| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Generic helpers for smp ipi calls |
| 4 | * |
| 5 | * (C) Jens Axboe <jens.axboe@oracle.com> 2008 |
| 6 | */ |
| 7 | |
| 8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 9 | |
| 10 | #include <linux/irq_work.h> |
| 11 | #include <linux/rcupdate.h> |
| 12 | #include <linux/rculist.h> |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/export.h> |
| 15 | #include <linux/percpu.h> |
| 16 | #include <linux/init.h> |
| 17 | #include <linux/interrupt.h> |
| 18 | #include <linux/gfp.h> |
| 19 | #include <linux/smp.h> |
| 20 | #include <linux/cpu.h> |
| 21 | #include <linux/sched.h> |
| 22 | #include <linux/sched/idle.h> |
| 23 | #include <linux/hypervisor.h> |
| 24 | #include <linux/sched/clock.h> |
| 25 | #include <linux/nmi.h> |
| 26 | #include <linux/sched/debug.h> |
| 27 | |
| 28 | #include "smpboot.h" |
| 29 | #include "sched/smp.h" |
| 30 | |
| 31 | #define CSD_TYPE(_csd) ((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK) |
| 32 | |
| 33 | struct call_function_data { |
| 34 | call_single_data_t __percpu *csd; |
| 35 | cpumask_var_t cpumask; |
| 36 | cpumask_var_t cpumask_ipi; |
| 37 | }; |
| 38 | |
| 39 | static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data); |
| 40 | |
| 41 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue); |
| 42 | |
| 43 | static void flush_smp_call_function_queue(bool warn_cpu_offline); |
| 44 | |
| 45 | int smpcfd_prepare_cpu(unsigned int cpu) |
| 46 | { |
| 47 | struct call_function_data *cfd = &per_cpu(cfd_data, cpu); |
| 48 | |
| 49 | if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL, |
| 50 | cpu_to_node(cpu))) |
| 51 | return -ENOMEM; |
| 52 | if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL, |
| 53 | cpu_to_node(cpu))) { |
| 54 | free_cpumask_var(cfd->cpumask); |
| 55 | return -ENOMEM; |
| 56 | } |
| 57 | cfd->csd = alloc_percpu(call_single_data_t); |
| 58 | if (!cfd->csd) { |
| 59 | free_cpumask_var(cfd->cpumask); |
| 60 | free_cpumask_var(cfd->cpumask_ipi); |
| 61 | return -ENOMEM; |
| 62 | } |
| 63 | |
| 64 | return 0; |
| 65 | } |
| 66 | |
| 67 | int smpcfd_dead_cpu(unsigned int cpu) |
| 68 | { |
| 69 | struct call_function_data *cfd = &per_cpu(cfd_data, cpu); |
| 70 | |
| 71 | free_cpumask_var(cfd->cpumask); |
| 72 | free_cpumask_var(cfd->cpumask_ipi); |
| 73 | free_percpu(cfd->csd); |
| 74 | return 0; |
| 75 | } |
| 76 | |
| 77 | int smpcfd_dying_cpu(unsigned int cpu) |
| 78 | { |
| 79 | /* |
| 80 | * The IPIs for the smp-call-function callbacks queued by other |
| 81 | * CPUs might arrive late, either due to hardware latencies or |
| 82 | * because this CPU disabled interrupts (inside stop-machine) |
| 83 | * before the IPIs were sent. So flush out any pending callbacks |
| 84 | * explicitly (without waiting for the IPIs to arrive), to |
| 85 | * ensure that the outgoing CPU doesn't go offline with work |
| 86 | * still pending. |
| 87 | */ |
| 88 | flush_smp_call_function_queue(false); |
| 89 | irq_work_run(); |
| 90 | return 0; |
| 91 | } |
| 92 | |
| 93 | void __init call_function_init(void) |
| 94 | { |
| 95 | int i; |
| 96 | |
| 97 | for_each_possible_cpu(i) |
| 98 | init_llist_head(&per_cpu(call_single_queue, i)); |
| 99 | |
| 100 | smpcfd_prepare_cpu(smp_processor_id()); |
| 101 | } |
| 102 | |
| 103 | #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG |
| 104 | |
| 105 | static DEFINE_PER_CPU(call_single_data_t *, cur_csd); |
| 106 | static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func); |
| 107 | static DEFINE_PER_CPU(void *, cur_csd_info); |
| 108 | |
| 109 | #define CSD_LOCK_TIMEOUT (5ULL * NSEC_PER_SEC) |
| 110 | static atomic_t csd_bug_count = ATOMIC_INIT(0); |
| 111 | |
| 112 | /* Record current CSD work for current CPU, NULL to erase. */ |
| 113 | static void csd_lock_record(call_single_data_t *csd) |
| 114 | { |
| 115 | if (!csd) { |
| 116 | smp_mb(); /* NULL cur_csd after unlock. */ |
| 117 | __this_cpu_write(cur_csd, NULL); |
| 118 | return; |
| 119 | } |
| 120 | __this_cpu_write(cur_csd_func, csd->func); |
| 121 | __this_cpu_write(cur_csd_info, csd->info); |
| 122 | smp_wmb(); /* func and info before csd. */ |
| 123 | __this_cpu_write(cur_csd, csd); |
| 124 | smp_mb(); /* Update cur_csd before function call. */ |
| 125 | /* Or before unlock, as the case may be. */ |
| 126 | } |
| 127 | |
| 128 | static __always_inline int csd_lock_wait_getcpu(call_single_data_t *csd) |
| 129 | { |
| 130 | unsigned int csd_type; |
| 131 | |
| 132 | csd_type = CSD_TYPE(csd); |
| 133 | if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC) |
| 134 | return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */ |
| 135 | return -1; |
| 136 | } |
| 137 | |
| 138 | /* |
| 139 | * Complain if too much time spent waiting. Note that only |
| 140 | * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU, |
| 141 | * so waiting on other types gets much less information. |
| 142 | */ |
| 143 | static __always_inline bool csd_lock_wait_toolong(call_single_data_t *csd, u64 ts0, u64 *ts1, int *bug_id) |
| 144 | { |
| 145 | int cpu = -1; |
| 146 | int cpux; |
| 147 | bool firsttime; |
| 148 | u64 ts2, ts_delta; |
| 149 | call_single_data_t *cpu_cur_csd; |
| 150 | unsigned int flags = READ_ONCE(csd->node.u_flags); |
| 151 | |
| 152 | if (!(flags & CSD_FLAG_LOCK)) { |
| 153 | if (!unlikely(*bug_id)) |
| 154 | return true; |
| 155 | cpu = csd_lock_wait_getcpu(csd); |
| 156 | pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n", |
| 157 | *bug_id, raw_smp_processor_id(), cpu); |
| 158 | return true; |
| 159 | } |
| 160 | |
| 161 | ts2 = sched_clock(); |
| 162 | ts_delta = ts2 - *ts1; |
| 163 | if (likely(ts_delta <= CSD_LOCK_TIMEOUT)) |
| 164 | return false; |
| 165 | |
| 166 | firsttime = !*bug_id; |
| 167 | if (firsttime) |
| 168 | *bug_id = atomic_inc_return(&csd_bug_count); |
| 169 | cpu = csd_lock_wait_getcpu(csd); |
| 170 | if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu)) |
| 171 | cpux = 0; |
| 172 | else |
| 173 | cpux = cpu; |
| 174 | cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */ |
| 175 | pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n", |
| 176 | firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0, |
| 177 | cpu, csd->func, csd->info); |
| 178 | if (cpu_cur_csd && csd != cpu_cur_csd) { |
| 179 | pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n", |
| 180 | *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)), |
| 181 | READ_ONCE(per_cpu(cur_csd_info, cpux))); |
| 182 | } else { |
| 183 | pr_alert("\tcsd: CSD lock (#%d) %s.\n", |
| 184 | *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request"); |
| 185 | } |
| 186 | if (cpu >= 0) { |
| 187 | if (!trigger_single_cpu_backtrace(cpu)) |
| 188 | dump_cpu_task(cpu); |
| 189 | if (!cpu_cur_csd) { |
| 190 | pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu); |
| 191 | arch_send_call_function_single_ipi(cpu); |
| 192 | } |
| 193 | } |
| 194 | dump_stack(); |
| 195 | *ts1 = ts2; |
| 196 | |
| 197 | return false; |
| 198 | } |
| 199 | |
| 200 | /* |
| 201 | * csd_lock/csd_unlock used to serialize access to per-cpu csd resources |
| 202 | * |
| 203 | * For non-synchronous ipi calls the csd can still be in use by the |
| 204 | * previous function call. For multi-cpu calls its even more interesting |
| 205 | * as we'll have to ensure no other cpu is observing our csd. |
| 206 | */ |
| 207 | static __always_inline void csd_lock_wait(call_single_data_t *csd) |
| 208 | { |
| 209 | int bug_id = 0; |
| 210 | u64 ts0, ts1; |
| 211 | |
| 212 | ts1 = ts0 = sched_clock(); |
| 213 | for (;;) { |
| 214 | if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id)) |
| 215 | break; |
| 216 | cpu_relax(); |
| 217 | } |
| 218 | smp_acquire__after_ctrl_dep(); |
| 219 | } |
| 220 | |
| 221 | #else |
| 222 | static void csd_lock_record(call_single_data_t *csd) |
| 223 | { |
| 224 | } |
| 225 | |
| 226 | static __always_inline void csd_lock_wait(call_single_data_t *csd) |
| 227 | { |
| 228 | smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK)); |
| 229 | } |
| 230 | #endif |
| 231 | |
| 232 | static __always_inline void csd_lock(call_single_data_t *csd) |
| 233 | { |
| 234 | csd_lock_wait(csd); |
| 235 | csd->node.u_flags |= CSD_FLAG_LOCK; |
| 236 | |
| 237 | /* |
| 238 | * prevent CPU from reordering the above assignment |
| 239 | * to ->flags with any subsequent assignments to other |
| 240 | * fields of the specified call_single_data_t structure: |
| 241 | */ |
| 242 | smp_wmb(); |
| 243 | } |
| 244 | |
| 245 | static __always_inline void csd_unlock(call_single_data_t *csd) |
| 246 | { |
| 247 | WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK)); |
| 248 | |
| 249 | /* |
| 250 | * ensure we're all done before releasing data: |
| 251 | */ |
| 252 | smp_store_release(&csd->node.u_flags, 0); |
| 253 | } |
| 254 | |
| 255 | static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data); |
| 256 | |
| 257 | void __smp_call_single_queue(int cpu, struct llist_node *node) |
| 258 | { |
| 259 | /* |
| 260 | * The list addition should be visible before sending the IPI |
| 261 | * handler locks the list to pull the entry off it because of |
| 262 | * normal cache coherency rules implied by spinlocks. |
| 263 | * |
| 264 | * If IPIs can go out of order to the cache coherency protocol |
| 265 | * in an architecture, sufficient synchronisation should be added |
| 266 | * to arch code to make it appear to obey cache coherency WRT |
| 267 | * locking and barrier primitives. Generic code isn't really |
| 268 | * equipped to do the right thing... |
| 269 | */ |
| 270 | if (llist_add(node, &per_cpu(call_single_queue, cpu))) |
| 271 | send_call_function_single_ipi(cpu); |
| 272 | } |
| 273 | |
| 274 | /* |
| 275 | * Insert a previously allocated call_single_data_t element |
| 276 | * for execution on the given CPU. data must already have |
| 277 | * ->func, ->info, and ->flags set. |
| 278 | */ |
| 279 | static int generic_exec_single(int cpu, call_single_data_t *csd) |
| 280 | { |
| 281 | if (cpu == smp_processor_id()) { |
| 282 | smp_call_func_t func = csd->func; |
| 283 | void *info = csd->info; |
| 284 | unsigned long flags; |
| 285 | |
| 286 | /* |
| 287 | * We can unlock early even for the synchronous on-stack case, |
| 288 | * since we're doing this from the same CPU.. |
| 289 | */ |
| 290 | csd_lock_record(csd); |
| 291 | csd_unlock(csd); |
| 292 | local_irq_save(flags); |
| 293 | func(info); |
| 294 | csd_lock_record(NULL); |
| 295 | local_irq_restore(flags); |
| 296 | return 0; |
| 297 | } |
| 298 | |
| 299 | if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) { |
| 300 | csd_unlock(csd); |
| 301 | return -ENXIO; |
| 302 | } |
| 303 | |
| 304 | __smp_call_single_queue(cpu, &csd->node.llist); |
| 305 | |
| 306 | return 0; |
| 307 | } |
| 308 | |
| 309 | /** |
| 310 | * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks |
| 311 | * |
| 312 | * Invoked by arch to handle an IPI for call function single. |
| 313 | * Must be called with interrupts disabled. |
| 314 | */ |
| 315 | void generic_smp_call_function_single_interrupt(void) |
| 316 | { |
| 317 | flush_smp_call_function_queue(true); |
| 318 | } |
| 319 | |
| 320 | /** |
| 321 | * flush_smp_call_function_queue - Flush pending smp-call-function callbacks |
| 322 | * |
| 323 | * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an |
| 324 | * offline CPU. Skip this check if set to 'false'. |
| 325 | * |
| 326 | * Flush any pending smp-call-function callbacks queued on this CPU. This is |
| 327 | * invoked by the generic IPI handler, as well as by a CPU about to go offline, |
| 328 | * to ensure that all pending IPI callbacks are run before it goes completely |
| 329 | * offline. |
| 330 | * |
| 331 | * Loop through the call_single_queue and run all the queued callbacks. |
| 332 | * Must be called with interrupts disabled. |
| 333 | */ |
| 334 | static void flush_smp_call_function_queue(bool warn_cpu_offline) |
| 335 | { |
| 336 | call_single_data_t *csd, *csd_next; |
| 337 | struct llist_node *entry, *prev; |
| 338 | struct llist_head *head; |
| 339 | static bool warned; |
| 340 | |
| 341 | lockdep_assert_irqs_disabled(); |
| 342 | |
| 343 | head = this_cpu_ptr(&call_single_queue); |
| 344 | entry = llist_del_all(head); |
| 345 | entry = llist_reverse_order(entry); |
| 346 | |
| 347 | /* There shouldn't be any pending callbacks on an offline CPU. */ |
| 348 | if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) && |
| 349 | !warned && !llist_empty(head))) { |
| 350 | warned = true; |
| 351 | WARN(1, "IPI on offline CPU %d\n", smp_processor_id()); |
| 352 | |
| 353 | /* |
| 354 | * We don't have to use the _safe() variant here |
| 355 | * because we are not invoking the IPI handlers yet. |
| 356 | */ |
| 357 | llist_for_each_entry(csd, entry, node.llist) { |
| 358 | switch (CSD_TYPE(csd)) { |
| 359 | case CSD_TYPE_ASYNC: |
| 360 | case CSD_TYPE_SYNC: |
| 361 | case CSD_TYPE_IRQ_WORK: |
| 362 | pr_warn("IPI callback %pS sent to offline CPU\n", |
| 363 | csd->func); |
| 364 | break; |
| 365 | |
| 366 | case CSD_TYPE_TTWU: |
| 367 | pr_warn("IPI task-wakeup sent to offline CPU\n"); |
| 368 | break; |
| 369 | |
| 370 | default: |
| 371 | pr_warn("IPI callback, unknown type %d, sent to offline CPU\n", |
| 372 | CSD_TYPE(csd)); |
| 373 | break; |
| 374 | } |
| 375 | } |
| 376 | } |
| 377 | |
| 378 | /* |
| 379 | * First; run all SYNC callbacks, people are waiting for us. |
| 380 | */ |
| 381 | prev = NULL; |
| 382 | llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { |
| 383 | /* Do we wait until *after* callback? */ |
| 384 | if (CSD_TYPE(csd) == CSD_TYPE_SYNC) { |
| 385 | smp_call_func_t func = csd->func; |
| 386 | void *info = csd->info; |
| 387 | |
| 388 | if (prev) { |
| 389 | prev->next = &csd_next->node.llist; |
| 390 | } else { |
| 391 | entry = &csd_next->node.llist; |
| 392 | } |
| 393 | |
| 394 | csd_lock_record(csd); |
| 395 | func(info); |
| 396 | csd_unlock(csd); |
| 397 | csd_lock_record(NULL); |
| 398 | } else { |
| 399 | prev = &csd->node.llist; |
| 400 | } |
| 401 | } |
| 402 | |
| 403 | if (!entry) |
| 404 | return; |
| 405 | |
| 406 | /* |
| 407 | * Second; run all !SYNC callbacks. |
| 408 | */ |
| 409 | prev = NULL; |
| 410 | llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { |
| 411 | int type = CSD_TYPE(csd); |
| 412 | |
| 413 | if (type != CSD_TYPE_TTWU) { |
| 414 | if (prev) { |
| 415 | prev->next = &csd_next->node.llist; |
| 416 | } else { |
| 417 | entry = &csd_next->node.llist; |
| 418 | } |
| 419 | |
| 420 | if (type == CSD_TYPE_ASYNC) { |
| 421 | smp_call_func_t func = csd->func; |
| 422 | void *info = csd->info; |
| 423 | |
| 424 | csd_lock_record(csd); |
| 425 | csd_unlock(csd); |
| 426 | func(info); |
| 427 | csd_lock_record(NULL); |
| 428 | } else if (type == CSD_TYPE_IRQ_WORK) { |
| 429 | irq_work_single(csd); |
| 430 | } |
| 431 | |
| 432 | } else { |
| 433 | prev = &csd->node.llist; |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | /* |
| 438 | * Third; only CSD_TYPE_TTWU is left, issue those. |
| 439 | */ |
| 440 | if (entry) |
| 441 | sched_ttwu_pending(entry); |
| 442 | } |
| 443 | |
| 444 | void flush_smp_call_function_from_idle(void) |
| 445 | { |
| 446 | unsigned long flags; |
| 447 | |
| 448 | if (llist_empty(this_cpu_ptr(&call_single_queue))) |
| 449 | return; |
| 450 | |
| 451 | local_irq_save(flags); |
| 452 | flush_smp_call_function_queue(true); |
| 453 | if (local_softirq_pending()) |
| 454 | do_softirq(); |
| 455 | |
| 456 | local_irq_restore(flags); |
| 457 | } |
| 458 | |
| 459 | /* |
| 460 | * smp_call_function_single - Run a function on a specific CPU |
| 461 | * @func: The function to run. This must be fast and non-blocking. |
| 462 | * @info: An arbitrary pointer to pass to the function. |
| 463 | * @wait: If true, wait until function has completed on other CPUs. |
| 464 | * |
| 465 | * Returns 0 on success, else a negative status code. |
| 466 | */ |
| 467 | int smp_call_function_single(int cpu, smp_call_func_t func, void *info, |
| 468 | int wait) |
| 469 | { |
| 470 | call_single_data_t *csd; |
| 471 | call_single_data_t csd_stack = { |
| 472 | .node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, }, |
| 473 | }; |
| 474 | int this_cpu; |
| 475 | int err; |
| 476 | |
| 477 | /* |
| 478 | * prevent preemption and reschedule on another processor, |
| 479 | * as well as CPU removal |
| 480 | */ |
| 481 | this_cpu = get_cpu(); |
| 482 | |
| 483 | /* |
| 484 | * Can deadlock when called with interrupts disabled. |
| 485 | * We allow cpu's that are not yet online though, as no one else can |
| 486 | * send smp call function interrupt to this cpu and as such deadlocks |
| 487 | * can't happen. |
| 488 | */ |
| 489 | WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() |
| 490 | && !oops_in_progress); |
| 491 | |
| 492 | /* |
| 493 | * When @wait we can deadlock when we interrupt between llist_add() and |
| 494 | * arch_send_call_function_ipi*(); when !@wait we can deadlock due to |
| 495 | * csd_lock() on because the interrupt context uses the same csd |
| 496 | * storage. |
| 497 | */ |
| 498 | WARN_ON_ONCE(!in_task()); |
| 499 | |
| 500 | csd = &csd_stack; |
| 501 | if (!wait) { |
| 502 | csd = this_cpu_ptr(&csd_data); |
| 503 | csd_lock(csd); |
| 504 | } |
| 505 | |
| 506 | csd->func = func; |
| 507 | csd->info = info; |
| 508 | #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG |
| 509 | csd->node.src = smp_processor_id(); |
| 510 | csd->node.dst = cpu; |
| 511 | #endif |
| 512 | |
| 513 | err = generic_exec_single(cpu, csd); |
| 514 | |
| 515 | if (wait) |
| 516 | csd_lock_wait(csd); |
| 517 | |
| 518 | put_cpu(); |
| 519 | |
| 520 | return err; |
| 521 | } |
| 522 | EXPORT_SYMBOL(smp_call_function_single); |
| 523 | |
| 524 | /** |
| 525 | * smp_call_function_single_async(): Run an asynchronous function on a |
| 526 | * specific CPU. |
| 527 | * @cpu: The CPU to run on. |
| 528 | * @csd: Pre-allocated and setup data structure |
| 529 | * |
| 530 | * Like smp_call_function_single(), but the call is asynchonous and |
| 531 | * can thus be done from contexts with disabled interrupts. |
| 532 | * |
| 533 | * The caller passes his own pre-allocated data structure |
| 534 | * (ie: embedded in an object) and is responsible for synchronizing it |
| 535 | * such that the IPIs performed on the @csd are strictly serialized. |
| 536 | * |
| 537 | * If the function is called with one csd which has not yet been |
| 538 | * processed by previous call to smp_call_function_single_async(), the |
| 539 | * function will return immediately with -EBUSY showing that the csd |
| 540 | * object is still in progress. |
| 541 | * |
| 542 | * NOTE: Be careful, there is unfortunately no current debugging facility to |
| 543 | * validate the correctness of this serialization. |
| 544 | */ |
| 545 | int smp_call_function_single_async(int cpu, call_single_data_t *csd) |
| 546 | { |
| 547 | int err = 0; |
| 548 | |
| 549 | preempt_disable(); |
| 550 | |
| 551 | if (csd->node.u_flags & CSD_FLAG_LOCK) { |
| 552 | err = -EBUSY; |
| 553 | goto out; |
| 554 | } |
| 555 | |
| 556 | csd->node.u_flags = CSD_FLAG_LOCK; |
| 557 | smp_wmb(); |
| 558 | |
| 559 | err = generic_exec_single(cpu, csd); |
| 560 | |
| 561 | out: |
| 562 | preempt_enable(); |
| 563 | |
| 564 | return err; |
| 565 | } |
| 566 | EXPORT_SYMBOL_GPL(smp_call_function_single_async); |
| 567 | |
| 568 | /* |
| 569 | * smp_call_function_any - Run a function on any of the given cpus |
| 570 | * @mask: The mask of cpus it can run on. |
| 571 | * @func: The function to run. This must be fast and non-blocking. |
| 572 | * @info: An arbitrary pointer to pass to the function. |
| 573 | * @wait: If true, wait until function has completed. |
| 574 | * |
| 575 | * Returns 0 on success, else a negative status code (if no cpus were online). |
| 576 | * |
| 577 | * Selection preference: |
| 578 | * 1) current cpu if in @mask |
| 579 | * 2) any cpu of current node if in @mask |
| 580 | * 3) any other online cpu in @mask |
| 581 | */ |
| 582 | int smp_call_function_any(const struct cpumask *mask, |
| 583 | smp_call_func_t func, void *info, int wait) |
| 584 | { |
| 585 | unsigned int cpu; |
| 586 | const struct cpumask *nodemask; |
| 587 | int ret; |
| 588 | |
| 589 | /* Try for same CPU (cheapest) */ |
| 590 | cpu = get_cpu(); |
| 591 | if (cpumask_test_cpu(cpu, mask)) |
| 592 | goto call; |
| 593 | |
| 594 | /* Try for same node. */ |
| 595 | nodemask = cpumask_of_node(cpu_to_node(cpu)); |
| 596 | for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids; |
| 597 | cpu = cpumask_next_and(cpu, nodemask, mask)) { |
| 598 | if (cpu_online(cpu)) |
| 599 | goto call; |
| 600 | } |
| 601 | |
| 602 | /* Any online will do: smp_call_function_single handles nr_cpu_ids. */ |
| 603 | cpu = cpumask_any_and(mask, cpu_online_mask); |
| 604 | call: |
| 605 | ret = smp_call_function_single(cpu, func, info, wait); |
| 606 | put_cpu(); |
| 607 | return ret; |
| 608 | } |
| 609 | EXPORT_SYMBOL_GPL(smp_call_function_any); |
| 610 | |
| 611 | static void smp_call_function_many_cond(const struct cpumask *mask, |
| 612 | smp_call_func_t func, void *info, |
| 613 | bool wait, smp_cond_func_t cond_func) |
| 614 | { |
| 615 | struct call_function_data *cfd; |
| 616 | int cpu, next_cpu, this_cpu = smp_processor_id(); |
| 617 | |
| 618 | /* |
| 619 | * Can deadlock when called with interrupts disabled. |
| 620 | * We allow cpu's that are not yet online though, as no one else can |
| 621 | * send smp call function interrupt to this cpu and as such deadlocks |
| 622 | * can't happen. |
| 623 | */ |
| 624 | WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() |
| 625 | && !oops_in_progress && !early_boot_irqs_disabled); |
| 626 | |
| 627 | /* |
| 628 | * When @wait we can deadlock when we interrupt between llist_add() and |
| 629 | * arch_send_call_function_ipi*(); when !@wait we can deadlock due to |
| 630 | * csd_lock() on because the interrupt context uses the same csd |
| 631 | * storage. |
| 632 | */ |
| 633 | WARN_ON_ONCE(!in_task()); |
| 634 | |
| 635 | /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */ |
| 636 | cpu = cpumask_first_and(mask, cpu_online_mask); |
| 637 | if (cpu == this_cpu) |
| 638 | cpu = cpumask_next_and(cpu, mask, cpu_online_mask); |
| 639 | |
| 640 | /* No online cpus? We're done. */ |
| 641 | if (cpu >= nr_cpu_ids) |
| 642 | return; |
| 643 | |
| 644 | /* Do we have another CPU which isn't us? */ |
| 645 | next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask); |
| 646 | if (next_cpu == this_cpu) |
| 647 | next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask); |
| 648 | |
| 649 | /* Fastpath: do that cpu by itself. */ |
| 650 | if (next_cpu >= nr_cpu_ids) { |
| 651 | if (!cond_func || cond_func(cpu, info)) |
| 652 | smp_call_function_single(cpu, func, info, wait); |
| 653 | return; |
| 654 | } |
| 655 | |
| 656 | cfd = this_cpu_ptr(&cfd_data); |
| 657 | |
| 658 | cpumask_and(cfd->cpumask, mask, cpu_online_mask); |
| 659 | __cpumask_clear_cpu(this_cpu, cfd->cpumask); |
| 660 | |
| 661 | /* Some callers race with other cpus changing the passed mask */ |
| 662 | if (unlikely(!cpumask_weight(cfd->cpumask))) |
| 663 | return; |
| 664 | |
| 665 | cpumask_clear(cfd->cpumask_ipi); |
| 666 | for_each_cpu(cpu, cfd->cpumask) { |
| 667 | call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu); |
| 668 | |
| 669 | if (cond_func && !cond_func(cpu, info)) |
| 670 | continue; |
| 671 | |
| 672 | csd_lock(csd); |
| 673 | if (wait) |
| 674 | csd->node.u_flags |= CSD_TYPE_SYNC; |
| 675 | csd->func = func; |
| 676 | csd->info = info; |
| 677 | #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG |
| 678 | csd->node.src = smp_processor_id(); |
| 679 | csd->node.dst = cpu; |
| 680 | #endif |
| 681 | if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) |
| 682 | __cpumask_set_cpu(cpu, cfd->cpumask_ipi); |
| 683 | } |
| 684 | |
| 685 | /* Send a message to all CPUs in the map */ |
| 686 | arch_send_call_function_ipi_mask(cfd->cpumask_ipi); |
| 687 | |
| 688 | if (wait) { |
| 689 | for_each_cpu(cpu, cfd->cpumask) { |
| 690 | call_single_data_t *csd; |
| 691 | |
| 692 | csd = per_cpu_ptr(cfd->csd, cpu); |
| 693 | csd_lock_wait(csd); |
| 694 | } |
| 695 | } |
| 696 | } |
| 697 | |
| 698 | /** |
| 699 | * smp_call_function_many(): Run a function on a set of other CPUs. |
| 700 | * @mask: The set of cpus to run on (only runs on online subset). |
| 701 | * @func: The function to run. This must be fast and non-blocking. |
| 702 | * @info: An arbitrary pointer to pass to the function. |
| 703 | * @wait: If true, wait (atomically) until function has completed |
| 704 | * on other CPUs. |
| 705 | * |
| 706 | * If @wait is true, then returns once @func has returned. |
| 707 | * |
| 708 | * You must not call this function with disabled interrupts or from a |
| 709 | * hardware interrupt handler or from a bottom half handler. Preemption |
| 710 | * must be disabled when calling this function. |
| 711 | */ |
| 712 | void smp_call_function_many(const struct cpumask *mask, |
| 713 | smp_call_func_t func, void *info, bool wait) |
| 714 | { |
| 715 | smp_call_function_many_cond(mask, func, info, wait, NULL); |
| 716 | } |
| 717 | EXPORT_SYMBOL(smp_call_function_many); |
| 718 | |
| 719 | /** |
| 720 | * smp_call_function(): Run a function on all other CPUs. |
| 721 | * @func: The function to run. This must be fast and non-blocking. |
| 722 | * @info: An arbitrary pointer to pass to the function. |
| 723 | * @wait: If true, wait (atomically) until function has completed |
| 724 | * on other CPUs. |
| 725 | * |
| 726 | * Returns 0. |
| 727 | * |
| 728 | * If @wait is true, then returns once @func has returned; otherwise |
| 729 | * it returns just before the target cpu calls @func. |
| 730 | * |
| 731 | * You must not call this function with disabled interrupts or from a |
| 732 | * hardware interrupt handler or from a bottom half handler. |
| 733 | */ |
| 734 | void smp_call_function(smp_call_func_t func, void *info, int wait) |
| 735 | { |
| 736 | preempt_disable(); |
| 737 | smp_call_function_many(cpu_online_mask, func, info, wait); |
| 738 | preempt_enable(); |
| 739 | } |
| 740 | EXPORT_SYMBOL(smp_call_function); |
| 741 | |
| 742 | /* Setup configured maximum number of CPUs to activate */ |
| 743 | unsigned int setup_max_cpus = NR_CPUS; |
| 744 | EXPORT_SYMBOL(setup_max_cpus); |
| 745 | |
| 746 | |
| 747 | /* |
| 748 | * Setup routine for controlling SMP activation |
| 749 | * |
| 750 | * Command-line option of "nosmp" or "maxcpus=0" will disable SMP |
| 751 | * activation entirely (the MPS table probe still happens, though). |
| 752 | * |
| 753 | * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer |
| 754 | * greater than 0, limits the maximum number of CPUs activated in |
| 755 | * SMP mode to <NUM>. |
| 756 | */ |
| 757 | |
| 758 | void __weak arch_disable_smp_support(void) { } |
| 759 | |
| 760 | static int __init nosmp(char *str) |
| 761 | { |
| 762 | setup_max_cpus = 0; |
| 763 | arch_disable_smp_support(); |
| 764 | |
| 765 | return 0; |
| 766 | } |
| 767 | |
| 768 | early_param("nosmp", nosmp); |
| 769 | |
| 770 | /* this is hard limit */ |
| 771 | static int __init nrcpus(char *str) |
| 772 | { |
| 773 | int nr_cpus; |
| 774 | |
| 775 | if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids) |
| 776 | nr_cpu_ids = nr_cpus; |
| 777 | |
| 778 | return 0; |
| 779 | } |
| 780 | |
| 781 | early_param("nr_cpus", nrcpus); |
| 782 | |
| 783 | static int __init maxcpus(char *str) |
| 784 | { |
| 785 | get_option(&str, &setup_max_cpus); |
| 786 | if (setup_max_cpus == 0) |
| 787 | arch_disable_smp_support(); |
| 788 | |
| 789 | return 0; |
| 790 | } |
| 791 | |
| 792 | early_param("maxcpus", maxcpus); |
| 793 | |
| 794 | /* Setup number of possible processor ids */ |
| 795 | unsigned int nr_cpu_ids __read_mostly = NR_CPUS; |
| 796 | EXPORT_SYMBOL(nr_cpu_ids); |
| 797 | |
| 798 | /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */ |
| 799 | void __init setup_nr_cpu_ids(void) |
| 800 | { |
| 801 | nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1; |
| 802 | } |
| 803 | |
| 804 | /* Called by boot processor to activate the rest. */ |
| 805 | void __init smp_init(void) |
| 806 | { |
| 807 | int num_nodes, num_cpus; |
| 808 | |
| 809 | idle_threads_init(); |
| 810 | cpuhp_threads_init(); |
| 811 | |
| 812 | pr_info("Bringing up secondary CPUs ...\n"); |
| 813 | |
| 814 | bringup_nonboot_cpus(setup_max_cpus); |
| 815 | |
| 816 | num_nodes = num_online_nodes(); |
| 817 | num_cpus = num_online_cpus(); |
| 818 | pr_info("Brought up %d node%s, %d CPU%s\n", |
| 819 | num_nodes, (num_nodes > 1 ? "s" : ""), |
| 820 | num_cpus, (num_cpus > 1 ? "s" : "")); |
| 821 | |
| 822 | /* Any cleanup work */ |
| 823 | smp_cpus_done(setup_max_cpus); |
| 824 | } |
| 825 | |
| 826 | /* |
| 827 | * Call a function on all processors. May be used during early boot while |
| 828 | * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead |
| 829 | * of local_irq_disable/enable(). |
| 830 | */ |
| 831 | void on_each_cpu(smp_call_func_t func, void *info, int wait) |
| 832 | { |
| 833 | unsigned long flags; |
| 834 | |
| 835 | preempt_disable(); |
| 836 | smp_call_function(func, info, wait); |
| 837 | local_irq_save(flags); |
| 838 | func(info); |
| 839 | local_irq_restore(flags); |
| 840 | preempt_enable(); |
| 841 | } |
| 842 | EXPORT_SYMBOL(on_each_cpu); |
| 843 | |
| 844 | /** |
| 845 | * on_each_cpu_mask(): Run a function on processors specified by |
| 846 | * cpumask, which may include the local processor. |
| 847 | * @mask: The set of cpus to run on (only runs on online subset). |
| 848 | * @func: The function to run. This must be fast and non-blocking. |
| 849 | * @info: An arbitrary pointer to pass to the function. |
| 850 | * @wait: If true, wait (atomically) until function has completed |
| 851 | * on other CPUs. |
| 852 | * |
| 853 | * If @wait is true, then returns once @func has returned. |
| 854 | * |
| 855 | * You must not call this function with disabled interrupts or from a |
| 856 | * hardware interrupt handler or from a bottom half handler. The |
| 857 | * exception is that it may be used during early boot while |
| 858 | * early_boot_irqs_disabled is set. |
| 859 | */ |
| 860 | void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func, |
| 861 | void *info, bool wait) |
| 862 | { |
| 863 | int cpu = get_cpu(); |
| 864 | |
| 865 | smp_call_function_many(mask, func, info, wait); |
| 866 | if (cpumask_test_cpu(cpu, mask)) { |
| 867 | unsigned long flags; |
| 868 | local_irq_save(flags); |
| 869 | func(info); |
| 870 | local_irq_restore(flags); |
| 871 | } |
| 872 | put_cpu(); |
| 873 | } |
| 874 | EXPORT_SYMBOL(on_each_cpu_mask); |
| 875 | |
| 876 | /* |
| 877 | * on_each_cpu_cond(): Call a function on each processor for which |
| 878 | * the supplied function cond_func returns true, optionally waiting |
| 879 | * for all the required CPUs to finish. This may include the local |
| 880 | * processor. |
| 881 | * @cond_func: A callback function that is passed a cpu id and |
| 882 | * the info parameter. The function is called |
| 883 | * with preemption disabled. The function should |
| 884 | * return a blooean value indicating whether to IPI |
| 885 | * the specified CPU. |
| 886 | * @func: The function to run on all applicable CPUs. |
| 887 | * This must be fast and non-blocking. |
| 888 | * @info: An arbitrary pointer to pass to both functions. |
| 889 | * @wait: If true, wait (atomically) until function has |
| 890 | * completed on other CPUs. |
| 891 | * |
| 892 | * Preemption is disabled to protect against CPUs going offline but not online. |
| 893 | * CPUs going online during the call will not be seen or sent an IPI. |
| 894 | * |
| 895 | * You must not call this function with disabled interrupts or |
| 896 | * from a hardware interrupt handler or from a bottom half handler. |
| 897 | */ |
| 898 | void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func, |
| 899 | void *info, bool wait, const struct cpumask *mask) |
| 900 | { |
| 901 | int cpu = get_cpu(); |
| 902 | |
| 903 | smp_call_function_many_cond(mask, func, info, wait, cond_func); |
| 904 | if (cpumask_test_cpu(cpu, mask) && cond_func(cpu, info)) { |
| 905 | unsigned long flags; |
| 906 | |
| 907 | local_irq_save(flags); |
| 908 | func(info); |
| 909 | local_irq_restore(flags); |
| 910 | } |
| 911 | put_cpu(); |
| 912 | } |
| 913 | EXPORT_SYMBOL(on_each_cpu_cond_mask); |
| 914 | |
| 915 | void on_each_cpu_cond(smp_cond_func_t cond_func, smp_call_func_t func, |
| 916 | void *info, bool wait) |
| 917 | { |
| 918 | on_each_cpu_cond_mask(cond_func, func, info, wait, cpu_online_mask); |
| 919 | } |
| 920 | EXPORT_SYMBOL(on_each_cpu_cond); |
| 921 | |
| 922 | static void do_nothing(void *unused) |
| 923 | { |
| 924 | } |
| 925 | |
| 926 | /** |
| 927 | * kick_all_cpus_sync - Force all cpus out of idle |
| 928 | * |
| 929 | * Used to synchronize the update of pm_idle function pointer. It's |
| 930 | * called after the pointer is updated and returns after the dummy |
| 931 | * callback function has been executed on all cpus. The execution of |
| 932 | * the function can only happen on the remote cpus after they have |
| 933 | * left the idle function which had been called via pm_idle function |
| 934 | * pointer. So it's guaranteed that nothing uses the previous pointer |
| 935 | * anymore. |
| 936 | */ |
| 937 | void kick_all_cpus_sync(void) |
| 938 | { |
| 939 | /* Make sure the change is visible before we kick the cpus */ |
| 940 | smp_mb(); |
| 941 | smp_call_function(do_nothing, NULL, 1); |
| 942 | } |
| 943 | EXPORT_SYMBOL_GPL(kick_all_cpus_sync); |
| 944 | |
| 945 | /** |
| 946 | * wake_up_all_idle_cpus - break all cpus out of idle |
| 947 | * wake_up_all_idle_cpus try to break all cpus which is in idle state even |
| 948 | * including idle polling cpus, for non-idle cpus, we will do nothing |
| 949 | * for them. |
| 950 | */ |
| 951 | void wake_up_all_idle_cpus(void) |
| 952 | { |
| 953 | int cpu; |
| 954 | |
| 955 | preempt_disable(); |
| 956 | for_each_online_cpu(cpu) { |
| 957 | if (cpu == smp_processor_id()) |
| 958 | continue; |
| 959 | |
| 960 | wake_up_if_idle(cpu); |
| 961 | } |
| 962 | preempt_enable(); |
| 963 | } |
| 964 | EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus); |
| 965 | |
| 966 | /** |
| 967 | * smp_call_on_cpu - Call a function on a specific cpu |
| 968 | * |
| 969 | * Used to call a function on a specific cpu and wait for it to return. |
| 970 | * Optionally make sure the call is done on a specified physical cpu via vcpu |
| 971 | * pinning in order to support virtualized environments. |
| 972 | */ |
| 973 | struct smp_call_on_cpu_struct { |
| 974 | struct work_struct work; |
| 975 | struct completion done; |
| 976 | int (*func)(void *); |
| 977 | void *data; |
| 978 | int ret; |
| 979 | int cpu; |
| 980 | }; |
| 981 | |
| 982 | static void smp_call_on_cpu_callback(struct work_struct *work) |
| 983 | { |
| 984 | struct smp_call_on_cpu_struct *sscs; |
| 985 | |
| 986 | sscs = container_of(work, struct smp_call_on_cpu_struct, work); |
| 987 | if (sscs->cpu >= 0) |
| 988 | hypervisor_pin_vcpu(sscs->cpu); |
| 989 | sscs->ret = sscs->func(sscs->data); |
| 990 | if (sscs->cpu >= 0) |
| 991 | hypervisor_pin_vcpu(-1); |
| 992 | |
| 993 | complete(&sscs->done); |
| 994 | } |
| 995 | |
| 996 | int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys) |
| 997 | { |
| 998 | struct smp_call_on_cpu_struct sscs = { |
| 999 | .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done), |
| 1000 | .func = func, |
| 1001 | .data = par, |
| 1002 | .cpu = phys ? cpu : -1, |
| 1003 | }; |
| 1004 | |
| 1005 | INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback); |
| 1006 | |
| 1007 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
| 1008 | return -ENXIO; |
| 1009 | |
| 1010 | queue_work_on(cpu, system_wq, &sscs.work); |
| 1011 | wait_for_completion(&sscs.done); |
| 1012 | |
| 1013 | return sscs.ret; |
| 1014 | } |
| 1015 | EXPORT_SYMBOL_GPL(smp_call_on_cpu); |