairing determined by caller's synchronization design. */ |
| 4603 | |
| 4604 | return t; |
| 4605 | } |
| 4606 | |
| 4607 | /** |
| 4608 | * wake_up_process - Wake up a specific process |
| 4609 | * @p: The process to be woken up. |
| 4610 | * |
| 4611 | * Attempt to wake up the nominated process and move it to the set of runnable |
| 4612 | * processes. |
| 4613 | * |
| 4614 | * Return: 1 if the process was woken up, 0 if it was already running. |
| 4615 | * |
| 4616 | * This function executes a full memory barrier before accessing the task state. |
| 4617 | */ |
| 4618 | int wake_up_process(struct task_struct *p) |
| 4619 | { |
| 4620 | return try_to_wake_up(p, TASK_NORMAL, 0); |
| 4621 | } |
| 4622 | EXPORT_SYMBOL(wake_up_process); |
| 4623 | |
| 4624 | int wake_up_state(struct task_struct *p, unsigned int state) |
| 4625 | { |
| 4626 | return try_to_wake_up(p, state, 0); |
| 4627 | } |
| 4628 | |
| 4629 | /* |
| 4630 | * Perform scheduler related setup for a newly forked process p. |
| 4631 | * p is forked by current. |
| 4632 | * |
| 4633 | * __sched_fork() is basic setup which is also used by sched_init() to |
| 4634 | * initialize the boot CPU's idle task. |
| 4635 | */ |
| 4636 | static void __sched_fork(u64 clone_flags, struct task_struct *p) |
| 4637 | { |
| 4638 | p->on_rq = 0; |
| 4639 | |
| 4640 | p->se.on_rq = 0; |
| 4641 | p->se.exec_start = 0; |
| 4642 | p->se.sum_exec_runtime = 0; |
| 4643 | p->se.prev_sum_exec_runtime = 0; |
| 4644 | p->se.nr_migrations = 0; |
| 4645 | p->se.vruntime = 0; |
| 4646 | p->se.vlag = 0; |
| 4647 | p->se.rel_deadline = 0; |
| 4648 | INIT_LIST_HEAD(&p->se.group_node); |
| 4649 | |
| 4650 | /* A delayed task cannot be in clone(). */ |
| 4651 | WARN_ON_ONCE(p->se.sched_delayed); |
| 4652 | WARN_ON_ONCE(p->is_blocked); |
| 4653 | |
| 4654 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 4655 | p->se.cfs_rq = NULL; |
| 4656 | #ifdef CONFIG_CFS_BANDWIDTH |
| 4657 | init_cfs_throttle_work(p); |
| 4658 | #endif |
| 4659 | #endif |
| 4660 | |
| 4661 | #ifdef CONFIG_SCHEDSTATS |
| 4662 | /* Even if schedstat is disabled, there should not be garbage */ |
| 4663 | memset(&p->stats, 0, sizeof(p->stats)); |
| 4664 | #endif |
| 4665 | |
| 4666 | init_dl_entity(&p->dl); |
| 4667 | |
| 4668 | INIT_LIST_HEAD(&p->rt.run_list); |
| 4669 | p->rt.timeout = 0; |
| 4670 | p->rt.time_slice = sched_rr_timeslice; |
| 4671 | p->rt.on_rq = 0; |
| 4672 | p->rt.on_list = 0; |
| 4673 | |
| 4674 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 4675 | init_scx_entity(&p->scx); |
| 4676 | #endif |
| 4677 | |
| 4678 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 4679 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
| 4680 | #endif |
| 4681 | |
| 4682 | #ifdef CONFIG_COMPACTION |
| 4683 | p->capture_control = NULL; |
| 4684 | #endif |
| 4685 | init_numa_balancing(clone_flags, p); |
| 4686 | p->wake_entry.u_flags = CSD_TYPE_TTWU; |
| 4687 | p->migration_pending = NULL; |
| 4688 | init_sched_mm(p); |
| 4689 | } |
| 4690 | |
| 4691 | DEFINE_STATIC_KEY_FALSE(sched_numa_balancing); |
| 4692 | |
| 4693 | #ifdef CONFIG_NUMA_BALANCING |
| 4694 | |
| 4695 | int sysctl_numa_balancing_mode; |
| 4696 | |
| 4697 | static void __set_numabalancing_state(bool enabled) |
| 4698 | { |
| 4699 | if (enabled) |
| 4700 | static_branch_enable(&sched_numa_balancing); |
| 4701 | else |
| 4702 | static_branch_disable(&sched_numa_balancing); |
| 4703 | } |
| 4704 | |
| 4705 | void set_numabalancing_state(bool enabled) |
| 4706 | { |
| 4707 | if (enabled) |
| 4708 | sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL; |
| 4709 | else |
| 4710 | sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED; |
| 4711 | __set_numabalancing_state(enabled); |
| 4712 | } |
| 4713 | |
| 4714 | #ifdef CONFIG_SYSCTL |
| 4715 | static void reset_memory_tiering(void) |
| 4716 | { |
| 4717 | struct pglist_data *pgdat; |
| 4718 | |
| 4719 | for_each_online_pgdat(pgdat) { |
| 4720 | pgdat->nbp_threshold = 0; |
| 4721 | pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); |
| 4722 | pgdat->nbp_th_start = jiffies_to_msecs(jiffies); |
| 4723 | } |
| 4724 | } |
| 4725 | |
| 4726 | static int sysctl_numa_balancing(const struct ctl_table *table, int write, |
| 4727 | void *buffer, size_t *lenp, loff_t *ppos) |
| 4728 | { |
| 4729 | struct ctl_table t; |
| 4730 | int err; |
| 4731 | int state = sysctl_numa_balancing_mode; |
| 4732 | |
| 4733 | if (write && !capable(CAP_SYS_ADMIN)) |
| 4734 | return -EPERM; |
| 4735 | |
| 4736 | t = *table; |
| 4737 | t.data = &state; |
| 4738 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); |
| 4739 | if (err < 0) |
| 4740 | return err; |
| 4741 | if (write) { |
| 4742 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && |
| 4743 | (state & NUMA_BALANCING_MEMORY_TIERING)) |
| 4744 | reset_memory_tiering(); |
| 4745 | sysctl_numa_balancing_mode = state; |
| 4746 | __set_numabalancing_state(state); |
| 4747 | } |
| 4748 | return err; |
| 4749 | } |
| 4750 | #endif /* CONFIG_SYSCTL */ |
| 4751 | #endif /* CONFIG_NUMA_BALANCING */ |
| 4752 | |
| 4753 | #ifdef CONFIG_SCHEDSTATS |
| 4754 | |
| 4755 | DEFINE_STATIC_KEY_FALSE(sched_schedstats); |
| 4756 | |
| 4757 | static void set_schedstats(bool enabled) |
| 4758 | { |
| 4759 | if (enabled) |
| 4760 | static_branch_enable(&sched_schedstats); |
| 4761 | else |
| 4762 | static_branch_disable(&sched_schedstats); |
| 4763 | } |
| 4764 | |
| 4765 | void force_schedstat_enabled(void) |
| 4766 | { |
| 4767 | if (!schedstat_enabled()) { |
| 4768 | pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n"); |
| 4769 | static_branch_enable(&sched_schedstats); |
| 4770 | } |
| 4771 | } |
| 4772 | |
| 4773 | static int __init setup_schedstats(char *str) |
| 4774 | { |
| 4775 | int ret = 0; |
| 4776 | if (!str) |
| 4777 | goto out; |
| 4778 | |
| 4779 | if (!strcmp(str, "enable")) { |
| 4780 | set_schedstats(true); |
| 4781 | ret = 1; |
| 4782 | } else if (!strcmp(str, "disable")) { |
| 4783 | set_schedstats(false); |
| 4784 | ret = 1; |
| 4785 | } |
| 4786 | out: |
| 4787 | if (!ret) |
| 4788 | pr_warn("Unable to parse schedstats=\n"); |
| 4789 | |
| 4790 | return ret; |
| 4791 | } |
| 4792 | __setup("schedstats=", setup_schedstats); |
| 4793 | |
| 4794 | #ifdef CONFIG_SYSCTL |
| 4795 | static int sysctl_schedstats(const struct ctl_table *table, int write, void *buffer, |
| 4796 | size_t *lenp, loff_t *ppos) |
| 4797 | { |
| 4798 | struct ctl_table t; |
| 4799 | int err; |
| 4800 | int state = static_branch_likely(&sched_schedstats); |
| 4801 | |
| 4802 | if (write && !capable(CAP_SYS_ADMIN)) |
| 4803 | return -EPERM; |
| 4804 | |
| 4805 | t = *table; |
| 4806 | t.data = &state; |
| 4807 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); |
| 4808 | if (err < 0) |
| 4809 | return err; |
| 4810 | if (write) |
| 4811 | set_schedstats(state); |
| 4812 | return err; |
| 4813 | } |
| 4814 | #endif /* CONFIG_SYSCTL */ |
| 4815 | #endif /* CONFIG_SCHEDSTATS */ |
| 4816 | |
| 4817 | #ifdef CONFIG_SYSCTL |
| 4818 | static const struct ctl_table sched_core_sysctls[] = { |
| 4819 | #ifdef CONFIG_SCHEDSTATS |
| 4820 | { |
| 4821 | .procname = "sched_schedstats", |
| 4822 | .data = NULL, |
| 4823 | .maxlen = sizeof(unsigned int), |
| 4824 | .mode = 0644, |
| 4825 | .proc_handler = sysctl_schedstats, |
| 4826 | .extra1 = SYSCTL_ZERO, |
| 4827 | .extra2 = SYSCTL_ONE, |
| 4828 | }, |
| 4829 | #endif /* CONFIG_SCHEDSTATS */ |
| 4830 | #ifdef CONFIG_UCLAMP_TASK |
| 4831 | { |
| 4832 | .procname = "sched_util_clamp_min", |
| 4833 | .data = &sysctl_sched_uclamp_util_min, |
| 4834 | .maxlen = sizeof(unsigned int), |
| 4835 | .mode = 0644, |
| 4836 | .proc_handler = sysctl_sched_uclamp_handler, |
| 4837 | }, |
| 4838 | { |
| 4839 | .procname = "sched_util_clamp_max", |
| 4840 | .data = &sysctl_sched_uclamp_util_max, |
| 4841 | .maxlen = sizeof(unsigned int), |
| 4842 | .mode = 0644, |
| 4843 | .proc_handler = sysctl_sched_uclamp_handler, |
| 4844 | }, |
| 4845 | { |
| 4846 | .procname = "sched_util_clamp_min_rt_default", |
| 4847 | .data = &sysctl_sched_uclamp_util_min_rt_default, |
| 4848 | .maxlen = sizeof(unsigned int), |
| 4849 | .mode = 0644, |
| 4850 | .proc_handler = sysctl_sched_uclamp_handler, |
| 4851 | }, |
| 4852 | #endif /* CONFIG_UCLAMP_TASK */ |
| 4853 | #ifdef CONFIG_NUMA_BALANCING |
| 4854 | { |
| 4855 | .procname = "numa_balancing", |
| 4856 | .data = NULL, /* filled in by handler */ |
| 4857 | .maxlen = sizeof(unsigned int), |
| 4858 | .mode = 0644, |
| 4859 | .proc_handler = sysctl_numa_balancing, |
| 4860 | .extra1 = SYSCTL_ZERO, |
| 4861 | .extra2 = SYSCTL_FOUR, |
| 4862 | }, |
| 4863 | #endif /* CONFIG_NUMA_BALANCING */ |
| 4864 | }; |
| 4865 | static int __init sched_core_sysctl_init(void) |
| 4866 | { |
| 4867 | register_sysctl_init("kernel", sched_core_sysctls); |
| 4868 | return 0; |
| 4869 | } |
| 4870 | late_initcall(sched_core_sysctl_init); |
| 4871 | #endif /* CONFIG_SYSCTL */ |
| 4872 | |
| 4873 | /* |
| 4874 | * fork()/clone()-time setup: |
| 4875 | */ |
| 4876 | int sched_fork(u64 clone_flags, struct task_struct *p) |
| 4877 | { |
| 4878 | __sched_fork(clone_flags, p); |
| 4879 | /* |
| 4880 | * We mark the process as NEW here. This guarantees that |
| 4881 | * nobody will actually run it, and a signal or other external |
| 4882 | * event cannot wake it up and insert it on the runqueue either. |
| 4883 | */ |
| 4884 | p->__state = TASK_NEW; |
| 4885 | |
| 4886 | /* |
| 4887 | * Make sure we do not leak PI boosting priority to the child. |
| 4888 | */ |
| 4889 | p->prio = current->normal_prio; |
| 4890 | |
| 4891 | uclamp_fork(p); |
| 4892 | |
| 4893 | /* |
| 4894 | * Revert to default priority/policy on fork if requested. |
| 4895 | */ |
| 4896 | if (unlikely(p->sched_reset_on_fork)) { |
| 4897 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { |
| 4898 | p->policy = SCHED_NORMAL; |
| 4899 | p->static_prio = NICE_TO_PRIO(0); |
| 4900 | p->rt_priority = 0; |
| 4901 | p->timer_slack_ns = p->default_timer_slack_ns; |
| 4902 | } else if (PRIO_TO_NICE(p->static_prio) < 0) |
| 4903 | p->static_prio = NICE_TO_PRIO(0); |
| 4904 | |
| 4905 | p->prio = p->normal_prio = p->static_prio; |
| 4906 | set_load_weight(p, false); |
| 4907 | p->se.custom_slice = 0; |
| 4908 | p->se.slice = sysctl_sched_base_slice; |
| 4909 | |
| 4910 | /* |
| 4911 | * We don't need the reset flag anymore after the fork. It has |
| 4912 | * fulfilled its duty: |
| 4913 | */ |
| 4914 | p->sched_reset_on_fork = 0; |
| 4915 | } |
| 4916 | |
| 4917 | if (dl_prio(p->prio)) |
| 4918 | return -EAGAIN; |
| 4919 | |
| 4920 | scx_pre_fork(p); |
| 4921 | |
| 4922 | if (rt_prio(p->prio)) { |
| 4923 | p->sched_class = &rt_sched_class; |
| 4924 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 4925 | } else if (task_should_scx(p->policy)) { |
| 4926 | p->sched_class = &ext_sched_class; |
| 4927 | #endif |
| 4928 | } else { |
| 4929 | p->sched_class = &fair_sched_class; |
| 4930 | } |
| 4931 | |
| 4932 | init_entity_runnable_average(&p->se); |
| 4933 | |
| 4934 | |
| 4935 | #ifdef CONFIG_SCHED_INFO |
| 4936 | if (likely(sched_info_on())) |
| 4937 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
| 4938 | #endif |
| 4939 | p->on_cpu = 0; |
| 4940 | init_task_preempt_count(p); |
| 4941 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
| 4942 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
| 4943 | |
| 4944 | return 0; |
| 4945 | } |
| 4946 | |
| 4947 | int sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs) |
| 4948 | { |
| 4949 | unsigned long flags; |
| 4950 | |
| 4951 | /* |
| 4952 | * Because we're not yet on the pid-hash, p->pi_lock isn't strictly |
| 4953 | * required yet, but lockdep gets upset if rules are violated. |
| 4954 | */ |
| 4955 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
| 4956 | #ifdef CONFIG_CGROUP_SCHED |
| 4957 | if (1) { |
| 4958 | struct task_group *tg; |
| 4959 | tg = container_of(kargs->cset->subsys[cpu_cgrp_id], |
| 4960 | struct task_group, css); |
| 4961 | tg = autogroup_task_group(p, tg); |
| 4962 | p->sched_task_group = tg; |
| 4963 | } |
| 4964 | #endif |
| 4965 | /* |
| 4966 | * We're setting the CPU for the first time, we don't migrate, |
| 4967 | * so use __set_task_cpu(). |
| 4968 | */ |
| 4969 | __set_task_cpu(p, smp_processor_id()); |
| 4970 | if (p->sched_class->task_fork) |
| 4971 | p->sched_class->task_fork(p); |
| 4972 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
| 4973 | |
| 4974 | return scx_fork(p, kargs); |
| 4975 | } |
| 4976 | |
| 4977 | void sched_cancel_fork(struct task_struct *p) |
| 4978 | { |
| 4979 | scx_cancel_fork(p); |
| 4980 | } |
| 4981 | |
| 4982 | static void sched_mm_cid_fork(struct task_struct *t); |
| 4983 | |
| 4984 | void sched_post_fork(struct task_struct *p) |
| 4985 | { |
| 4986 | sched_mm_cid_fork(p); |
| 4987 | uclamp_post_fork(p); |
| 4988 | scx_post_fork(p); |
| 4989 | } |
| 4990 | |
| 4991 | u64 to_ratio(u64 period, u64 runtime) |
| 4992 | { |
| 4993 | if (runtime == RUNTIME_INF) |
| 4994 | return BW_UNIT; |
| 4995 | |
| 4996 | /* |
| 4997 | * Doing this here saves a lot of checks in all |
| 4998 | * the calling paths, and returning zero seems |
| 4999 | * safe for them anyway. |
| 5000 | */ |
| 5001 | if (period == 0) |
| 5002 | return 0; |
| 5003 | |
| 5004 | return div64_u64(runtime << BW_SHIFT, period); |
| 5005 | } |
| 5006 | |
| 5007 | /* |
| 5008 | * wake_up_new_task - wake up a newly created task for the first time. |
| 5009 | * |
| 5010 | * This function will do some initial scheduler statistics housekeeping |
| 5011 | * that must be done for every newly created context, then puts the task |
| 5012 | * on the runqueue and wakes it. |
| 5013 | */ |
| 5014 | void wake_up_new_task(struct task_struct *p) |
| 5015 | { |
| 5016 | struct rq_flags rf; |
| 5017 | struct rq *rq; |
| 5018 | int wake_flags = WF_FORK; |
| 5019 | |
| 5020 | raw_spin_lock_irqsave(&p->pi_lock, rf.flags); |
| 5021 | WRITE_ONCE(p->__state, TASK_RUNNING); |
| 5022 | /* |
| 5023 | * Fork balancing, do it here and not earlier because: |
| 5024 | * - cpus_ptr can change in the fork path |
| 5025 | * - any previously selected CPU might disappear through hotplug |
| 5026 | * |
| 5027 | * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq, |
| 5028 | * as we're not fully set-up yet. |
| 5029 | */ |
| 5030 | p->recent_used_cpu = task_cpu(p); |
| 5031 | __set_task_cpu(p, select_task_rq(p, task_cpu(p), &wake_flags)); |
| 5032 | rq = __task_rq_lock(p, &rf); |
| 5033 | update_rq_clock(rq); |
| 5034 | post_init_entity_util_avg(p); |
| 5035 | |
| 5036 | activate_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_INITIAL); |
| 5037 | trace_sched_wakeup_new(p); |
| 5038 | wakeup_preempt(rq, p, wake_flags); |
| 5039 | if (p->sched_class->task_woken) { |
| 5040 | /* |
| 5041 | * Nothing relies on rq->lock after this, so it's fine to |
| 5042 | * drop it. |
| 5043 | */ |
| 5044 | rq_unpin_lock(rq, &rf); |
| 5045 | p->sched_class->task_woken(rq, p); |
| 5046 | rq_repin_lock(rq, &rf); |
| 5047 | } |
| 5048 | task_rq_unlock(rq, p, &rf); |
| 5049 | } |
| 5050 | |
| 5051 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
| 5052 | |
| 5053 | static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key); |
| 5054 | |
| 5055 | void preempt_notifier_inc(void) |
| 5056 | { |
| 5057 | static_branch_inc(&preempt_notifier_key); |
| 5058 | } |
| 5059 | EXPORT_SYMBOL_GPL(preempt_notifier_inc); |
| 5060 | |
| 5061 | void preempt_notifier_dec(void) |
| 5062 | { |
| 5063 | static_branch_dec(&preempt_notifier_key); |
| 5064 | } |
| 5065 | EXPORT_SYMBOL_GPL(preempt_notifier_dec); |
| 5066 | |
| 5067 | /** |
| 5068 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
| 5069 | * @notifier: notifier struct to register |
| 5070 | */ |
| 5071 | void preempt_notifier_register(struct preempt_notifier *notifier) |
| 5072 | { |
| 5073 | if (!static_branch_unlikely(&preempt_notifier_key)) |
| 5074 | WARN(1, "registering preempt_notifier while notifiers disabled\n"); |
| 5075 | |
| 5076 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
| 5077 | } |
| 5078 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
| 5079 | |
| 5080 | /** |
| 5081 | * preempt_notifier_unregister - no longer interested in preemption notifications |
| 5082 | * @notifier: notifier struct to unregister |
| 5083 | * |
| 5084 | * This is *not* safe to call from within a preemption notifier. |
| 5085 | */ |
| 5086 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
| 5087 | { |
| 5088 | hlist_del(¬ifier->link); |
| 5089 | } |
| 5090 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
| 5091 | |
| 5092 | static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 5093 | { |
| 5094 | struct preempt_notifier *notifier; |
| 5095 | |
| 5096 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
| 5097 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
| 5098 | } |
| 5099 | |
| 5100 | static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 5101 | { |
| 5102 | if (static_branch_unlikely(&preempt_notifier_key)) |
| 5103 | __fire_sched_in_preempt_notifiers(curr); |
| 5104 | } |
| 5105 | |
| 5106 | static void |
| 5107 | __fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 5108 | struct task_struct *next) |
| 5109 | { |
| 5110 | struct preempt_notifier *notifier; |
| 5111 | |
| 5112 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
| 5113 | notifier->ops->sched_out(notifier, next); |
| 5114 | } |
| 5115 | |
| 5116 | static __always_inline void |
| 5117 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 5118 | struct task_struct *next) |
| 5119 | { |
| 5120 | if (static_branch_unlikely(&preempt_notifier_key)) |
| 5121 | __fire_sched_out_preempt_notifiers(curr, next); |
| 5122 | } |
| 5123 | |
| 5124 | #else /* !CONFIG_PREEMPT_NOTIFIERS: */ |
| 5125 | |
| 5126 | static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
| 5127 | { |
| 5128 | } |
| 5129 | |
| 5130 | static inline void |
| 5131 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
| 5132 | struct task_struct *next) |
| 5133 | { |
| 5134 | } |
| 5135 | |
| 5136 | #endif /* !CONFIG_PREEMPT_NOTIFIERS */ |
| 5137 | |
| 5138 | static inline void prepare_task(struct task_struct *next) |
| 5139 | { |
| 5140 | /* |
| 5141 | * Claim the task as running, we do this before switching to it |
| 5142 | * such that any running task will have this set. |
| 5143 | * |
| 5144 | * See the smp_load_acquire(&p->on_cpu) case in ttwu() and |
| 5145 | * its ordering comment. |
| 5146 | */ |
| 5147 | WRITE_ONCE(next->on_cpu, 1); |
| 5148 | } |
| 5149 | |
| 5150 | static inline void finish_task(struct task_struct *prev) |
| 5151 | { |
| 5152 | /* |
| 5153 | * This must be the very last reference to @prev from this CPU. After |
| 5154 | * p->on_cpu is cleared, the task can be moved to a different CPU. We |
| 5155 | * must ensure this doesn't happen until the switch is completely |
| 5156 | * finished. |
| 5157 | * |
| 5158 | * In particular, the load of prev->state in finish_task_switch() must |
| 5159 | * happen before this. |
| 5160 | * |
| 5161 | * Pairs with the smp_cond_load_acquire() in try_to_wake_up(). |
| 5162 | */ |
| 5163 | smp_store_release(&prev->on_cpu, 0); |
| 5164 | } |
| 5165 | |
| 5166 | /* |
| 5167 | * Only called from __schedule context |
| 5168 | * |
| 5169 | * There are some cases where we are going to re-do the action |
| 5170 | * that added the balance callbacks. We may not be in a state |
| 5171 | * where we can run them, so just zap them so they can be |
| 5172 | * properly re-added on the next time around. This is similar |
| 5173 | * handling to running the callbacks, except we just don't call |
| 5174 | * them. |
| 5175 | */ |
| 5176 | static void zap_balance_callbacks(struct rq *rq) |
| 5177 | { |
| 5178 | struct balance_callback *next, *head; |
| 5179 | bool found = false; |
| 5180 | |
| 5181 | lockdep_assert_rq_held(rq); |
| 5182 | |
| 5183 | head = rq->balance_callback; |
| 5184 | while (head) { |
| 5185 | if (head == &balance_push_callback) |
| 5186 | found = true; |
| 5187 | next = head->next; |
| 5188 | head->next = NULL; |
| 5189 | head = next; |
| 5190 | } |
| 5191 | rq->balance_callback = found ? &balance_push_callback : NULL; |
| 5192 | } |
| 5193 | |
| 5194 | static void do_balance_callbacks(struct rq *rq, struct balance_callback *head) |
| 5195 | { |
| 5196 | void (*func)(struct rq *rq); |
| 5197 | struct balance_callback *next; |
| 5198 | |
| 5199 | lockdep_assert_rq_held(rq); |
| 5200 | |
| 5201 | while (head) { |
| 5202 | func = head->func; |
| 5203 | next = head->next; |
| 5204 | head->next = NULL; |
| 5205 | head = next; |
| 5206 | |
| 5207 | func(rq); |
| 5208 | } |
| 5209 | } |
| 5210 | |
| 5211 | static void balance_push(struct rq *rq); |
| 5212 | |
| 5213 | /* |
| 5214 | * balance_push_callback is a right abuse of the callback interface and plays |
| 5215 | * by significantly different rules. |
| 5216 | * |
| 5217 | * Where the normal balance_callback's purpose is to be ran in the same context |
| 5218 | * that queued it (only later, when it's safe to drop rq->lock again), |
| 5219 | * balance_push_callback is specifically targeted at __schedule(). |
| 5220 | * |
| 5221 | * This abuse is tolerated because it places all the unlikely/odd cases behind |
| 5222 | * a single test, namely: rq->balance_callback == NULL. |
| 5223 | */ |
| 5224 | struct balance_callback balance_push_callback = { |
| 5225 | .next = NULL, |
| 5226 | .func = balance_push, |
| 5227 | }; |
| 5228 | |
| 5229 | static inline struct balance_callback * |
| 5230 | __splice_balance_callbacks(struct rq *rq, bool split) |
| 5231 | { |
| 5232 | struct balance_callback *head = rq->balance_callback; |
| 5233 | |
| 5234 | if (likely(!head)) |
| 5235 | return NULL; |
| 5236 | |
| 5237 | lockdep_assert_rq_held(rq); |
| 5238 | /* |
| 5239 | * Must not take balance_push_callback off the list when |
| 5240 | * splice_balance_callbacks() and balance_callbacks() are not |
| 5241 | * in the same rq->lock section. |
| 5242 | * |
| 5243 | * In that case it would be possible for __schedule() to interleave |
| 5244 | * and observe the list empty. |
| 5245 | */ |
| 5246 | if (split && head == &balance_push_callback) |
| 5247 | head = NULL; |
| 5248 | else |
| 5249 | rq->balance_callback = NULL; |
| 5250 | |
| 5251 | return head; |
| 5252 | } |
| 5253 | |
| 5254 | struct balance_callback *splice_balance_callbacks(struct rq *rq) |
| 5255 | { |
| 5256 | return __splice_balance_callbacks(rq, true); |
| 5257 | } |
| 5258 | |
| 5259 | void __balance_callbacks(struct rq *rq, struct rq_flags *rf) |
| 5260 | { |
| 5261 | if (rf) |
| 5262 | rq_unpin_lock(rq, rf); |
| 5263 | do_balance_callbacks(rq, __splice_balance_callbacks(rq, false)); |
| 5264 | if (rf) |
| 5265 | rq_repin_lock(rq, rf); |
| 5266 | } |
| 5267 | |
| 5268 | void balance_callbacks(struct rq *rq, struct balance_callback *head) |
| 5269 | { |
| 5270 | unsigned long flags; |
| 5271 | |
| 5272 | if (unlikely(head)) { |
| 5273 | raw_spin_rq_lock_irqsave(rq, flags); |
| 5274 | do_balance_callbacks(rq, head); |
| 5275 | raw_spin_rq_unlock_irqrestore(rq, flags); |
| 5276 | } |
| 5277 | } |
| 5278 | |
| 5279 | static inline void |
| 5280 | prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf) |
| 5281 | __releases(__rq_lockp(rq)) |
| 5282 | __acquires(__rq_lockp(this_rq())) |
| 5283 | { |
| 5284 | /* |
| 5285 | * Since the runqueue lock will be released by the next |
| 5286 | * task (which is an invalid locking op but in the case |
| 5287 | * of the scheduler it's an obvious special-case), so we |
| 5288 | * do an early lockdep release here: |
| 5289 | */ |
| 5290 | rq_unpin_lock(rq, rf); |
| 5291 | spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_); |
| 5292 | #ifdef CONFIG_DEBUG_SPINLOCK |
| 5293 | /* this is a valid case when another task releases the spinlock */ |
| 5294 | rq_lockp(rq)->owner = next; |
| 5295 | #endif |
| 5296 | /* |
| 5297 | * Model the rq reference switcheroo. |
| 5298 | */ |
| 5299 | __release(__rq_lockp(rq)); |
| 5300 | __acquire(__rq_lockp(this_rq())); |
| 5301 | } |
| 5302 | |
| 5303 | static inline void finish_lock_switch(struct rq *rq) |
| 5304 | __releases(__rq_lockp(rq)) |
| 5305 | { |
| 5306 | /* |
| 5307 | * If we are tracking spinlock dependencies then we have to |
| 5308 | * fix up the runqueue lock - which gets 'carried over' from |
| 5309 | * prev into current: |
| 5310 | */ |
| 5311 | spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_); |
| 5312 | __balance_callbacks(rq, NULL); |
| 5313 | hrtick_schedule_exit(rq); |
| 5314 | raw_spin_rq_unlock_irq(rq); |
| 5315 | } |
| 5316 | |
| 5317 | /* |
| 5318 | * NOP if the arch has not defined these: |
| 5319 | */ |
| 5320 | |
| 5321 | #ifndef prepare_arch_switch |
| 5322 | # define prepare_arch_switch(next) do { } while (0) |
| 5323 | #endif |
| 5324 | |
| 5325 | #ifndef finish_arch_post_lock_switch |
| 5326 | # define finish_arch_post_lock_switch() do { } while (0) |
| 5327 | #endif |
| 5328 | |
| 5329 | static inline void kmap_local_sched_out(void) |
| 5330 | { |
| 5331 | #ifdef CONFIG_KMAP_LOCAL |
| 5332 | if (unlikely(current->kmap_ctrl.idx)) |
| 5333 | __kmap_local_sched_out(); |
| 5334 | #endif |
| 5335 | } |
| 5336 | |
| 5337 | static inline void kmap_local_sched_in(void) |
| 5338 | { |
| 5339 | #ifdef CONFIG_KMAP_LOCAL |
| 5340 | if (unlikely(current->kmap_ctrl.idx)) |
| 5341 | __kmap_local_sched_in(); |
| 5342 | #endif |
| 5343 | } |
| 5344 | |
| 5345 | /** |
| 5346 | * prepare_task_switch - prepare to switch tasks |
| 5347 | * @rq: the runqueue preparing to switch |
| 5348 | * @prev: the current task that is being switched out |
| 5349 | * @next: the task we are going to switch to. |
| 5350 | * |
| 5351 | * This is called with the rq lock held and interrupts off. It must |
| 5352 | * be paired with a subsequent finish_task_switch after the context |
| 5353 | * switch. |
| 5354 | * |
| 5355 | * prepare_task_switch sets up locking and calls architecture specific |
| 5356 | * hooks. |
| 5357 | */ |
| 5358 | static inline void |
| 5359 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
| 5360 | struct task_struct *next) |
| 5361 | __must_hold(__rq_lockp(rq)) |
| 5362 | { |
| 5363 | kcov_prepare_switch(prev); |
| 5364 | sched_info_switch(rq, prev, next); |
| 5365 | perf_event_task_sched_out(prev, next); |
| 5366 | fire_sched_out_preempt_notifiers(prev, next); |
| 5367 | kmap_local_sched_out(); |
| 5368 | prepare_task(next); |
| 5369 | prepare_arch_switch(next); |
| 5370 | } |
| 5371 | |
| 5372 | /** |
| 5373 | * finish_task_switch - clean up after a task-switch |
| 5374 | * @prev: the thread we just switched away from. |
| 5375 | * |
| 5376 | * finish_task_switch must be called after the context switch, paired |
| 5377 | * with a prepare_task_switch call before the context switch. |
| 5378 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
| 5379 | * and do any other architecture-specific cleanup actions. |
| 5380 | * |
| 5381 | * Note that we may have delayed dropping an mm in context_switch(). If |
| 5382 | * so, we finish that here outside of the runqueue lock. (Doing it |
| 5383 | * with the lock held can cause deadlocks; see schedule() for |
| 5384 | * details.) |
| 5385 | * |
| 5386 | * The context switch have flipped the stack from under us and restored the |
| 5387 | * local variables which were saved when this task called schedule() in the |
| 5388 | * past. 'prev == current' is still correct but we need to recalculate this_rq |
| 5389 | * because prev may have moved to another CPU. |
| 5390 | */ |
| 5391 | static struct rq *finish_task_switch(struct task_struct *prev) |
| 5392 | __releases(__rq_lockp(this_rq())) |
| 5393 | { |
| 5394 | struct rq *rq = this_rq(); |
| 5395 | struct mm_struct *mm = rq->prev_mm; |
| 5396 | unsigned int prev_state; |
| 5397 | |
| 5398 | /* |
| 5399 | * The previous task will have left us with a preempt_count of 2 |
| 5400 | * because it left us after: |
| 5401 | * |
| 5402 | * schedule() |
| 5403 | * preempt_disable(); // 1 |
| 5404 | * __schedule() |
| 5405 | * raw_spin_lock_irq(&rq->lock) // 2 |
| 5406 | * |
| 5407 | * Also, see FORK_PREEMPT_COUNT. |
| 5408 | */ |
| 5409 | if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET, |
| 5410 | "corrupted preempt_count: %s/%d/0x%x\n", |
| 5411 | current->comm, current->pid, preempt_count())) |
| 5412 | preempt_count_set(FORK_PREEMPT_COUNT); |
| 5413 | |
| 5414 | rq->prev_mm = NULL; |
| 5415 | |
| 5416 | /* |
| 5417 | * A task struct has one reference for the use as "current". |
| 5418 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
| 5419 | * schedule one last time. The schedule call will never return, and |
| 5420 | * the scheduled task must drop that reference. |
| 5421 | * |
| 5422 | * We must observe prev->state before clearing prev->on_cpu (in |
| 5423 | * finish_task), otherwise a concurrent wakeup can get prev |
| 5424 | * running on another CPU and we could rave with its RUNNING -> DEAD |
| 5425 | * transition, resulting in a double drop. |
| 5426 | */ |
| 5427 | prev_state = READ_ONCE(prev->__state); |
| 5428 | vtime_task_switch(prev); |
| 5429 | perf_event_task_sched_in(prev, current); |
| 5430 | finish_task(prev); |
| 5431 | tick_nohz_task_switch(); |
| 5432 | finish_lock_switch(rq); |
| 5433 | finish_arch_post_lock_switch(); |
| 5434 | kcov_finish_switch(current); |
| 5435 | /* |
| 5436 | * kmap_local_sched_out() is invoked with rq::lock held and |
| 5437 | * interrupts disabled. There is no requirement for that, but the |
| 5438 | * sched out code does not have an interrupt enabled section. |
| 5439 | * Restoring the maps on sched in does not require interrupts being |
| 5440 | * disabled either. |
| 5441 | */ |
| 5442 | kmap_local_sched_in(); |
| 5443 | |
| 5444 | /* |
| 5445 | * Any cached block-layer timestamp (plug->cur_ktime) is stale now, |
| 5446 | * invalidate it. |
| 5447 | */ |
| 5448 | blk_plug_invalidate_ts(); |
| 5449 | |
| 5450 | fire_sched_in_preempt_notifiers(current); |
| 5451 | /* |
| 5452 | * When switching through a kernel thread, the loop in |
| 5453 | * membarrier_{private,global}_expedited() may have observed that |
| 5454 | * kernel thread and not issued an IPI. It is therefore possible to |
| 5455 | * schedule between user->kernel->user threads without passing though |
| 5456 | * switch_mm(). Membarrier requires a barrier after storing to |
| 5457 | * rq->curr, before returning to userspace, so provide them here: |
| 5458 | * |
| 5459 | * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly |
| 5460 | * provided by mmdrop_lazy_tlb(), |
| 5461 | * - a sync_core for SYNC_CORE. |
| 5462 | */ |
| 5463 | if (mm) { |
| 5464 | membarrier_mm_sync_core_before_usermode(mm); |
| 5465 | mmdrop_lazy_tlb_sched(mm); |
| 5466 | } |
| 5467 | |
| 5468 | if (unlikely(prev_state == TASK_DEAD)) { |
| 5469 | if (prev->sched_class->task_dead) |
| 5470 | prev->sched_class->task_dead(prev); |
| 5471 | |
| 5472 | /* |
| 5473 | * sched_ext_dead() must come before cgroup_task_dead() to |
| 5474 | * prevent cgroups from being removed while its member tasks are |
| 5475 | * visible to SCX schedulers. |
| 5476 | */ |
| 5477 | sched_ext_dead(prev); |
| 5478 | cgroup_task_dead(prev); |
| 5479 | |
| 5480 | /* Task is done with its stack. */ |
| 5481 | put_task_stack(prev); |
| 5482 | |
| 5483 | put_task_struct_rcu_user(prev); |
| 5484 | } |
| 5485 | |
| 5486 | return rq; |
| 5487 | } |
| 5488 | |
| 5489 | /** |
| 5490 | * schedule_tail - first thing a freshly forked thread must call. |
| 5491 | * @prev: the thread we just switched away from. |
| 5492 | */ |
| 5493 | asmlinkage __visible void schedule_tail(struct task_struct *prev) |
| 5494 | __releases(__rq_lockp(this_rq())) |
| 5495 | { |
| 5496 | /* |
| 5497 | * New tasks start with FORK_PREEMPT_COUNT, see there and |
| 5498 | * finish_task_switch() for details. |
| 5499 | * |
| 5500 | * finish_task_switch() will drop rq->lock() and lower preempt_count |
| 5501 | * and the preempt_enable() will end up enabling preemption (on |
| 5502 | * PREEMPT_COUNT kernels). |
| 5503 | */ |
| 5504 | |
| 5505 | finish_task_switch(prev); |
| 5506 | /* |
| 5507 | * This is a special case: the newly created task has just |
| 5508 | * switched the context for the first time. It is returning from |
| 5509 | * schedule for the first time in this path. |
| 5510 | */ |
| 5511 | trace_sched_exit_tp(true); |
| 5512 | preempt_enable(); |
| 5513 | |
| 5514 | if (current->set_child_tid) |
| 5515 | put_user(task_pid_vnr(current), current->set_child_tid); |
| 5516 | |
| 5517 | calculate_sigpending(); |
| 5518 | } |
| 5519 | |
| 5520 | /* |
| 5521 | * context_switch - switch to the new MM and the new thread's register state. |
| 5522 | */ |
| 5523 | static __always_inline struct rq * |
| 5524 | context_switch(struct rq *rq, struct task_struct *prev, |
| 5525 | struct task_struct *next, struct rq_flags *rf) |
| 5526 | __releases(__rq_lockp(rq)) |
| 5527 | { |
| 5528 | prepare_task_switch(rq, prev, next); |
| 5529 | |
| 5530 | /* |
| 5531 | * For paravirt, this is coupled with an exit in switch_to to |
| 5532 | * combine the page table reload and the switch backend into |
| 5533 | * one hypercall. |
| 5534 | */ |
| 5535 | arch_start_context_switch(prev); |
| 5536 | |
| 5537 | /* |
| 5538 | * kernel -> kernel lazy + transfer active |
| 5539 | * user -> kernel lazy + mmgrab_lazy_tlb() active |
| 5540 | * |
| 5541 | * kernel -> user switch + mmdrop_lazy_tlb() active |
| 5542 | * user -> user switch |
| 5543 | */ |
| 5544 | if (!next->mm) { // to kernel |
| 5545 | enter_lazy_tlb(prev->active_mm, next); |
| 5546 | |
| 5547 | next->active_mm = prev->active_mm; |
| 5548 | if (prev->mm) // from user |
| 5549 | mmgrab_lazy_tlb(prev->active_mm); |
| 5550 | else |
| 5551 | prev->active_mm = NULL; |
| 5552 | } else { // to user |
| 5553 | membarrier_switch_mm(rq, prev->active_mm, next->mm); |
| 5554 | /* |
| 5555 | * sys_membarrier() requires an smp_mb() between setting |
| 5556 | * rq->curr / membarrier_switch_mm() and returning to userspace. |
| 5557 | * |
| 5558 | * The below provides this either through switch_mm(), or in |
| 5559 | * case 'prev->active_mm == next->mm' through |
| 5560 | * finish_task_switch()'s mmdrop(). |
| 5561 | */ |
| 5562 | switch_mm_irqs_off(prev->active_mm, next->mm, next); |
| 5563 | lru_gen_use_mm(next->mm); |
| 5564 | |
| 5565 | if (!prev->mm) { // from kernel |
| 5566 | /* will mmdrop_lazy_tlb() in finish_task_switch(). */ |
| 5567 | rq->prev_mm = prev->active_mm; |
| 5568 | prev->active_mm = NULL; |
| 5569 | } |
| 5570 | } |
| 5571 | |
| 5572 | mm_cid_switch_to(prev, next); |
| 5573 | |
| 5574 | /* |
| 5575 | * Tell rseq that the task was scheduled in. Must be after |
| 5576 | * switch_mm_cid() to get the TIF flag set. |
| 5577 | */ |
| 5578 | rseq_sched_switch_event(next); |
| 5579 | |
| 5580 | prepare_lock_switch(rq, next, rf); |
| 5581 | |
| 5582 | /* Here we just switch the register state and the stack. */ |
| 5583 | switch_to(prev, next, prev); |
| 5584 | barrier(); |
| 5585 | |
| 5586 | return finish_task_switch(prev); |
| 5587 | } |
| 5588 | |
| 5589 | /* |
| 5590 | * nr_running and nr_context_switches: |
| 5591 | * |
| 5592 | * externally visible scheduler statistics: current number of runnable |
| 5593 | * threads, total number of context switches performed since bootup. |
| 5594 | */ |
| 5595 | unsigned int nr_running(void) |
| 5596 | { |
| 5597 | unsigned int i, sum = 0; |
| 5598 | |
| 5599 | for_each_online_cpu(i) |
| 5600 | sum += cpu_rq(i)->nr_running; |
| 5601 | |
| 5602 | return sum; |
| 5603 | } |
| 5604 | |
| 5605 | /* |
| 5606 | * Check if only the current task is running on the CPU. |
| 5607 | * |
| 5608 | * Caution: this function does not check that the caller has disabled |
| 5609 | * preemption, thus the result might have a time-of-check-to-time-of-use |
| 5610 | * race. The caller is responsible to use it correctly, for example: |
| 5611 | * |
| 5612 | * - from a non-preemptible section (of course) |
| 5613 | * |
| 5614 | * - from a thread that is bound to a single CPU |
| 5615 | * |
| 5616 | * - in a loop with very short iterations (e.g. a polling loop) |
| 5617 | */ |
| 5618 | bool single_task_running(void) |
| 5619 | { |
| 5620 | return raw_rq()->nr_running == 1; |
| 5621 | } |
| 5622 | EXPORT_SYMBOL(single_task_running); |
| 5623 | |
| 5624 | unsigned long long nr_context_switches_cpu(int cpu) |
| 5625 | { |
| 5626 | return cpu_rq(cpu)->nr_switches; |
| 5627 | } |
| 5628 | |
| 5629 | unsigned long long nr_context_switches(void) |
| 5630 | { |
| 5631 | int i; |
| 5632 | unsigned long long sum = 0; |
| 5633 | |
| 5634 | for_each_possible_cpu(i) |
| 5635 | sum += cpu_rq(i)->nr_switches; |
| 5636 | |
| 5637 | return sum; |
| 5638 | } |
| 5639 | |
| 5640 | /* |
| 5641 | * Return whether any tasks are accounted as I/O waiters on this CPU. |
| 5642 | * They may wake up on a different CPU. |
| 5643 | */ |
| 5644 | bool cpu_has_iowait(int cpu) |
| 5645 | { |
| 5646 | return atomic_read(&cpu_rq(cpu)->nr_iowait) != 0; |
| 5647 | } |
| 5648 | |
| 5649 | /* |
| 5650 | * IO-wait accounting, and how it's mostly bollocks (on SMP). |
| 5651 | * |
| 5652 | * The idea behind IO-wait account is to account the idle time that we could |
| 5653 | * have spend running if it were not for IO. That is, if we were to improve the |
| 5654 | * storage performance, we'd have a proportional reduction in IO-wait time. |
| 5655 | * |
| 5656 | * This all works nicely on UP, where, when a task blocks on IO, we account |
| 5657 | * idle time as IO-wait, because if the storage were faster, it could've been |
| 5658 | * running and we'd not be idle. |
| 5659 | * |
| 5660 | * This has been extended to SMP, by doing the same for each CPU. This however |
| 5661 | * is broken. |
| 5662 | * |
| 5663 | * Imagine for instance the case where two tasks block on one CPU, only the one |
| 5664 | * CPU will have IO-wait accounted, while the other has regular idle. Even |
| 5665 | * though, if the storage were faster, both could've ran at the same time, |
| 5666 | * utilising both CPUs. |
| 5667 | * |
| 5668 | * This means, that when looking globally, the current IO-wait accounting on |
| 5669 | * SMP is a lower bound, by reason of under accounting. |
| 5670 | * |
| 5671 | * Worse, since the numbers are provided per CPU, they are sometimes |
| 5672 | * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly |
| 5673 | * associated with any one particular CPU, it can wake to another CPU than it |
| 5674 | * blocked on. This means the per CPU IO-wait number is meaningless. |
| 5675 | * |
| 5676 | * Task CPU affinities can make all that even more 'interesting'. |
| 5677 | */ |
| 5678 | |
| 5679 | unsigned int nr_iowait(void) |
| 5680 | { |
| 5681 | unsigned int i, sum = 0; |
| 5682 | |
| 5683 | for_each_possible_cpu(i) |
| 5684 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
| 5685 | |
| 5686 | return sum; |
| 5687 | } |
| 5688 | |
| 5689 | /* |
| 5690 | * sched_exec - execve() is a valuable balancing opportunity, because at |
| 5691 | * this point the task has the smallest effective memory and cache footprint. |
| 5692 | */ |
| 5693 | void sched_exec(void) |
| 5694 | { |
| 5695 | struct task_struct *p = current; |
| 5696 | struct migration_arg arg; |
| 5697 | int dest_cpu; |
| 5698 | |
| 5699 | scoped_guard (raw_spinlock_irqsave, &p->pi_lock) { |
| 5700 | dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC); |
| 5701 | if (dest_cpu == smp_processor_id()) |
| 5702 | return; |
| 5703 | |
| 5704 | if (unlikely(!cpu_active(dest_cpu))) |
| 5705 | return; |
| 5706 | |
| 5707 | arg = (struct migration_arg){ p, dest_cpu }; |
| 5708 | } |
| 5709 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); |
| 5710 | } |
| 5711 | |
| 5712 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
| 5713 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat) = { |
| 5714 | #ifdef CONFIG_NO_HZ_COMMON |
| 5715 | .idle_sleeptime_seq = SEQCNT_ZERO(kernel_cpustat.idle_sleeptime_seq) |
| 5716 | #endif |
| 5717 | }; |
| 5718 | |
| 5719 | EXPORT_PER_CPU_SYMBOL(kstat); |
| 5720 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); |
| 5721 | |
| 5722 | /* |
| 5723 | * The function fair_sched_class.update_curr accesses the struct curr |
| 5724 | * and its field curr->exec_start; when called from task_sched_runtime(), |
| 5725 | * we observe a high rate of cache misses in practice. |
| 5726 | * Prefetching this data results in improved performance. |
| 5727 | */ |
| 5728 | static inline void prefetch_curr_exec_start(struct task_struct *p) |
| 5729 | { |
| 5730 | struct sched_entity *curr = task_rq(p)->cfs.curr; |
| 5731 | |
| 5732 | prefetch(curr); |
| 5733 | prefetch(&curr->exec_start); |
| 5734 | } |
| 5735 | |
| 5736 | /* |
| 5737 | * Return accounted runtime for the task. |
| 5738 | * In case the task is currently running, return the runtime plus current's |
| 5739 | * pending runtime that have not been accounted yet. |
| 5740 | */ |
| 5741 | unsigned long long task_sched_runtime(struct task_struct *p) |
| 5742 | { |
| 5743 | struct rq_flags rf; |
| 5744 | struct rq *rq; |
| 5745 | u64 ns; |
| 5746 | |
| 5747 | #ifdef CONFIG_64BIT |
| 5748 | /* |
| 5749 | * 64-bit doesn't need locks to atomically read a 64-bit value. |
| 5750 | * So we have a optimization chance when the task's delta_exec is 0. |
| 5751 | * Reading ->on_cpu is racy, but this is OK. |
| 5752 | * |
| 5753 | * If we race with it leaving CPU, we'll take a lock. So we're correct. |
| 5754 | * If we race with it entering CPU, unaccounted time is 0. This is |
| 5755 | * indistinguishable from the read occurring a few cycles earlier. |
| 5756 | * If we see ->on_cpu without ->on_rq, the task is leaving, and has |
| 5757 | * been accounted, so we're correct here as well. |
| 5758 | */ |
| 5759 | if (!p->on_cpu || !task_on_rq_queued(p)) |
| 5760 | return p->se.sum_exec_runtime; |
| 5761 | #endif |
| 5762 | |
| 5763 | rq = task_rq_lock(p, &rf); |
| 5764 | /* |
| 5765 | * Must be ->curr _and_ ->on_rq. If dequeued, we would |
| 5766 | * project cycles that may never be accounted to this |
| 5767 | * thread, breaking clock_gettime(). |
| 5768 | */ |
| 5769 | if (task_current_donor(rq, p) && task_on_rq_queued(p)) { |
| 5770 | prefetch_curr_exec_start(p); |
| 5771 | update_rq_clock(rq); |
| 5772 | p->sched_class->update_curr(rq); |
| 5773 | } |
| 5774 | ns = p->se.sum_exec_runtime; |
| 5775 | task_rq_unlock(rq, p, &rf); |
| 5776 | |
| 5777 | return ns; |
| 5778 | } |
| 5779 | |
| 5780 | static u64 cpu_resched_latency(struct rq *rq) |
| 5781 | { |
| 5782 | int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms); |
| 5783 | u64 resched_latency, now = rq_clock(rq); |
| 5784 | static bool warned_once; |
| 5785 | |
| 5786 | if (sysctl_resched_latency_warn_once && warned_once) |
| 5787 | return 0; |
| 5788 | |
| 5789 | if (!need_resched() || !latency_warn_ms) |
| 5790 | return 0; |
| 5791 | |
| 5792 | if (system_state == SYSTEM_BOOTING) |
| 5793 | return 0; |
| 5794 | |
| 5795 | if (!rq->last_seen_need_resched_ns) { |
| 5796 | rq->last_seen_need_resched_ns = now; |
| 5797 | rq->ticks_without_resched = 0; |
| 5798 | return 0; |
| 5799 | } |
| 5800 | |
| 5801 | rq->ticks_without_resched++; |
| 5802 | resched_latency = now - rq->last_seen_need_resched_ns; |
| 5803 | if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC) |
| 5804 | return 0; |
| 5805 | |
| 5806 | warned_once = true; |
| 5807 | |
| 5808 | return resched_latency; |
| 5809 | } |
| 5810 | |
| 5811 | static int __init setup_resched_latency_warn_ms(char *str) |
| 5812 | { |
| 5813 | long val; |
| 5814 | |
| 5815 | if ((kstrtol(str, 0, &val))) { |
| 5816 | pr_warn("Unable to set resched_latency_warn_ms\n"); |
| 5817 | return 1; |
| 5818 | } |
| 5819 | |
| 5820 | sysctl_resched_latency_warn_ms = val; |
| 5821 | return 1; |
| 5822 | } |
| 5823 | __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms); |
| 5824 | |
| 5825 | /* |
| 5826 | * This function gets called by the timer code, with HZ frequency. |
| 5827 | * We call it with interrupts disabled. |
| 5828 | */ |
| 5829 | void sched_tick(void) |
| 5830 | { |
| 5831 | int cpu = smp_processor_id(); |
| 5832 | struct rq *rq = cpu_rq(cpu); |
| 5833 | /* scheduler accounting goes to the donor task */ |
| 5834 | struct task_struct *curr, *donor; |
| 5835 | struct rq_flags rf; |
| 5836 | unsigned long hw_pressure; |
| 5837 | u64 resched_latency; |
| 5838 | |
| 5839 | if (!cpu_preferred(cpu)) |
| 5840 | sched_push_current_non_preferred_cpu(rq); |
| 5841 | |
| 5842 | if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) |
| 5843 | arch_scale_freq_tick(); |
| 5844 | |
| 5845 | sched_clock_tick(); |
| 5846 | |
| 5847 | rq_lock(rq, &rf); |
| 5848 | curr = rq->curr; |
| 5849 | donor = rq->donor; |
| 5850 | |
| 5851 | psi_account_irqtime(rq, curr, NULL); |
| 5852 | |
| 5853 | update_rq_clock(rq); |
| 5854 | hw_pressure = arch_scale_hw_pressure(cpu_of(rq)); |
| 5855 | update_hw_load_avg(rq_clock_task(rq), rq, hw_pressure); |
| 5856 | |
| 5857 | if (dynamic_preempt_lazy() && tif_test_bit(TIF_NEED_RESCHED_LAZY)) |
| 5858 | resched_curr(rq); |
| 5859 | |
| 5860 | donor->sched_class->task_tick(rq, donor, 0); |
| 5861 | if (sched_feat(LATENCY_WARN)) |
| 5862 | resched_latency = cpu_resched_latency(rq); |
| 5863 | calc_global_load_tick(rq); |
| 5864 | sched_core_tick(rq); |
| 5865 | scx_tick(rq); |
| 5866 | |
| 5867 | rq_unlock(rq, &rf); |
| 5868 | |
| 5869 | if (sched_feat(LATENCY_WARN) && resched_latency) |
| 5870 | resched_latency_warn(cpu, resched_latency); |
| 5871 | |
| 5872 | perf_event_task_tick(); |
| 5873 | |
| 5874 | if (curr->flags & PF_WQ_WORKER) |
| 5875 | wq_worker_tick(curr); |
| 5876 | |
| 5877 | if (!scx_switched_all()) { |
| 5878 | rq->idle_balance = idle_cpu(cpu); |
| 5879 | sched_balance_trigger(rq); |
| 5880 | } |
| 5881 | } |
| 5882 | |
| 5883 | #ifdef CONFIG_NO_HZ_FULL |
| 5884 | |
| 5885 | struct tick_work { |
| 5886 | int cpu; |
| 5887 | atomic_t state; |
| 5888 | struct delayed_work work; |
| 5889 | }; |
| 5890 | /* Values for ->state, see diagram below. */ |
| 5891 | #define TICK_SCHED_REMOTE_OFFLINE 0 |
| 5892 | #define TICK_SCHED_REMOTE_OFFLINING 1 |
| 5893 | #define TICK_SCHED_REMOTE_RUNNING 2 |
| 5894 | |
| 5895 | /* |
| 5896 | * State diagram for ->state: |
| 5897 | * |
| 5898 | * |
| 5899 | * TICK_SCHED_REMOTE_OFFLINE |
| 5900 | * | ^ |
| 5901 | * | | |
| 5902 | * | | sched_tick_remote() |
| 5903 | * | | |
| 5904 | * | | |
| 5905 | * +--TICK_SCHED_REMOTE_OFFLINING |
| 5906 | * | ^ |
| 5907 | * | | |
| 5908 | * sched_tick_start() | | sched_tick_stop() |
| 5909 | * | | |
| 5910 | * V | |
| 5911 | * TICK_SCHED_REMOTE_RUNNING |
| 5912 | * |
| 5913 | * |
| 5914 | * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote() |
| 5915 | * and sched_tick_start() are happy to leave the state in RUNNING. |
| 5916 | */ |
| 5917 | |
| 5918 | static struct tick_work __percpu *tick_work_cpu; |
| 5919 | |
| 5920 | static void sched_tick_remote(struct work_struct *work) |
| 5921 | { |
| 5922 | struct delayed_work *dwork = to_delayed_work(work); |
| 5923 | struct tick_work *twork = container_of(dwork, struct tick_work, work); |
| 5924 | int cpu = twork->cpu; |
| 5925 | struct rq *rq = cpu_rq(cpu); |
| 5926 | int os; |
| 5927 | |
| 5928 | /* |
| 5929 | * Handle the tick only if it appears the remote CPU is running in full |
| 5930 | * dynticks mode. The check is racy by nature, but missing a tick or |
| 5931 | * having one too much is no big deal because the scheduler tick updates |
| 5932 | * statistics and checks timeslices in a time-independent way, regardless |
| 5933 | * of when exactly it is running. |
| 5934 | */ |
| 5935 | if (tick_nohz_tick_stopped_cpu(cpu)) { |
| 5936 | guard(rq_lock_irq)(rq); |
| 5937 | struct task_struct *curr = rq->curr; |
| 5938 | |
| 5939 | if (cpu_online(cpu)) { |
| 5940 | /* |
| 5941 | * Since this is a remote tick for full dynticks mode, |
| 5942 | * we are always sure that there is no proxy (only a |
| 5943 | * single task is running). |
| 5944 | */ |
| 5945 | WARN_ON_ONCE(rq->curr != rq->donor); |
| 5946 | update_rq_clock(rq); |
| 5947 | |
| 5948 | if (!is_idle_task(curr)) { |
| 5949 | /* |
| 5950 | * Make sure the next tick runs within a |
| 5951 | * reasonable amount of time. |
| 5952 | */ |
| 5953 | u64 delta = rq_clock_task(rq) - curr->se.exec_start; |
| 5954 | WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 30); |
| 5955 | } |
| 5956 | curr->sched_class->task_tick(rq, curr, 0); |
| 5957 | |
| 5958 | calc_load_nohz_remote(rq); |
| 5959 | } |
| 5960 | } |
| 5961 | |
| 5962 | /* |
| 5963 | * Run the remote tick once per second (1Hz). This arbitrary |
| 5964 | * frequency is large enough to avoid overload but short enough |
| 5965 | * to keep scheduler internal stats reasonably up to date. But |
| 5966 | * first update state to reflect hotplug activity if required. |
| 5967 | */ |
| 5968 | os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING); |
| 5969 | WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE); |
| 5970 | if (os == TICK_SCHED_REMOTE_RUNNING) |
| 5971 | queue_delayed_work(system_dfl_wq, dwork, HZ); |
| 5972 | } |
| 5973 | |
| 5974 | static void sched_tick_start(int cpu) |
| 5975 | { |
| 5976 | int os; |
| 5977 | struct tick_work *twork; |
| 5978 | |
| 5979 | if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) |
| 5980 | return; |
| 5981 | |
| 5982 | WARN_ON_ONCE(!tick_work_cpu); |
| 5983 | |
| 5984 | twork = per_cpu_ptr(tick_work_cpu, cpu); |
| 5985 | os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING); |
| 5986 | WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING); |
| 5987 | if (os == TICK_SCHED_REMOTE_OFFLINE) { |
| 5988 | twork->cpu = cpu; |
| 5989 | INIT_DELAYED_WORK(&twork->work, sched_tick_remote); |
| 5990 | queue_delayed_work(system_dfl_wq, &twork->work, HZ); |
| 5991 | } |
| 5992 | } |
| 5993 | |
| 5994 | #ifdef CONFIG_HOTPLUG_CPU |
| 5995 | static void sched_tick_stop(int cpu) |
| 5996 | { |
| 5997 | struct tick_work *twork; |
| 5998 | int os; |
| 5999 | |
| 6000 | if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) |
| 6001 | return; |
| 6002 | |
| 6003 | WARN_ON_ONCE(!tick_work_cpu); |
| 6004 | |
| 6005 | twork = per_cpu_ptr(tick_work_cpu, cpu); |
| 6006 | /* There cannot be competing actions, but don't rely on stop-machine. */ |
| 6007 | os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING); |
| 6008 | WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING); |
| 6009 | /* Don't cancel, as this would mess up the state machine. */ |
| 6010 | } |
| 6011 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 6012 | |
| 6013 | int __init sched_tick_offload_init(void) |
| 6014 | { |
| 6015 | tick_work_cpu = alloc_percpu(struct tick_work); |
| 6016 | BUG_ON(!tick_work_cpu); |
| 6017 | return 0; |
| 6018 | } |
| 6019 | |
| 6020 | #else /* !CONFIG_NO_HZ_FULL: */ |
| 6021 | static inline void sched_tick_start(int cpu) { } |
| 6022 | static inline void sched_tick_stop(int cpu) { } |
| 6023 | #endif /* !CONFIG_NO_HZ_FULL */ |
| 6024 | |
| 6025 | #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
| 6026 | defined(CONFIG_TRACE_PREEMPT_TOGGLE)) |
| 6027 | /* |
| 6028 | * If the value passed in is equal to the current preempt count |
| 6029 | * then we just disabled preemption. Start timing the latency. |
| 6030 | */ |
| 6031 | static inline void preempt_latency_start(int val) |
| 6032 | { |
| 6033 | if (preempt_count() == val) { |
| 6034 | unsigned long ip = get_lock_parent_ip(); |
| 6035 | #ifdef CONFIG_DEBUG_PREEMPT |
| 6036 | current->preempt_disable_ip = ip; |
| 6037 | #endif |
| 6038 | trace_preempt_off(CALLER_ADDR0, ip); |
| 6039 | } |
| 6040 | } |
| 6041 | |
| 6042 | void preempt_count_add(int val) |
| 6043 | { |
| 6044 | #ifdef CONFIG_DEBUG_PREEMPT |
| 6045 | /* |
| 6046 | * Underflow? |
| 6047 | * |
| 6048 | * Cannot detect underflow based on the current preempt_count() value |
| 6049 | * if using HAS_SEPARATE_PREEMPT_RESCHED_BITS because preempt count takes all 32 |
| 6050 | * bits. |
| 6051 | */ |
| 6052 | if (!IS_ENABLED(CONFIG_HAS_SEPARATE_PREEMPT_RESCHED_BITS) && |
| 6053 | DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
| 6054 | return; |
| 6055 | #endif |
| 6056 | __preempt_count_add(val); |
| 6057 | #ifdef CONFIG_DEBUG_PREEMPT |
| 6058 | /* |
| 6059 | * Spinlock count overflowing soon? |
| 6060 | */ |
| 6061 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
| 6062 | PREEMPT_MASK - 10); |
| 6063 | #endif |
| 6064 | preempt_latency_start(val); |
| 6065 | } |
| 6066 | EXPORT_SYMBOL(preempt_count_add); |
| 6067 | NOKPROBE_SYMBOL(preempt_count_add); |
| 6068 | |
| 6069 | /* |
| 6070 | * If the value passed in equals to the current preempt count |
| 6071 | * then we just enabled preemption. Stop timing the latency. |
| 6072 | */ |
| 6073 | static inline void preempt_latency_stop(int val) |
| 6074 | { |
| 6075 | if (preempt_count() == val) |
| 6076 | trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); |
| 6077 | } |
| 6078 | |
| 6079 | void preempt_count_sub(int val) |
| 6080 | { |
| 6081 | #ifdef CONFIG_DEBUG_PREEMPT |
| 6082 | /* |
| 6083 | * Underflow? |
| 6084 | */ |
| 6085 | unsigned int uval = val; |
| 6086 | unsigned int pc = preempt_count(); |
| 6087 | |
| 6088 | if (DEBUG_LOCKS_WARN_ON(pc - uval > pc)) |
| 6089 | return; |
| 6090 | /* |
| 6091 | * Is the spinlock portion underflowing? |
| 6092 | */ |
| 6093 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
| 6094 | !(preempt_count() & PREEMPT_MASK))) |
| 6095 | return; |
| 6096 | #endif |
| 6097 | |
| 6098 | preempt_latency_stop(val); |
| 6099 | __preempt_count_sub(val); |
| 6100 | } |
| 6101 | EXPORT_SYMBOL(preempt_count_sub); |
| 6102 | NOKPROBE_SYMBOL(preempt_count_sub); |
| 6103 | |
| 6104 | #else |
| 6105 | static inline void preempt_latency_start(int val) { } |
| 6106 | static inline void preempt_latency_stop(int val) { } |
| 6107 | #endif |
| 6108 | |
| 6109 | static inline unsigned long get_preempt_disable_ip(struct task_struct *p) |
| 6110 | { |
| 6111 | #ifdef CONFIG_DEBUG_PREEMPT |
| 6112 | return p->preempt_disable_ip; |
| 6113 | #else |
| 6114 | return 0; |
| 6115 | #endif |
| 6116 | } |
| 6117 | |
| 6118 | /* |
| 6119 | * Print scheduling while atomic bug: |
| 6120 | */ |
| 6121 | static noinline void __schedule_bug(struct task_struct *prev) |
| 6122 | { |
| 6123 | /* Save this before calling printk(), since that will clobber it */ |
| 6124 | unsigned long preempt_disable_ip = get_preempt_disable_ip(current); |
| 6125 | |
| 6126 | if (oops_in_progress) |
| 6127 | return; |
| 6128 | |
| 6129 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
| 6130 | prev->comm, prev->pid, preempt_count()); |
| 6131 | |
| 6132 | debug_show_held_locks(prev); |
| 6133 | print_modules(); |
| 6134 | if (irqs_disabled()) |
| 6135 | print_irqtrace_events(prev); |
| 6136 | if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) { |
| 6137 | pr_err("Preemption disabled at:"); |
| 6138 | print_ip_sym(KERN_ERR, preempt_disable_ip); |
| 6139 | } |
| 6140 | check_panic_on_warn("scheduling while atomic"); |
| 6141 | |
| 6142 | dump_stack(); |
| 6143 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
| 6144 | } |
| 6145 | |
| 6146 | /* |
| 6147 | * Various schedule()-time debugging checks and statistics: |
| 6148 | */ |
| 6149 | static inline void schedule_debug(struct task_struct *prev, bool preempt) |
| 6150 | { |
| 6151 | #ifdef CONFIG_SCHED_STACK_END_CHECK |
| 6152 | if (task_stack_end_corrupted(prev)) |
| 6153 | panic("corrupted stack end detected inside scheduler\n"); |
| 6154 | |
| 6155 | if (task_scs_end_corrupted(prev)) |
| 6156 | panic("corrupted shadow stack detected inside scheduler\n"); |
| 6157 | #endif |
| 6158 | |
| 6159 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP |
| 6160 | if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) { |
| 6161 | printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n", |
| 6162 | prev->comm, prev->pid, prev->non_block_count); |
| 6163 | dump_stack(); |
| 6164 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
| 6165 | } |
| 6166 | #endif |
| 6167 | |
| 6168 | if (unlikely(in_atomic_preempt_off())) { |
| 6169 | __schedule_bug(prev); |
| 6170 | preempt_count_set(PREEMPT_DISABLED); |
| 6171 | } |
| 6172 | rcu_sleep_check(); |
| 6173 | WARN_ON_ONCE(ct_state() == CT_STATE_USER); |
| 6174 | |
| 6175 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
| 6176 | |
| 6177 | schedstat_inc(this_rq()->sched_count); |
| 6178 | } |
| 6179 | |
| 6180 | static void prev_balance(struct rq *rq, struct rq_flags *rf) |
| 6181 | { |
| 6182 | const struct sched_class *start_class = rq->donor->sched_class; |
| 6183 | const struct sched_class *class; |
| 6184 | |
| 6185 | /* |
| 6186 | * We must do the balancing pass before put_prev_task(), such |
| 6187 | * that when we release the rq->lock the task is in the same |
| 6188 | * state as before we took rq->lock. |
| 6189 | * |
| 6190 | * We can terminate the balance pass as soon as we know there is |
| 6191 | * a runnable task of @class priority or higher. |
| 6192 | */ |
| 6193 | for_active_class_range(class, start_class, &idle_sched_class) { |
| 6194 | if (class->balance && class->balance(rq, rf)) |
| 6195 | break; |
| 6196 | } |
| 6197 | } |
| 6198 | |
| 6199 | /* |
| 6200 | * Pick up the highest-prio task: |
| 6201 | */ |
| 6202 | static inline struct task_struct * |
| 6203 | __pick_next_task(struct rq *rq, struct rq_flags *rf) |
| 6204 | __must_hold(__rq_lockp(rq)) |
| 6205 | { |
| 6206 | const struct sched_class *class; |
| 6207 | struct task_struct *p; |
| 6208 | |
| 6209 | rq->dl_server = NULL; |
| 6210 | |
| 6211 | if (scx_enabled()) |
| 6212 | goto restart; |
| 6213 | |
| 6214 | /* |
| 6215 | * Optimization: we know that if all tasks are in the fair class we can |
| 6216 | * call that function directly, but only if the @prev task wasn't of a |
| 6217 | * higher scheduling class, because otherwise those lose the |
| 6218 | * opportunity to pull in more work from other CPUs. |
| 6219 | */ |
| 6220 | if (likely(!sched_class_above(rq->donor->sched_class, &fair_sched_class) && |
| 6221 | rq->nr_running == rq->cfs.h_nr_queued)) { |
| 6222 | |
| 6223 | p = pick_task_fair(rq, rf); |
| 6224 | if (unlikely(p == RETRY_TASK)) |
| 6225 | goto restart; |
| 6226 | |
| 6227 | /* Assume the next prioritized class is idle_sched_class */ |
| 6228 | if (!p) |
| 6229 | p = pick_task_idle(rq, rf); |
| 6230 | |
| 6231 | put_prev_set_next_task(rq, rq->donor, p); |
| 6232 | return p; |
| 6233 | } |
| 6234 | |
| 6235 | restart: |
| 6236 | prev_balance(rq, rf); |
| 6237 | |
| 6238 | for_each_active_class(class) { |
| 6239 | p = class->pick_task(rq, rf); |
| 6240 | if (unlikely(p == RETRY_TASK)) |
| 6241 | goto restart; |
| 6242 | if (p) { |
| 6243 | put_prev_set_next_task(rq, rq->donor, p); |
| 6244 | return p; |
| 6245 | } |
| 6246 | } |
| 6247 | |
| 6248 | BUG(); /* The idle class should always have a runnable task. */ |
| 6249 | } |
| 6250 | |
| 6251 | #ifdef CONFIG_SCHED_CORE |
| 6252 | static inline bool is_task_rq_idle(struct task_struct *t) |
| 6253 | { |
| 6254 | return (task_rq(t)->idle == t); |
| 6255 | } |
| 6256 | |
| 6257 | static inline bool cookie_equals(struct task_struct *a, unsigned long cookie) |
| 6258 | { |
| 6259 | return is_task_rq_idle(a) || (a->core_cookie == cookie); |
| 6260 | } |
| 6261 | |
| 6262 | static inline bool cookie_match(struct task_struct *a, struct task_struct *b) |
| 6263 | { |
| 6264 | if (is_task_rq_idle(a) || is_task_rq_idle(b)) |
| 6265 | return true; |
| 6266 | |
| 6267 | return a->core_cookie == b->core_cookie; |
| 6268 | } |
| 6269 | |
| 6270 | /* |
| 6271 | * Careful; this can return RETRY_TASK, it does not include the retry-loop |
| 6272 | * itself due to the whole SMT pick retry thing below. |
| 6273 | */ |
| 6274 | static inline struct task_struct *pick_task(struct rq *rq, struct rq_flags *rf) |
| 6275 | { |
| 6276 | const struct sched_class *class; |
| 6277 | struct task_struct *p; |
| 6278 | |
| 6279 | rq->dl_server = NULL; |
| 6280 | |
| 6281 | for_each_active_class(class) { |
| 6282 | p = class->pick_task(rq, rf); |
| 6283 | if (p) |
| 6284 | return p; |
| 6285 | } |
| 6286 | |
| 6287 | BUG(); /* The idle class should always have a runnable task. */ |
| 6288 | } |
| 6289 | |
| 6290 | extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi); |
| 6291 | |
| 6292 | static void queue_core_balance(struct rq *rq); |
| 6293 | |
| 6294 | static struct task_struct * |
| 6295 | pick_next_task(struct rq *rq, struct rq_flags *rf) |
| 6296 | __must_hold(__rq_lockp(rq)) |
| 6297 | { |
| 6298 | struct task_struct *next, *p, *max; |
| 6299 | const struct cpumask *smt_mask; |
| 6300 | bool fi_before = false; |
| 6301 | bool core_clock_updated = (rq == rq->core); |
| 6302 | unsigned long cookie; |
| 6303 | int i, cpu, occ = 0; |
| 6304 | struct rq *rq_i; |
| 6305 | bool need_sync = false; |
| 6306 | |
| 6307 | if (!sched_core_enabled(rq)) |
| 6308 | return __pick_next_task(rq, rf); |
| 6309 | |
| 6310 | cpu = cpu_of(rq); |
| 6311 | |
| 6312 | /* Stopper task is switching into idle, no need core-wide selection. */ |
| 6313 | if (cpu_is_offline(cpu)) { |
| 6314 | /* |
| 6315 | * Reset core_pick so that we don't enter the fastpath when |
| 6316 | * coming online. core_pick would already be migrated to |
| 6317 | * another cpu during offline. |
| 6318 | */ |
| 6319 | rq->core_pick = NULL; |
| 6320 | rq->core_dl_server = NULL; |
| 6321 | return __pick_next_task(rq, rf); |
| 6322 | } |
| 6323 | |
| 6324 | rq->core->core_pick_in_flight++; |
| 6325 | |
| 6326 | /* |
| 6327 | * If there were no {en,de}queues since we picked (IOW, the task |
| 6328 | * pointers are all still valid), and we haven't scheduled the last |
| 6329 | * pick yet, do so now. |
| 6330 | * |
| 6331 | * rq->core_pick can be NULL if no selection was made for a CPU because |
| 6332 | * it was either offline or went offline during a sibling's core-wide |
| 6333 | * selection. In this case, do a core-wide selection. |
| 6334 | */ |
| 6335 | if (rq->core->core_pick_seq == rq->core->core_task_seq && |
| 6336 | rq->core_pick) { |
| 6337 | next = rq->core_pick; |
| 6338 | rq->dl_server = rq->core_dl_server; |
| 6339 | rq->core_pick = NULL; |
| 6340 | rq->core_dl_server = NULL; |
| 6341 | goto out_set_next; |
| 6342 | } |
| 6343 | |
| 6344 | prev_balance(rq, rf); |
| 6345 | |
| 6346 | smt_mask = cpu_smt_mask(cpu); |
| 6347 | |
| 6348 | restart: |
| 6349 | need_sync |= !!rq->core->core_cookie; |
| 6350 | |
| 6351 | /* reset state */ |
| 6352 | rq->core->core_cookie = 0UL; |
| 6353 | if (rq->core->core_forceidle_count) { |
| 6354 | if (!core_clock_updated) { |
| 6355 | update_rq_clock(rq->core); |
| 6356 | core_clock_updated = true; |
| 6357 | } |
| 6358 | sched_core_account_forceidle(rq); |
| 6359 | /* reset after accounting force idle */ |
| 6360 | rq->core->core_forceidle_start = 0; |
| 6361 | rq->core->core_forceidle_count = 0; |
| 6362 | rq->core->core_forceidle_occupation = 0; |
| 6363 | need_sync = true; |
| 6364 | fi_before = true; |
| 6365 | } |
| 6366 | |
| 6367 | /* |
| 6368 | * core->core_task_seq, core->core_pick_seq |
| 6369 | * |
| 6370 | * @task_seq guards the task state ({en,de}queues) |
| 6371 | * @pick_seq is the @task_seq we did a selection on |
| 6372 | * |
| 6373 | * Once a core-wide selection is committed, a non-NULL core_pick denotes |
| 6374 | * a pick which still needs to be consumed on this CPU. |
| 6375 | * |
| 6376 | * However, preemptions can cause multiple picks on the same task set. |
| 6377 | * 'Fix' this by also increasing @task_seq for every pick. |
| 6378 | */ |
| 6379 | rq->core->core_task_seq++; |
| 6380 | |
| 6381 | /* |
| 6382 | * Optimize for common case where this CPU has no cookies |
| 6383 | * and there are no cookied tasks running on siblings. |
| 6384 | */ |
| 6385 | if (!need_sync) { |
| 6386 | next = pick_task(rq, rf); |
| 6387 | if (unlikely(next == RETRY_TASK)) { |
| 6388 | /* rq lock may have been dropped, clocks invalidated */ |
| 6389 | core_clock_updated = false; |
| 6390 | if (!(rq->clock_update_flags & RQCF_UPDATED)) |
| 6391 | update_rq_clock(rq); |
| 6392 | goto restart; |
| 6393 | } |
| 6394 | |
| 6395 | if (!next->core_cookie) { |
| 6396 | rq->core_pick = NULL; |
| 6397 | rq->core_dl_server = NULL; |
| 6398 | /* |
| 6399 | * For robustness, update the min_vruntime_fi for |
| 6400 | * unconstrained picks as well. |
| 6401 | */ |
| 6402 | WARN_ON_ONCE(fi_before); |
| 6403 | task_vruntime_update(rq, next, false); |
| 6404 | goto out_set_next; |
| 6405 | } |
| 6406 | } |
| 6407 | |
| 6408 | /* |
| 6409 | * For each thread: do the regular task pick and find the max prio task |
| 6410 | * amongst them. |
| 6411 | * |
| 6412 | * Tie-break prio towards the current CPU |
| 6413 | */ |
| 6414 | max = NULL; |
| 6415 | for_each_cpu_wrap(i, smt_mask, cpu) { |
| 6416 | rq_i = cpu_rq(i); |
| 6417 | |
| 6418 | /* |
| 6419 | * Current cpu always has its clock updated on entrance to |
| 6420 | * pick_next_task(). If the current cpu is not the core, |
| 6421 | * the core may also have been updated above. |
| 6422 | */ |
| 6423 | if (i != cpu && (rq_i != rq->core || !core_clock_updated)) |
| 6424 | update_rq_clock(rq_i); |
| 6425 | |
| 6426 | p = pick_task(rq_i, rf); |
| 6427 | if (unlikely(p == RETRY_TASK)) { |
| 6428 | /* rq lock may have been dropped, clocks invalidated */ |
| 6429 | core_clock_updated = false; |
| 6430 | if (!(rq->clock_update_flags & RQCF_UPDATED)) |
| 6431 | update_rq_clock(rq); |
| 6432 | goto restart; |
| 6433 | } |
| 6434 | |
| 6435 | rq_i->core_pick = p; |
| 6436 | rq_i->core_dl_server = rq_i->dl_server; |
| 6437 | |
| 6438 | if (!max || prio_less(max, p, fi_before)) |
| 6439 | max = p; |
| 6440 | } |
| 6441 | |
| 6442 | cookie = rq->core->core_cookie = max->core_cookie; |
| 6443 | |
| 6444 | /* |
| 6445 | * For each thread: try and find a runnable task that matches @max or |
| 6446 | * force idle. |
| 6447 | */ |
| 6448 | for_each_cpu(i, smt_mask) { |
| 6449 | rq_i = cpu_rq(i); |
| 6450 | p = rq_i->core_pick; |
| 6451 | |
| 6452 | if (!cookie_equals(p, cookie)) { |
| 6453 | p = NULL; |
| 6454 | if (cookie) |
| 6455 | p = sched_core_find(rq_i, cookie); |
| 6456 | if (!p) |
| 6457 | p = idle_sched_class.pick_task(rq_i, rf); |
| 6458 | } |
| 6459 | |
| 6460 | rq_i->core_pick = p; |
| 6461 | rq_i->core_dl_server = NULL; |
| 6462 | |
| 6463 | if (p == rq_i->idle) { |
| 6464 | if (rq_i->nr_running) { |
| 6465 | rq->core->core_forceidle_count++; |
| 6466 | if (!fi_before) |
| 6467 | rq->core->core_forceidle_seq++; |
| 6468 | } |
| 6469 | } else { |
| 6470 | occ++; |
| 6471 | } |
| 6472 | } |
| 6473 | |
| 6474 | if (schedstat_enabled() && rq->core->core_forceidle_count) { |
| 6475 | rq->core->core_forceidle_start = rq_clock(rq->core); |
| 6476 | rq->core->core_forceidle_occupation = occ; |
| 6477 | } |
| 6478 | |
| 6479 | rq->core->core_pick_seq = rq->core->core_task_seq; |
| 6480 | next = rq->core_pick; |
| 6481 | |
| 6482 | /* Something should have been selected for current CPU */ |
| 6483 | WARN_ON_ONCE(!next); |
| 6484 | |
| 6485 | /* |
| 6486 | * Reschedule siblings |
| 6487 | * |
| 6488 | * NOTE: L1TF -- at this point we're no longer running the old task and |
| 6489 | * sending an IPI (below) ensures the sibling will no longer be running |
| 6490 | * their task. This ensures there is no inter-sibling overlap between |
| 6491 | * non-matching user state. |
| 6492 | */ |
| 6493 | for_each_cpu(i, smt_mask) { |
| 6494 | rq_i = cpu_rq(i); |
| 6495 | |
| 6496 | /* |
| 6497 | * An online sibling might have gone offline before a task |
| 6498 | * could be picked for it, or it might be offline but later |
| 6499 | * happen to come online, but its too late and nothing was |
| 6500 | * picked for it. That's Ok - it will pick tasks for itself, |
| 6501 | * so ignore it. |
| 6502 | */ |
| 6503 | if (!rq_i->core_pick) |
| 6504 | continue; |
| 6505 | |
| 6506 | /* |
| 6507 | * Update for new !FI->FI transitions, or if continuing to be in !FI: |
| 6508 | * fi_before fi update? |
| 6509 | * 0 0 1 |
| 6510 | * 0 1 1 |
| 6511 | * 1 0 1 |
| 6512 | * 1 1 0 |
| 6513 | */ |
| 6514 | if (!(fi_before && rq->core->core_forceidle_count)) |
| 6515 | task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count); |
| 6516 | |
| 6517 | rq_i->core_pick->core_occupation = occ; |
| 6518 | |
| 6519 | if (i == cpu) { |
| 6520 | rq_i->core_pick = NULL; |
| 6521 | rq_i->core_dl_server = NULL; |
| 6522 | continue; |
| 6523 | } |
| 6524 | |
| 6525 | /* Did we break L1TF mitigation requirements? */ |
| 6526 | WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick)); |
| 6527 | |
| 6528 | if (rq_i->curr == rq_i->core_pick) { |
| 6529 | rq_i->core_pick = NULL; |
| 6530 | rq_i->core_dl_server = NULL; |
| 6531 | continue; |
| 6532 | } |
| 6533 | |
| 6534 | resched_curr(rq_i); |
| 6535 | } |
| 6536 | |
| 6537 | out_set_next: |
| 6538 | rq->core->core_pick_in_flight--; |
| 6539 | put_prev_set_next_task(rq, rq->donor, next); |
| 6540 | if (rq->core->core_forceidle_count && next == rq->idle) |
| 6541 | queue_core_balance(rq); |
| 6542 | |
| 6543 | return next; |
| 6544 | } |
| 6545 | |
| 6546 | static bool try_steal_cookie(int this, int that) |
| 6547 | { |
| 6548 | struct rq *dst = cpu_rq(this), *src = cpu_rq(that); |
| 6549 | struct task_struct *p; |
| 6550 | unsigned long cookie; |
| 6551 | bool success = false; |
| 6552 | |
| 6553 | guard(irq)(); |
| 6554 | guard(double_rq_lock)(dst, src); |
| 6555 | |
| 6556 | cookie = dst->core->core_cookie; |
| 6557 | if (!cookie) |
| 6558 | return false; |
| 6559 | |
| 6560 | if (dst->curr != dst->idle) |
| 6561 | return false; |
| 6562 | |
| 6563 | p = sched_core_find(src, cookie); |
| 6564 | if (!p) |
| 6565 | return false; |
| 6566 | |
| 6567 | do { |
| 6568 | if (p == src->core_pick || p == src->curr || p == src->donor) |
| 6569 | goto next; |
| 6570 | |
| 6571 | if (task_is_blocked(p)) |
| 6572 | goto next; |
| 6573 | |
| 6574 | if (!is_cpu_allowed(p, this)) |
| 6575 | goto next; |
| 6576 | |
| 6577 | if (p->core_occupation > dst->idle->core_occupation) |
| 6578 | goto next; |
| 6579 | /* |
| 6580 | * sched_core_find() and sched_core_next() will ensure |
| 6581 | * that task @p is not throttled now, we also need to |
| 6582 | * check whether the runqueue of the destination CPU is |
| 6583 | * being throttled. |
| 6584 | */ |
| 6585 | if (sched_task_is_throttled(p, this)) |
| 6586 | goto next; |
| 6587 | |
| 6588 | move_queued_task_locked(src, dst, p); |
| 6589 | resched_curr(dst); |
| 6590 | |
| 6591 | success = true; |
| 6592 | break; |
| 6593 | |
| 6594 | next: |
| 6595 | p = sched_core_next(p, cookie); |
| 6596 | } while (p); |
| 6597 | |
| 6598 | return success; |
| 6599 | } |
| 6600 | |
| 6601 | static bool steal_cookie_task(int cpu, struct sched_domain *sd) |
| 6602 | { |
| 6603 | int i; |
| 6604 | |
| 6605 | for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) { |
| 6606 | if (i == cpu) |
| 6607 | continue; |
| 6608 | |
| 6609 | if (need_resched()) |
| 6610 | break; |
| 6611 | |
| 6612 | if (try_steal_cookie(cpu, i)) |
| 6613 | return true; |
| 6614 | } |
| 6615 | |
| 6616 | return false; |
| 6617 | } |
| 6618 | |
| 6619 | static void sched_core_balance(struct rq *rq) |
| 6620 | __must_hold(__rq_lockp(rq)) |
| 6621 | { |
| 6622 | struct sched_domain *sd; |
| 6623 | int cpu = cpu_of(rq); |
| 6624 | |
| 6625 | guard(preempt)(); |
| 6626 | guard(rcu)(); |
| 6627 | |
| 6628 | raw_spin_rq_unlock_irq(rq); |
| 6629 | for_each_domain(cpu, sd) { |
| 6630 | if (need_resched()) |
| 6631 | break; |
| 6632 | |
| 6633 | if (steal_cookie_task(cpu, sd)) |
| 6634 | break; |
| 6635 | } |
| 6636 | raw_spin_rq_lock_irq(rq); |
| 6637 | } |
| 6638 | |
| 6639 | static DEFINE_PER_CPU(struct balance_callback, core_balance_head); |
| 6640 | |
| 6641 | static void queue_core_balance(struct rq *rq) |
| 6642 | { |
| 6643 | if (!sched_core_enabled(rq)) |
| 6644 | return; |
| 6645 | |
| 6646 | if (!rq->core->core_cookie) |
| 6647 | return; |
| 6648 | |
| 6649 | if (!rq->nr_running) /* not forced idle */ |
| 6650 | return; |
| 6651 | |
| 6652 | queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance); |
| 6653 | } |
| 6654 | |
| 6655 | DEFINE_LOCK_GUARD_1(core_lock, int, |
| 6656 | sched_core_lock(*_T->lock, &_T->flags), |
| 6657 | sched_core_unlock(*_T->lock, &_T->flags), |
| 6658 | unsigned long flags) |
| 6659 | |
| 6660 | static void sched_core_cpu_starting(unsigned int cpu) |
| 6661 | { |
| 6662 | const struct cpumask *smt_mask = cpu_smt_mask(cpu); |
| 6663 | struct rq *rq = cpu_rq(cpu), *core_rq = NULL; |
| 6664 | int t; |
| 6665 | |
| 6666 | guard(core_lock)(&cpu); |
| 6667 | |
| 6668 | WARN_ON_ONCE(rq->core != rq); |
| 6669 | |
| 6670 | /* if we're the first, we'll be our own leader */ |
| 6671 | if (cpumask_weight(smt_mask) == 1) |
| 6672 | return; |
| 6673 | |
| 6674 | /* find the leader */ |
| 6675 | for_each_cpu(t, smt_mask) { |
| 6676 | if (t == cpu) |
| 6677 | continue; |
| 6678 | rq = cpu_rq(t); |
| 6679 | if (rq->core == rq) { |
| 6680 | core_rq = rq; |
| 6681 | break; |
| 6682 | } |
| 6683 | } |
| 6684 | |
| 6685 | if (WARN_ON_ONCE(!core_rq)) /* whoopsie */ |
| 6686 | return; |
| 6687 | |
| 6688 | /* install and validate core_rq */ |
| 6689 | for_each_cpu(t, smt_mask) { |
| 6690 | rq = cpu_rq(t); |
| 6691 | |
| 6692 | if (t == cpu) |
| 6693 | rq->core = core_rq; |
| 6694 | |
| 6695 | WARN_ON_ONCE(rq->core != core_rq); |
| 6696 | } |
| 6697 | } |
| 6698 | |
| 6699 | static void sched_core_cpu_deactivate(unsigned int cpu) |
| 6700 | { |
| 6701 | const struct cpumask *smt_mask = cpu_smt_mask(cpu); |
| 6702 | struct rq *rq = cpu_rq(cpu), *core_rq = NULL; |
| 6703 | int t; |
| 6704 | |
| 6705 | guard(core_lock)(&cpu); |
| 6706 | |
| 6707 | /* if we're the last man standing, nothing to do */ |
| 6708 | if (cpumask_weight(smt_mask) == 1) { |
| 6709 | WARN_ON_ONCE(rq->core != rq); |
| 6710 | return; |
| 6711 | } |
| 6712 | |
| 6713 | /* if we're not the leader, nothing to do */ |
| 6714 | if (rq->core != rq) |
| 6715 | return; |
| 6716 | |
| 6717 | /* find a new leader */ |
| 6718 | for_each_cpu(t, smt_mask) { |
| 6719 | if (t == cpu) |
| 6720 | continue; |
| 6721 | core_rq = cpu_rq(t); |
| 6722 | break; |
| 6723 | } |
| 6724 | |
| 6725 | if (WARN_ON_ONCE(!core_rq)) /* impossible */ |
| 6726 | return; |
| 6727 | |
| 6728 | /* copy the shared state to the new leader */ |
| 6729 | core_rq->core_task_seq = rq->core_task_seq; |
| 6730 | core_rq->core_pick_seq = rq->core_pick_seq; |
| 6731 | core_rq->core_cookie = rq->core_cookie; |
| 6732 | core_rq->core_forceidle_count = rq->core_forceidle_count; |
| 6733 | core_rq->core_forceidle_seq = rq->core_forceidle_seq; |
| 6734 | core_rq->core_forceidle_occupation = rq->core_forceidle_occupation; |
| 6735 | |
| 6736 | /* |
| 6737 | * A stale leftover would bias the count forever if this CPU later |
| 6738 | * returns as its own leader. Move, don't copy. |
| 6739 | */ |
| 6740 | core_rq->core_pick_in_flight = rq->core_pick_in_flight; |
| 6741 | rq->core_pick_in_flight = 0; |
| 6742 | |
| 6743 | /* |
| 6744 | * Accounting edge for forced idle is handled in pick_next_task(). |
| 6745 | * Don't need another one here, since the hotplug thread shouldn't |
| 6746 | * have a cookie. |
| 6747 | */ |
| 6748 | core_rq->core_forceidle_start = 0; |
| 6749 | |
| 6750 | /* install new leader */ |
| 6751 | for_each_cpu(t, smt_mask) { |
| 6752 | rq = cpu_rq(t); |
| 6753 | rq->core = core_rq; |
| 6754 | } |
| 6755 | } |
| 6756 | |
| 6757 | static inline void sched_core_cpu_dying(unsigned int cpu) |
| 6758 | { |
| 6759 | struct rq *rq = cpu_rq(cpu); |
| 6760 | |
| 6761 | if (rq->core != rq) |
| 6762 | rq->core = rq; |
| 6763 | } |
| 6764 | |
| 6765 | #else /* !CONFIG_SCHED_CORE: */ |
| 6766 | |
| 6767 | static inline void sched_core_cpu_starting(unsigned int cpu) {} |
| 6768 | static inline void sched_core_cpu_deactivate(unsigned int cpu) {} |
| 6769 | static inline void sched_core_cpu_dying(unsigned int cpu) {} |
| 6770 | |
| 6771 | static struct task_struct * |
| 6772 | pick_next_task(struct rq *rq, struct rq_flags *rf) |
| 6773 | __must_hold(__rq_lockp(rq)) |
| 6774 | { |
| 6775 | return __pick_next_task(rq, rf); |
| 6776 | } |
| 6777 | |
| 6778 | #endif /* !CONFIG_SCHED_CORE */ |
| 6779 | |
| 6780 | /* |
| 6781 | * Constants for the sched_mode argument of __schedule(). |
| 6782 | * |
| 6783 | * The mode argument allows RT enabled kernels to differentiate a |
| 6784 | * preemption from blocking on an 'sleeping' spin/rwlock. |
| 6785 | */ |
| 6786 | #define SM_IDLE (-1) |
| 6787 | #define SM_NONE 0 |
| 6788 | #define SM_PREEMPT 1 |
| 6789 | #define SM_RTLOCK_WAIT 2 |
| 6790 | |
| 6791 | /* |
| 6792 | * Helper function for __schedule() |
| 6793 | * |
| 6794 | * Tries to deactivate the task, unless the should_block arg |
| 6795 | * is false or if a signal is pending. In the case a signal |
| 6796 | * is pending, marks the task's __state as RUNNING (and clear |
| 6797 | * blocked_on). |
| 6798 | */ |
| 6799 | static bool try_to_block_task(struct rq *rq, struct task_struct *p, |
| 6800 | unsigned long *task_state_p, bool should_block) |
| 6801 | { |
| 6802 | unsigned long task_state = *task_state_p; |
| 6803 | |
| 6804 | WARN_ON_ONCE(p->is_blocked); |
| 6805 | |
| 6806 | if (signal_pending_state(task_state, p)) { |
| 6807 | WRITE_ONCE(p->__state, TASK_RUNNING); |
| 6808 | *task_state_p = TASK_RUNNING; |
| 6809 | clear_task_blocked_on(p, NULL); |
| 6810 | |
| 6811 | return false; |
| 6812 | } |
| 6813 | |
| 6814 | p->is_blocked = 1; |
| 6815 | |
| 6816 | /* |
| 6817 | * We check should_block after signal_pending because we |
| 6818 | * will want to wake the task in that case. But if |
| 6819 | * should_block is false, its likely due to the task being |
| 6820 | * blocked on a mutex, and we want to keep it on the runqueue |
| 6821 | * to be selectable for proxy-execution. |
| 6822 | */ |
| 6823 | if (!should_block) |
| 6824 | return false; |
| 6825 | |
| 6826 | block_task(rq, p, task_state); |
| 6827 | return true; |
| 6828 | } |
| 6829 | |
| 6830 | #ifdef CONFIG_SCHED_PROXY_EXEC |
| 6831 | static inline void proxy_set_task_cpu(struct task_struct *p, int cpu) |
| 6832 | { |
| 6833 | unsigned int wake_cpu; |
| 6834 | |
| 6835 | /* |
| 6836 | * Since we are enqueuing a blocked task on a cpu it may |
| 6837 | * not be able to run on, preserve wake_cpu when we |
| 6838 | * __set_task_cpu so we can return the task to where it |
| 6839 | * was previously runnable. |
| 6840 | */ |
| 6841 | wake_cpu = p->wake_cpu; |
| 6842 | __set_task_cpu(p, cpu); |
| 6843 | p->wake_cpu = wake_cpu; |
| 6844 | } |
| 6845 | |
| 6846 | static inline struct task_struct *proxy_resched_idle(struct rq *rq) |
| 6847 | { |
| 6848 | put_prev_set_next_task(rq, rq->donor, rq->idle); |
| 6849 | rq->next_class = &idle_sched_class; |
| 6850 | rq_set_donor(rq, rq->idle); |
| 6851 | set_tsk_need_resched(rq->idle); |
| 6852 | return rq->idle; |
| 6853 | } |
| 6854 | |
| 6855 | static void proxy_deactivate(struct rq *rq, struct task_struct *donor) |
| 6856 | { |
| 6857 | unsigned long state = READ_ONCE(donor->__state); |
| 6858 | |
| 6859 | WARN_ON_ONCE(state == TASK_RUNNING); |
| 6860 | WARN_ON_ONCE(donor->blocked_on); |
| 6861 | /* |
| 6862 | * Because we got donor from pick_next_task(), it is *crucial* |
| 6863 | * that we call proxy_resched_idle() before we deactivate it. |
| 6864 | * As once we deactivate donor, donor->on_rq is set to zero, |
| 6865 | * which allows ttwu() to immediately try to wake the task on |
| 6866 | * another rq. So we cannot use *any* references to donor |
| 6867 | * after that point. So things like cfs_rq->curr or rq->donor |
| 6868 | * need to be changed from next *before* we deactivate. |
| 6869 | */ |
| 6870 | proxy_resched_idle(rq); |
| 6871 | block_task(rq, donor, state); |
| 6872 | } |
| 6873 | |
| 6874 | /* |
| 6875 | * Remove a retained proxy donor before changing its scheduler ownership. |
| 6876 | * The caller holds p->pi_lock, so p cannot wake and migrate if block_task() |
| 6877 | * drops it from the runqueue. If DELAY_DEQUEUE keeps a blocked fair task |
| 6878 | * queued, switching_from_fair() completes the dequeue in the immediately |
| 6879 | * following sched_change_begin(). |
| 6880 | */ |
| 6881 | void sched_proxy_block_task(struct rq *rq, struct task_struct *p) |
| 6882 | { |
| 6883 | unsigned long state = READ_ONCE(p->__state); |
| 6884 | |
| 6885 | lockdep_assert_held(&p->pi_lock); |
| 6886 | lockdep_assert_rq_held(rq); |
| 6887 | |
| 6888 | if (!p->is_blocked || !task_on_rq_queued(p)) |
| 6889 | return; |
| 6890 | if (WARN_ON_ONCE(state == TASK_RUNNING)) |
| 6891 | return; |
| 6892 | |
| 6893 | if (task_current_donor(rq, p)) |
| 6894 | proxy_reset_donor(rq); |
| 6895 | |
| 6896 | if (!p->se.sched_delayed) |
| 6897 | block_task(rq, p, state); |
| 6898 | |
| 6899 | WARN_ON_ONCE(task_on_rq_queued(p) && !p->se.sched_delayed); |
| 6900 | } |
| 6901 | |
| 6902 | static inline void proxy_release_rq_lock(struct rq *rq, struct rq_flags *rf) |
| 6903 | __releases(__rq_lockp(rq)) |
| 6904 | { |
| 6905 | /* |
| 6906 | * The class scheduler may have queued a balance callback |
| 6907 | * from pick_next_task() called earlier. |
| 6908 | * |
| 6909 | * So here we have to zap callbacks before unlocking the rq |
| 6910 | * as another CPU may jump in and call sched_balance_rq |
| 6911 | * which can trip the warning in rq_pin_lock() if we |
| 6912 | * leave callbacks set. |
| 6913 | * |
| 6914 | * After we later reaquire the rq lock, we will force __schedule() |
| 6915 | * to pick_again, so the callbacks will get re-established. |
| 6916 | */ |
| 6917 | zap_balance_callbacks(rq); |
| 6918 | rq_unpin_lock(rq, rf); |
| 6919 | raw_spin_rq_unlock(rq); |
| 6920 | } |
| 6921 | |
| 6922 | static inline void proxy_reacquire_rq_lock(struct rq *rq, struct rq_flags *rf) |
| 6923 | __acquires(__rq_lockp(rq)) |
| 6924 | { |
| 6925 | raw_spin_rq_lock(rq); |
| 6926 | rq_repin_lock(rq, rf); |
| 6927 | update_rq_clock(rq); |
| 6928 | } |
| 6929 | |
| 6930 | /* |
| 6931 | * If the blocked-on relationship crosses CPUs, migrate @p to the |
| 6932 | * owner's CPU. |
| 6933 | * |
| 6934 | * This is because we must respect the CPU affinity of execution |
| 6935 | * contexts (owner) but we can ignore affinity for scheduling |
| 6936 | * contexts (@p). So we have to move scheduling contexts towards |
| 6937 | * potential execution contexts. |
| 6938 | * |
| 6939 | * Note: The owner can disappear, but simply migrate to @target_cpu |
| 6940 | * and leave that CPU to sort things out. |
| 6941 | */ |
| 6942 | static void proxy_migrate_task(struct rq *rq, struct rq_flags *rf, |
| 6943 | struct task_struct *p, int target_cpu) |
| 6944 | __must_hold(__rq_lockp(rq)) |
| 6945 | { |
| 6946 | struct rq *target_rq = cpu_rq(target_cpu); |
| 6947 | LIST_HEAD(migrate_list); |
| 6948 | |
| 6949 | lockdep_assert_rq_held(rq); |
| 6950 | /* |
| 6951 | * Since we are migrating a blocked donor, it could be rq->donor, |
| 6952 | * and we want to make sure there aren't any references from this |
| 6953 | * rq to it before we drop the lock. This avoids another cpu |
| 6954 | * jumping in and grabbing the rq lock and referencing rq->donor |
| 6955 | * or cfs_rq->curr, etc after we have migrated it to another cpu, |
| 6956 | * and before we pick_again in __schedule. |
| 6957 | * |
| 6958 | * So call proxy_resched_idle() to drop the rq->donor references |
| 6959 | * before we release the lock. |
| 6960 | */ |
| 6961 | proxy_resched_idle(rq); |
| 6962 | for (; p; p = p->blocked_donor) { |
| 6963 | WARN_ON(p == rq->curr); |
| 6964 | deactivate_task(rq, p, DEQUEUE_NOCLOCK); |
| 6965 | proxy_set_task_cpu(p, target_cpu); |
| 6966 | /* |
| 6967 | * We can re-use se.group_node to migrate the thing, |
| 6968 | * because @p is deactivated (won't be balanced) and |
| 6969 | * we hold the rq_lock. |
| 6970 | */ |
| 6971 | list_add(&p->se.group_node, &migrate_list); |
| 6972 | } |
| 6973 | proxy_release_rq_lock(rq, rf); |
| 6974 | |
| 6975 | __attach_tasks(target_rq, &migrate_list); |
| 6976 | |
| 6977 | proxy_reacquire_rq_lock(rq, rf); |
| 6978 | } |
| 6979 | |
| 6980 | /* |
| 6981 | * Find runnable lock owner to proxy for mutex blocked donor |
| 6982 | * |
| 6983 | * Follow the blocked-on relation: |
| 6984 | * |
| 6985 | * ,-> task |
| 6986 | * | | blocked-on |
| 6987 | * | v |
| 6988 | * blocked_donor | mutex |
| 6989 | * | | owner |
| 6990 | * | v |
| 6991 | * `-- task |
| 6992 | * |
| 6993 | * and set the blocked_donor relation, this latter is used by the mutex |
| 6994 | * code to find which (blocked) task to hand-off to. |
| 6995 | * |
| 6996 | * Lock order: |
| 6997 | * |
| 6998 | * p->pi_lock |
| 6999 | * rq->lock |
| 7000 | * mutex->wait_lock |
| 7001 | * p->blocked_lock |
| 7002 | * |
| 7003 | * Returns the task that is going to be used as execution context (the one |
| 7004 | * that is actually going to be run on cpu_of(rq)). |
| 7005 | */ |
| 7006 | static struct task_struct * |
| 7007 | find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf) |
| 7008 | __must_hold(__rq_lockp(rq)) |
| 7009 | { |
| 7010 | struct task_struct *owner = NULL; |
| 7011 | bool curr_in_chain = false; |
| 7012 | int this_cpu = cpu_of(rq); |
| 7013 | struct task_struct *p; |
| 7014 | int owner_cpu; |
| 7015 | |
| 7016 | /* Follow blocked_on chain. */ |
| 7017 | for (p = donor; p->is_blocked; p = owner) { |
| 7018 | /* if its PROXY_WAKING, do return migration or run if current */ |
| 7019 | struct mutex *mutex = p->blocked_on; |
| 7020 | if (!mutex) { |
| 7021 | clear_task_blocked_on(p, mutex); |
| 7022 | if (task_current(rq, p)) { |
| 7023 | p->is_blocked = 0; |
| 7024 | return p; |
| 7025 | } |
| 7026 | goto deactivate; |
| 7027 | } |
| 7028 | |
| 7029 | /* |
| 7030 | * By taking mutex->wait_lock we hold off concurrent mutex_unlock() |
| 7031 | * and ensure @owner sticks around. |
| 7032 | */ |
| 7033 | guard(raw_spinlock)(&mutex->wait_lock); |
| 7034 | guard(raw_spinlock)(&p->blocked_lock); |
| 7035 | |
| 7036 | /* Check again that p is blocked with blocked_lock held */ |
| 7037 | if (mutex != __get_task_blocked_on(p)) { |
| 7038 | /* |
| 7039 | * Something changed in the blocked_on chain and |
| 7040 | * we don't know if only at this level. So, let's |
| 7041 | * just bail out completely and let __schedule() |
| 7042 | * figure things out (pick_again loop). |
| 7043 | */ |
| 7044 | return NULL; |
| 7045 | } |
| 7046 | |
| 7047 | if (task_current(rq, p)) |
| 7048 | curr_in_chain = true; |
| 7049 | |
| 7050 | owner = __mutex_owner(mutex); |
| 7051 | if (!owner) { |
| 7052 | /* |
| 7053 | * If there is no owner, either clear blocked_on |
| 7054 | * and return p (if it is current and safe to |
| 7055 | * just run on this rq), or return-migrate the task. |
| 7056 | */ |
| 7057 | __clear_task_blocked_on(p, NULL); |
| 7058 | if (task_current(rq, p)) { |
| 7059 | p->is_blocked = 0; |
| 7060 | return p; |
| 7061 | } |
| 7062 | goto deactivate; |
| 7063 | } |
| 7064 | |
| 7065 | if (!READ_ONCE(owner->on_rq) || owner->se.sched_delayed) { |
| 7066 | /* XXX Don't handle blocked owners/delayed dequeue yet */ |
| 7067 | if (curr_in_chain) |
| 7068 | goto resched_idle; |
| 7069 | __clear_task_blocked_on(p, NULL); |
| 7070 | goto deactivate; |
| 7071 | } |
| 7072 | |
| 7073 | owner_cpu = task_cpu(owner); |
| 7074 | if (owner_cpu != this_cpu) { |
| 7075 | /* |
| 7076 | * @owner can disappear, simply migrate to @owner_cpu |
| 7077 | * and leave that CPU to sort things out. |
| 7078 | */ |
| 7079 | if (curr_in_chain) |
| 7080 | goto resched_idle; |
| 7081 | goto migrate_task; |
| 7082 | } |
| 7083 | |
| 7084 | if (task_on_rq_migrating(owner)) { |
| 7085 | /* |
| 7086 | * One of the chain of mutex owners is currently migrating to this |
| 7087 | * CPU, but has not yet been enqueued because we are holding the |
| 7088 | * rq lock. As a simple solution, just schedule rq->idle to give |
| 7089 | * the migration a chance to complete. Much like the migrate_task |
| 7090 | * case we should end up back in find_proxy_task(), this time |
| 7091 | * hopefully with all relevant tasks already enqueued. |
| 7092 | */ |
| 7093 | goto resched_idle; |
| 7094 | } |
| 7095 | |
| 7096 | /* |
| 7097 | * Its possible to race where after we check owner->on_rq |
| 7098 | * but before we check (owner_cpu != this_cpu) that the |
| 7099 | * task on another cpu was migrated back to this cpu. In |
| 7100 | * that case it could slip by our checks. So double check |
| 7101 | * we are still on this cpu and not migrating. If we get |
| 7102 | * inconsistent results, try again. |
| 7103 | */ |
| 7104 | if (!task_on_rq_queued(owner) || task_cpu(owner) != this_cpu) |
| 7105 | return NULL; |
| 7106 | |
| 7107 | if (owner == p) { |
| 7108 | /* |
| 7109 | * It's possible we interleave with mutex_unlock like: |
| 7110 | * |
| 7111 | * lock(&rq->lock); |
| 7112 | * find_proxy_task() |
| 7113 | * mutex_unlock() |
| 7114 | * lock(&wait_lock); |
| 7115 | * donor(owner) = current->blocked_donor; |
| 7116 | * unlock(&wait_lock); |
| 7117 | * |
| 7118 | * wake_up_q(); |
| 7119 | * ... |
| 7120 | * ttwu_runnable() |
| 7121 | * __task_rq_lock() |
| 7122 | * lock(&wait_lock); |
| 7123 | * owner == p |
| 7124 | * |
| 7125 | * Which leaves us to finish the ttwu_runnable() and make it go. |
| 7126 | * |
| 7127 | * So schedule rq->idle so that ttwu_runnable() can get the rq |
| 7128 | * lock and mark owner as running. |
| 7129 | */ |
| 7130 | goto resched_idle; |
| 7131 | } |
| 7132 | /* |
| 7133 | * OK, now we're absolutely sure @owner is on this |
| 7134 | * rq, therefore holding @rq->lock is sufficient to |
| 7135 | * guarantee its existence, as per ttwu_remote(). |
| 7136 | */ |
| 7137 | owner->blocked_donor = p; |
| 7138 | } |
| 7139 | WARN_ON_ONCE(owner && !owner->on_rq); |
| 7140 | |
| 7141 | if (owner && !sched_cpu_cookie_match(rq, owner)) { |
| 7142 | if (curr_in_chain) |
| 7143 | return proxy_resched_idle(rq); |
| 7144 | p = donor; /* Deactivate the donor, not the runnable owner */ |
| 7145 | clear_task_blocked_on(p, NULL); |
| 7146 | goto deactivate; |
| 7147 | } |
| 7148 | return owner; |
| 7149 | |
| 7150 | resched_idle: |
| 7151 | return proxy_resched_idle(rq); |
| 7152 | deactivate: |
| 7153 | proxy_deactivate(rq, p); |
| 7154 | return NULL; |
| 7155 | migrate_task: |
| 7156 | proxy_migrate_task(rq, rf, p, owner_cpu); |
| 7157 | return NULL; |
| 7158 | } |
| 7159 | #else /* SCHED_PROXY_EXEC */ |
| 7160 | static struct task_struct * |
| 7161 | find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf) |
| 7162 | { |
| 7163 | WARN_ONCE(1, "This should never be called in the !SCHED_PROXY_EXEC case\n"); |
| 7164 | return donor; |
| 7165 | } |
| 7166 | #endif /* SCHED_PROXY_EXEC */ |
| 7167 | |
| 7168 | /* |
| 7169 | * __schedule() is the main scheduler function. |
| 7170 | * |
| 7171 | * The main means of driving the scheduler and thus entering this function are: |
| 7172 | * |
| 7173 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. |
| 7174 | * |
| 7175 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return |
| 7176 | * paths. For example, see arch/x86/entry_64.S. |
| 7177 | * |
| 7178 | * To drive preemption between tasks, the scheduler sets the flag in timer |
| 7179 | * interrupt handler sched_tick(). |
| 7180 | * |
| 7181 | * 3. Wakeups don't really cause entry into schedule(). They add a |
| 7182 | * task to the run-queue and that's it. |
| 7183 | * |
| 7184 | * Now, if the new task added to the run-queue preempts the current |
| 7185 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets |
| 7186 | * called on the nearest possible occasion: |
| 7187 | * |
| 7188 | * - If the kernel is preemptible (CONFIG_PREEMPTION=y): |
| 7189 | * |
| 7190 | * - in syscall or exception context, at the next outmost |
| 7191 | * preempt_enable(). (this might be as soon as the wake_up()'s |
| 7192 | * spin_unlock()!) |
| 7193 | * |
| 7194 | * - in IRQ context, return from interrupt-handler to |
| 7195 | * preemptible context |
| 7196 | * |
| 7197 | * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set) |
| 7198 | * then at the next: |
| 7199 | * |
| 7200 | * - cond_resched() call |
| 7201 | * - explicit schedule() call |
| 7202 | * - return from syscall or exception to user-space |
| 7203 | * - return from interrupt-handler to user-space |
| 7204 | * |
| 7205 | * WARNING: must be called with preemption disabled! |
| 7206 | */ |
| 7207 | static void __sched notrace __schedule(int sched_mode) |
| 7208 | { |
| 7209 | struct task_struct *prev, *next; |
| 7210 | /* |
| 7211 | * On PREEMPT_RT kernel, SM_RTLOCK_WAIT is noted |
| 7212 | * as a preemption by schedule_debug() and RCU. |
| 7213 | */ |
| 7214 | bool preempt = sched_mode > SM_NONE; |
| 7215 | bool is_switch = false; |
| 7216 | unsigned long *switch_count; |
| 7217 | unsigned long prev_state; |
| 7218 | struct rq_flags rf; |
| 7219 | struct rq *rq; |
| 7220 | int cpu; |
| 7221 | |
| 7222 | /* Trace preemptions consistently with task switches */ |
| 7223 | trace_sched_entry_tp(sched_mode == SM_PREEMPT); |
| 7224 | |
| 7225 | cpu = smp_processor_id(); |
| 7226 | rq = cpu_rq(cpu); |
| 7227 | prev = rq->curr; |
| 7228 | |
| 7229 | schedule_debug(prev, preempt); |
| 7230 | |
| 7231 | klp_sched_try_switch(prev); |
| 7232 | |
| 7233 | local_irq_disable(); |
| 7234 | rcu_note_context_switch(preempt); |
| 7235 | migrate_disable_switch(rq, prev); |
| 7236 | |
| 7237 | /* |
| 7238 | * Make sure that signal_pending_state()->signal_pending() below |
| 7239 | * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) |
| 7240 | * done by the caller to avoid the race with signal_wake_up(): |
| 7241 | * |
| 7242 | * __set_current_state(@state) signal_wake_up() |
| 7243 | * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING) |
| 7244 | * wake_up_state(p, state) |
| 7245 | * LOCK rq->lock LOCK p->pi_state |
| 7246 | * smp_mb__after_spinlock() smp_mb__after_spinlock() |
| 7247 | * if (signal_pending_state()) if (p->state & @state) |
| 7248 | * |
| 7249 | * Also, the membarrier system call requires a full memory barrier |
| 7250 | * after coming from user-space, before storing to rq->curr; this |
| 7251 | * barrier matches a full barrier in the proximity of the membarrier |
| 7252 | * system call exit. |
| 7253 | */ |
| 7254 | rq_lock(rq, &rf); |
| 7255 | smp_mb__after_spinlock(); |
| 7256 | |
| 7257 | hrtick_schedule_enter(rq); |
| 7258 | |
| 7259 | /* Promote REQ to ACT */ |
| 7260 | rq->clock_update_flags <<= 1; |
| 7261 | update_rq_clock(rq); |
| 7262 | rq->clock_update_flags = RQCF_UPDATED; |
| 7263 | |
| 7264 | switch_count = &prev->nivcsw; |
| 7265 | |
| 7266 | /* Task state changes only considers SM_PREEMPT as preemption */ |
| 7267 | preempt = sched_mode == SM_PREEMPT; |
| 7268 | |
| 7269 | /* |
| 7270 | * We must load prev->state once (task_struct::state is volatile), such |
| 7271 | * that we form a control dependency vs deactivate_task() below. |
| 7272 | */ |
| 7273 | prev_state = READ_ONCE(prev->__state); |
| 7274 | if (sched_mode == SM_IDLE) { |
| 7275 | /* SCX must consult the BPF scheduler to tell if rq is empty */ |
| 7276 | if (!rq->nr_running && !scx_enabled()) { |
| 7277 | next = prev; |
| 7278 | rq->next_class = &idle_sched_class; |
| 7279 | goto picked; |
| 7280 | } |
| 7281 | } else if (!preempt && prev_state) { |
| 7282 | /* |
| 7283 | * Keep mutex-blocked tasks on the runqueue for proxy execution |
| 7284 | * only when their scheduling class allows it. Without proxy |
| 7285 | * execution, task_is_blocked() always returns false. |
| 7286 | */ |
| 7287 | try_to_block_task(rq, prev, &prev_state, |
| 7288 | !task_is_blocked(prev) || !scx_allow_proxy_exec(prev)); |
| 7289 | switch_count = &prev->nvcsw; |
| 7290 | } |
| 7291 | |
| 7292 | pick_again: |
| 7293 | assert_balance_callbacks_empty(rq); |
| 7294 | next = pick_next_task(rq, &rf); |
| 7295 | rq->next_class = next->sched_class; |
| 7296 | if (sched_proxy_exec()) { |
| 7297 | struct task_struct *prev_donor = rq->donor; |
| 7298 | |
| 7299 | rq_set_donor(rq, next); |
| 7300 | next->blocked_donor = NULL; |
| 7301 | if (unlikely(next->is_blocked)) { |
| 7302 | next = find_proxy_task(rq, next, &rf); |
| 7303 | if (!next) { |
| 7304 | zap_balance_callbacks(rq); |
| 7305 | goto pick_again; |
| 7306 | } |
| 7307 | if (next == rq->idle) { |
| 7308 | zap_balance_callbacks(rq); |
| 7309 | scx_proxy_reenqueue_retry(rq, next); |
| 7310 | goto keep_resched; |
| 7311 | } |
| 7312 | } |
| 7313 | if (rq->donor == prev_donor && prev != next) { |
| 7314 | struct task_struct *donor = rq->donor; |
| 7315 | /* |
| 7316 | * When transitioning like: |
| 7317 | * |
| 7318 | * prev next |
| 7319 | * donor: B B |
| 7320 | * curr: A B or C |
| 7321 | * |
| 7322 | * then put_prev_set_next_task() will not have done |
| 7323 | * anything, since B == B. However, A might have |
| 7324 | * missed a RT/DL balance opportunity due to being |
| 7325 | * on_cpu. |
| 7326 | */ |
| 7327 | donor->sched_class->put_prev_task(rq, donor, donor); |
| 7328 | donor->sched_class->set_next_task(rq, donor, SNT_PICK); |
| 7329 | } |
| 7330 | scx_proxy_donor_start(rq); |
| 7331 | scx_proxy_reenqueue_retry(rq, next); |
| 7332 | } else { |
| 7333 | rq_set_donor(rq, next); |
| 7334 | } |
| 7335 | |
| 7336 | picked: |
| 7337 | clear_tsk_need_resched(prev); |
| 7338 | clear_preempt_need_resched(); |
| 7339 | keep_resched: |
| 7340 | rq->last_seen_need_resched_ns = 0; |
| 7341 | |
| 7342 | is_switch = prev != next; |
| 7343 | if (likely(is_switch)) { |
| 7344 | rq->nr_switches++; |
| 7345 | /* |
| 7346 | * RCU users of rcu_dereference(rq->curr) may not see |
| 7347 | * changes to task_struct made by pick_next_task(). |
| 7348 | */ |
| 7349 | RCU_INIT_POINTER(rq->curr, next); |
| 7350 | |
| 7351 | /* |
| 7352 | * The membarrier system call requires each architecture |
| 7353 | * to have a full memory barrier after updating |
| 7354 | * rq->curr, before returning to user-space. |
| 7355 | * |
| 7356 | * Here are the schemes providing that barrier on the |
| 7357 | * various architectures: |
| 7358 | * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC, |
| 7359 | * RISC-V. switch_mm() relies on membarrier_arch_switch_mm() |
| 7360 | * on PowerPC and on RISC-V. |
| 7361 | * - finish_lock_switch() for weakly-ordered |
| 7362 | * architectures where spin_unlock is a full barrier, |
| 7363 | * - switch_to() for arm64 (weakly-ordered, spin_unlock |
| 7364 | * is a RELEASE barrier), |
| 7365 | * |
| 7366 | * The barrier matches a full barrier in the proximity of |
| 7367 | * the membarrier system call entry. |
| 7368 | * |
| 7369 | * On RISC-V, this barrier pairing is also needed for the |
| 7370 | * SYNC_CORE command when switching between processes, cf. |
| 7371 | * the inline comments in membarrier_arch_switch_mm(). |
| 7372 | */ |
| 7373 | ++*switch_count; |
| 7374 | |
| 7375 | psi_account_irqtime(rq, prev, next); |
| 7376 | psi_sched_switch(prev, next, !task_on_rq_queued(prev) || |
| 7377 | prev->se.sched_delayed); |
| 7378 | |
| 7379 | trace_sched_switch(preempt, prev, next, prev_state); |
| 7380 | |
| 7381 | /* Also unlocks the rq: */ |
| 7382 | rq = context_switch(rq, prev, next, &rf); |
| 7383 | } else { |
| 7384 | rq_unpin_lock(rq, &rf); |
| 7385 | __balance_callbacks(rq, NULL); |
| 7386 | hrtick_schedule_exit(rq); |
| 7387 | raw_spin_rq_unlock_irq(rq); |
| 7388 | } |
| 7389 | trace_sched_exit_tp(is_switch); |
| 7390 | } |
| 7391 | |
| 7392 | void __noreturn do_task_dead(void) |
| 7393 | { |
| 7394 | /* Causes final put_task_struct in finish_task_switch(): */ |
| 7395 | set_special_state(TASK_DEAD); |
| 7396 | |
| 7397 | /* Tell freezer to ignore us: */ |
| 7398 | current->flags |= PF_NOFREEZE; |
| 7399 | |
| 7400 | __schedule(SM_NONE); |
| 7401 | BUG(); |
| 7402 | |
| 7403 | /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */ |
| 7404 | for (;;) |
| 7405 | cpu_relax(); |
| 7406 | } |
| 7407 | |
| 7408 | static inline void sched_submit_work(struct task_struct *tsk) |
| 7409 | { |
| 7410 | static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG); |
| 7411 | unsigned int task_flags; |
| 7412 | |
| 7413 | /* |
| 7414 | * Establish LD_WAIT_CONFIG context to ensure none of the code called |
| 7415 | * will use a blocking primitive -- which would lead to recursion. |
| 7416 | */ |
| 7417 | lock_map_acquire_try(&sched_map); |
| 7418 | |
| 7419 | task_flags = tsk->flags; |
| 7420 | /* |
| 7421 | * If a worker goes to sleep, notify and ask workqueue whether it |
| 7422 | * wants to wake up a task to maintain concurrency. |
| 7423 | */ |
| 7424 | if (task_flags & PF_WQ_WORKER) |
| 7425 | wq_worker_sleeping(tsk); |
| 7426 | else if (task_flags & PF_IO_WORKER) |
| 7427 | io_wq_worker_sleeping(tsk); |
| 7428 | |
| 7429 | /* |
| 7430 | * spinlock and rwlock must not flush block requests. This will |
| 7431 | * deadlock if the callback attempts to acquire a lock which is |
| 7432 | * already acquired. |
| 7433 | */ |
| 7434 | WARN_ON_ONCE(current->__state & TASK_RTLOCK_WAIT); |
| 7435 | |
| 7436 | /* |
| 7437 | * If we are going to sleep and we have plugged IO queued, |
| 7438 | * make sure to submit it to avoid deadlocks. |
| 7439 | */ |
| 7440 | blk_flush_plug(tsk->plug, true); |
| 7441 | |
| 7442 | lock_map_release(&sched_map); |
| 7443 | } |
| 7444 | |
| 7445 | static void sched_update_worker(struct task_struct *tsk) |
| 7446 | { |
| 7447 | if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) { |
| 7448 | if (tsk->flags & PF_WQ_WORKER) |
| 7449 | wq_worker_running(tsk); |
| 7450 | else |
| 7451 | io_wq_worker_running(tsk); |
| 7452 | } |
| 7453 | } |
| 7454 | |
| 7455 | static __always_inline void __schedule_loop(int sched_mode) |
| 7456 | { |
| 7457 | do { |
| 7458 | preempt_disable(); |
| 7459 | __schedule(sched_mode); |
| 7460 | sched_preempt_enable_no_resched(); |
| 7461 | } while (need_resched()); |
| 7462 | } |
| 7463 | |
| 7464 | asmlinkage __visible void __sched schedule(void) |
| 7465 | { |
| 7466 | struct task_struct *tsk = current; |
| 7467 | |
| 7468 | #ifdef CONFIG_RT_MUTEXES |
| 7469 | lockdep_assert(!tsk->sched_rt_mutex); |
| 7470 | #endif |
| 7471 | |
| 7472 | if (!task_is_running(tsk)) |
| 7473 | sched_submit_work(tsk); |
| 7474 | __schedule_loop(SM_NONE); |
| 7475 | sched_update_worker(tsk); |
| 7476 | } |
| 7477 | EXPORT_SYMBOL(schedule); |
| 7478 | |
| 7479 | /* |
| 7480 | * synchronize_rcu_tasks() makes sure that no task is stuck in preempted |
| 7481 | * state (have scheduled out non-voluntarily) by making sure that all |
| 7482 | * tasks have either left the run queue or have gone into user space. |
| 7483 | * As idle tasks do not do either, they must not ever be preempted |
| 7484 | * (schedule out non-voluntarily). |
| 7485 | * |
| 7486 | * schedule_idle() is similar to schedule_preempt_disable() except that it |
| 7487 | * never enables preemption because it does not call sched_submit_work(). |
| 7488 | */ |
| 7489 | void __sched schedule_idle(void) |
| 7490 | { |
| 7491 | /* |
| 7492 | * As this skips calling sched_submit_work(), which the idle task does |
| 7493 | * regardless because that function is a NOP when the task is in a |
| 7494 | * TASK_RUNNING state, make sure this isn't used someplace that the |
| 7495 | * current task can be in any other state. Note, idle is always in the |
| 7496 | * TASK_RUNNING state. |
| 7497 | */ |
| 7498 | WARN_ON_ONCE(current->__state); |
| 7499 | do { |
| 7500 | __schedule(SM_IDLE); |
| 7501 | } while (need_resched()); |
| 7502 | } |
| 7503 | |
| 7504 | #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK) |
| 7505 | asmlinkage __visible void __sched schedule_user(void) |
| 7506 | { |
| 7507 | /* |
| 7508 | * If we come here after a random call to set_need_resched(), |
| 7509 | * or we have been woken up remotely but the IPI has not yet arrived, |
| 7510 | * we haven't yet exited the RCU idle mode. Do it here manually until |
| 7511 | * we find a better solution. |
| 7512 | * |
| 7513 | * NB: There are buggy callers of this function. Ideally we |
| 7514 | * should warn if prev_state != CT_STATE_USER, but that will trigger |
| 7515 | * too frequently to make sense yet. |
| 7516 | */ |
| 7517 | enum ctx_state prev_state = exception_enter(); |
| 7518 | schedule(); |
| 7519 | exception_exit(prev_state); |
| 7520 | } |
| 7521 | #endif |
| 7522 | |
| 7523 | /** |
| 7524 | * schedule_preempt_disabled - called with preemption disabled |
| 7525 | * |
| 7526 | * Returns with preemption disabled. Note: preempt_count must be 1 |
| 7527 | */ |
| 7528 | void __sched schedule_preempt_disabled(void) |
| 7529 | { |
| 7530 | sched_preempt_enable_no_resched(); |
| 7531 | schedule(); |
| 7532 | preempt_disable(); |
| 7533 | } |
| 7534 | |
| 7535 | #ifdef CONFIG_PREEMPT_RT |
| 7536 | void __sched notrace schedule_rtlock(void) |
| 7537 | { |
| 7538 | __schedule_loop(SM_RTLOCK_WAIT); |
| 7539 | } |
| 7540 | NOKPROBE_SYMBOL(schedule_rtlock); |
| 7541 | #endif |
| 7542 | |
| 7543 | static void __sched notrace preempt_schedule_common(void) |
| 7544 | { |
| 7545 | do { |
| 7546 | /* |
| 7547 | * Because the function tracer can trace preempt_count_sub() |
| 7548 | * and it also uses preempt_enable/disable_notrace(), if |
| 7549 | * NEED_RESCHED is set, the preempt_enable_notrace() called |
| 7550 | * by the function tracer will call this function again and |
| 7551 | * cause infinite recursion. |
| 7552 | * |
| 7553 | * Preemption must be disabled here before the function |
| 7554 | * tracer can trace. Break up preempt_disable() into two |
| 7555 | * calls. One to disable preemption without fear of being |
| 7556 | * traced. The other to still record the preemption latency, |
| 7557 | * which can also be traced by the function tracer. |
| 7558 | */ |
| 7559 | preempt_disable_notrace(); |
| 7560 | preempt_latency_start(1); |
| 7561 | __schedule(SM_PREEMPT); |
| 7562 | preempt_latency_stop(1); |
| 7563 | preempt_enable_no_resched_notrace(); |
| 7564 | |
| 7565 | /* |
| 7566 | * Check again in case we missed a preemption opportunity |
| 7567 | * between schedule and now. |
| 7568 | */ |
| 7569 | } while (need_resched()); |
| 7570 | } |
| 7571 | |
| 7572 | #ifdef CONFIG_PREEMPTION |
| 7573 | /* |
| 7574 | * This is the entry point to schedule() from in-kernel preemption |
| 7575 | * off of preempt_enable. |
| 7576 | */ |
| 7577 | asmlinkage __visible void __sched notrace preempt_schedule(void) |
| 7578 | { |
| 7579 | /* |
| 7580 | * If there is a non-zero preempt_count or interrupts are disabled, |
| 7581 | * we do not want to preempt the current task. Just return.. |
| 7582 | */ |
| 7583 | if (likely(!preemptible())) |
| 7584 | return; |
| 7585 | preempt_schedule_common(); |
| 7586 | } |
| 7587 | NOKPROBE_SYMBOL(preempt_schedule); |
| 7588 | EXPORT_SYMBOL(preempt_schedule); |
| 7589 | |
| 7590 | /** |
| 7591 | * preempt_schedule_notrace - preempt_schedule called by tracing |
| 7592 | * |
| 7593 | * The tracing infrastructure uses preempt_enable_notrace to prevent |
| 7594 | * recursion and tracing preempt enabling caused by the tracing |
| 7595 | * infrastructure itself. But as tracing can happen in areas coming |
| 7596 | * from userspace or just about to enter userspace, a preempt enable |
| 7597 | * can occur before user_exit() is called. This will cause the scheduler |
| 7598 | * to be called when the system is still in usermode. |
| 7599 | * |
| 7600 | * To prevent this, the preempt_enable_notrace will use this function |
| 7601 | * instead of preempt_schedule() to exit user context if needed before |
| 7602 | * calling the scheduler. |
| 7603 | */ |
| 7604 | asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) |
| 7605 | { |
| 7606 | enum ctx_state prev_ctx; |
| 7607 | |
| 7608 | if (likely(!preemptible())) |
| 7609 | return; |
| 7610 | |
| 7611 | do { |
| 7612 | /* |
| 7613 | * Because the function tracer can trace preempt_count_sub() |
| 7614 | * and it also uses preempt_enable/disable_notrace(), if |
| 7615 | * NEED_RESCHED is set, the preempt_enable_notrace() called |
| 7616 | * by the function tracer will call this function again and |
| 7617 | * cause infinite recursion. |
| 7618 | * |
| 7619 | * Preemption must be disabled here before the function |
| 7620 | * tracer can trace. Break up preempt_disable() into two |
| 7621 | * calls. One to disable preemption without fear of being |
| 7622 | * traced. The other to still record the preemption latency, |
| 7623 | * which can also be traced by the function tracer. |
| 7624 | */ |
| 7625 | preempt_disable_notrace(); |
| 7626 | preempt_latency_start(1); |
| 7627 | /* |
| 7628 | * Needs preempt disabled in case user_exit() is traced |
| 7629 | * and the tracer calls preempt_enable_notrace() causing |
| 7630 | * an infinite recursion. |
| 7631 | */ |
| 7632 | prev_ctx = exception_enter(); |
| 7633 | __schedule(SM_PREEMPT); |
| 7634 | exception_exit(prev_ctx); |
| 7635 | |
| 7636 | preempt_latency_stop(1); |
| 7637 | preempt_enable_no_resched_notrace(); |
| 7638 | } while (need_resched()); |
| 7639 | } |
| 7640 | EXPORT_SYMBOL_GPL(preempt_schedule_notrace); |
| 7641 | |
| 7642 | #endif /* CONFIG_PREEMPTION */ |
| 7643 | |
| 7644 | /* |
| 7645 | * This is the entry point to schedule() from kernel preemption |
| 7646 | * off of IRQ context. |
| 7647 | * Note, that this is called and return with IRQs disabled. This will |
| 7648 | * protect us against recursive calling from IRQ contexts. |
| 7649 | */ |
| 7650 | asmlinkage __visible void __sched preempt_schedule_irq(void) |
| 7651 | { |
| 7652 | enum ctx_state prev_state; |
| 7653 | |
| 7654 | /* Catch callers which need to be fixed */ |
| 7655 | BUG_ON(preempt_count() || !irqs_disabled()); |
| 7656 | |
| 7657 | prev_state = exception_enter(); |
| 7658 | |
| 7659 | do { |
| 7660 | preempt_disable(); |
| 7661 | local_irq_enable(); |
| 7662 | __schedule(SM_PREEMPT); |
| 7663 | local_irq_disable(); |
| 7664 | sched_preempt_enable_no_resched(); |
| 7665 | } while (need_resched()); |
| 7666 | |
| 7667 | exception_exit(prev_state); |
| 7668 | } |
| 7669 | |
| 7670 | int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags, |
| 7671 | void *key) |
| 7672 | { |
| 7673 | WARN_ON_ONCE(wake_flags & ~(WF_SYNC|WF_CURRENT_CPU)); |
| 7674 | return try_to_wake_up(curr->private, mode, wake_flags); |
| 7675 | } |
| 7676 | EXPORT_SYMBOL(default_wake_function); |
| 7677 | |
| 7678 | const struct sched_class *__setscheduler_class(int policy, int prio) |
| 7679 | { |
| 7680 | if (dl_prio(prio)) |
| 7681 | return &dl_sched_class; |
| 7682 | |
| 7683 | if (rt_prio(prio)) |
| 7684 | return &rt_sched_class; |
| 7685 | |
| 7686 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 7687 | if (task_should_scx(policy)) |
| 7688 | return &ext_sched_class; |
| 7689 | #endif |
| 7690 | |
| 7691 | return &fair_sched_class; |
| 7692 | } |
| 7693 | |
| 7694 | #ifdef CONFIG_RT_MUTEXES |
| 7695 | |
| 7696 | /* |
| 7697 | * Would be more useful with typeof()/auto_type but they don't mix with |
| 7698 | * bit-fields. Since it's a local thing, use int. Keep the generic sounding |
| 7699 | * name such that if someone were to implement this function we get to compare |
| 7700 | * notes. |
| 7701 | */ |
| 7702 | #define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; }) |
| 7703 | |
| 7704 | void rt_mutex_pre_schedule(void) |
| 7705 | { |
| 7706 | lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1)); |
| 7707 | sched_submit_work(current); |
| 7708 | } |
| 7709 | |
| 7710 | /* |
| 7711 | * Used within the futex syscall context, skips sched_submit_work() because none |
| 7712 | * its work will be done. Asserts ensure that it is indeed the case. |
| 7713 | */ |
| 7714 | void rt_mutex_futex_pre_schedule(void) |
| 7715 | { |
| 7716 | lockdep_assert(!(current->flags & (PF_WQ_WORKER | PF_IO_WORKER))); |
| 7717 | lockdep_assert(!current->plug); |
| 7718 | lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1)); |
| 7719 | } |
| 7720 | |
| 7721 | void rt_mutex_schedule(void) |
| 7722 | { |
| 7723 | lockdep_assert(current->sched_rt_mutex); |
| 7724 | __schedule_loop(SM_NONE); |
| 7725 | } |
| 7726 | |
| 7727 | void rt_mutex_post_schedule(void) |
| 7728 | { |
| 7729 | sched_update_worker(current); |
| 7730 | lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0)); |
| 7731 | } |
| 7732 | |
| 7733 | void rt_mutex_futex_post_schedule(void) |
| 7734 | { |
| 7735 | lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0)); |
| 7736 | } |
| 7737 | |
| 7738 | /* |
| 7739 | * rt_mutex_setprio - set the current priority of a task |
| 7740 | * @p: task to boost |
| 7741 | * @pi_task: donor task |
| 7742 | * |
| 7743 | * This function changes the 'effective' priority of a task. It does |
| 7744 | * not touch ->normal_prio like __setscheduler(). |
| 7745 | * |
| 7746 | * Used by the rt_mutex code to implement priority inheritance |
| 7747 | * logic. Call site only calls if the priority of the task changed. |
| 7748 | */ |
| 7749 | void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task) |
| 7750 | { |
| 7751 | int prio, oldprio, queue_flag = |
| 7752 | DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; |
| 7753 | const struct sched_class *prev_class, *next_class; |
| 7754 | struct rq_flags rf; |
| 7755 | struct rq *rq; |
| 7756 | |
| 7757 | /* XXX used to be waiter->prio, not waiter->task->prio */ |
| 7758 | prio = __rt_effective_prio(pi_task, p->normal_prio); |
| 7759 | |
| 7760 | /* |
| 7761 | * If nothing changed; bail early. |
| 7762 | */ |
| 7763 | if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio)) |
| 7764 | return; |
| 7765 | |
| 7766 | rq = __task_rq_lock(p, &rf); |
| 7767 | update_rq_clock(rq); |
| 7768 | /* |
| 7769 | * Set under pi_lock && rq->lock, such that the value can be used under |
| 7770 | * either lock. |
| 7771 | * |
| 7772 | * Note that there is loads of tricky to make this pointer cache work |
| 7773 | * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to |
| 7774 | * ensure a task is de-boosted (pi_task is set to NULL) before the |
| 7775 | * task is allowed to run again (and can exit). This ensures the pointer |
| 7776 | * points to a blocked task -- which guarantees the task is present. |
| 7777 | */ |
| 7778 | p->pi_top_task = pi_task; |
| 7779 | |
| 7780 | /* |
| 7781 | * For FIFO/RR we only need to set prio, if that matches we're done. |
| 7782 | */ |
| 7783 | if (prio == p->prio && !dl_prio(prio)) |
| 7784 | goto out_unlock; |
| 7785 | |
| 7786 | /* |
| 7787 | * Idle task boosting is a no-no in general. There is one |
| 7788 | * exception, when PREEMPT_RT and NOHZ is active: |
| 7789 | * |
| 7790 | * The idle task calls get_next_timer_interrupt() and holds |
| 7791 | * the timer wheel base->lock on the CPU and another CPU wants |
| 7792 | * to access the timer (probably to cancel it). We can safely |
| 7793 | * ignore the boosting request, as the idle CPU runs this code |
| 7794 | * with interrupts disabled and will complete the lock |
| 7795 | * protected section without being interrupted. So there is no |
| 7796 | * real need to boost. |
| 7797 | */ |
| 7798 | if (unlikely(p == rq->idle)) { |
| 7799 | WARN_ON(p != rq->curr); |
| 7800 | WARN_ON(p->pi_blocked_on); |
| 7801 | goto out_unlock; |
| 7802 | } |
| 7803 | |
| 7804 | trace_sched_pi_setprio(p, pi_task); |
| 7805 | oldprio = p->prio; |
| 7806 | |
| 7807 | if (oldprio == prio && !dl_prio(prio)) |
| 7808 | queue_flag &= ~DEQUEUE_MOVE; |
| 7809 | |
| 7810 | prev_class = p->sched_class; |
| 7811 | next_class = __setscheduler_class(p->policy, prio); |
| 7812 | |
| 7813 | if (prev_class != next_class) |
| 7814 | queue_flag |= DEQUEUE_CLASS; |
| 7815 | |
| 7816 | scoped_guard (sched_change, p, queue_flag) { |
| 7817 | /* |
| 7818 | * Boosting condition are: |
| 7819 | * 1. -rt task is running and holds mutex A |
| 7820 | * --> -dl task blocks on mutex A |
| 7821 | * |
| 7822 | * 2. -dl task is running and holds mutex A |
| 7823 | * --> -dl task blocks on mutex A and could preempt the |
| 7824 | * running task |
| 7825 | */ |
| 7826 | if (dl_prio(prio)) { |
| 7827 | if (!dl_prio(p->normal_prio) || |
| 7828 | (pi_task && dl_prio(pi_task->prio) && |
| 7829 | dl_entity_preempt(&pi_task->dl, &p->dl))) { |
| 7830 | p->dl.pi_se = pi_task->dl.pi_se; |
| 7831 | scope->flags |= ENQUEUE_REPLENISH; |
| 7832 | } else { |
| 7833 | p->dl.pi_se = &p->dl; |
| 7834 | } |
| 7835 | } else if (rt_prio(prio)) { |
| 7836 | if (dl_prio(oldprio)) |
| 7837 | p->dl.pi_se = &p->dl; |
| 7838 | if (oldprio < prio) |
| 7839 | scope->flags |= ENQUEUE_HEAD; |
| 7840 | } else { |
| 7841 | if (dl_prio(oldprio)) |
| 7842 | p->dl.pi_se = &p->dl; |
| 7843 | if (rt_prio(oldprio)) |
| 7844 | p->rt.timeout = 0; |
| 7845 | } |
| 7846 | |
| 7847 | p->sched_class = next_class; |
| 7848 | p->prio = prio; |
| 7849 | } |
| 7850 | out_unlock: |
| 7851 | /* Caller holds task_struct::pi_lock, IRQs are still disabled */ |
| 7852 | |
| 7853 | __balance_callbacks(rq, &rf); |
| 7854 | __task_rq_unlock(rq, p, &rf); |
| 7855 | } |
| 7856 | #endif /* CONFIG_RT_MUTEXES */ |
| 7857 | |
| 7858 | #if !defined(CONFIG_PREEMPTION) |
| 7859 | int __sched __cond_resched(void) |
| 7860 | { |
| 7861 | if (should_resched(0) && !irqs_disabled()) { |
| 7862 | preempt_schedule_common(); |
| 7863 | return 1; |
| 7864 | } |
| 7865 | /* |
| 7866 | * In PREEMPT_RCU kernels, ->rcu_read_lock_nesting tells the tick |
| 7867 | * whether the current CPU is in an RCU read-side critical section, |
| 7868 | * so the tick can report quiescent states even for CPUs looping |
| 7869 | * in kernel context. In contrast, in non-preemptible kernels, |
| 7870 | * RCU readers leave no in-memory hints, which means that CPU-bound |
| 7871 | * processes executing in kernel context might never report an |
| 7872 | * RCU quiescent state. Therefore, the following code causes |
| 7873 | * cond_resched() to report a quiescent state, but only when RCU |
| 7874 | * is in urgent need of one. |
| 7875 | * A third case, preemptible, but non-PREEMPT_RCU provides for |
| 7876 | * urgently needed quiescent states via rcu_flavor_sched_clock_irq(). |
| 7877 | */ |
| 7878 | #ifndef CONFIG_PREEMPT_RCU |
| 7879 | rcu_all_qs(); |
| 7880 | #endif |
| 7881 | return 0; |
| 7882 | } |
| 7883 | EXPORT_SYMBOL(__cond_resched); |
| 7884 | #endif |
| 7885 | |
| 7886 | /* |
| 7887 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
| 7888 | * call schedule, and on return reacquire the lock. |
| 7889 | * |
| 7890 | * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level |
| 7891 | * operations here to prevent schedule() from being called twice (once via |
| 7892 | * spin_unlock(), once by hand). |
| 7893 | */ |
| 7894 | int __cond_resched_lock(spinlock_t *lock) |
| 7895 | { |
| 7896 | int resched = should_resched(PREEMPT_LOCK_OFFSET); |
| 7897 | int ret = 0; |
| 7898 | |
| 7899 | lockdep_assert_held(lock); |
| 7900 | |
| 7901 | if (spin_needbreak(lock) || resched) { |
| 7902 | spin_unlock(lock); |
| 7903 | if (!_cond_resched()) |
| 7904 | cpu_relax(); |
| 7905 | ret = 1; |
| 7906 | spin_lock(lock); |
| 7907 | } |
| 7908 | return ret; |
| 7909 | } |
| 7910 | EXPORT_SYMBOL(__cond_resched_lock); |
| 7911 | |
| 7912 | int __cond_resched_rwlock_read(rwlock_t *lock) |
| 7913 | { |
| 7914 | int resched = should_resched(PREEMPT_LOCK_OFFSET); |
| 7915 | int ret = 0; |
| 7916 | |
| 7917 | lockdep_assert_held_read(lock); |
| 7918 | |
| 7919 | if (rwlock_needbreak(lock) || resched) { |
| 7920 | read_unlock(lock); |
| 7921 | if (!_cond_resched()) |
| 7922 | cpu_relax(); |
| 7923 | ret = 1; |
| 7924 | read_lock(lock); |
| 7925 | } |
| 7926 | return ret; |
| 7927 | } |
| 7928 | EXPORT_SYMBOL(__cond_resched_rwlock_read); |
| 7929 | |
| 7930 | int __cond_resched_rwlock_write(rwlock_t *lock) |
| 7931 | { |
| 7932 | int resched = should_resched(PREEMPT_LOCK_OFFSET); |
| 7933 | int ret = 0; |
| 7934 | |
| 7935 | lockdep_assert_held_write(lock); |
| 7936 | |
| 7937 | if (rwlock_needbreak(lock) || resched) { |
| 7938 | write_unlock(lock); |
| 7939 | if (!_cond_resched()) |
| 7940 | cpu_relax(); |
| 7941 | ret = 1; |
| 7942 | write_lock(lock); |
| 7943 | } |
| 7944 | return ret; |
| 7945 | } |
| 7946 | EXPORT_SYMBOL(__cond_resched_rwlock_write); |
| 7947 | |
| 7948 | #ifdef CONFIG_PREEMPT_DYNAMIC |
| 7949 | |
| 7950 | # ifdef CONFIG_GENERIC_IRQ_ENTRY |
| 7951 | # include <linux/irq-entry-common.h> |
| 7952 | # endif |
| 7953 | |
| 7954 | /* |
| 7955 | * NONE: |
| 7956 | * (unselectable) |
| 7957 | * |
| 7958 | * VOLUNTARY: |
| 7959 | * (unselectable) |
| 7960 | * |
| 7961 | * FULL: |
| 7962 | * dynamic_preempt_lazy <- false |
| 7963 | * |
| 7964 | * LAZY: |
| 7965 | * dynamic_preempt_lazy <- true |
| 7966 | */ |
| 7967 | |
| 7968 | enum { |
| 7969 | preempt_dynamic_undefined = -1, |
| 7970 | preempt_dynamic_full, |
| 7971 | preempt_dynamic_lazy, |
| 7972 | }; |
| 7973 | |
| 7974 | int preempt_dynamic_mode = preempt_dynamic_undefined; |
| 7975 | |
| 7976 | int sched_dynamic_mode(const char *str) |
| 7977 | { |
| 7978 | if (!strcmp(str, "full")) |
| 7979 | return preempt_dynamic_full; |
| 7980 | |
| 7981 | if (!strcmp(str, "lazy")) |
| 7982 | return preempt_dynamic_lazy; |
| 7983 | |
| 7984 | return -EINVAL; |
| 7985 | } |
| 7986 | |
| 7987 | # define preempt_dynamic_key_enable(f) static_key_enable(&sk_dynamic_##f.key) |
| 7988 | # define preempt_dynamic_key_disable(f) static_key_disable(&sk_dynamic_##f.key) |
| 7989 | |
| 7990 | static DEFINE_MUTEX(sched_dynamic_mutex); |
| 7991 | |
| 7992 | static void __sched_dynamic_update(int mode) |
| 7993 | { |
| 7994 | switch (mode) { |
| 7995 | case preempt_dynamic_full: |
| 7996 | preempt_dynamic_key_disable(preempt_lazy); |
| 7997 | if (mode != preempt_dynamic_mode) |
| 7998 | pr_info("Dynamic Preempt: full\n"); |
| 7999 | break; |
| 8000 | |
| 8001 | case preempt_dynamic_lazy: |
| 8002 | preempt_dynamic_key_enable(preempt_lazy); |
| 8003 | if (mode != preempt_dynamic_mode) |
| 8004 | pr_info("Dynamic Preempt: lazy\n"); |
| 8005 | break; |
| 8006 | } |
| 8007 | |
| 8008 | WRITE_ONCE(preempt_dynamic_mode, mode); |
| 8009 | } |
| 8010 | |
| 8011 | void sched_dynamic_update(int mode) |
| 8012 | { |
| 8013 | mutex_lock(&sched_dynamic_mutex); |
| 8014 | __sched_dynamic_update(mode); |
| 8015 | mutex_unlock(&sched_dynamic_mutex); |
| 8016 | } |
| 8017 | |
| 8018 | static int __init setup_preempt_mode(char *str) |
| 8019 | { |
| 8020 | int mode = sched_dynamic_mode(str); |
| 8021 | if (mode < 0) { |
| 8022 | pr_warn("Dynamic Preempt: unsupported mode: %s\n", str); |
| 8023 | return 0; |
| 8024 | } |
| 8025 | |
| 8026 | sched_dynamic_update(mode); |
| 8027 | return 1; |
| 8028 | } |
| 8029 | __setup("preempt=", setup_preempt_mode); |
| 8030 | |
| 8031 | static void __init preempt_dynamic_init(void) |
| 8032 | { |
| 8033 | if (preempt_dynamic_mode == preempt_dynamic_undefined) { |
| 8034 | if (IS_ENABLED(CONFIG_PREEMPT_LAZY)) { |
| 8035 | sched_dynamic_update(preempt_dynamic_lazy); |
| 8036 | } else { |
| 8037 | /* Default static call setting, nothing to do */ |
| 8038 | WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT)); |
| 8039 | preempt_dynamic_mode = preempt_dynamic_full; |
| 8040 | pr_info("Dynamic Preempt: full\n"); |
| 8041 | } |
| 8042 | } |
| 8043 | } |
| 8044 | |
| 8045 | # define PREEMPT_MODEL_ACCESSOR(mode) \ |
| 8046 | bool preempt_model_##mode(void) \ |
| 8047 | { \ |
| 8048 | int mode = READ_ONCE(preempt_dynamic_mode); \ |
| 8049 | WARN_ON_ONCE(mode == preempt_dynamic_undefined); \ |
| 8050 | return mode == preempt_dynamic_##mode; \ |
| 8051 | } \ |
| 8052 | EXPORT_SYMBOL_GPL(preempt_model_##mode) |
| 8053 | |
| 8054 | PREEMPT_MODEL_ACCESSOR(full); |
| 8055 | PREEMPT_MODEL_ACCESSOR(lazy); |
| 8056 | |
| 8057 | #else /* !CONFIG_PREEMPT_DYNAMIC: */ |
| 8058 | |
| 8059 | #define preempt_dynamic_mode -1 |
| 8060 | |
| 8061 | static inline void preempt_dynamic_init(void) { } |
| 8062 | |
| 8063 | #endif /* CONFIG_PREEMPT_DYNAMIC */ |
| 8064 | |
| 8065 | const char *preempt_modes[] = { |
| 8066 | "full", "lazy", NULL, |
| 8067 | }; |
| 8068 | |
| 8069 | const char *preempt_model_str(void) |
| 8070 | { |
| 8071 | bool brace = IS_ENABLED(CONFIG_PREEMPT_RT) && |
| 8072 | (IS_ENABLED(CONFIG_PREEMPT_DYNAMIC) || |
| 8073 | IS_ENABLED(CONFIG_PREEMPT_LAZY)); |
| 8074 | static char buf[128]; |
| 8075 | |
| 8076 | if (IS_ENABLED(CONFIG_PREEMPT_BUILD)) { |
| 8077 | struct seq_buf s; |
| 8078 | |
| 8079 | seq_buf_init(&s, buf, sizeof(buf)); |
| 8080 | seq_buf_puts(&s, "PREEMPT"); |
| 8081 | |
| 8082 | if (IS_ENABLED(CONFIG_PREEMPT_RT)) |
| 8083 | seq_buf_printf(&s, "%sRT%s", |
| 8084 | brace ? "_{" : "_", |
| 8085 | brace ? "," : ""); |
| 8086 | |
| 8087 | if (IS_ENABLED(CONFIG_PREEMPT_DYNAMIC)) { |
| 8088 | seq_buf_printf(&s, "(%s)%s", |
| 8089 | preempt_dynamic_mode >= 0 ? |
| 8090 | preempt_modes[preempt_dynamic_mode] : "undef", |
| 8091 | brace ? "}" : ""); |
| 8092 | return seq_buf_str(&s); |
| 8093 | } |
| 8094 | |
| 8095 | if (IS_ENABLED(CONFIG_PREEMPT_LAZY)) { |
| 8096 | seq_buf_printf(&s, "LAZY%s", |
| 8097 | brace ? "}" : ""); |
| 8098 | return seq_buf_str(&s); |
| 8099 | } |
| 8100 | |
| 8101 | return seq_buf_str(&s); |
| 8102 | } |
| 8103 | |
| 8104 | if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY_BUILD)) |
| 8105 | return "VOLUNTARY"; |
| 8106 | |
| 8107 | return "NONE"; |
| 8108 | } |
| 8109 | |
| 8110 | int io_schedule_prepare(void) |
| 8111 | { |
| 8112 | int old_iowait = current->in_iowait; |
| 8113 | |
| 8114 | current->in_iowait = 1; |
| 8115 | blk_flush_plug(current->plug, true); |
| 8116 | return old_iowait; |
| 8117 | } |
| 8118 | |
| 8119 | void io_schedule_finish(int token) |
| 8120 | { |
| 8121 | current->in_iowait = token; |
| 8122 | } |
| 8123 | |
| 8124 | /* |
| 8125 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
| 8126 | * that process accounting knows that this is a task in IO wait state. |
| 8127 | */ |
| 8128 | long __sched io_schedule_timeout(long timeout) |
| 8129 | { |
| 8130 | int token; |
| 8131 | long ret; |
| 8132 | |
| 8133 | token = io_schedule_prepare(); |
| 8134 | ret = schedule_timeout(timeout); |
| 8135 | io_schedule_finish(token); |
| 8136 | |
| 8137 | return ret; |
| 8138 | } |
| 8139 | EXPORT_SYMBOL(io_schedule_timeout); |
| 8140 | |
| 8141 | void __sched io_schedule(void) |
| 8142 | { |
| 8143 | int token; |
| 8144 | |
| 8145 | token = io_schedule_prepare(); |
| 8146 | schedule(); |
| 8147 | io_schedule_finish(token); |
| 8148 | } |
| 8149 | EXPORT_SYMBOL(io_schedule); |
| 8150 | |
| 8151 | void sched_show_task(struct task_struct *p) |
| 8152 | { |
| 8153 | unsigned long free; |
| 8154 | int ppid; |
| 8155 | |
| 8156 | if (!try_get_task_stack(p)) |
| 8157 | return; |
| 8158 | |
| 8159 | pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p)); |
| 8160 | |
| 8161 | if (task_is_running(p)) |
| 8162 | pr_cont(" running task "); |
| 8163 | free = stack_not_used(p); |
| 8164 | ppid = 0; |
| 8165 | rcu_read_lock(); |
| 8166 | if (pid_alive(p)) |
| 8167 | ppid = task_pid_nr(rcu_dereference(p->real_parent)); |
| 8168 | rcu_read_unlock(); |
| 8169 | pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d task_flags:0x%04x flags:0x%08lx\n", |
| 8170 | free, task_pid_nr(p), task_tgid_nr(p), |
| 8171 | ppid, p->flags, read_task_thread_flags(p)); |
| 8172 | |
| 8173 | print_worker_info(KERN_INFO, p); |
| 8174 | print_stop_info(KERN_INFO, p); |
| 8175 | print_scx_info(KERN_INFO, p); |
| 8176 | show_stack(p, NULL, KERN_INFO); |
| 8177 | put_task_stack(p); |
| 8178 | } |
| 8179 | EXPORT_SYMBOL_GPL(sched_show_task); |
| 8180 | |
| 8181 | static inline bool |
| 8182 | state_filter_match(unsigned long state_filter, struct task_struct *p) |
| 8183 | { |
| 8184 | unsigned int state = READ_ONCE(p->__state); |
| 8185 | |
| 8186 | /* no filter, everything matches */ |
| 8187 | if (!state_filter) |
| 8188 | return true; |
| 8189 | |
| 8190 | /* filter, but doesn't match */ |
| 8191 | if (!(state & state_filter)) |
| 8192 | return false; |
| 8193 | |
| 8194 | /* |
| 8195 | * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows |
| 8196 | * TASK_KILLABLE). |
| 8197 | */ |
| 8198 | if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD)) |
| 8199 | return false; |
| 8200 | |
| 8201 | return true; |
| 8202 | } |
| 8203 | |
| 8204 | |
| 8205 | void show_state_filter(unsigned int state_filter) |
| 8206 | { |
| 8207 | struct task_struct *g, *p; |
| 8208 | |
| 8209 | rcu_read_lock(); |
| 8210 | for_each_process_thread(g, p) { |
| 8211 | /* |
| 8212 | * reset the NMI-timeout, listing all files on a slow |
| 8213 | * console might take a lot of time: |
| 8214 | * Also, reset softlockup watchdogs on all CPUs, because |
| 8215 | * another CPU might be blocked waiting for us to process |
| 8216 | * an IPI. |
| 8217 | */ |
| 8218 | touch_nmi_watchdog(); |
| 8219 | touch_all_softlockup_watchdogs(); |
| 8220 | if (state_filter_match(state_filter, p)) |
| 8221 | sched_show_task(p); |
| 8222 | } |
| 8223 | |
| 8224 | if (!state_filter) |
| 8225 | sysrq_sched_debug_show(); |
| 8226 | |
| 8227 | rcu_read_unlock(); |
| 8228 | /* |
| 8229 | * Only show locks if all tasks are dumped: |
| 8230 | */ |
| 8231 | if (!state_filter) |
| 8232 | debug_show_all_locks(); |
| 8233 | } |
| 8234 | |
| 8235 | /** |
| 8236 | * init_idle - set up an idle thread for a given CPU |
| 8237 | * @idle: task in question |
| 8238 | * @cpu: CPU the idle task belongs to |
| 8239 | * |
| 8240 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
| 8241 | * flag, to make booting more robust. |
| 8242 | */ |
| 8243 | void __init init_idle(struct task_struct *idle, int cpu) |
| 8244 | { |
| 8245 | struct affinity_context ac = (struct affinity_context) { |
| 8246 | .new_mask = cpumask_of(cpu), |
| 8247 | .flags = 0, |
| 8248 | }; |
| 8249 | struct rq *rq = cpu_rq(cpu); |
| 8250 | unsigned long flags; |
| 8251 | |
| 8252 | raw_spin_lock_irqsave(&idle->pi_lock, flags); |
| 8253 | raw_spin_rq_lock(rq); |
| 8254 | |
| 8255 | idle->__state = TASK_RUNNING; |
| 8256 | idle->se.exec_start = sched_clock(); |
| 8257 | /* |
| 8258 | * PF_KTHREAD should already be set at this point; regardless, make it |
| 8259 | * look like a proper per-CPU kthread. |
| 8260 | */ |
| 8261 | idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY; |
| 8262 | kthread_set_per_cpu(idle, cpu); |
| 8263 | |
| 8264 | /* |
| 8265 | * No validation and serialization required at boot time and for |
| 8266 | * setting up the idle tasks of not yet online CPUs. |
| 8267 | */ |
| 8268 | set_cpus_allowed_common(idle, &ac); |
| 8269 | /* |
| 8270 | * We're having a chicken and egg problem, even though we are |
| 8271 | * holding rq->lock, the CPU isn't yet set to this CPU so the |
| 8272 | * lockdep check in task_group() will fail. |
| 8273 | * |
| 8274 | * Similar case to sched_fork(). / Alternatively we could |
| 8275 | * use task_rq_lock() here and obtain the other rq->lock. |
| 8276 | * |
| 8277 | * Silence PROVE_RCU |
| 8278 | */ |
| 8279 | rcu_read_lock(); |
| 8280 | __set_task_cpu(idle, cpu); |
| 8281 | rcu_read_unlock(); |
| 8282 | |
| 8283 | rq->idle = idle; |
| 8284 | rq_set_donor(rq, idle); |
| 8285 | rcu_assign_pointer(rq->curr, idle); |
| 8286 | idle->on_rq = TASK_ON_RQ_QUEUED; |
| 8287 | idle->on_cpu = 1; |
| 8288 | raw_spin_rq_unlock(rq); |
| 8289 | raw_spin_unlock_irqrestore(&idle->pi_lock, flags); |
| 8290 | |
| 8291 | /* Set the preempt count _outside_ the spinlocks! */ |
| 8292 | init_idle_preempt_count(idle, cpu); |
| 8293 | |
| 8294 | /* |
| 8295 | * The idle tasks have their own, simple scheduling class: |
| 8296 | */ |
| 8297 | idle->sched_class = &idle_sched_class; |
| 8298 | ftrace_graph_init_idle_task(idle, cpu); |
| 8299 | vtime_init_idle(idle, cpu); |
| 8300 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); |
| 8301 | } |
| 8302 | |
| 8303 | int cpuset_cpumask_can_shrink(const struct cpumask *cur, |
| 8304 | const struct cpumask *trial) |
| 8305 | { |
| 8306 | int ret = 1; |
| 8307 | |
| 8308 | if (cpumask_empty(cur)) |
| 8309 | return ret; |
| 8310 | |
| 8311 | ret = dl_cpuset_cpumask_can_shrink(cur, trial); |
| 8312 | |
| 8313 | return ret; |
| 8314 | } |
| 8315 | |
| 8316 | int task_can_attach(struct task_struct *p) |
| 8317 | { |
| 8318 | int ret = 0; |
| 8319 | |
| 8320 | /* |
| 8321 | * Kthreads which disallow setaffinity shouldn't be moved |
| 8322 | * to a new cpuset; we don't want to change their CPU |
| 8323 | * affinity and isolating such threads by their set of |
| 8324 | * allowed nodes is unnecessary. Thus, cpusets are not |
| 8325 | * applicable for such threads. This prevents checking for |
| 8326 | * success of set_cpus_allowed_ptr() on all attached tasks |
| 8327 | * before cpus_mask may be changed. |
| 8328 | */ |
| 8329 | if (p->flags & PF_NO_SETAFFINITY) |
| 8330 | ret = -EINVAL; |
| 8331 | |
| 8332 | return ret; |
| 8333 | } |
| 8334 | |
| 8335 | bool sched_smp_initialized __read_mostly; |
| 8336 | |
| 8337 | #ifdef CONFIG_NUMA_BALANCING |
| 8338 | /* Migrate current task p to target_cpu */ |
| 8339 | int migrate_task_to(struct task_struct *p, int target_cpu) |
| 8340 | { |
| 8341 | struct migration_arg arg = { p, target_cpu }; |
| 8342 | int curr_cpu = task_cpu(p); |
| 8343 | |
| 8344 | if (curr_cpu == target_cpu) |
| 8345 | return 0; |
| 8346 | |
| 8347 | if (!cpumask_test_cpu(target_cpu, p->cpus_ptr)) |
| 8348 | return -EINVAL; |
| 8349 | |
| 8350 | /* TODO: This is not properly updating schedstats */ |
| 8351 | |
| 8352 | trace_sched_move_numa(p, curr_cpu, target_cpu); |
| 8353 | return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); |
| 8354 | } |
| 8355 | |
| 8356 | /* |
| 8357 | * Requeue a task on a given node and accurately track the number of NUMA |
| 8358 | * tasks on the runqueues |
| 8359 | */ |
| 8360 | void sched_setnuma(struct task_struct *p, int nid) |
| 8361 | { |
| 8362 | guard(task_rq_lock)(p); |
| 8363 | scoped_guard (sched_change, p, DEQUEUE_SAVE) |
| 8364 | p->numa_preferred_nid = nid; |
| 8365 | } |
| 8366 | #endif /* CONFIG_NUMA_BALANCING */ |
| 8367 | |
| 8368 | #ifdef CONFIG_HOTPLUG_CPU |
| 8369 | /* |
| 8370 | * Invoked on the outgoing CPU in context of the CPU hotplug thread |
| 8371 | * after ensuring that there are no user space tasks left on the CPU. |
| 8372 | * |
| 8373 | * If there is a lazy mm in use on the hotplug thread, drop it and |
| 8374 | * switch to init_mm. |
| 8375 | * |
| 8376 | * The reference count on init_mm is dropped in finish_cpu(). |
| 8377 | */ |
| 8378 | static void sched_force_init_mm(void) |
| 8379 | { |
| 8380 | struct mm_struct *mm = current->active_mm; |
| 8381 | |
| 8382 | if (mm != &init_mm) { |
| 8383 | mmgrab_lazy_tlb(&init_mm); |
| 8384 | local_irq_disable(); |
| 8385 | current->active_mm = &init_mm; |
| 8386 | switch_mm_irqs_off(mm, &init_mm, current); |
| 8387 | local_irq_enable(); |
| 8388 | finish_arch_post_lock_switch(); |
| 8389 | mmdrop_lazy_tlb(mm); |
| 8390 | } |
| 8391 | |
| 8392 | /* finish_cpu(), as ran on the BP, will clean up the active_mm state */ |
| 8393 | } |
| 8394 | |
| 8395 | static int __balance_push_cpu_stop(void *arg) |
| 8396 | { |
| 8397 | struct task_struct *p = arg; |
| 8398 | struct rq *rq = this_rq(); |
| 8399 | struct rq_flags rf; |
| 8400 | int cpu; |
| 8401 | |
| 8402 | scoped_guard (raw_spinlock_irq, &p->pi_lock) { |
| 8403 | /* |
| 8404 | * We may change the underlying rq, but the locks held will |
| 8405 | * appropriately be "transferred" when switching. |
| 8406 | */ |
| 8407 | context_unsafe_alias(rq); |
| 8408 | |
| 8409 | cpu = select_fallback_rq(rq->cpu, p); |
| 8410 | |
| 8411 | rq_lock(rq, &rf); |
| 8412 | update_rq_clock(rq); |
| 8413 | if (task_rq(p) == rq && task_on_rq_queued(p)) |
| 8414 | rq = __migrate_task(rq, &rf, p, cpu); |
| 8415 | rq_unlock(rq, &rf); |
| 8416 | } |
| 8417 | |
| 8418 | put_task_struct(p); |
| 8419 | |
| 8420 | return 0; |
| 8421 | } |
| 8422 | |
| 8423 | static DEFINE_PER_CPU(struct cpu_stop_work, push_work); |
| 8424 | |
| 8425 | /* |
| 8426 | * Ensure we only run per-cpu kthreads once the CPU goes !active. |
| 8427 | * |
| 8428 | * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only |
| 8429 | * effective when the hotplug motion is down. |
| 8430 | */ |
| 8431 | static void balance_push(struct rq *rq) |
| 8432 | __must_hold(__rq_lockp(rq)) |
| 8433 | { |
| 8434 | struct task_struct *push_task = rq->curr; |
| 8435 | |
| 8436 | lockdep_assert_rq_held(rq); |
| 8437 | |
| 8438 | /* |
| 8439 | * Ensure the thing is persistent until balance_push_set(.on = false); |
| 8440 | */ |
| 8441 | rq->balance_callback = &balance_push_callback; |
| 8442 | |
| 8443 | /* |
| 8444 | * Only active while going offline and when invoked on the outgoing |
| 8445 | * CPU. |
| 8446 | */ |
| 8447 | if (!cpu_dying(rq->cpu) || rq != this_rq()) |
| 8448 | return; |
| 8449 | |
| 8450 | /* |
| 8451 | * Both the cpu-hotplug and stop task are in this case and are |
| 8452 | * required to complete the hotplug process. |
| 8453 | */ |
| 8454 | if (kthread_is_per_cpu(push_task) || |
| 8455 | is_migration_disabled(push_task)) { |
| 8456 | |
| 8457 | /* |
| 8458 | * If this is the idle task on the outgoing CPU try to wake |
| 8459 | * up the hotplug control thread which might wait for the |
| 8460 | * last task to vanish. The rcuwait_active() check is |
| 8461 | * accurate here because the waiter is pinned on this CPU |
| 8462 | * and can't obviously be running in parallel. |
| 8463 | * |
| 8464 | * On RT kernels this also has to check whether there are |
| 8465 | * pinned and scheduled out tasks on the runqueue. They |
| 8466 | * need to leave the migrate disabled section first. |
| 8467 | */ |
| 8468 | if (!rq->nr_running && !rq_has_pinned_tasks(rq) && |
| 8469 | rcuwait_active(&rq->hotplug_wait)) { |
| 8470 | raw_spin_rq_unlock(rq); |
| 8471 | rcuwait_wake_up(&rq->hotplug_wait); |
| 8472 | raw_spin_rq_lock(rq); |
| 8473 | } |
| 8474 | return; |
| 8475 | } |
| 8476 | |
| 8477 | get_task_struct(push_task); |
| 8478 | /* |
| 8479 | * Temporarily drop rq->lock such that we can wake-up the stop task. |
| 8480 | * Both preemption and IRQs are still disabled. |
| 8481 | */ |
| 8482 | preempt_disable(); |
| 8483 | raw_spin_rq_unlock(rq); |
| 8484 | stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task, |
| 8485 | this_cpu_ptr(&push_work)); |
| 8486 | preempt_enable(); |
| 8487 | /* |
| 8488 | * At this point need_resched() is true and we'll take the loop in |
| 8489 | * schedule(). The next pick is obviously going to be the stop task |
| 8490 | * which kthread_is_per_cpu() and will push this task away. |
| 8491 | */ |
| 8492 | raw_spin_rq_lock(rq); |
| 8493 | } |
| 8494 | |
| 8495 | static void balance_push_set(int cpu, bool on) |
| 8496 | { |
| 8497 | struct rq *rq = cpu_rq(cpu); |
| 8498 | struct rq_flags rf; |
| 8499 | |
| 8500 | rq_lock_irqsave(rq, &rf); |
| 8501 | if (on) { |
| 8502 | WARN_ON_ONCE(rq->balance_callback); |
| 8503 | rq->balance_callback = &balance_push_callback; |
| 8504 | } else if (rq->balance_callback == &balance_push_callback) { |
| 8505 | rq->balance_callback = NULL; |
| 8506 | } |
| 8507 | rq_unlock_irqrestore(rq, &rf); |
| 8508 | } |
| 8509 | |
| 8510 | /* |
| 8511 | * Invoked from a CPUs hotplug control thread after the CPU has been marked |
| 8512 | * inactive. All tasks which are not per CPU kernel threads are either |
| 8513 | * pushed off this CPU now via balance_push() or placed on a different CPU |
| 8514 | * during wakeup. Wait until the CPU is quiescent. |
| 8515 | */ |
| 8516 | static void balance_hotplug_wait(void) |
| 8517 | { |
| 8518 | struct rq *rq = this_rq(); |
| 8519 | |
| 8520 | rcuwait_wait_event(&rq->hotplug_wait, |
| 8521 | rq->nr_running == 1 && !rq_has_pinned_tasks(rq), |
| 8522 | TASK_UNINTERRUPTIBLE); |
| 8523 | } |
| 8524 | |
| 8525 | #else /* !CONFIG_HOTPLUG_CPU: */ |
| 8526 | |
| 8527 | static inline void balance_push(struct rq *rq) |
| 8528 | { |
| 8529 | } |
| 8530 | |
| 8531 | static inline void balance_push_set(int cpu, bool on) |
| 8532 | { |
| 8533 | } |
| 8534 | |
| 8535 | static inline void balance_hotplug_wait(void) |
| 8536 | { |
| 8537 | } |
| 8538 | |
| 8539 | #endif /* !CONFIG_HOTPLUG_CPU */ |
| 8540 | |
| 8541 | void set_rq_online(struct rq *rq) |
| 8542 | { |
| 8543 | if (!rq->online) { |
| 8544 | const struct sched_class *class; |
| 8545 | |
| 8546 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
| 8547 | rq->online = 1; |
| 8548 | |
| 8549 | for_each_class(class) { |
| 8550 | if (class->rq_online) |
| 8551 | class->rq_online(rq); |
| 8552 | } |
| 8553 | } |
| 8554 | } |
| 8555 | |
| 8556 | void set_rq_offline(struct rq *rq) |
| 8557 | { |
| 8558 | if (rq->online) { |
| 8559 | const struct sched_class *class; |
| 8560 | |
| 8561 | update_rq_clock(rq); |
| 8562 | for_each_class(class) { |
| 8563 | if (class->rq_offline) |
| 8564 | class->rq_offline(rq); |
| 8565 | } |
| 8566 | |
| 8567 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
| 8568 | rq->online = 0; |
| 8569 | } |
| 8570 | } |
| 8571 | |
| 8572 | static inline void sched_set_rq_online(struct rq *rq, int cpu) |
| 8573 | { |
| 8574 | struct rq_flags rf; |
| 8575 | |
| 8576 | rq_lock_irqsave(rq, &rf); |
| 8577 | if (rq->rd) { |
| 8578 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 8579 | set_rq_online(rq); |
| 8580 | } |
| 8581 | rq_unlock_irqrestore(rq, &rf); |
| 8582 | } |
| 8583 | |
| 8584 | static inline void sched_set_rq_offline(struct rq *rq, int cpu) |
| 8585 | { |
| 8586 | struct rq_flags rf; |
| 8587 | |
| 8588 | rq_lock_irqsave(rq, &rf); |
| 8589 | if (rq->rd) { |
| 8590 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
| 8591 | set_rq_offline(rq); |
| 8592 | } |
| 8593 | rq_unlock_irqrestore(rq, &rf); |
| 8594 | } |
| 8595 | |
| 8596 | /* |
| 8597 | * used to mark begin/end of suspend/resume: |
| 8598 | */ |
| 8599 | static int num_cpus_frozen; |
| 8600 | |
| 8601 | /* |
| 8602 | * Update cpusets according to cpu_active mask. If cpusets are |
| 8603 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper |
| 8604 | * around partition_sched_domains(). |
| 8605 | * |
| 8606 | * If we come here as part of a suspend/resume, don't touch cpusets because we |
| 8607 | * want to restore it back to its original state upon resume anyway. |
| 8608 | */ |
| 8609 | static void cpuset_cpu_active(void) |
| 8610 | { |
| 8611 | if (cpuhp_tasks_frozen) { |
| 8612 | /* |
| 8613 | * num_cpus_frozen tracks how many CPUs are involved in suspend |
| 8614 | * resume sequence. As long as this is not the last online |
| 8615 | * operation in the resume sequence, just build a single sched |
| 8616 | * domain, ignoring cpusets. |
| 8617 | */ |
| 8618 | cpuset_reset_sched_domains(); |
| 8619 | if (--num_cpus_frozen) |
| 8620 | return; |
| 8621 | /* |
| 8622 | * This is the last CPU online operation. So fall through and |
| 8623 | * restore the original sched domains by considering the |
| 8624 | * cpuset configurations. |
| 8625 | */ |
| 8626 | cpuset_force_rebuild(); |
| 8627 | } |
| 8628 | cpuset_update_active_cpus(); |
| 8629 | } |
| 8630 | |
| 8631 | static void cpuset_cpu_inactive(unsigned int cpu) |
| 8632 | { |
| 8633 | if (!cpuhp_tasks_frozen) { |
| 8634 | cpuset_update_active_cpus(); |
| 8635 | } else { |
| 8636 | num_cpus_frozen++; |
| 8637 | cpuset_reset_sched_domains(); |
| 8638 | } |
| 8639 | } |
| 8640 | |
| 8641 | static inline void sched_smt_present_inc(int cpu) |
| 8642 | { |
| 8643 | if (cpumask_weight(cpu_smt_mask(cpu)) == 2) |
| 8644 | static_branch_inc_cpuslocked(&sched_smt_present); |
| 8645 | } |
| 8646 | |
| 8647 | static inline void sched_smt_present_dec(int cpu) |
| 8648 | { |
| 8649 | if (cpumask_weight(cpu_smt_mask(cpu)) == 2) |
| 8650 | static_branch_dec_cpuslocked(&sched_smt_present); |
| 8651 | } |
| 8652 | |
| 8653 | int sched_cpu_activate(unsigned int cpu) |
| 8654 | { |
| 8655 | struct rq *rq = cpu_rq(cpu); |
| 8656 | |
| 8657 | /* |
| 8658 | * Clear the balance_push callback and prepare to schedule |
| 8659 | * regular tasks. |
| 8660 | */ |
| 8661 | balance_push_set(cpu, false); |
| 8662 | |
| 8663 | /* |
| 8664 | * When going up, increment the number of cores with SMT present. |
| 8665 | */ |
| 8666 | sched_smt_present_inc(cpu); |
| 8667 | set_cpu_active(cpu, true); |
| 8668 | |
| 8669 | if (sched_smp_initialized) { |
| 8670 | sched_update_numa(cpu, true); |
| 8671 | sched_domains_numa_masks_set(cpu); |
| 8672 | cpuset_cpu_active(); |
| 8673 | } |
| 8674 | |
| 8675 | scx_rq_activate(rq); |
| 8676 | |
| 8677 | /* |
| 8678 | * Put the rq online, if not already. This happens: |
| 8679 | * |
| 8680 | * 1) In the early boot process, because we build the real domains |
| 8681 | * after all CPUs have been brought up. |
| 8682 | * |
| 8683 | * 2) At runtime, if cpuset_cpu_active() fails to rebuild the |
| 8684 | * domains. |
| 8685 | */ |
| 8686 | sched_set_rq_online(rq, cpu); |
| 8687 | |
| 8688 | /* preferred is subset of active and follows its state */ |
| 8689 | set_cpu_preferred(cpu, true); |
| 8690 | |
| 8691 | return 0; |
| 8692 | } |
| 8693 | |
| 8694 | int sched_cpu_deactivate(unsigned int cpu) |
| 8695 | { |
| 8696 | struct rq *rq = cpu_rq(cpu); |
| 8697 | int ret; |
| 8698 | |
| 8699 | ret = dl_bw_deactivate(cpu); |
| 8700 | |
| 8701 | if (ret) |
| 8702 | return ret; |
| 8703 | |
| 8704 | set_cpu_preferred(cpu, false); |
| 8705 | |
| 8706 | /* |
| 8707 | * Remove CPU from nohz.idle_cpus_mask to prevent participating in |
| 8708 | * load balancing when not active |
| 8709 | */ |
| 8710 | scoped_guard (rcu) |
| 8711 | nohz_balance_exit_idle(rq); |
| 8712 | |
| 8713 | set_cpu_active(cpu, false); |
| 8714 | |
| 8715 | /* |
| 8716 | * From this point forward, this CPU will refuse to run any task that |
| 8717 | * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively |
| 8718 | * push those tasks away until this gets cleared, see |
| 8719 | * sched_cpu_dying(). |
| 8720 | */ |
| 8721 | balance_push_set(cpu, true); |
| 8722 | |
| 8723 | /* |
| 8724 | * We've cleared cpu_active_mask / set balance_push, wait for all |
| 8725 | * preempt-disabled and RCU users of this state to go away such that |
| 8726 | * all new such users will observe it. |
| 8727 | * |
| 8728 | * Specifically, we rely on ttwu to no longer target this CPU, see |
| 8729 | * ttwu_queue_cond() and is_cpu_allowed(). |
| 8730 | * |
| 8731 | * Do sync before park smpboot threads to take care the RCU boost case. |
| 8732 | */ |
| 8733 | synchronize_rcu(); |
| 8734 | |
| 8735 | sched_domains_free_llc_id(cpu); |
| 8736 | |
| 8737 | sched_set_rq_offline(rq, cpu); |
| 8738 | |
| 8739 | scx_rq_deactivate(rq); |
| 8740 | |
| 8741 | /* |
| 8742 | * When going down, decrement the number of cores with SMT present. |
| 8743 | */ |
| 8744 | sched_smt_present_dec(cpu); |
| 8745 | |
| 8746 | sched_core_cpu_deactivate(cpu); |
| 8747 | |
| 8748 | if (!sched_smp_initialized) |
| 8749 | return 0; |
| 8750 | |
| 8751 | sched_update_numa(cpu, false); |
| 8752 | cpuset_cpu_inactive(cpu); |
| 8753 | sched_domains_numa_masks_clear(cpu); |
| 8754 | return 0; |
| 8755 | } |
| 8756 | |
| 8757 | static void sched_rq_cpu_starting(unsigned int cpu) |
| 8758 | { |
| 8759 | struct rq *rq = cpu_rq(cpu); |
| 8760 | |
| 8761 | rq->calc_load_update = calc_load_update; |
| 8762 | update_max_interval(); |
| 8763 | } |
| 8764 | |
| 8765 | int sched_cpu_starting(unsigned int cpu) |
| 8766 | { |
| 8767 | sched_core_cpu_starting(cpu); |
| 8768 | sched_rq_cpu_starting(cpu); |
| 8769 | sched_tick_start(cpu); |
| 8770 | return 0; |
| 8771 | } |
| 8772 | |
| 8773 | #ifdef CONFIG_HOTPLUG_CPU |
| 8774 | |
| 8775 | /* |
| 8776 | * Invoked immediately before the stopper thread is invoked to bring the |
| 8777 | * CPU down completely. At this point all per CPU kthreads except the |
| 8778 | * hotplug thread (current) and the stopper thread (inactive) have been |
| 8779 | * either parked or have been unbound from the outgoing CPU. Ensure that |
| 8780 | * any of those which might be on the way out are gone. |
| 8781 | * |
| 8782 | * If after this point a bound task is being woken on this CPU then the |
| 8783 | * responsible hotplug callback has failed to do it's job. |
| 8784 | * sched_cpu_dying() will catch it with the appropriate fireworks. |
| 8785 | */ |
| 8786 | int sched_cpu_wait_empty(unsigned int cpu) |
| 8787 | { |
| 8788 | balance_hotplug_wait(); |
| 8789 | sched_force_init_mm(); |
| 8790 | return 0; |
| 8791 | } |
| 8792 | |
| 8793 | /* |
| 8794 | * Since this CPU is going 'away' for a while, fold any nr_active delta we |
| 8795 | * might have. Called from the CPU stopper task after ensuring that the |
| 8796 | * stopper is the last running task on the CPU, so nr_active count is |
| 8797 | * stable. We need to take the tear-down thread which is calling this into |
| 8798 | * account, so we hand in adjust = 1 to the load calculation. |
| 8799 | * |
| 8800 | * Also see the comment "Global load-average calculations". |
| 8801 | */ |
| 8802 | static void calc_load_migrate(struct rq *rq) |
| 8803 | { |
| 8804 | long delta = calc_load_fold_active(rq, 1); |
| 8805 | |
| 8806 | if (delta) |
| 8807 | atomic_long_add(delta, &calc_load_tasks); |
| 8808 | } |
| 8809 | |
| 8810 | static void dump_rq_tasks(struct rq *rq, const char *loglvl) |
| 8811 | { |
| 8812 | struct task_struct *g, *p; |
| 8813 | int cpu = cpu_of(rq); |
| 8814 | |
| 8815 | lockdep_assert_rq_held(rq); |
| 8816 | |
| 8817 | printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running); |
| 8818 | for_each_process_thread(g, p) { |
| 8819 | if (task_cpu(p) != cpu) |
| 8820 | continue; |
| 8821 | |
| 8822 | if (!task_on_rq_queued(p)) |
| 8823 | continue; |
| 8824 | |
| 8825 | printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm); |
| 8826 | } |
| 8827 | } |
| 8828 | |
| 8829 | int sched_cpu_dying(unsigned int cpu) |
| 8830 | { |
| 8831 | struct rq *rq = cpu_rq(cpu); |
| 8832 | struct rq_flags rf; |
| 8833 | |
| 8834 | /* Handle pending wakeups and then migrate everything off */ |
| 8835 | sched_tick_stop(cpu); |
| 8836 | |
| 8837 | rq_lock_irqsave(rq, &rf); |
| 8838 | update_rq_clock(rq); |
| 8839 | if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) { |
| 8840 | WARN(true, "Dying CPU not properly vacated!"); |
| 8841 | dump_rq_tasks(rq, KERN_WARNING); |
| 8842 | } |
| 8843 | dl_server_stop(&rq->fair_server); |
| 8844 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 8845 | dl_server_stop(&rq->ext_server); |
| 8846 | #endif |
| 8847 | rq_unlock_irqrestore(rq, &rf); |
| 8848 | |
| 8849 | calc_load_migrate(rq); |
| 8850 | update_max_interval(); |
| 8851 | hrtick_clear(rq); |
| 8852 | sched_core_cpu_dying(cpu); |
| 8853 | return 0; |
| 8854 | } |
| 8855 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 8856 | |
| 8857 | void __init sched_init_smp(void) |
| 8858 | { |
| 8859 | sched_init_numa(NUMA_NO_NODE); |
| 8860 | |
| 8861 | prandom_init_once(&sched_rnd_state); |
| 8862 | |
| 8863 | /* |
| 8864 | * There's no userspace yet to cause hotplug operations; hence all the |
| 8865 | * CPU masks are stable and all blatant races in the below code cannot |
| 8866 | * happen. |
| 8867 | */ |
| 8868 | sched_domains_mutex_lock(); |
| 8869 | sched_init_domains(cpu_active_mask); |
| 8870 | sched_domains_mutex_unlock(); |
| 8871 | |
| 8872 | /* Move init over to a non-isolated CPU */ |
| 8873 | if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0) |
| 8874 | BUG(); |
| 8875 | current->flags &= ~PF_NO_SETAFFINITY; |
| 8876 | sched_init_granularity(); |
| 8877 | |
| 8878 | init_sched_rt_class(); |
| 8879 | init_sched_dl_class(); |
| 8880 | |
| 8881 | sched_init_dl_servers(); |
| 8882 | |
| 8883 | sched_smp_initialized = true; |
| 8884 | } |
| 8885 | |
| 8886 | static int __init migration_init(void) |
| 8887 | { |
| 8888 | sched_cpu_starting(smp_processor_id()); |
| 8889 | return 0; |
| 8890 | } |
| 8891 | early_initcall(migration_init); |
| 8892 | |
| 8893 | int in_sched_functions(unsigned long addr) |
| 8894 | { |
| 8895 | return in_lock_functions(addr) || |
| 8896 | (addr >= (unsigned long)__sched_text_start |
| 8897 | && addr < (unsigned long)__sched_text_end); |
| 8898 | } |
| 8899 | |
| 8900 | #ifdef CONFIG_CGROUP_SCHED |
| 8901 | /* |
| 8902 | * Default task group. |
| 8903 | * Every task in system belongs to this group at bootup. |
| 8904 | */ |
| 8905 | struct task_group root_task_group; |
| 8906 | LIST_HEAD(task_groups); |
| 8907 | |
| 8908 | /* Cacheline aligned slab cache for task_group */ |
| 8909 | static struct kmem_cache *task_group_cache __ro_after_init; |
| 8910 | #endif |
| 8911 | |
| 8912 | void __init sched_init(void) |
| 8913 | { |
| 8914 | unsigned long __maybe_unused ptr = 0; |
| 8915 | int i; |
| 8916 | |
| 8917 | /* Make sure the linker didn't screw up */ |
| 8918 | BUG_ON(!sched_class_above(&stop_sched_class, &dl_sched_class)); |
| 8919 | BUG_ON(!sched_class_above(&dl_sched_class, &rt_sched_class)); |
| 8920 | BUG_ON(!sched_class_above(&rt_sched_class, &fair_sched_class)); |
| 8921 | BUG_ON(!sched_class_above(&fair_sched_class, &idle_sched_class)); |
| 8922 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 8923 | BUG_ON(!sched_class_above(&fair_sched_class, &ext_sched_class)); |
| 8924 | BUG_ON(!sched_class_above(&ext_sched_class, &idle_sched_class)); |
| 8925 | #endif |
| 8926 | |
| 8927 | wait_bit_init(); |
| 8928 | |
| 8929 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 8930 | root_task_group.cfs_rq = &runqueues.cfs; |
| 8931 | |
| 8932 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; |
| 8933 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL); |
| 8934 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 8935 | #ifdef CONFIG_EXT_GROUP_SCHED |
| 8936 | scx_tg_init(&root_task_group); |
| 8937 | #endif /* CONFIG_EXT_GROUP_SCHED */ |
| 8938 | #ifdef CONFIG_RT_GROUP_SCHED |
| 8939 | ptr += 2 * nr_cpu_ids * sizeof(void **); |
| 8940 | ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT); |
| 8941 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
| 8942 | ptr += nr_cpu_ids * sizeof(void **); |
| 8943 | |
| 8944 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
| 8945 | ptr += nr_cpu_ids * sizeof(void **); |
| 8946 | |
| 8947 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 8948 | |
| 8949 | init_defrootdomain(); |
| 8950 | |
| 8951 | #ifdef CONFIG_RT_GROUP_SCHED |
| 8952 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
| 8953 | global_rt_period(), global_rt_runtime()); |
| 8954 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 8955 | |
| 8956 | #ifdef CONFIG_CGROUP_SCHED |
| 8957 | task_group_cache = KMEM_CACHE(task_group, 0); |
| 8958 | |
| 8959 | list_add(&root_task_group.list, &task_groups); |
| 8960 | INIT_LIST_HEAD(&root_task_group.children); |
| 8961 | INIT_LIST_HEAD(&root_task_group.siblings); |
| 8962 | autogroup_init(&init_task); |
| 8963 | #endif /* CONFIG_CGROUP_SCHED */ |
| 8964 | |
| 8965 | for_each_possible_cpu(i) { |
| 8966 | struct rq *rq; |
| 8967 | |
| 8968 | rq = cpu_rq(i); |
| 8969 | raw_spin_lock_init(&rq->__lock); |
| 8970 | rq->nr_running = 0; |
| 8971 | rq->calc_load_active = 0; |
| 8972 | rq->calc_load_update = jiffies + LOAD_FREQ; |
| 8973 | init_cfs_rq(&rq->cfs); |
| 8974 | init_rt_rq(&rq->rt); |
| 8975 | init_dl_rq(&rq->dl); |
| 8976 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 8977 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
| 8978 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; |
| 8979 | /* |
| 8980 | * How much CPU bandwidth does root_task_group get? |
| 8981 | * |
| 8982 | * In case of task-groups formed through the cgroup filesystem, it |
| 8983 | * gets 100% of the CPU resources in the system. This overall |
| 8984 | * system CPU resource is divided among the tasks of |
| 8985 | * root_task_group and its child task-groups in a fair manner, |
| 8986 | * based on each entity's (task or task-group's) weight |
| 8987 | * (se->load.weight). |
| 8988 | * |
| 8989 | * In other words, if root_task_group has 10 tasks of weight |
| 8990 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
| 8991 | * then A0's share of the CPU resource is: |
| 8992 | * |
| 8993 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
| 8994 | * |
| 8995 | * We achieve this by letting root_task_group's tasks sit |
| 8996 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). |
| 8997 | */ |
| 8998 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); |
| 8999 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 9000 | |
| 9001 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9002 | /* |
| 9003 | * This is required for init cpu because rt.c:__enable_runtime() |
| 9004 | * starts working after scheduler_running, which is not the case |
| 9005 | * yet. |
| 9006 | */ |
| 9007 | rq->rt.rt_runtime = global_rt_runtime(); |
| 9008 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); |
| 9009 | #endif |
| 9010 | rq->next_class = &idle_sched_class; |
| 9011 | |
| 9012 | rq->sd = NULL; |
| 9013 | rq->rd = NULL; |
| 9014 | rq->cpu_capacity = SCHED_CAPACITY_SCALE; |
| 9015 | rq->balance_callback = &balance_push_callback; |
| 9016 | rq->active_balance = 0; |
| 9017 | rq->next_balance = jiffies; |
| 9018 | rq->push_cpu = 0; |
| 9019 | rq->cpu = i; |
| 9020 | rq->online = 0; |
| 9021 | rq->idle_stamp = 0; |
| 9022 | rq->avg_idle = 2*sysctl_sched_migration_cost; |
| 9023 | rq->max_idle_balance_cost = sysctl_sched_migration_cost; |
| 9024 | |
| 9025 | INIT_LIST_HEAD(&rq->cfs_tasks); |
| 9026 | |
| 9027 | rq_attach_root(rq, &def_root_domain); |
| 9028 | #ifdef CONFIG_NO_HZ_COMMON |
| 9029 | rq->last_blocked_load_update_tick = jiffies; |
| 9030 | atomic_set(&rq->nohz_flags, 0); |
| 9031 | |
| 9032 | INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq); |
| 9033 | #endif |
| 9034 | #ifdef CONFIG_HOTPLUG_CPU |
| 9035 | rcuwait_init(&rq->hotplug_wait); |
| 9036 | #endif |
| 9037 | hrtick_rq_init(rq); |
| 9038 | atomic_set(&rq->nr_iowait, 0); |
| 9039 | fair_server_init(rq); |
| 9040 | #ifdef CONFIG_SCHED_CLASS_EXT |
| 9041 | ext_server_init(rq); |
| 9042 | #endif |
| 9043 | |
| 9044 | #ifdef CONFIG_SCHED_CORE |
| 9045 | rq->core = rq; |
| 9046 | rq->core_pick = NULL; |
| 9047 | rq->core_dl_server = NULL; |
| 9048 | rq->core_enabled = 0; |
| 9049 | rq->core_tree = RB_ROOT; |
| 9050 | rq->core_forceidle_count = 0; |
| 9051 | rq->core_forceidle_occupation = 0; |
| 9052 | rq->core_forceidle_start = 0; |
| 9053 | rq->core_pick_in_flight = 0; |
| 9054 | |
| 9055 | rq->core_cookie = 0UL; |
| 9056 | #endif |
| 9057 | #ifdef CONFIG_SCHED_CACHE |
| 9058 | raw_spin_lock_init(&rq->cpu_epoch_lock); |
| 9059 | rq->cpu_epoch_next = jiffies; |
| 9060 | #endif |
| 9061 | |
| 9062 | zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i)); |
| 9063 | } |
| 9064 | |
| 9065 | set_load_weight(&init_task, false); |
| 9066 | init_task.se.slice = sysctl_sched_base_slice, |
| 9067 | |
| 9068 | /* |
| 9069 | * The boot idle thread does lazy MMU switching as well: |
| 9070 | */ |
| 9071 | mmgrab_lazy_tlb(&init_mm); |
| 9072 | enter_lazy_tlb(&init_mm, current); |
| 9073 | |
| 9074 | /* |
| 9075 | * The idle task doesn't need the kthread struct to function, but it |
| 9076 | * is dressed up as a per-CPU kthread and thus needs to play the part |
| 9077 | * if we want to avoid special-casing it in code that deals with per-CPU |
| 9078 | * kthreads. |
| 9079 | */ |
| 9080 | WARN_ON(!set_kthread_struct(current)); |
| 9081 | |
| 9082 | /* |
| 9083 | * Make us the idle thread. Technically, schedule() should not be |
| 9084 | * called from this thread, however somewhere below it might be, |
| 9085 | * but because we are the idle thread, we just pick up running again |
| 9086 | * when this runqueue becomes "idle". |
| 9087 | */ |
| 9088 | __sched_fork(0, current); |
| 9089 | init_idle(current, smp_processor_id()); |
| 9090 | |
| 9091 | calc_load_update = jiffies + LOAD_FREQ; |
| 9092 | |
| 9093 | idle_thread_set_boot_cpu(); |
| 9094 | |
| 9095 | balance_push_set(smp_processor_id(), false); |
| 9096 | init_sched_fair_class(); |
| 9097 | init_sched_ext_class(); |
| 9098 | |
| 9099 | psi_init(); |
| 9100 | |
| 9101 | init_uclamp(); |
| 9102 | |
| 9103 | preempt_dynamic_init(); |
| 9104 | |
| 9105 | scheduler_running = 1; |
| 9106 | } |
| 9107 | |
| 9108 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP |
| 9109 | |
| 9110 | void __might_sleep(const char *file, int line) |
| 9111 | { |
| 9112 | unsigned int state = get_current_state(); |
| 9113 | /* |
| 9114 | * Blocking primitives will set (and therefore destroy) current->state, |
| 9115 | * since we will exit with TASK_RUNNING make sure we enter with it, |
| 9116 | * otherwise we will destroy state. |
| 9117 | */ |
| 9118 | WARN_ONCE(state != TASK_RUNNING && current->task_state_change, |
| 9119 | "do not call blocking ops when !TASK_RUNNING; " |
| 9120 | "state=%x set at [<%p>] %pS\n", state, |
| 9121 | (void *)current->task_state_change, |
| 9122 | (void *)current->task_state_change); |
| 9123 | |
| 9124 | __might_resched(file, line, 0); |
| 9125 | } |
| 9126 | EXPORT_SYMBOL(__might_sleep); |
| 9127 | |
| 9128 | static void print_preempt_disable_ip(int preempt_offset, unsigned long ip) |
| 9129 | { |
| 9130 | if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT)) |
| 9131 | return; |
| 9132 | |
| 9133 | if (preempt_count() == preempt_offset) |
| 9134 | return; |
| 9135 | |
| 9136 | pr_err("Preemption disabled at:"); |
| 9137 | print_ip_sym(KERN_ERR, ip); |
| 9138 | } |
| 9139 | |
| 9140 | static inline bool resched_offsets_ok(unsigned int offsets) |
| 9141 | { |
| 9142 | unsigned int nested = preempt_count(); |
| 9143 | |
| 9144 | nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT; |
| 9145 | |
| 9146 | return nested == offsets; |
| 9147 | } |
| 9148 | |
| 9149 | void __might_resched(const char *file, int line, unsigned int offsets) |
| 9150 | { |
| 9151 | /* Ratelimiting timestamp: */ |
| 9152 | static unsigned long prev_jiffy; |
| 9153 | |
| 9154 | unsigned long preempt_disable_ip; |
| 9155 | |
| 9156 | /* WARN_ON_ONCE() by default, no rate limit required: */ |
| 9157 | rcu_sleep_check(); |
| 9158 | |
| 9159 | if ((resched_offsets_ok(offsets) && !irqs_disabled() && |
| 9160 | !is_idle_task(current) && !current->non_block_count) || |
| 9161 | system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING || |
| 9162 | oops_in_progress) |
| 9163 | return; |
| 9164 | |
| 9165 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 9166 | return; |
| 9167 | prev_jiffy = jiffies; |
| 9168 | |
| 9169 | /* Save this before calling printk(), since that will clobber it: */ |
| 9170 | preempt_disable_ip = get_preempt_disable_ip(current); |
| 9171 | |
| 9172 | pr_err("BUG: sleeping function called from invalid context at %s:%d\n", |
| 9173 | file, line); |
| 9174 | pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n", |
| 9175 | in_atomic(), irqs_disabled(), current->non_block_count, |
| 9176 | current->pid, current->comm); |
| 9177 | pr_err("preempt_count: %x, expected: %x\n", preempt_count(), |
| 9178 | offsets & MIGHT_RESCHED_PREEMPT_MASK); |
| 9179 | |
| 9180 | if (IS_ENABLED(CONFIG_PREEMPT_RCU)) { |
| 9181 | pr_err("RCU nest depth: %d, expected: %u\n", |
| 9182 | rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT); |
| 9183 | } |
| 9184 | |
| 9185 | if (task_stack_end_corrupted(current)) |
| 9186 | pr_emerg("Thread overran stack, or stack corrupted\n"); |
| 9187 | |
| 9188 | debug_show_held_locks(current); |
| 9189 | if (irqs_disabled()) |
| 9190 | print_irqtrace_events(current); |
| 9191 | |
| 9192 | print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK, |
| 9193 | preempt_disable_ip); |
| 9194 | |
| 9195 | dump_stack(); |
| 9196 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
| 9197 | } |
| 9198 | EXPORT_SYMBOL(__might_resched); |
| 9199 | |
| 9200 | void __cant_sleep(const char *file, int line) |
| 9201 | { |
| 9202 | static unsigned long prev_jiffy; |
| 9203 | |
| 9204 | if (irqs_disabled()) |
| 9205 | return; |
| 9206 | |
| 9207 | if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) |
| 9208 | return; |
| 9209 | |
| 9210 | if (preempt_count()) |
| 9211 | return; |
| 9212 | |
| 9213 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 9214 | return; |
| 9215 | prev_jiffy = jiffies; |
| 9216 | |
| 9217 | printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line); |
| 9218 | printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", |
| 9219 | in_atomic(), irqs_disabled(), |
| 9220 | current->pid, current->comm); |
| 9221 | |
| 9222 | debug_show_held_locks(current); |
| 9223 | dump_stack(); |
| 9224 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
| 9225 | } |
| 9226 | EXPORT_SYMBOL_GPL(__cant_sleep); |
| 9227 | |
| 9228 | # ifdef CONFIG_SMP |
| 9229 | void __cant_migrate(const char *file, int line) |
| 9230 | { |
| 9231 | static unsigned long prev_jiffy; |
| 9232 | |
| 9233 | if (irqs_disabled()) |
| 9234 | return; |
| 9235 | |
| 9236 | if (is_migration_disabled(current)) |
| 9237 | return; |
| 9238 | |
| 9239 | if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) |
| 9240 | return; |
| 9241 | |
| 9242 | if (preempt_count()) |
| 9243 | return; |
| 9244 | |
| 9245 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
| 9246 | return; |
| 9247 | prev_jiffy = jiffies; |
| 9248 | |
| 9249 | pr_err("BUG: assuming non migratable context at %s:%d\n", file, line); |
| 9250 | pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n", |
| 9251 | in_atomic(), irqs_disabled(), is_migration_disabled(current), |
| 9252 | current->pid, current->comm); |
| 9253 | |
| 9254 | debug_show_held_locks(current); |
| 9255 | dump_stack(); |
| 9256 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
| 9257 | } |
| 9258 | EXPORT_SYMBOL_GPL(__cant_migrate); |
| 9259 | # endif /* CONFIG_SMP */ |
| 9260 | #endif /* CONFIG_DEBUG_ATOMIC_SLEEP */ |
| 9261 | |
| 9262 | #ifdef CONFIG_MAGIC_SYSRQ |
| 9263 | void normalize_rt_tasks(void) |
| 9264 | { |
| 9265 | struct task_struct *g, *p; |
| 9266 | struct sched_attr attr = { |
| 9267 | .sched_policy = SCHED_NORMAL, |
| 9268 | }; |
| 9269 | |
| 9270 | read_lock(&tasklist_lock); |
| 9271 | for_each_process_thread(g, p) { |
| 9272 | /* |
| 9273 | * Only normalize user tasks: |
| 9274 | */ |
| 9275 | if (p->flags & PF_KTHREAD) |
| 9276 | continue; |
| 9277 | |
| 9278 | p->se.exec_start = 0; |
| 9279 | schedstat_set(p->stats.wait_start, 0); |
| 9280 | schedstat_set(p->stats.sleep_start, 0); |
| 9281 | schedstat_set(p->stats.block_start, 0); |
| 9282 | |
| 9283 | if (!rt_or_dl_task(p)) { |
| 9284 | /* |
| 9285 | * Renice negative nice level userspace |
| 9286 | * tasks back to 0: |
| 9287 | */ |
| 9288 | if (task_nice(p) < 0) |
| 9289 | set_user_nice(p, 0); |
| 9290 | continue; |
| 9291 | } |
| 9292 | |
| 9293 | __sched_setscheduler(p, &attr, false, false); |
| 9294 | } |
| 9295 | read_unlock(&tasklist_lock); |
| 9296 | } |
| 9297 | |
| 9298 | #endif /* CONFIG_MAGIC_SYSRQ */ |
| 9299 | |
| 9300 | #ifdef CONFIG_KGDB_KDB |
| 9301 | /* |
| 9302 | * These functions are only useful for KDB. |
| 9303 | * |
| 9304 | * They can only be called when the whole system has been |
| 9305 | * stopped - every CPU needs to be quiescent, and no scheduling |
| 9306 | * activity can take place. Using them for anything else would |
| 9307 | * be a serious bug, and as a result, they aren't even visible |
| 9308 | * under any other configuration. |
| 9309 | */ |
| 9310 | |
| 9311 | /** |
| 9312 | * curr_task - return the current task for a given CPU. |
| 9313 | * @cpu: the processor in question. |
| 9314 | * |
| 9315 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
| 9316 | * |
| 9317 | * Return: The current task for @cpu. |
| 9318 | */ |
| 9319 | struct task_struct *curr_task(int cpu) |
| 9320 | { |
| 9321 | return cpu_curr(cpu); |
| 9322 | } |
| 9323 | |
| 9324 | #endif /* CONFIG_KGDB_KDB */ |
| 9325 | |
| 9326 | #ifdef CONFIG_CGROUP_SCHED |
| 9327 | /* task_group_lock serializes the addition/removal of task groups */ |
| 9328 | static DEFINE_SPINLOCK(task_group_lock); |
| 9329 | |
| 9330 | static inline void alloc_uclamp_sched_group(struct task_group *tg, |
| 9331 | struct task_group *parent) |
| 9332 | { |
| 9333 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 9334 | enum uclamp_id clamp_id; |
| 9335 | |
| 9336 | for_each_clamp_id(clamp_id) { |
| 9337 | uclamp_se_set(&tg->uclamp_req[clamp_id], |
| 9338 | uclamp_none(clamp_id), false); |
| 9339 | tg->uclamp[clamp_id] = parent->uclamp[clamp_id]; |
| 9340 | } |
| 9341 | #endif |
| 9342 | } |
| 9343 | |
| 9344 | static void sched_free_group(struct task_group *tg) |
| 9345 | { |
| 9346 | free_fair_sched_group(tg); |
| 9347 | free_rt_sched_group(tg); |
| 9348 | autogroup_free(tg); |
| 9349 | kmem_cache_free(task_group_cache, tg); |
| 9350 | } |
| 9351 | |
| 9352 | static void sched_free_group_rcu(struct rcu_head *rcu) |
| 9353 | { |
| 9354 | sched_free_group(container_of(rcu, struct task_group, rcu)); |
| 9355 | } |
| 9356 | |
| 9357 | static void sched_unregister_group(struct task_group *tg) |
| 9358 | { |
| 9359 | unregister_fair_sched_group(tg); |
| 9360 | unregister_rt_sched_group(tg); |
| 9361 | /* |
| 9362 | * We have to wait for yet another RCU grace period to expire, as |
| 9363 | * print_cfs_stats() might run concurrently. |
| 9364 | */ |
| 9365 | call_rcu(&tg->rcu, sched_free_group_rcu); |
| 9366 | } |
| 9367 | |
| 9368 | /* allocate runqueue etc for a new task group */ |
| 9369 | struct task_group *sched_create_group(struct task_group *parent) |
| 9370 | { |
| 9371 | struct task_group *tg; |
| 9372 | |
| 9373 | tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO); |
| 9374 | if (!tg) |
| 9375 | return ERR_PTR(-ENOMEM); |
| 9376 | |
| 9377 | if (!alloc_fair_sched_group(tg, parent)) |
| 9378 | goto err; |
| 9379 | |
| 9380 | if (!alloc_rt_sched_group(tg, parent)) |
| 9381 | goto err; |
| 9382 | |
| 9383 | scx_tg_init(tg); |
| 9384 | alloc_uclamp_sched_group(tg, parent); |
| 9385 | |
| 9386 | return tg; |
| 9387 | |
| 9388 | err: |
| 9389 | sched_free_group(tg); |
| 9390 | return ERR_PTR(-ENOMEM); |
| 9391 | } |
| 9392 | |
| 9393 | void sched_online_group(struct task_group *tg, struct task_group *parent) |
| 9394 | { |
| 9395 | unsigned long flags; |
| 9396 | |
| 9397 | spin_lock_irqsave(&task_group_lock, flags); |
| 9398 | list_add_tail_rcu(&tg->list, &task_groups); |
| 9399 | |
| 9400 | /* Root should already exist: */ |
| 9401 | WARN_ON(!parent); |
| 9402 | |
| 9403 | tg->parent = parent; |
| 9404 | INIT_LIST_HEAD(&tg->children); |
| 9405 | list_add_rcu(&tg->siblings, &parent->children); |
| 9406 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 9407 | |
| 9408 | online_fair_sched_group(tg); |
| 9409 | } |
| 9410 | |
| 9411 | /* RCU callback to free various structures associated with a task group */ |
| 9412 | static void sched_unregister_group_rcu(struct rcu_head *rhp) |
| 9413 | { |
| 9414 | /* Now it should be safe to free those cfs_rqs: */ |
| 9415 | sched_unregister_group(container_of(rhp, struct task_group, rcu)); |
| 9416 | } |
| 9417 | |
| 9418 | void sched_destroy_group(struct task_group *tg) |
| 9419 | { |
| 9420 | /* Wait for possible concurrent references to cfs_rqs complete: */ |
| 9421 | call_rcu(&tg->rcu, sched_unregister_group_rcu); |
| 9422 | } |
| 9423 | |
| 9424 | void sched_release_group(struct task_group *tg) |
| 9425 | { |
| 9426 | unsigned long flags; |
| 9427 | |
| 9428 | /* |
| 9429 | * Unlink first, to avoid walk_tg_tree_from() from finding us (via |
| 9430 | * sched_cfs_period_timer()). |
| 9431 | * |
| 9432 | * For this to be effective, we have to wait for all pending users of |
| 9433 | * this task group to leave their RCU critical section to ensure no new |
| 9434 | * user will see our dying task group any more. Specifically ensure |
| 9435 | * that tg_unthrottle_up() won't add decayed cfs_rq's to it. |
| 9436 | * |
| 9437 | * We therefore defer calling unregister_fair_sched_group() to |
| 9438 | * sched_unregister_group() which is guarantied to get called only after the |
| 9439 | * current RCU grace period has expired. |
| 9440 | */ |
| 9441 | spin_lock_irqsave(&task_group_lock, flags); |
| 9442 | list_del_rcu(&tg->list); |
| 9443 | list_del_rcu(&tg->siblings); |
| 9444 | spin_unlock_irqrestore(&task_group_lock, flags); |
| 9445 | } |
| 9446 | |
| 9447 | static void sched_change_group(struct task_struct *tsk) |
| 9448 | { |
| 9449 | struct task_group *tg; |
| 9450 | |
| 9451 | /* |
| 9452 | * All callers are synchronized by task_rq_lock(); we do not use RCU |
| 9453 | * which is pointless here. Thus, we pass "true" to task_css_check() |
| 9454 | * to prevent lockdep warnings. |
| 9455 | */ |
| 9456 | tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), |
| 9457 | struct task_group, css); |
| 9458 | tg = autogroup_task_group(tsk, tg); |
| 9459 | tsk->sched_task_group = tg; |
| 9460 | |
| 9461 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9462 | if (tsk->sched_class->task_change_group) |
| 9463 | tsk->sched_class->task_change_group(tsk); |
| 9464 | else |
| 9465 | #endif |
| 9466 | set_task_rq(tsk, task_cpu(tsk)); |
| 9467 | } |
| 9468 | |
| 9469 | /* |
| 9470 | * Change task's runqueue when it moves between groups. |
| 9471 | * |
| 9472 | * The caller of this function should have put the task in its new group by |
| 9473 | * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect |
| 9474 | * its new group. |
| 9475 | */ |
| 9476 | void sched_move_task(struct task_struct *tsk, bool for_autogroup) |
| 9477 | { |
| 9478 | unsigned int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE; |
| 9479 | bool resched = false; |
| 9480 | bool queued = false; |
| 9481 | struct rq *rq; |
| 9482 | |
| 9483 | CLASS(task_rq_lock, rq_guard)(tsk); |
| 9484 | rq = rq_guard.rq; |
| 9485 | |
| 9486 | scoped_guard (sched_change, tsk, queue_flags) { |
| 9487 | sched_change_group(tsk); |
| 9488 | if (!for_autogroup) |
| 9489 | scx_cgroup_move_task(tsk); |
| 9490 | if (scope->running) |
| 9491 | resched = true; |
| 9492 | queued = scope->queued; |
| 9493 | } |
| 9494 | |
| 9495 | if (resched) |
| 9496 | resched_curr(rq); |
| 9497 | else if (queued) |
| 9498 | wakeup_preempt(rq, tsk, 0); |
| 9499 | |
| 9500 | __balance_callbacks(rq, &rq_guard.rf); |
| 9501 | } |
| 9502 | |
| 9503 | static struct cgroup_subsys_state * |
| 9504 | cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) |
| 9505 | { |
| 9506 | struct task_group *parent = css_tg(parent_css); |
| 9507 | struct task_group *tg; |
| 9508 | |
| 9509 | if (!parent) { |
| 9510 | /* This is early initialization for the top cgroup */ |
| 9511 | return &root_task_group.css; |
| 9512 | } |
| 9513 | |
| 9514 | tg = sched_create_group(parent); |
| 9515 | if (IS_ERR(tg)) |
| 9516 | return ERR_PTR(-ENOMEM); |
| 9517 | |
| 9518 | return &tg->css; |
| 9519 | } |
| 9520 | |
| 9521 | /* Expose task group only after completing cgroup initialization */ |
| 9522 | static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) |
| 9523 | { |
| 9524 | struct task_group *tg = css_tg(css); |
| 9525 | struct task_group *parent = css_tg(css->parent); |
| 9526 | int ret; |
| 9527 | |
| 9528 | ret = scx_tg_online(tg); |
| 9529 | if (ret) |
| 9530 | return ret; |
| 9531 | |
| 9532 | if (parent) |
| 9533 | sched_online_group(tg, parent); |
| 9534 | |
| 9535 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 9536 | /* Propagate the effective uclamp value for the new group */ |
| 9537 | guard(mutex)(&uclamp_mutex); |
| 9538 | guard(rcu)(); |
| 9539 | cpu_util_update_eff(css); |
| 9540 | #endif |
| 9541 | |
| 9542 | return 0; |
| 9543 | } |
| 9544 | |
| 9545 | static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) |
| 9546 | { |
| 9547 | struct task_group *tg = css_tg(css); |
| 9548 | |
| 9549 | scx_tg_offline(tg); |
| 9550 | } |
| 9551 | |
| 9552 | static void cpu_cgroup_css_released(struct cgroup_subsys_state *css) |
| 9553 | { |
| 9554 | struct task_group *tg = css_tg(css); |
| 9555 | |
| 9556 | sched_release_group(tg); |
| 9557 | } |
| 9558 | |
| 9559 | static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) |
| 9560 | { |
| 9561 | struct task_group *tg = css_tg(css); |
| 9562 | |
| 9563 | /* |
| 9564 | * Relies on the RCU grace period between css_released() and this. |
| 9565 | */ |
| 9566 | sched_unregister_group(tg); |
| 9567 | } |
| 9568 | |
| 9569 | static int cpu_cgroup_can_attach(struct cgroup_taskset *tset) |
| 9570 | { |
| 9571 | #ifdef CONFIG_RT_GROUP_SCHED |
| 9572 | struct task_struct *task; |
| 9573 | struct cgroup_subsys_state *css; |
| 9574 | |
| 9575 | if (!rt_group_sched_enabled()) |
| 9576 | goto scx_check; |
| 9577 | |
| 9578 | cgroup_taskset_for_each(task, css, tset) { |
| 9579 | if (!sched_rt_can_attach(css_tg(css), task)) |
| 9580 | return -EINVAL; |
| 9581 | } |
| 9582 | scx_check: |
| 9583 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 9584 | return scx_cgroup_can_attach(tset); |
| 9585 | } |
| 9586 | |
| 9587 | static void cpu_cgroup_attach(struct cgroup_taskset *tset) |
| 9588 | { |
| 9589 | struct task_struct *task; |
| 9590 | struct cgroup_subsys_state *css; |
| 9591 | |
| 9592 | cgroup_taskset_for_each(task, css, tset) |
| 9593 | sched_move_task(task, false); |
| 9594 | } |
| 9595 | |
| 9596 | static void cpu_cgroup_cancel_attach(struct cgroup_taskset *tset) |
| 9597 | { |
| 9598 | scx_cgroup_cancel_attach(tset); |
| 9599 | } |
| 9600 | |
| 9601 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 9602 | static void cpu_util_update_eff(struct cgroup_subsys_state *css) |
| 9603 | { |
| 9604 | struct cgroup_subsys_state *top_css = css; |
| 9605 | struct uclamp_se *uc_parent = NULL; |
| 9606 | struct uclamp_se *uc_se = NULL; |
| 9607 | unsigned int eff[UCLAMP_CNT]; |
| 9608 | enum uclamp_id clamp_id; |
| 9609 | unsigned int clamps; |
| 9610 | |
| 9611 | lockdep_assert_held(&uclamp_mutex); |
| 9612 | WARN_ON_ONCE(!rcu_read_lock_held()); |
| 9613 | |
| 9614 | css_for_each_descendant_pre(css, top_css) { |
| 9615 | uc_parent = css_tg(css)->parent |
| 9616 | ? css_tg(css)->parent->uclamp : NULL; |
| 9617 | |
| 9618 | for_each_clamp_id(clamp_id) { |
| 9619 | /* Assume effective clamps matches requested clamps */ |
| 9620 | eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value; |
| 9621 | /* Cap effective clamps with parent's effective clamps */ |
| 9622 | if (uc_parent && |
| 9623 | eff[clamp_id] > uc_parent[clamp_id].value) { |
| 9624 | eff[clamp_id] = uc_parent[clamp_id].value; |
| 9625 | } |
| 9626 | } |
| 9627 | /* Ensure protection is always capped by limit */ |
| 9628 | eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]); |
| 9629 | |
| 9630 | /* Propagate most restrictive effective clamps */ |
| 9631 | clamps = 0x0; |
| 9632 | uc_se = css_tg(css)->uclamp; |
| 9633 | for_each_clamp_id(clamp_id) { |
| 9634 | if (eff[clamp_id] == uc_se[clamp_id].value) |
| 9635 | continue; |
| 9636 | uc_se[clamp_id].value = eff[clamp_id]; |
| 9637 | uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]); |
| 9638 | clamps |= (0x1 << clamp_id); |
| 9639 | } |
| 9640 | if (!clamps) { |
| 9641 | css = css_rightmost_descendant(css); |
| 9642 | continue; |
| 9643 | } |
| 9644 | |
| 9645 | /* Immediately update descendants RUNNABLE tasks */ |
| 9646 | uclamp_update_active_tasks(css); |
| 9647 | } |
| 9648 | } |
| 9649 | |
| 9650 | /* |
| 9651 | * Integer 10^N with a given N exponent by casting to integer the literal "1eN" |
| 9652 | * C expression. Since there is no way to convert a macro argument (N) into a |
| 9653 | * character constant, use two levels of macros. |
| 9654 | */ |
| 9655 | #define _POW10(exp) ((unsigned int)1e##exp) |
| 9656 | #define POW10(exp) _POW10(exp) |
| 9657 | |
| 9658 | struct uclamp_request { |
| 9659 | #define UCLAMP_PERCENT_SHIFT 2 |
| 9660 | #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT)) |
| 9661 | s64 percent; |
| 9662 | u64 util; |
| 9663 | int ret; |
| 9664 | }; |
| 9665 | |
| 9666 | static inline struct uclamp_request |
| 9667 | capacity_from_percent(char *buf) |
| 9668 | { |
| 9669 | struct uclamp_request req = { |
| 9670 | .percent = UCLAMP_PERCENT_SCALE, |
| 9671 | .util = SCHED_CAPACITY_SCALE, |
| 9672 | .ret = 0, |
| 9673 | }; |
| 9674 | |
| 9675 | buf = strim(buf); |
| 9676 | if (strcmp(buf, "max")) { |
| 9677 | req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT, |
| 9678 | &req.percent); |
| 9679 | if (req.ret) |
| 9680 | return req; |
| 9681 | if ((u64)req.percent > UCLAMP_PERCENT_SCALE) { |
| 9682 | req.ret = -ERANGE; |
| 9683 | return req; |
| 9684 | } |
| 9685 | |
| 9686 | req.util = req.percent << SCHED_CAPACITY_SHIFT; |
| 9687 | req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE); |
| 9688 | } |
| 9689 | |
| 9690 | return req; |
| 9691 | } |
| 9692 | |
| 9693 | static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf, |
| 9694 | size_t nbytes, loff_t off, |
| 9695 | enum uclamp_id clamp_id) |
| 9696 | { |
| 9697 | struct uclamp_request req; |
| 9698 | struct task_group *tg; |
| 9699 | |
| 9700 | req = capacity_from_percent(buf); |
| 9701 | if (req.ret) |
| 9702 | return req.ret; |
| 9703 | |
| 9704 | sched_uclamp_enable(); |
| 9705 | |
| 9706 | guard(mutex)(&uclamp_mutex); |
| 9707 | guard(rcu)(); |
| 9708 | |
| 9709 | tg = css_tg(of_css(of)); |
| 9710 | if (tg->uclamp_req[clamp_id].value != req.util) |
| 9711 | uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false); |
| 9712 | |
| 9713 | /* |
| 9714 | * Because of not recoverable conversion rounding we keep track of the |
| 9715 | * exact requested value |
| 9716 | */ |
| 9717 | tg->uclamp_pct[clamp_id] = req.percent; |
| 9718 | |
| 9719 | /* Update effective clamps to track the most restrictive value */ |
| 9720 | cpu_util_update_eff(of_css(of)); |
| 9721 | |
| 9722 | return nbytes; |
| 9723 | } |
| 9724 | |
| 9725 | static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of, |
| 9726 | char *buf, size_t nbytes, |
| 9727 | loff_t off) |
| 9728 | { |
| 9729 | return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN); |
| 9730 | } |
| 9731 | |
| 9732 | static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of, |
| 9733 | char *buf, size_t nbytes, |
| 9734 | loff_t off) |
| 9735 | { |
| 9736 | return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX); |
| 9737 | } |
| 9738 | |
| 9739 | static inline void cpu_uclamp_print(struct seq_file *sf, |
| 9740 | enum uclamp_id clamp_id) |
| 9741 | { |
| 9742 | struct task_group *tg; |
| 9743 | u64 util_clamp; |
| 9744 | u64 percent; |
| 9745 | u32 rem; |
| 9746 | |
| 9747 | scoped_guard (rcu) { |
| 9748 | tg = css_tg(seq_css(sf)); |
| 9749 | util_clamp = tg->uclamp_req[clamp_id].value; |
| 9750 | } |
| 9751 | |
| 9752 | if (util_clamp == SCHED_CAPACITY_SCALE) { |
| 9753 | seq_puts(sf, "max\n"); |
| 9754 | return; |
| 9755 | } |
| 9756 | |
| 9757 | percent = tg->uclamp_pct[clamp_id]; |
| 9758 | percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem); |
| 9759 | seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem); |
| 9760 | } |
| 9761 | |
| 9762 | static int cpu_uclamp_min_show(struct seq_file *sf, void *v) |
| 9763 | { |
| 9764 | cpu_uclamp_print(sf, UCLAMP_MIN); |
| 9765 | return 0; |
| 9766 | } |
| 9767 | |
| 9768 | static int cpu_uclamp_max_show(struct seq_file *sf, void *v) |
| 9769 | { |
| 9770 | cpu_uclamp_print(sf, UCLAMP_MAX); |
| 9771 | return 0; |
| 9772 | } |
| 9773 | #endif /* CONFIG_UCLAMP_TASK_GROUP */ |
| 9774 | |
| 9775 | #ifdef CONFIG_GROUP_SCHED_WEIGHT |
| 9776 | static unsigned long tg_weight(struct task_group *tg) |
| 9777 | { |
| 9778 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 9779 | return scale_load_down(tg->shares); |
| 9780 | #else |
| 9781 | return sched_weight_from_cgroup(tg->scx.weight); |
| 9782 | #endif |
| 9783 | } |
| 9784 | |
| 9785 | static int cpu_shares_write_u64(struct cgroup_subsys_state *css, |
| 9786 | struct cftype *cftype, u64 shareval) |
| 9787 | { |
| 9788 | int ret; |
| 9789 | |
| 9790 | if (shareval > scale_load_down(ULONG_MAX)) |
| 9791 | shareval = MAX_SHARES; |
| 9792 | ret = sched_group_set_shares(css_tg(css), scale_load(shareval)); |
| 9793 | if (!ret) |
| 9794 | scx_group_set_weight(css_tg(css), |
| 9795 | sched_weight_to_cgroup(shareval)); |
| 9796 | return ret; |
| 9797 | } |
| 9798 | |
| 9799 | static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, |
| 9800 | struct cftype *cft) |
| 9801 | { |
| 9802 | return tg_weight(css_tg(css)); |
| 9803 | } |
| 9804 | #endif /* CONFIG_GROUP_SCHED_WEIGHT */ |
| 9805 | |
| 9806 | #ifdef CONFIG_CFS_BANDWIDTH |
| 9807 | static DEFINE_MUTEX(cfs_constraints_mutex); |
| 9808 | |
| 9809 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); |
| 9810 | |
| 9811 | static int tg_set_cfs_bandwidth(struct task_group *tg, |
| 9812 | u64 period_us, u64 quota_us, u64 burst_us) |
| 9813 | { |
| 9814 | int i, ret = 0, runtime_enabled, runtime_was_enabled; |
| 9815 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 9816 | u64 period, quota, burst; |
| 9817 | |
| 9818 | period = (u64)period_us * NSEC_PER_USEC; |
| 9819 | |
| 9820 | if (quota_us == RUNTIME_INF) |
| 9821 | quota = RUNTIME_INF; |
| 9822 | else |
| 9823 | quota = (u64)quota_us * NSEC_PER_USEC; |
| 9824 | |
| 9825 | burst = (u64)burst_us * NSEC_PER_USEC; |
| 9826 | |
| 9827 | /* |
| 9828 | * Prevent race between setting of cfs_rq->runtime_enabled and |
| 9829 | * unthrottle_offline_cfs_rqs(). |
| 9830 | */ |
| 9831 | guard(cpus_read_lock)(); |
| 9832 | guard(mutex)(&cfs_constraints_mutex); |
| 9833 | |
| 9834 | ret = __cfs_schedulable(tg, period, quota); |
| 9835 | if (ret) |
| 9836 | return ret; |
| 9837 | |
| 9838 | runtime_enabled = quota != RUNTIME_INF; |
| 9839 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; |
| 9840 | /* |
| 9841 | * If we need to toggle cfs_bandwidth_used, off->on must occur |
| 9842 | * before making related changes, and on->off must occur afterwards |
| 9843 | */ |
| 9844 | if (runtime_enabled && !runtime_was_enabled) |
| 9845 | cfs_bandwidth_usage_inc(); |
| 9846 | |
| 9847 | scoped_guard (raw_spinlock_irq, &cfs_b->lock) { |
| 9848 | cfs_b->period = ns_to_ktime(period); |
| 9849 | cfs_b->quota = quota; |
| 9850 | cfs_b->burst = burst; |
| 9851 | |
| 9852 | __refill_cfs_bandwidth_runtime(cfs_b); |
| 9853 | |
| 9854 | /* |
| 9855 | * Restart the period timer (if active) to handle new |
| 9856 | * period expiry: |
| 9857 | */ |
| 9858 | if (runtime_enabled) |
| 9859 | start_cfs_bandwidth(cfs_b); |
| 9860 | } |
| 9861 | |
| 9862 | for_each_online_cpu(i) { |
| 9863 | struct cfs_rq *cfs_rq = tg_cfs_rq(tg, i); |
| 9864 | struct rq *rq = cfs_rq->rq; |
| 9865 | |
| 9866 | guard(rq_lock_irq)(rq); |
| 9867 | |
| 9868 | cfs_rq->runtime_enabled = runtime_enabled; |
| 9869 | cfs_rq->runtime_remaining = 1; |
| 9870 | |
| 9871 | if (cfs_rq->throttled) { |
| 9872 | update_rq_clock(rq); |
| 9873 | unthrottle_cfs_rq(cfs_rq); |
| 9874 | } |
| 9875 | } |
| 9876 | |
| 9877 | if (runtime_was_enabled && !runtime_enabled) |
| 9878 | cfs_bandwidth_usage_dec(); |
| 9879 | |
| 9880 | return 0; |
| 9881 | } |
| 9882 | |
| 9883 | static u64 tg_get_cfs_period(struct task_group *tg) |
| 9884 | { |
| 9885 | u64 cfs_period_us; |
| 9886 | |
| 9887 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); |
| 9888 | do_div(cfs_period_us, NSEC_PER_USEC); |
| 9889 | |
| 9890 | return cfs_period_us; |
| 9891 | } |
| 9892 | |
| 9893 | static u64 tg_get_cfs_quota(struct task_group *tg) |
| 9894 | { |
| 9895 | u64 quota_us; |
| 9896 | |
| 9897 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) |
| 9898 | return RUNTIME_INF; |
| 9899 | |
| 9900 | quota_us = tg->cfs_bandwidth.quota; |
| 9901 | do_div(quota_us, NSEC_PER_USEC); |
| 9902 | |
| 9903 | return quota_us; |
| 9904 | } |
| 9905 | |
| 9906 | static u64 tg_get_cfs_burst(struct task_group *tg) |
| 9907 | { |
| 9908 | u64 burst_us; |
| 9909 | |
| 9910 | burst_us = tg->cfs_bandwidth.burst; |
| 9911 | do_div(burst_us, NSEC_PER_USEC); |
| 9912 | |
| 9913 | return burst_us; |
| 9914 | } |
| 9915 | |
| 9916 | struct cfs_schedulable_data { |
| 9917 | struct task_group *tg; |
| 9918 | u64 period, quota; |
| 9919 | }; |
| 9920 | |
| 9921 | /* |
| 9922 | * normalize group quota/period to be quota/max_period |
| 9923 | * note: units are usecs |
| 9924 | */ |
| 9925 | static u64 normalize_cfs_quota(struct task_group *tg, |
| 9926 | struct cfs_schedulable_data *d) |
| 9927 | { |
| 9928 | u64 quota, period; |
| 9929 | |
| 9930 | if (tg == d->tg) { |
| 9931 | period = d->period; |
| 9932 | quota = d->quota; |
| 9933 | } else { |
| 9934 | period = tg_get_cfs_period(tg); |
| 9935 | quota = tg_get_cfs_quota(tg); |
| 9936 | } |
| 9937 | |
| 9938 | /* note: these should typically be equivalent */ |
| 9939 | if (quota == RUNTIME_INF || quota == -1) |
| 9940 | return RUNTIME_INF; |
| 9941 | |
| 9942 | return to_ratio(period, quota); |
| 9943 | } |
| 9944 | |
| 9945 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) |
| 9946 | { |
| 9947 | struct cfs_schedulable_data *d = data; |
| 9948 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 9949 | s64 quota = 0, parent_quota = -1; |
| 9950 | |
| 9951 | if (!tg->parent) { |
| 9952 | quota = RUNTIME_INF; |
| 9953 | } else { |
| 9954 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; |
| 9955 | |
| 9956 | quota = normalize_cfs_quota(tg, d); |
| 9957 | parent_quota = parent_b->hierarchical_quota; |
| 9958 | |
| 9959 | /* |
| 9960 | * Ensure max(child_quota) <= parent_quota. On cgroup2, |
| 9961 | * always take the non-RUNTIME_INF min. On cgroup1, only |
| 9962 | * inherit when no limit is set. In both cases this is used |
| 9963 | * by the scheduler to determine if a given CFS task has a |
| 9964 | * bandwidth constraint at some higher level. |
| 9965 | */ |
| 9966 | if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) { |
| 9967 | if (quota == RUNTIME_INF) |
| 9968 | quota = parent_quota; |
| 9969 | else if (parent_quota != RUNTIME_INF) |
| 9970 | quota = min(quota, parent_quota); |
| 9971 | } else { |
| 9972 | if (quota == RUNTIME_INF) |
| 9973 | quota = parent_quota; |
| 9974 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) |
| 9975 | return -EINVAL; |
| 9976 | } |
| 9977 | } |
| 9978 | cfs_b->hierarchical_quota = quota; |
| 9979 | |
| 9980 | return 0; |
| 9981 | } |
| 9982 | |
| 9983 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) |
| 9984 | { |
| 9985 | struct cfs_schedulable_data data = { |
| 9986 | .tg = tg, |
| 9987 | .period = period, |
| 9988 | .quota = quota, |
| 9989 | }; |
| 9990 | |
| 9991 | if (quota != RUNTIME_INF) { |
| 9992 | do_div(data.period, NSEC_PER_USEC); |
| 9993 | do_div(data.quota, NSEC_PER_USEC); |
| 9994 | } |
| 9995 | |
| 9996 | guard(rcu)(); |
| 9997 | return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); |
| 9998 | } |
| 9999 | |
| 10000 | static int cpu_cfs_stat_show(struct seq_file *sf, void *v) |
| 10001 | { |
| 10002 | struct task_group *tg = css_tg(seq_css(sf)); |
| 10003 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 10004 | |
| 10005 | seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); |
| 10006 | seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); |
| 10007 | seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); |
| 10008 | |
| 10009 | if (schedstat_enabled() && tg != &root_task_group) { |
| 10010 | struct sched_statistics *stats; |
| 10011 | u64 ws = 0; |
| 10012 | int i; |
| 10013 | |
| 10014 | for_each_possible_cpu(i) { |
| 10015 | stats = __schedstats_from_se(tg_se(tg, i)); |
| 10016 | ws += schedstat_val(stats->wait_sum); |
| 10017 | } |
| 10018 | |
| 10019 | seq_printf(sf, "wait_sum %llu\n", ws); |
| 10020 | } |
| 10021 | |
| 10022 | seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst); |
| 10023 | seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time); |
| 10024 | |
| 10025 | return 0; |
| 10026 | } |
| 10027 | |
| 10028 | static u64 throttled_time_self(struct task_group *tg) |
| 10029 | { |
| 10030 | int i; |
| 10031 | u64 total = 0; |
| 10032 | |
| 10033 | for_each_possible_cpu(i) { |
| 10034 | total += READ_ONCE(tg_cfs_rq(tg, i)->throttled_clock_self_time); |
| 10035 | } |
| 10036 | |
| 10037 | return total; |
| 10038 | } |
| 10039 | |
| 10040 | static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v) |
| 10041 | { |
| 10042 | struct task_group *tg = css_tg(seq_css(sf)); |
| 10043 | |
| 10044 | seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg)); |
| 10045 | |
| 10046 | return 0; |
| 10047 | } |
| 10048 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10049 | |
| 10050 | #ifdef CONFIG_GROUP_SCHED_BANDWIDTH |
| 10051 | const u64 max_bw_quota_period_us = 1 * USEC_PER_SEC; /* 1s */ |
| 10052 | static const u64 min_bw_quota_period_us = 1 * USEC_PER_MSEC; /* 1ms */ |
| 10053 | /* More than 203 days if BW_SHIFT equals 20. */ |
| 10054 | static const u64 max_bw_runtime_us = MAX_BW; |
| 10055 | |
| 10056 | static void tg_bandwidth(struct task_group *tg, |
| 10057 | u64 *period_us_p, u64 *quota_us_p, u64 *burst_us_p) |
| 10058 | { |
| 10059 | #ifdef CONFIG_CFS_BANDWIDTH |
| 10060 | if (period_us_p) |
| 10061 | *period_us_p = tg_get_cfs_period(tg); |
| 10062 | if (quota_us_p) |
| 10063 | *quota_us_p = tg_get_cfs_quota(tg); |
| 10064 | if (burst_us_p) |
| 10065 | *burst_us_p = tg_get_cfs_burst(tg); |
| 10066 | #else /* !CONFIG_CFS_BANDWIDTH */ |
| 10067 | if (period_us_p) |
| 10068 | *period_us_p = tg->scx.bw_period_us; |
| 10069 | if (quota_us_p) |
| 10070 | *quota_us_p = tg->scx.bw_quota_us; |
| 10071 | if (burst_us_p) |
| 10072 | *burst_us_p = tg->scx.bw_burst_us; |
| 10073 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10074 | } |
| 10075 | |
| 10076 | static u64 cpu_period_read_u64(struct cgroup_subsys_state *css, |
| 10077 | struct cftype *cft) |
| 10078 | { |
| 10079 | u64 period_us; |
| 10080 | |
| 10081 | tg_bandwidth(css_tg(css), &period_us, NULL, NULL); |
| 10082 | return period_us; |
| 10083 | } |
| 10084 | |
| 10085 | static int tg_set_bandwidth(struct task_group *tg, |
| 10086 | u64 period_us, u64 quota_us, u64 burst_us) |
| 10087 | { |
| 10088 | const u64 max_usec = U64_MAX / NSEC_PER_USEC; |
| 10089 | int ret = 0; |
| 10090 | |
| 10091 | if (tg == &root_task_group) |
| 10092 | return -EINVAL; |
| 10093 | |
| 10094 | /* Values should survive translation to nsec */ |
| 10095 | if (period_us > max_usec || |
| 10096 | (quota_us != RUNTIME_INF && quota_us > max_usec) || |
| 10097 | burst_us > max_usec) |
| 10098 | return -EINVAL; |
| 10099 | |
| 10100 | /* |
| 10101 | * Ensure we have some amount of bandwidth every period. This is to |
| 10102 | * prevent reaching a state of large arrears when throttled via |
| 10103 | * entity_tick() resulting in prolonged exit starvation. |
| 10104 | */ |
| 10105 | if (quota_us < min_bw_quota_period_us || |
| 10106 | period_us < min_bw_quota_period_us) |
| 10107 | return -EINVAL; |
| 10108 | |
| 10109 | /* |
| 10110 | * Likewise, bound things on the other side by preventing insane quota |
| 10111 | * periods. This also allows us to normalize in computing quota |
| 10112 | * feasibility. |
| 10113 | */ |
| 10114 | if (period_us > max_bw_quota_period_us) |
| 10115 | return -EINVAL; |
| 10116 | |
| 10117 | /* |
| 10118 | * Bound quota to defend quota against overflow during bandwidth shift. |
| 10119 | */ |
| 10120 | if (quota_us != RUNTIME_INF && quota_us > max_bw_runtime_us) |
| 10121 | return -EINVAL; |
| 10122 | |
| 10123 | if (quota_us != RUNTIME_INF && (burst_us > quota_us || |
| 10124 | burst_us + quota_us > max_bw_runtime_us)) |
| 10125 | return -EINVAL; |
| 10126 | |
| 10127 | #ifdef CONFIG_CFS_BANDWIDTH |
| 10128 | ret = tg_set_cfs_bandwidth(tg, period_us, quota_us, burst_us); |
| 10129 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10130 | if (!ret) |
| 10131 | scx_group_set_bandwidth(tg, period_us, quota_us, burst_us); |
| 10132 | return ret; |
| 10133 | } |
| 10134 | |
| 10135 | static s64 cpu_quota_read_s64(struct cgroup_subsys_state *css, |
| 10136 | struct cftype *cft) |
| 10137 | { |
| 10138 | u64 quota_us; |
| 10139 | |
| 10140 | tg_bandwidth(css_tg(css), NULL, "a_us, NULL); |
| 10141 | return quota_us; /* (s64)RUNTIME_INF becomes -1 */ |
| 10142 | } |
| 10143 | |
| 10144 | static u64 cpu_burst_read_u64(struct cgroup_subsys_state *css, |
| 10145 | struct cftype *cft) |
| 10146 | { |
| 10147 | u64 burst_us; |
| 10148 | |
| 10149 | tg_bandwidth(css_tg(css), NULL, NULL, &burst_us); |
| 10150 | return burst_us; |
| 10151 | } |
| 10152 | |
| 10153 | static int cpu_period_write_u64(struct cgroup_subsys_state *css, |
| 10154 | struct cftype *cftype, u64 period_us) |
| 10155 | { |
| 10156 | struct task_group *tg = css_tg(css); |
| 10157 | u64 quota_us, burst_us; |
| 10158 | |
| 10159 | tg_bandwidth(tg, NULL, "a_us, &burst_us); |
| 10160 | return tg_set_bandwidth(tg, period_us, quota_us, burst_us); |
| 10161 | } |
| 10162 | |
| 10163 | static int cpu_quota_write_s64(struct cgroup_subsys_state *css, |
| 10164 | struct cftype *cftype, s64 quota_us) |
| 10165 | { |
| 10166 | struct task_group *tg = css_tg(css); |
| 10167 | u64 period_us, burst_us; |
| 10168 | |
| 10169 | if (quota_us < 0) |
| 10170 | quota_us = RUNTIME_INF; |
| 10171 | |
| 10172 | tg_bandwidth(tg, &period_us, NULL, &burst_us); |
| 10173 | return tg_set_bandwidth(tg, period_us, quota_us, burst_us); |
| 10174 | } |
| 10175 | |
| 10176 | static int cpu_burst_write_u64(struct cgroup_subsys_state *css, |
| 10177 | struct cftype *cftype, u64 burst_us) |
| 10178 | { |
| 10179 | struct task_group *tg = css_tg(css); |
| 10180 | u64 period_us, quota_us; |
| 10181 | |
| 10182 | tg_bandwidth(tg, &period_us, "a_us, NULL); |
| 10183 | return tg_set_bandwidth(tg, period_us, quota_us, burst_us); |
| 10184 | } |
| 10185 | #endif /* CONFIG_GROUP_SCHED_BANDWIDTH */ |
| 10186 | |
| 10187 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10188 | static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, |
| 10189 | struct cftype *cft, s64 val) |
| 10190 | { |
| 10191 | return sched_group_set_rt_runtime(css_tg(css), val); |
| 10192 | } |
| 10193 | |
| 10194 | static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, |
| 10195 | struct cftype *cft) |
| 10196 | { |
| 10197 | return sched_group_rt_runtime(css_tg(css)); |
| 10198 | } |
| 10199 | |
| 10200 | static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, |
| 10201 | struct cftype *cftype, u64 rt_period_us) |
| 10202 | { |
| 10203 | return sched_group_set_rt_period(css_tg(css), rt_period_us); |
| 10204 | } |
| 10205 | |
| 10206 | static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, |
| 10207 | struct cftype *cft) |
| 10208 | { |
| 10209 | return sched_group_rt_period(css_tg(css)); |
| 10210 | } |
| 10211 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 10212 | |
| 10213 | #ifdef CONFIG_GROUP_SCHED_WEIGHT |
| 10214 | static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css, |
| 10215 | struct cftype *cft) |
| 10216 | { |
| 10217 | return css_tg(css)->idle; |
| 10218 | } |
| 10219 | |
| 10220 | static int cpu_idle_write_s64(struct cgroup_subsys_state *css, |
| 10221 | struct cftype *cft, s64 idle) |
| 10222 | { |
| 10223 | int ret; |
| 10224 | |
| 10225 | ret = sched_group_set_idle(css_tg(css), idle); |
| 10226 | if (!ret) |
| 10227 | scx_group_set_idle(css_tg(css), idle); |
| 10228 | return ret; |
| 10229 | } |
| 10230 | #endif /* CONFIG_GROUP_SCHED_WEIGHT */ |
| 10231 | |
| 10232 | static struct cftype cpu_legacy_files[] = { |
| 10233 | #ifdef CONFIG_GROUP_SCHED_WEIGHT |
| 10234 | { |
| 10235 | .name = "shares", |
| 10236 | .read_u64 = cpu_shares_read_u64, |
| 10237 | .write_u64 = cpu_shares_write_u64, |
| 10238 | }, |
| 10239 | { |
| 10240 | .name = "idle", |
| 10241 | .read_s64 = cpu_idle_read_s64, |
| 10242 | .write_s64 = cpu_idle_write_s64, |
| 10243 | }, |
| 10244 | #endif |
| 10245 | #ifdef CONFIG_GROUP_SCHED_BANDWIDTH |
| 10246 | { |
| 10247 | .name = "cfs_period_us", |
| 10248 | .read_u64 = cpu_period_read_u64, |
| 10249 | .write_u64 = cpu_period_write_u64, |
| 10250 | }, |
| 10251 | { |
| 10252 | .name = "cfs_quota_us", |
| 10253 | .read_s64 = cpu_quota_read_s64, |
| 10254 | .write_s64 = cpu_quota_write_s64, |
| 10255 | }, |
| 10256 | { |
| 10257 | .name = "cfs_burst_us", |
| 10258 | .read_u64 = cpu_burst_read_u64, |
| 10259 | .write_u64 = cpu_burst_write_u64, |
| 10260 | }, |
| 10261 | #endif |
| 10262 | #ifdef CONFIG_CFS_BANDWIDTH |
| 10263 | { |
| 10264 | .name = "stat", |
| 10265 | .seq_show = cpu_cfs_stat_show, |
| 10266 | }, |
| 10267 | { |
| 10268 | .name = "stat.local", |
| 10269 | .seq_show = cpu_cfs_local_stat_show, |
| 10270 | }, |
| 10271 | #endif |
| 10272 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 10273 | { |
| 10274 | .name = "uclamp.min", |
| 10275 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10276 | .seq_show = cpu_uclamp_min_show, |
| 10277 | .write = cpu_uclamp_min_write, |
| 10278 | }, |
| 10279 | { |
| 10280 | .name = "uclamp.max", |
| 10281 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10282 | .seq_show = cpu_uclamp_max_show, |
| 10283 | .write = cpu_uclamp_max_write, |
| 10284 | }, |
| 10285 | #endif |
| 10286 | { } /* Terminate */ |
| 10287 | }; |
| 10288 | |
| 10289 | #ifdef CONFIG_RT_GROUP_SCHED |
| 10290 | static struct cftype rt_group_files[] = { |
| 10291 | { |
| 10292 | .name = "rt_runtime_us", |
| 10293 | .read_s64 = cpu_rt_runtime_read, |
| 10294 | .write_s64 = cpu_rt_runtime_write, |
| 10295 | }, |
| 10296 | { |
| 10297 | .name = "rt_period_us", |
| 10298 | .read_u64 = cpu_rt_period_read_uint, |
| 10299 | .write_u64 = cpu_rt_period_write_uint, |
| 10300 | }, |
| 10301 | { } /* Terminate */ |
| 10302 | }; |
| 10303 | |
| 10304 | # ifdef CONFIG_RT_GROUP_SCHED_DEFAULT_DISABLED |
| 10305 | DEFINE_STATIC_KEY_FALSE(rt_group_sched); |
| 10306 | # else |
| 10307 | DEFINE_STATIC_KEY_TRUE(rt_group_sched); |
| 10308 | # endif |
| 10309 | |
| 10310 | static int __init setup_rt_group_sched(char *str) |
| 10311 | { |
| 10312 | long val; |
| 10313 | |
| 10314 | if (kstrtol(str, 0, &val) || val < 0 || val > 1) { |
| 10315 | pr_warn("Unable to set rt_group_sched\n"); |
| 10316 | return 1; |
| 10317 | } |
| 10318 | if (val) |
| 10319 | static_branch_enable(&rt_group_sched); |
| 10320 | else |
| 10321 | static_branch_disable(&rt_group_sched); |
| 10322 | |
| 10323 | return 1; |
| 10324 | } |
| 10325 | __setup("rt_group_sched=", setup_rt_group_sched); |
| 10326 | |
| 10327 | static int __init cpu_rt_group_init(void) |
| 10328 | { |
| 10329 | if (!rt_group_sched_enabled()) |
| 10330 | return 0; |
| 10331 | |
| 10332 | WARN_ON(cgroup_add_legacy_cftypes(&cpu_cgrp_subsys, rt_group_files)); |
| 10333 | return 0; |
| 10334 | } |
| 10335 | subsys_initcall(cpu_rt_group_init); |
| 10336 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 10337 | |
| 10338 | static int cpu_extra_stat_show(struct seq_file *sf, |
| 10339 | struct cgroup_subsys_state *css) |
| 10340 | { |
| 10341 | #ifdef CONFIG_CFS_BANDWIDTH |
| 10342 | { |
| 10343 | struct task_group *tg = css_tg(css); |
| 10344 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
| 10345 | u64 throttled_usec, burst_usec; |
| 10346 | |
| 10347 | throttled_usec = cfs_b->throttled_time; |
| 10348 | do_div(throttled_usec, NSEC_PER_USEC); |
| 10349 | burst_usec = cfs_b->burst_time; |
| 10350 | do_div(burst_usec, NSEC_PER_USEC); |
| 10351 | |
| 10352 | seq_printf(sf, "nr_periods %d\n" |
| 10353 | "nr_throttled %d\n" |
| 10354 | "throttled_usec %llu\n" |
| 10355 | "nr_bursts %d\n" |
| 10356 | "burst_usec %llu\n", |
| 10357 | cfs_b->nr_periods, cfs_b->nr_throttled, |
| 10358 | throttled_usec, cfs_b->nr_burst, burst_usec); |
| 10359 | } |
| 10360 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10361 | return 0; |
| 10362 | } |
| 10363 | |
| 10364 | static int cpu_local_stat_show(struct seq_file *sf, |
| 10365 | struct cgroup_subsys_state *css) |
| 10366 | { |
| 10367 | #ifdef CONFIG_CFS_BANDWIDTH |
| 10368 | { |
| 10369 | struct task_group *tg = css_tg(css); |
| 10370 | u64 throttled_self_usec; |
| 10371 | |
| 10372 | throttled_self_usec = throttled_time_self(tg); |
| 10373 | do_div(throttled_self_usec, NSEC_PER_USEC); |
| 10374 | |
| 10375 | seq_printf(sf, "throttled_usec %llu\n", |
| 10376 | throttled_self_usec); |
| 10377 | } |
| 10378 | #endif |
| 10379 | return 0; |
| 10380 | } |
| 10381 | |
| 10382 | #ifdef CONFIG_GROUP_SCHED_WEIGHT |
| 10383 | |
| 10384 | static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css, |
| 10385 | struct cftype *cft) |
| 10386 | { |
| 10387 | return sched_weight_to_cgroup(tg_weight(css_tg(css))); |
| 10388 | } |
| 10389 | |
| 10390 | static int cpu_weight_write_u64(struct cgroup_subsys_state *css, |
| 10391 | struct cftype *cft, u64 cgrp_weight) |
| 10392 | { |
| 10393 | unsigned long weight; |
| 10394 | int ret; |
| 10395 | |
| 10396 | if (cgrp_weight < CGROUP_WEIGHT_MIN || cgrp_weight > CGROUP_WEIGHT_MAX) |
| 10397 | return -ERANGE; |
| 10398 | |
| 10399 | weight = sched_weight_from_cgroup(cgrp_weight); |
| 10400 | |
| 10401 | ret = sched_group_set_shares(css_tg(css), scale_load(weight)); |
| 10402 | if (!ret) |
| 10403 | scx_group_set_weight(css_tg(css), cgrp_weight); |
| 10404 | return ret; |
| 10405 | } |
| 10406 | |
| 10407 | static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css, |
| 10408 | struct cftype *cft) |
| 10409 | { |
| 10410 | unsigned long weight = tg_weight(css_tg(css)); |
| 10411 | int last_delta = INT_MAX; |
| 10412 | int prio, delta; |
| 10413 | |
| 10414 | /* find the closest nice value to the current weight */ |
| 10415 | for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) { |
| 10416 | delta = abs(sched_prio_to_weight[prio] - weight); |
| 10417 | if (delta >= last_delta) |
| 10418 | break; |
| 10419 | last_delta = delta; |
| 10420 | } |
| 10421 | |
| 10422 | return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO); |
| 10423 | } |
| 10424 | |
| 10425 | static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css, |
| 10426 | struct cftype *cft, s64 nice) |
| 10427 | { |
| 10428 | unsigned long weight; |
| 10429 | int idx, ret; |
| 10430 | |
| 10431 | if (nice < MIN_NICE || nice > MAX_NICE) |
| 10432 | return -ERANGE; |
| 10433 | |
| 10434 | idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO; |
| 10435 | idx = array_index_nospec(idx, 40); |
| 10436 | weight = sched_prio_to_weight[idx]; |
| 10437 | |
| 10438 | ret = sched_group_set_shares(css_tg(css), scale_load(weight)); |
| 10439 | if (!ret) |
| 10440 | scx_group_set_weight(css_tg(css), |
| 10441 | sched_weight_to_cgroup(weight)); |
| 10442 | return ret; |
| 10443 | } |
| 10444 | #endif /* CONFIG_GROUP_SCHED_WEIGHT */ |
| 10445 | |
| 10446 | static void __maybe_unused cpu_period_quota_print(struct seq_file *sf, |
| 10447 | long period, long quota) |
| 10448 | { |
| 10449 | if (quota < 0) |
| 10450 | seq_puts(sf, "max"); |
| 10451 | else |
| 10452 | seq_printf(sf, "%ld", quota); |
| 10453 | |
| 10454 | seq_printf(sf, " %ld\n", period); |
| 10455 | } |
| 10456 | |
| 10457 | /* caller should put the current value in *@periodp before calling */ |
| 10458 | static int __maybe_unused cpu_period_quota_parse(char *buf, u64 *period_us_p, |
| 10459 | u64 *quota_us_p) |
| 10460 | { |
| 10461 | char tok[21]; /* U64_MAX */ |
| 10462 | |
| 10463 | if (sscanf(buf, "%20s %llu", tok, period_us_p) < 1) |
| 10464 | return -EINVAL; |
| 10465 | |
| 10466 | if (sscanf(tok, "%llu", quota_us_p) < 1) { |
| 10467 | if (!strcmp(tok, "max")) |
| 10468 | *quota_us_p = RUNTIME_INF; |
| 10469 | else |
| 10470 | return -EINVAL; |
| 10471 | } |
| 10472 | |
| 10473 | return 0; |
| 10474 | } |
| 10475 | |
| 10476 | #ifdef CONFIG_GROUP_SCHED_BANDWIDTH |
| 10477 | static int cpu_max_show(struct seq_file *sf, void *v) |
| 10478 | { |
| 10479 | struct task_group *tg = css_tg(seq_css(sf)); |
| 10480 | u64 period_us, quota_us; |
| 10481 | |
| 10482 | tg_bandwidth(tg, &period_us, "a_us, NULL); |
| 10483 | cpu_period_quota_print(sf, period_us, quota_us); |
| 10484 | return 0; |
| 10485 | } |
| 10486 | |
| 10487 | static ssize_t cpu_max_write(struct kernfs_open_file *of, |
| 10488 | char *buf, size_t nbytes, loff_t off) |
| 10489 | { |
| 10490 | struct task_group *tg = css_tg(of_css(of)); |
| 10491 | u64 period_us, quota_us, burst_us; |
| 10492 | int ret; |
| 10493 | |
| 10494 | tg_bandwidth(tg, &period_us, NULL, &burst_us); |
| 10495 | ret = cpu_period_quota_parse(buf, &period_us, "a_us); |
| 10496 | if (!ret) |
| 10497 | ret = tg_set_bandwidth(tg, period_us, quota_us, burst_us); |
| 10498 | return ret ?: nbytes; |
| 10499 | } |
| 10500 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10501 | |
| 10502 | static struct cftype cpu_files[] = { |
| 10503 | #ifdef CONFIG_GROUP_SCHED_WEIGHT |
| 10504 | { |
| 10505 | .name = "weight", |
| 10506 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10507 | .read_u64 = cpu_weight_read_u64, |
| 10508 | .write_u64 = cpu_weight_write_u64, |
| 10509 | }, |
| 10510 | { |
| 10511 | .name = "weight.nice", |
| 10512 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10513 | .read_s64 = cpu_weight_nice_read_s64, |
| 10514 | .write_s64 = cpu_weight_nice_write_s64, |
| 10515 | }, |
| 10516 | { |
| 10517 | .name = "idle", |
| 10518 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10519 | .read_s64 = cpu_idle_read_s64, |
| 10520 | .write_s64 = cpu_idle_write_s64, |
| 10521 | }, |
| 10522 | #endif |
| 10523 | #ifdef CONFIG_GROUP_SCHED_BANDWIDTH |
| 10524 | { |
| 10525 | .name = "max", |
| 10526 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10527 | .seq_show = cpu_max_show, |
| 10528 | .write = cpu_max_write, |
| 10529 | }, |
| 10530 | { |
| 10531 | .name = "max.burst", |
| 10532 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10533 | .read_u64 = cpu_burst_read_u64, |
| 10534 | .write_u64 = cpu_burst_write_u64, |
| 10535 | }, |
| 10536 | #endif /* CONFIG_CFS_BANDWIDTH */ |
| 10537 | #ifdef CONFIG_UCLAMP_TASK_GROUP |
| 10538 | { |
| 10539 | .name = "uclamp.min", |
| 10540 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10541 | .seq_show = cpu_uclamp_min_show, |
| 10542 | .write = cpu_uclamp_min_write, |
| 10543 | }, |
| 10544 | { |
| 10545 | .name = "uclamp.max", |
| 10546 | .flags = CFTYPE_NOT_ON_ROOT, |
| 10547 | .seq_show = cpu_uclamp_max_show, |
| 10548 | .write = cpu_uclamp_max_write, |
| 10549 | }, |
| 10550 | #endif /* CONFIG_UCLAMP_TASK_GROUP */ |
| 10551 | { } /* terminate */ |
| 10552 | }; |
| 10553 | |
| 10554 | struct cgroup_subsys cpu_cgrp_subsys = { |
| 10555 | .css_alloc = cpu_cgroup_css_alloc, |
| 10556 | .css_online = cpu_cgroup_css_online, |
| 10557 | .css_offline = cpu_cgroup_css_offline, |
| 10558 | .css_released = cpu_cgroup_css_released, |
| 10559 | .css_free = cpu_cgroup_css_free, |
| 10560 | .css_extra_stat_show = cpu_extra_stat_show, |
| 10561 | .css_local_stat_show = cpu_local_stat_show, |
| 10562 | .can_attach = cpu_cgroup_can_attach, |
| 10563 | .attach = cpu_cgroup_attach, |
| 10564 | .cancel_attach = cpu_cgroup_cancel_attach, |
| 10565 | .legacy_cftypes = cpu_legacy_files, |
| 10566 | .dfl_cftypes = cpu_files, |
| 10567 | .early_init = true, |
| 10568 | .threaded = true, |
| 10569 | }; |
| 10570 | |
| 10571 | #endif /* CONFIG_CGROUP_SCHED */ |
| 10572 | |
| 10573 | void dump_cpu_task(int cpu) |
| 10574 | { |
| 10575 | if (in_hardirq() && cpu == smp_processor_id()) { |
| 10576 | struct pt_regs *regs; |
| 10577 | |
| 10578 | regs = get_irq_regs(); |
| 10579 | if (regs) { |
| 10580 | show_regs(regs); |
| 10581 | return; |
| 10582 | } |
| 10583 | } |
| 10584 | |
| 10585 | if (trigger_single_cpu_backtrace(cpu)) |
| 10586 | return; |
| 10587 | |
| 10588 | pr_info("Task dump for CPU %d:\n", cpu); |
| 10589 | sched_show_task(cpu_curr(cpu)); |
| 10590 | } |
| 10591 | |
| 10592 | /* |
| 10593 | * Nice levels are multiplicative, with a gentle 10% change for every |
| 10594 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to |
| 10595 | * nice 1, it will get ~10% less CPU time than another CPU-bound task |
| 10596 | * that remained on nice 0. |
| 10597 | * |
| 10598 | * The "10% effect" is relative and cumulative: from _any_ nice level, |
| 10599 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level |
| 10600 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
| 10601 | * If a task goes up by ~10% and another task goes down by ~10% then |
| 10602 | * the relative distance between them is ~25%.) |
| 10603 | */ |
| 10604 | const int sched_prio_to_weight[40] = { |
| 10605 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
| 10606 | /* -15 */ 29154, 23254, 18705, 14949, 11916, |
| 10607 | /* -10 */ 9548, 7620, 6100, 4904, 3906, |
| 10608 | /* -5 */ 3121, 2501, 1991, 1586, 1277, |
| 10609 | /* 0 */ 1024, 820, 655, 526, 423, |
| 10610 | /* 5 */ 335, 272, 215, 172, 137, |
| 10611 | /* 10 */ 110, 87, 70, 56, 45, |
| 10612 | /* 15 */ 36, 29, 23, 18, 15, |
| 10613 | }; |
| 10614 | |
| 10615 | /* |
| 10616 | * Inverse (2^32/x) values of the sched_prio_to_weight[] array, pre-calculated. |
| 10617 | * |
| 10618 | * In cases where the weight does not change often, we can use the |
| 10619 | * pre-calculated inverse to speed up arithmetics by turning divisions |
| 10620 | * into multiplications: |
| 10621 | */ |
| 10622 | const u32 sched_prio_to_wmult[40] = { |
| 10623 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
| 10624 | /* -15 */ 147320, 184698, 229616, 287308, 360437, |
| 10625 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, |
| 10626 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, |
| 10627 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, |
| 10628 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, |
| 10629 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, |
| 10630 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, |
| 10631 | }; |
| 10632 | |
| 10633 | void call_trace_sched_update_nr_running(struct rq *rq, int count) |
| 10634 | { |
| 10635 | trace_sched_update_nr_running_tp(rq, count); |
| 10636 | } |
| 10637 | |
| 10638 | #ifdef CONFIG_SCHED_MM_CID |
| 10639 | /* |
| 10640 | * Concurrency IDentifier management |
| 10641 | * |
| 10642 | * Serialization rules: |
| 10643 | * |
| 10644 | * mm::mm_cid::mutex: Serializes fork() and exit() and therefore |
| 10645 | * protects mm::mm_cid::users and mode switch |
| 10646 | * transitions |
| 10647 | * |
| 10648 | * mm::mm_cid::lock: Serializes mm_update_max_cids() and |
| 10649 | * mm_update_cpus_allowed(). Nests in mm_cid::mutex |
| 10650 | * and runqueue lock. |
| 10651 | * |
| 10652 | * The mm_cidmask bitmap is not protected by any of the mm::mm_cid locks |
| 10653 | * and can only be modified with atomic operations. |
| 10654 | * |
| 10655 | * The mm::mm_cid:pcpu per CPU storage is protected by the CPUs runqueue |
| 10656 | * lock. |
| 10657 | * |
| 10658 | * CID ownership: |
| 10659 | * |
| 10660 | * A CID is either owned by a task (stored in task_struct::mm_cid.cid) or |
| 10661 | * by a CPU (stored in mm::mm_cid.pcpu::cid). CIDs owned by CPUs have the |
| 10662 | * MM_CID_ONCPU bit set. |
| 10663 | * |
| 10664 | * During the transition of ownership mode, the MM_CID_TRANSIT bit is set |
| 10665 | * on the CIDs. When this bit is set the tasks drop the CID back into the |
| 10666 | * pool when scheduling out. |
| 10667 | * |
| 10668 | * Both bits (ONCPU and TRANSIT) are filtered out by task_cid() when the |
| 10669 | * CID is actually handed over to user space in the RSEQ memory. |
| 10670 | * |
| 10671 | * Mode switching: |
| 10672 | * |
| 10673 | * The ownership mode is per process and stored in mm:mm_cid::mode with the |
| 10674 | * following possible states: |
| 10675 | * |
| 10676 | * 0: Per task ownership |
| 10677 | * 0 | MM_CID_TRANSIT: Transition from per CPU to per task |
| 10678 | * MM_CID_ONCPU: Per CPU ownership |
| 10679 | * MM_CID_ONCPU | MM_CID_TRANSIT: Transition from per task to per CPU |
| 10680 | * |
| 10681 | * All transitions of ownership mode happen in two phases: |
| 10682 | * |
| 10683 | * 1) mm:mm_cid::mode has the MM_CID_TRANSIT bit set. This is OR'ed on the |
| 10684 | * CIDs and denotes that the CID is only temporarily owned by a |
| 10685 | * task. When the task schedules out it drops the CID back into the |
| 10686 | * pool if this bit is set. |
| 10687 | * |
| 10688 | * 2) The initiating context walks the per CPU space or the tasks to fixup |
| 10689 | * or drop the CIDs and after completion it clears MM_CID_TRANSIT in |
| 10690 | * mm:mm_cid::mode. After that point the CIDs are strictly task or CPU |
| 10691 | * owned again. |
| 10692 | * |
| 10693 | * This two phase transition is required to prevent CID space exhaustion |
| 10694 | * during the transition as a direct transfer of ownership would fail: |
| 10695 | * |
| 10696 | * - On task to CPU mode switch if a task is scheduled in on one CPU and |
| 10697 | * then migrated to another CPU before the fixup freed enough per task |
| 10698 | * CIDs. |
| 10699 | * |
| 10700 | * - On CPU to task mode switch if two tasks are scheduled in on the same |
| 10701 | * CPU before the fixup freed per CPU CIDs. |
| 10702 | * |
| 10703 | * Both scenarios can result in a live lock because sched_in() is invoked |
| 10704 | * with runqueue lock held and loops in search of a CID and the fixup |
| 10705 | * thread can't make progress freeing them up because it is stuck on the |
| 10706 | * same runqueue lock. |
| 10707 | * |
| 10708 | * While MM_CID_TRANSIT is active during the transition phase the MM_CID |
| 10709 | * bitmap can be contended, but that's a temporary contention bound to the |
| 10710 | * transition period. After that everything goes back into steady state and |
| 10711 | * nothing except fork() and exit() will touch the bitmap. This is an |
| 10712 | * acceptable tradeoff as it completely avoids complex serialization, |
| 10713 | * memory barriers and atomic operations for the common case. |
| 10714 | * |
| 10715 | * Aside of that this mechanism also ensures RT compability: |
| 10716 | * |
| 10717 | * - The task which runs the fixup is fully preemptible except for the |
| 10718 | * short runqueue lock held sections. |
| 10719 | * |
| 10720 | * - The transient impact of the bitmap contention is only problematic |
| 10721 | * when there is a thundering herd scenario of tasks scheduling in and |
| 10722 | * out concurrently. There is not much which can be done about that |
| 10723 | * except for avoiding mode switching by a proper overall system |
| 10724 | * configuration. |
| 10725 | * |
| 10726 | * Switching to per CPU mode happens when the user count becomes greater |
| 10727 | * than the maximum number of CIDs, which is calculated by: |
| 10728 | * |
| 10729 | * opt_cids = min(mm_cid::nr_cpus_allowed, mm_cid::users); |
| 10730 | * max_cids = min(1.25 * opt_cids, num_possible_cpus()); |
| 10731 | * |
| 10732 | * The +25% allowance is useful for tight CPU masks in scenarios where only |
| 10733 | * a few threads are created and destroyed to avoid frequent mode |
| 10734 | * switches. Though this allowance shrinks, the closer opt_cids becomes to |
| 10735 | * num_possible_cpus(), which is the (unfortunate) hard ABI limit. |
| 10736 | * |
| 10737 | * At the point of switching to per CPU mode the new user is not yet |
| 10738 | * visible in the system, so the task which initiated the fork() runs the |
| 10739 | * fixup function. mm_cid_fixup_tasks_to_cpu() walks the thread list and |
| 10740 | * either marks each task owned CID with MM_CID_TRANSIT if the task is |
| 10741 | * running on a CPU or drops it into the CID pool if a task is not on a |
| 10742 | * CPU. Tasks which schedule in before the task walk reaches them do the |
| 10743 | * handover in mm_cid_schedin(). When mm_cid_fixup_tasks_to_cpus() |
| 10744 | * completes it is guaranteed that no task related to that MM owns a CID |
| 10745 | * anymore. |
| 10746 | * |
| 10747 | * Switching back to task mode happens when the user count goes below the |
| 10748 | * threshold which was recorded on the per CPU mode switch: |
| 10749 | * |
| 10750 | * pcpu_thrs = min(opt_cids - (opt_cids / 4), num_possible_cpus() / 2); |
| 10751 | * |
| 10752 | * This threshold is updated when a affinity change increases the number of |
| 10753 | * allowed CPUs for the MM, which might cause a switch back to per task |
| 10754 | * mode. |
| 10755 | * |
| 10756 | * If the switch back was initiated by a exiting task, then that task runs |
| 10757 | * the fixup function. If it was initiated by a affinity change, then it's |
| 10758 | * run either in the deferred update function in context of a workqueue or |
| 10759 | * by a task which forks a new one or by a task which exits. Whatever |
| 10760 | * happens first. mm_cid_fixup_cpus_to_task() walks through the possible |
| 10761 | * CPUs and either marks the CPU owned CIDs with MM_CID_TRANSIT if a |
| 10762 | * related task is running on the CPU or drops it into the pool. Tasks |
| 10763 | * which are scheduled in before the fixup covered them do the handover |
| 10764 | * themself. When mm_cid_fixup_cpus_to_tasks() completes it is guaranteed |
| 10765 | * that no CID related to that MM is owned by a CPU anymore. |
| 10766 | */ |
| 10767 | |
| 10768 | /* |
| 10769 | * Update the CID range properties when the constraints change. Invoked via |
| 10770 | * fork(), exit() and affinity changes |
| 10771 | */ |
| 10772 | static void __mm_update_max_cids(struct mm_mm_cid *mc) |
| 10773 | { |
| 10774 | unsigned int opt_cids, max_cids; |
| 10775 | |
| 10776 | /* Calculate the new optimal constraint */ |
| 10777 | opt_cids = min(mc->nr_cpus_allowed, mc->users); |
| 10778 | |
| 10779 | /* Adjust the maximum CIDs to +25% limited by the number of possible CPUs */ |
| 10780 | max_cids = min(opt_cids + (opt_cids / 4), num_possible_cpus()); |
| 10781 | WRITE_ONCE(mc->max_cids, max_cids); |
| 10782 | } |
| 10783 | |
| 10784 | static inline unsigned int mm_cid_calc_pcpu_thrs(struct mm_mm_cid *mc) |
| 10785 | { |
| 10786 | unsigned int opt_cids; |
| 10787 | |
| 10788 | opt_cids = min(mc->nr_cpus_allowed, mc->users); |
| 10789 | /* Has to be at least 1 because 0 indicates PCPU mode off */ |
| 10790 | return max(min(opt_cids - opt_cids / 4, num_possible_cpus() / 2), 1); |
| 10791 | } |
| 10792 | |
| 10793 | static bool mm_update_max_cids(struct mm_struct *mm) |
| 10794 | { |
| 10795 | struct mm_mm_cid *mc = &mm->mm_cid; |
| 10796 | bool percpu = cid_on_cpu(mc->mode); |
| 10797 | |
| 10798 | lockdep_assert_held(&mm->mm_cid.lock); |
| 10799 | |
| 10800 | /* Clear deferred mode switch flag. A change is handled by the caller */ |
| 10801 | mc->update_deferred = false; |
| 10802 | __mm_update_max_cids(mc); |
| 10803 | |
| 10804 | /* Check whether owner mode must be changed */ |
| 10805 | if (!percpu) { |
| 10806 | /* Enable per CPU mode when the number of users is above max_cids */ |
| 10807 | if (mc->users > mc->max_cids) |
| 10808 | mc->pcpu_thrs = mm_cid_calc_pcpu_thrs(mc); |
| 10809 | } else { |
| 10810 | /* Switch back to per task if user count under threshold */ |
| 10811 | if (mc->users < mc->pcpu_thrs) |
| 10812 | mc->pcpu_thrs = 0; |
| 10813 | } |
| 10814 | |
| 10815 | /* Mode change required? */ |
| 10816 | if (percpu == !!mc->pcpu_thrs) |
| 10817 | return false; |
| 10818 | |
| 10819 | /* Flip the mode and set the transition flag to bridge the transfer */ |
| 10820 | WRITE_ONCE(mc->mode, mc->mode ^ (MM_CID_TRANSIT | MM_CID_ONCPU)); |
| 10821 | /* |
| 10822 | * Order the store against the subsequent fixups so that |
| 10823 | * acquire(rq::lock) cannot be reordered by the CPU before the |
| 10824 | * store. |
| 10825 | */ |
| 10826 | smp_mb(); |
| 10827 | return true; |
| 10828 | } |
| 10829 | |
| 10830 | static inline void mm_update_cpus_allowed(struct mm_struct *mm, const struct cpumask *affmsk) |
| 10831 | { |
| 10832 | struct cpumask *mm_allowed; |
| 10833 | struct mm_mm_cid *mc; |
| 10834 | unsigned int weight; |
| 10835 | |
| 10836 | if (!mm || !READ_ONCE(mm->mm_cid.users)) |
| 10837 | return; |
| 10838 | /* |
| 10839 | * mm::mm_cid::mm_cpus_allowed is the superset of each threads |
| 10840 | * allowed CPUs mask which means it can only grow. |
| 10841 | */ |
| 10842 | mc = &mm->mm_cid; |
| 10843 | guard(raw_spinlock)(&mc->lock); |
| 10844 | mm_allowed = mm_cpus_allowed(mm); |
| 10845 | weight = cpumask_weighted_or(mm_allowed, mm_allowed, affmsk); |
| 10846 | if (weight == mc->nr_cpus_allowed) |
| 10847 | return; |
| 10848 | |
| 10849 | WRITE_ONCE(mc->nr_cpus_allowed, weight); |
| 10850 | __mm_update_max_cids(mc); |
| 10851 | if (!cid_on_cpu(mc->mode)) |
| 10852 | return; |
| 10853 | |
| 10854 | /* Adjust the threshold to the wider set */ |
| 10855 | mc->pcpu_thrs = mm_cid_calc_pcpu_thrs(mc); |
| 10856 | /* Switch back to per task mode? */ |
| 10857 | if (mc->users >= mc->pcpu_thrs) |
| 10858 | return; |
| 10859 | |
| 10860 | /* Don't queue twice */ |
| 10861 | if (mc->update_deferred) |
| 10862 | return; |
| 10863 | |
| 10864 | /* Queue the irq work, which schedules the real work */ |
| 10865 | mc->update_deferred = true; |
| 10866 | irq_work_queue(&mc->irq_work); |
| 10867 | } |
| 10868 | |
| 10869 | static inline void mm_cid_complete_transit(struct mm_struct *mm, unsigned int mode) |
| 10870 | { |
| 10871 | /* |
| 10872 | * Ensure that the store removing the TRANSIT bit cannot be |
| 10873 | * reordered by the CPU before the fixups have been completed. |
| 10874 | */ |
| 10875 | smp_mb(); |
| 10876 | WRITE_ONCE(mm->mm_cid.mode, mode); |
| 10877 | } |
| 10878 | |
| 10879 | static inline void mm_cid_transit_to_task(struct task_struct *t, struct mm_cid_pcpu *pcp) |
| 10880 | { |
| 10881 | if (cid_on_cpu(t->mm_cid.cid)) { |
| 10882 | unsigned int cid = cpu_cid_to_cid(t->mm_cid.cid); |
| 10883 | |
| 10884 | t->mm_cid.cid = cid_to_transit_cid(cid); |
| 10885 | pcp->cid = t->mm_cid.cid; |
| 10886 | } |
| 10887 | } |
| 10888 | |
| 10889 | static void mm_cid_fixup_cpus_to_tasks(struct mm_struct *mm) |
| 10890 | { |
| 10891 | unsigned int cpu; |
| 10892 | |
| 10893 | /* Walk the CPUs and fixup all stale CIDs */ |
| 10894 | for_each_possible_cpu(cpu) { |
| 10895 | struct mm_cid_pcpu *pcp = per_cpu_ptr(mm->mm_cid.pcpu, cpu); |
| 10896 | struct rq *rq = cpu_rq(cpu); |
| 10897 | |
| 10898 | /* Remote access to mm::mm_cid::pcpu requires rq_lock */ |
| 10899 | guard(rq_lock_irq)(rq); |
| 10900 | /* Is the CID still owned by the CPU? */ |
| 10901 | if (cid_on_cpu(pcp->cid)) { |
| 10902 | /* |
| 10903 | * If rq->curr has @mm, transfer it with the |
| 10904 | * transition bit set. Otherwise drop it. |
| 10905 | */ |
| 10906 | if (rq->curr->mm == mm && rq->curr->mm_cid.active) |
| 10907 | mm_cid_transit_to_task(rq->curr, pcp); |
| 10908 | else |
| 10909 | mm_drop_cid_on_cpu(mm, pcp); |
| 10910 | |
| 10911 | } else if (rq->curr->mm == mm && rq->curr->mm_cid.active) { |
| 10912 | unsigned int cid = rq->curr->mm_cid.cid; |
| 10913 | |
| 10914 | /* |
| 10915 | * Set the transition bit only on a genuine task-owned |
| 10916 | * CID. A running active task can legitimately have |
| 10917 | * MM_CID_UNSET here: in per-CPU mode CIDs are assigned |
| 10918 | * lazily on schedule-in, so the fork()/execve() window |
| 10919 | * leaves the task active with no owned CID. Setting the |
| 10920 | * transition bit on MM_CID_UNSET would later feed |
| 10921 | * clear_bit() an out-of-bounds bit number via |
| 10922 | * mm_cid_schedout(), so exclude it. A CPU-owned |
| 10923 | * (MM_CID_ONCPU) CID is handled by the cid_on_cpu() |
| 10924 | * branch above and never reaches here. |
| 10925 | */ |
| 10926 | if (cid != MM_CID_UNSET && !cid_in_transit(cid)) { |
| 10927 | cid = cid_to_transit_cid(cid); |
| 10928 | rq->curr->mm_cid.cid = cid; |
| 10929 | pcp->cid = cid; |
| 10930 | } |
| 10931 | } |
| 10932 | } |
| 10933 | mm_cid_complete_transit(mm, 0); |
| 10934 | } |
| 10935 | |
| 10936 | static inline void mm_cid_transit_to_cpu(struct task_struct *t, struct mm_cid_pcpu *pcp) |
| 10937 | { |
| 10938 | if (cid_on_task(t->mm_cid.cid)) { |
| 10939 | t->mm_cid.cid = cid_to_transit_cid(t->mm_cid.cid); |
| 10940 | pcp->cid = t->mm_cid.cid; |
| 10941 | } |
| 10942 | } |
| 10943 | |
| 10944 | static void mm_cid_fixup_task_to_cpu(struct task_struct *t, struct mm_struct *mm) |
| 10945 | { |
| 10946 | /* Remote access to mm::mm_cid::pcpu requires rq_lock */ |
| 10947 | guard(task_rq_lock)(t); |
| 10948 | if (cid_on_task(t->mm_cid.cid)) { |
| 10949 | /* If running on the CPU, put the CID in transit mode, otherwise drop it */ |
| 10950 | if (task_rq(t)->curr == t) |
| 10951 | mm_cid_transit_to_cpu(t, per_cpu_ptr(mm->mm_cid.pcpu, task_cpu(t))); |
| 10952 | else |
| 10953 | mm_unset_cid_on_task(t); |
| 10954 | } |
| 10955 | } |
| 10956 | |
| 10957 | static void mm_cid_fixup_tasks_to_cpus(void) |
| 10958 | { |
| 10959 | struct mm_struct *mm = current->mm; |
| 10960 | struct task_struct *t; |
| 10961 | |
| 10962 | lockdep_assert_held(&mm->mm_cid.mutex); |
| 10963 | |
| 10964 | hlist_for_each_entry(t, &mm->mm_cid.user_list, mm_cid.node) { |
| 10965 | /* Current has already transferred before invoking the fixup. */ |
| 10966 | if (t != current) |
| 10967 | mm_cid_fixup_task_to_cpu(t, mm); |
| 10968 | } |
| 10969 | |
| 10970 | mm_cid_complete_transit(mm, MM_CID_ONCPU); |
| 10971 | } |
| 10972 | |
| 10973 | static bool sched_mm_cid_add_user(struct task_struct *t, struct mm_struct *mm) |
| 10974 | { |
| 10975 | lockdep_assert_held(&mm->mm_cid.lock); |
| 10976 | |
| 10977 | t->mm_cid.active = 1; |
| 10978 | hlist_add_head(&t->mm_cid.node, &mm->mm_cid.user_list); |
| 10979 | mm->mm_cid.users++; |
| 10980 | return mm_update_max_cids(mm); |
| 10981 | } |
| 10982 | |
| 10983 | static void sched_mm_cid_fork(struct task_struct *t) |
| 10984 | { |
| 10985 | struct mm_struct *mm = t->mm; |
| 10986 | bool percpu; |
| 10987 | |
| 10988 | if (!mm) |
| 10989 | return; |
| 10990 | |
| 10991 | WARN_ON_ONCE(t->mm_cid.cid != MM_CID_UNSET); |
| 10992 | |
| 10993 | guard(mutex)(&mm->mm_cid.mutex); |
| 10994 | scoped_guard(raw_spinlock_irq, &mm->mm_cid.lock) { |
| 10995 | struct mm_cid_pcpu *pcp = this_cpu_ptr(mm->mm_cid.pcpu); |
| 10996 | |
| 10997 | /* First user ? */ |
| 10998 | if (!mm->mm_cid.users) { |
| 10999 | sched_mm_cid_add_user(t, mm); |
| 11000 | t->mm_cid.cid = mm_get_cid(mm); |
| 11001 | /* Required for execve() */ |
| 11002 | pcp->cid = t->mm_cid.cid; |
| 11003 | return; |
| 11004 | } |
| 11005 | |
| 11006 | if (!sched_mm_cid_add_user(t, mm)) { |
| 11007 | if (!cid_on_cpu(mm->mm_cid.mode)) |
| 11008 | t->mm_cid.cid = mm_get_cid(mm); |
| 11009 | return; |
| 11010 | } |
| 11011 | |
| 11012 | /* Handle the mode change and transfer current's CID */ |
| 11013 | percpu = cid_on_cpu(mm->mm_cid.mode); |
| 11014 | if (!percpu) |
| 11015 | mm_cid_transit_to_task(current, pcp); |
| 11016 | else |
| 11017 | mm_cid_transit_to_cpu(current, pcp); |
| 11018 | } |
| 11019 | |
| 11020 | if (percpu) { |
| 11021 | mm_cid_fixup_tasks_to_cpus(); |
| 11022 | } else { |
| 11023 | mm_cid_fixup_cpus_to_tasks(mm); |
| 11024 | t->mm_cid.cid = mm_get_cid(mm); |
| 11025 | } |
| 11026 | } |
| 11027 | |
| 11028 | static bool sched_mm_cid_remove_user(struct task_struct *t) |
| 11029 | { |
| 11030 | lockdep_assert_held(&t->mm->mm_cid.lock); |
| 11031 | |
| 11032 | t->mm_cid.active = 0; |
| 11033 | /* Clear the transition bit */ |
| 11034 | t->mm_cid.cid = cid_from_transit_cid(t->mm_cid.cid); |
| 11035 | mm_unset_cid_on_task(t); |
| 11036 | hlist_del_init(&t->mm_cid.node); |
| 11037 | t->mm->mm_cid.users--; |
| 11038 | return mm_update_max_cids(t->mm); |
| 11039 | } |
| 11040 | |
| 11041 | static bool __sched_mm_cid_exit(struct task_struct *t) |
| 11042 | { |
| 11043 | struct mm_struct *mm = t->mm; |
| 11044 | |
| 11045 | if (!sched_mm_cid_remove_user(t)) |
| 11046 | return false; |
| 11047 | /* |
| 11048 | * Contrary to fork() this only deals with a switch back to per |
| 11049 | * task mode either because the above decreased users or an |
| 11050 | * affinity change increased the number of allowed CPUs and the |
| 11051 | * deferred fixup did not run yet. |
| 11052 | */ |
| 11053 | if (WARN_ON_ONCE(cid_on_cpu(mm->mm_cid.mode))) |
| 11054 | return false; |
| 11055 | /* |
| 11056 | * A failed fork(2) cleanup never gets here, so @current must have |
| 11057 | * the same MM as @t. That's true for exit() and the failed |
| 11058 | * pthread_create() cleanup case. |
| 11059 | */ |
| 11060 | if (WARN_ON_ONCE(current->mm != mm)) |
| 11061 | return false; |
| 11062 | return true; |
| 11063 | } |
| 11064 | |
| 11065 | /* |
| 11066 | * When a task exits, the MM CID held by the task is not longer required as |
| 11067 | * the task cannot return to user space. |
| 11068 | */ |
| 11069 | void sched_mm_cid_exit(struct task_struct *t) |
| 11070 | { |
| 11071 | struct mm_struct *mm = t->mm; |
| 11072 | |
| 11073 | if (!mm || !t->mm_cid.active) |
| 11074 | return; |
| 11075 | /* |
| 11076 | * Ensure that only one instance is doing MM CID operations within |
| 11077 | * a MM. The common case is uncontended. The rare fixup case adds |
| 11078 | * some overhead. |
| 11079 | */ |
| 11080 | scoped_guard(mutex, &mm->mm_cid.mutex) { |
| 11081 | /* mm_cid::mutex is sufficient to protect mm_cid::users */ |
| 11082 | if (likely(mm->mm_cid.users > 1)) { |
| 11083 | scoped_guard(raw_spinlock_irq, &mm->mm_cid.lock) { |
| 11084 | if (!__sched_mm_cid_exit(t)) |
| 11085 | return; |
| 11086 | /* |
| 11087 | * Mode change. The task has the CID unset |
| 11088 | * already and dealt with an eventually set |
| 11089 | * TRANSIT bit. If the CID is owned by the CPU |
| 11090 | * then drop it. |
| 11091 | */ |
| 11092 | mm_drop_cid_on_cpu(mm, this_cpu_ptr(mm->mm_cid.pcpu)); |
| 11093 | } |
| 11094 | mm_cid_fixup_cpus_to_tasks(mm); |
| 11095 | return; |
| 11096 | } |
| 11097 | /* Last user */ |
| 11098 | scoped_guard(raw_spinlock_irq, &mm->mm_cid.lock) { |
| 11099 | /* Required across execve() */ |
| 11100 | if (t == current) |
| 11101 | mm_cid_transit_to_task(t, this_cpu_ptr(mm->mm_cid.pcpu)); |
| 11102 | /* Ignore mode change. There is nothing to do. */ |
| 11103 | sched_mm_cid_remove_user(t); |
| 11104 | } |
| 11105 | } |
| 11106 | |
| 11107 | /* |
| 11108 | * As this is the last user (execve(), process exit or failed |
| 11109 | * fork(2)) there is no concurrency anymore. |
| 11110 | * |
| 11111 | * Synchronize eventually pending work to ensure that there are no |
| 11112 | * dangling references left. @t->mm_cid.users is zero so nothing |
| 11113 | * can queue this work anymore. |
| 11114 | */ |
| 11115 | irq_work_sync(&mm->mm_cid.irq_work); |
| 11116 | cancel_work_sync(&mm->mm_cid.work); |
| 11117 | } |
| 11118 | |
| 11119 | /* Deactivate MM CID allocation across execve() */ |
| 11120 | void sched_mm_cid_before_execve(struct task_struct *t) |
| 11121 | { |
| 11122 | sched_mm_cid_exit(t); |
| 11123 | } |
| 11124 | |
| 11125 | /* Reactivate MM CID after execve() */ |
| 11126 | void sched_mm_cid_after_execve(struct task_struct *t) |
| 11127 | { |
| 11128 | if (t->mm) |
| 11129 | sched_mm_cid_fork(t); |
| 11130 | } |
| 11131 | |
| 11132 | static void mm_cid_work_fn(struct work_struct *work) |
| 11133 | { |
| 11134 | struct mm_struct *mm = container_of(work, struct mm_struct, mm_cid.work); |
| 11135 | |
| 11136 | guard(mutex)(&mm->mm_cid.mutex); |
| 11137 | /* Did the last user task exit already? */ |
| 11138 | if (!mm->mm_cid.users) |
| 11139 | return; |
| 11140 | |
| 11141 | scoped_guard(raw_spinlock_irq, &mm->mm_cid.lock) { |
| 11142 | /* Have fork() or exit() handled it already? */ |
| 11143 | if (!mm->mm_cid.update_deferred) |
| 11144 | return; |
| 11145 | /* This clears mm_cid::update_deferred */ |
| 11146 | if (!mm_update_max_cids(mm)) |
| 11147 | return; |
| 11148 | /* Affinity changes can only switch back to task mode */ |
| 11149 | if (WARN_ON_ONCE(cid_on_cpu(mm->mm_cid.mode))) |
| 11150 | return; |
| 11151 | } |
| 11152 | mm_cid_fixup_cpus_to_tasks(mm); |
| 11153 | } |
| 11154 | |
| 11155 | static void mm_cid_irq_work(struct irq_work *work) |
| 11156 | { |
| 11157 | struct mm_struct *mm = container_of(work, struct mm_struct, mm_cid.irq_work); |
| 11158 | |
| 11159 | /* |
| 11160 | * Needs to be unconditional because mm_cid::lock cannot be held |
| 11161 | * when scheduling work as mm_update_cpus_allowed() nests inside |
| 11162 | * rq::lock and schedule_work() might end up in wakeup... |
| 11163 | */ |
| 11164 | schedule_work(&mm->mm_cid.work); |
| 11165 | } |
| 11166 | |
| 11167 | void mm_init_cid(struct mm_struct *mm, struct task_struct *p) |
| 11168 | { |
| 11169 | mm->mm_cid.max_cids = 0; |
| 11170 | mm->mm_cid.mode = 0; |
| 11171 | mm->mm_cid.nr_cpus_allowed = p->nr_cpus_allowed; |
| 11172 | mm->mm_cid.users = 0; |
| 11173 | mm->mm_cid.pcpu_thrs = 0; |
| 11174 | mm->mm_cid.update_deferred = 0; |
| 11175 | raw_spin_lock_init(&mm->mm_cid.lock); |
| 11176 | mutex_init(&mm->mm_cid.mutex); |
| 11177 | mm->mm_cid.irq_work = IRQ_WORK_INIT_HARD(mm_cid_irq_work); |
| 11178 | INIT_WORK(&mm->mm_cid.work, mm_cid_work_fn); |
| 11179 | INIT_HLIST_HEAD(&mm->mm_cid.user_list); |
| 11180 | cpumask_copy(mm_cpus_allowed(mm), &p->cpus_mask); |
| 11181 | bitmap_zero(mm_cidmask(mm), num_possible_cpus()); |
| 11182 | } |
| 11183 | #else /* CONFIG_SCHED_MM_CID */ |
| 11184 | static inline void mm_update_cpus_allowed(struct mm_struct *mm, const struct cpumask *affmsk) { } |
| 11185 | static inline void sched_mm_cid_fork(struct task_struct *t) { } |
| 11186 | #endif /* !CONFIG_SCHED_MM_CID */ |
| 11187 | |
| 11188 | static DEFINE_PER_CPU(struct sched_change_ctx, sched_change_ctx); |
| 11189 | |
| 11190 | struct sched_change_ctx *sched_change_begin(struct task_struct *p, unsigned int flags) |
| 11191 | { |
| 11192 | struct sched_change_ctx *ctx = this_cpu_ptr(&sched_change_ctx); |
| 11193 | struct rq *rq = task_rq(p); |
| 11194 | |
| 11195 | /* |
| 11196 | * Must exclusively use matched flags since this is both dequeue and |
| 11197 | * enqueue. |
| 11198 | */ |
| 11199 | WARN_ON_ONCE(flags & 0xFFFF0000); |
| 11200 | |
| 11201 | lockdep_assert_rq_held(rq); |
| 11202 | |
| 11203 | if (!(flags & DEQUEUE_NOCLOCK)) { |
| 11204 | update_rq_clock(rq); |
| 11205 | flags |= DEQUEUE_NOCLOCK; |
| 11206 | } |
| 11207 | |
| 11208 | if ((flags & DEQUEUE_CLASS) && p->sched_class->switching_from) |
| 11209 | p->sched_class->switching_from(rq, p); |
| 11210 | |
| 11211 | *ctx = (struct sched_change_ctx){ |
| 11212 | .p = p, |
| 11213 | .class = p->sched_class, |
| 11214 | .flags = flags, |
| 11215 | .queued = task_on_rq_queued(p), |
| 11216 | .running = task_current_donor(rq, p), |
| 11217 | }; |
| 11218 | |
| 11219 | if (!(flags & DEQUEUE_CLASS)) { |
| 11220 | if (p->sched_class->get_prio) |
| 11221 | ctx->prio = p->sched_class->get_prio(rq, p); |
| 11222 | else |
| 11223 | ctx->prio = p->prio; |
| 11224 | } |
| 11225 | |
| 11226 | if (ctx->queued) |
| 11227 | dequeue_task(rq, p, flags); |
| 11228 | if (ctx->running) |
| 11229 | put_prev_task(rq, p); |
| 11230 | |
| 11231 | if ((flags & DEQUEUE_CLASS) && p->sched_class->switched_from) |
| 11232 | p->sched_class->switched_from(rq, p); |
| 11233 | |
| 11234 | return ctx; |
| 11235 | } |
| 11236 | |
| 11237 | void sched_change_end(struct sched_change_ctx *ctx) |
| 11238 | { |
| 11239 | struct task_struct *p = ctx->p; |
| 11240 | struct rq *rq = task_rq(p); |
| 11241 | |
| 11242 | lockdep_assert_rq_held(rq); |
| 11243 | |
| 11244 | /* |
| 11245 | * Changing class without *QUEUE_CLASS is bad. |
| 11246 | */ |
| 11247 | WARN_ON_ONCE(p->sched_class != ctx->class && !(ctx->flags & ENQUEUE_CLASS)); |
| 11248 | |
| 11249 | if ((ctx->flags & ENQUEUE_CLASS) && p->sched_class->switching_to) |
| 11250 | p->sched_class->switching_to(rq, p); |
| 11251 | |
| 11252 | if (ctx->queued) |
| 11253 | enqueue_task(rq, p, ctx->flags); |
| 11254 | if (ctx->running) |
| 11255 | set_next_task(rq, p); |
| 11256 | |
| 11257 | if (ctx->flags & ENQUEUE_CLASS) { |
| 11258 | if (p->sched_class->switched_to) |
| 11259 | p->sched_class->switched_to(rq, p); |
| 11260 | |
| 11261 | /* |
| 11262 | * Check if the task can preempt the current in case the class |
| 11263 | * was upgraded. Note that none of the switch*_from() methods |
| 11264 | * know the new class and none of the switch*_to() methods know |
| 11265 | * the old class so we must resched here. |
| 11266 | */ |
| 11267 | if (ctx->queued) |
| 11268 | wakeup_preempt(rq, p, 0); |
| 11269 | |
| 11270 | if (ctx->running) { |
| 11271 | /* |
| 11272 | * If this was a degradation in class; make sure to |
| 11273 | * reschedule. |
| 11274 | */ |
| 11275 | if (sched_class_above(ctx->class, p->sched_class)) |
| 11276 | resched_curr(rq); |
| 11277 | } |
| 11278 | } else { |
| 11279 | p->sched_class->prio_changed(rq, p, ctx->prio); |
| 11280 | } |
| 11281 | } |
| 11282 | |
| 11283 | #ifdef CONFIG_PREFERRED_CPU |
| 11284 | static DEFINE_PER_CPU(struct cpu_stop_work, npc_push_task_work); |
| 11285 | |
| 11286 | static int sched_non_preferred_cpu_push_stop(void *arg) |
| 11287 | { |
| 11288 | struct task_struct *p = arg; |
| 11289 | struct rq *rq = this_rq(); |
| 11290 | struct rq_flags rf; |
| 11291 | int cpu; |
| 11292 | |
| 11293 | if (cpu_preferred(rq->cpu)) { |
| 11294 | scoped_guard(rq_lock_irqsave, rq) |
| 11295 | rq->npc_push_work_pending = false; |
| 11296 | put_task_struct(p); |
| 11297 | return 0; |
| 11298 | } |
| 11299 | |
| 11300 | scoped_guard (raw_spinlock_irq, &p->pi_lock) { |
| 11301 | /* |
| 11302 | * select_fallback_rq() may acquire the rq lock in case of |
| 11303 | * fallback. So call it before grabbing rq lock. If the task |
| 11304 | * migrates to another CPU before the rq lock is acquired, |
| 11305 | * subsequent validation of task's current rq will help to |
| 11306 | * safely bail out. |
| 11307 | */ |
| 11308 | cpu = select_fallback_rq(rq->cpu, p); |
| 11309 | context_unsafe_alias(rq); |
| 11310 | rq_lock(rq, &rf); |
| 11311 | rq->npc_push_work_pending = false; |
| 11312 | update_rq_clock(rq); |
| 11313 | |
| 11314 | if (task_rq(p) == rq && task_on_rq_queued(p)) { |
| 11315 | struct rq *dest_rq = __migrate_task(rq, &rf, p, cpu); |
| 11316 | |
| 11317 | if (rq != dest_rq) |
| 11318 | schedstat_inc(p->stats.nr_migrations_cpu_non_preferred); |
| 11319 | rq = dest_rq; |
| 11320 | } |
| 11321 | rq_unlock(rq, &rf); |
| 11322 | } |
| 11323 | |
| 11324 | put_task_struct(p); |
| 11325 | return 0; |
| 11326 | } |
| 11327 | |
| 11328 | /* |
| 11329 | * Push the current task running on non-preferred CPU(npc). |
| 11330 | * Using this non preferred CPU will lead to more contention |
| 11331 | * in the host. So it is better not to use this CPU. |
| 11332 | * |
| 11333 | * Since task is running, call a stopper to push the task out. This is |
| 11334 | * similar to how task moves during hotplug. In select_fallback_rq() a |
| 11335 | * preferred CPU will be chosen and henceforth task shouldn't come back to |
| 11336 | * this CPU again. |
| 11337 | * |
| 11338 | * Works for FAIR class only. |
| 11339 | * |
| 11340 | * If task is affined only on non-preferred CPUs, no point in moving it out. |
| 11341 | */ |
| 11342 | void sched_push_current_non_preferred_cpu(struct rq *rq) |
| 11343 | { |
| 11344 | struct task_struct *push_task = rq->curr; |
| 11345 | |
| 11346 | scoped_guard(rq_lock, rq) { |
| 11347 | /* Push the task if its explicit affinity allows */ |
| 11348 | if (!task_can_migrate_to_preferred(push_task, rq->cpu)) |
| 11349 | return; |
| 11350 | |
| 11351 | /* There is already a stopper thread. Don't race with it. */ |
| 11352 | if (rq->npc_push_work_pending) |
| 11353 | return; |
| 11354 | |
| 11355 | if (is_migration_disabled(push_task)) |
| 11356 | return; |
| 11357 | |
| 11358 | rq->npc_push_work_pending = true; |
| 11359 | } |
| 11360 | |
| 11361 | /* sched_tick runs with interrupts disabled. */ |
| 11362 | get_task_struct(push_task); |
| 11363 | stop_one_cpu_nowait(rq->cpu, sched_non_preferred_cpu_push_stop, |
| 11364 | push_task, this_cpu_ptr(&npc_push_task_work)); |
| 11365 | } |
| 11366 | #endif |