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 #include <linux/jump_label.h> 28 29 #include "smpboot.h" 30 #include "sched/smp.h" 31 32 #define CSD_TYPE(_csd) ((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK) 33 34 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 35 union cfd_seq_cnt { 36 u64 val; 37 struct { 38 u64 src:16; 39 u64 dst:16; 40 #define CFD_SEQ_NOCPU 0xffff 41 u64 type:4; 42 #define CFD_SEQ_QUEUE 0 43 #define CFD_SEQ_IPI 1 44 #define CFD_SEQ_NOIPI 2 45 #define CFD_SEQ_PING 3 46 #define CFD_SEQ_PINGED 4 47 #define CFD_SEQ_HANDLE 5 48 #define CFD_SEQ_DEQUEUE 6 49 #define CFD_SEQ_IDLE 7 50 #define CFD_SEQ_GOTIPI 8 51 #define CFD_SEQ_HDLEND 9 52 u64 cnt:28; 53 } u; 54 }; 55 56 static char *seq_type[] = { 57 [CFD_SEQ_QUEUE] = "queue", 58 [CFD_SEQ_IPI] = "ipi", 59 [CFD_SEQ_NOIPI] = "noipi", 60 [CFD_SEQ_PING] = "ping", 61 [CFD_SEQ_PINGED] = "pinged", 62 [CFD_SEQ_HANDLE] = "handle", 63 [CFD_SEQ_DEQUEUE] = "dequeue (src CPU 0 == empty)", 64 [CFD_SEQ_IDLE] = "idle", 65 [CFD_SEQ_GOTIPI] = "gotipi", 66 [CFD_SEQ_HDLEND] = "hdlend (src CPU 0 == early)", 67 }; 68 69 struct cfd_seq_local { 70 u64 ping; 71 u64 pinged; 72 u64 handle; 73 u64 dequeue; 74 u64 idle; 75 u64 gotipi; 76 u64 hdlend; 77 }; 78 #endif 79 80 struct cfd_percpu { 81 call_single_data_t csd; 82 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 83 u64 seq_queue; 84 u64 seq_ipi; 85 u64 seq_noipi; 86 #endif 87 }; 88 89 struct call_function_data { 90 struct cfd_percpu __percpu *pcpu; 91 cpumask_var_t cpumask; 92 cpumask_var_t cpumask_ipi; 93 }; 94 95 static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data); 96 97 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue); 98 99 static void __flush_smp_call_function_queue(bool warn_cpu_offline); 100 101 int smpcfd_prepare_cpu(unsigned int cpu) 102 { 103 struct call_function_data *cfd = &per_cpu(cfd_data, cpu); 104 105 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL, 106 cpu_to_node(cpu))) 107 return -ENOMEM; 108 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL, 109 cpu_to_node(cpu))) { 110 free_cpumask_var(cfd->cpumask); 111 return -ENOMEM; 112 } 113 cfd->pcpu = alloc_percpu(struct cfd_percpu); 114 if (!cfd->pcpu) { 115 free_cpumask_var(cfd->cpumask); 116 free_cpumask_var(cfd->cpumask_ipi); 117 return -ENOMEM; 118 } 119 120 return 0; 121 } 122 123 int smpcfd_dead_cpu(unsigned int cpu) 124 { 125 struct call_function_data *cfd = &per_cpu(cfd_data, cpu); 126 127 free_cpumask_var(cfd->cpumask); 128 free_cpumask_var(cfd->cpumask_ipi); 129 free_percpu(cfd->pcpu); 130 return 0; 131 } 132 133 int smpcfd_dying_cpu(unsigned int cpu) 134 { 135 /* 136 * The IPIs for the smp-call-function callbacks queued by other 137 * CPUs might arrive late, either due to hardware latencies or 138 * because this CPU disabled interrupts (inside stop-machine) 139 * before the IPIs were sent. So flush out any pending callbacks 140 * explicitly (without waiting for the IPIs to arrive), to 141 * ensure that the outgoing CPU doesn't go offline with work 142 * still pending. 143 */ 144 __flush_smp_call_function_queue(false); 145 irq_work_run(); 146 return 0; 147 } 148 149 void __init call_function_init(void) 150 { 151 int i; 152 153 for_each_possible_cpu(i) 154 init_llist_head(&per_cpu(call_single_queue, i)); 155 156 smpcfd_prepare_cpu(smp_processor_id()); 157 } 158 159 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 160 161 static DEFINE_STATIC_KEY_FALSE(csdlock_debug_enabled); 162 static DEFINE_STATIC_KEY_FALSE(csdlock_debug_extended); 163 164 static int __init csdlock_debug(char *str) 165 { 166 unsigned int val = 0; 167 168 if (str && !strcmp(str, "ext")) { 169 val = 1; 170 static_branch_enable(&csdlock_debug_extended); 171 } else 172 get_option(&str, &val); 173 174 if (val) 175 static_branch_enable(&csdlock_debug_enabled); 176 177 return 1; 178 } 179 __setup("csdlock_debug=", csdlock_debug); 180 181 static DEFINE_PER_CPU(call_single_data_t *, cur_csd); 182 static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func); 183 static DEFINE_PER_CPU(void *, cur_csd_info); 184 static DEFINE_PER_CPU(struct cfd_seq_local, cfd_seq_local); 185 186 static ulong csd_lock_timeout = 5000; /* CSD lock timeout in milliseconds. */ 187 module_param(csd_lock_timeout, ulong, 0444); 188 189 static atomic_t csd_bug_count = ATOMIC_INIT(0); 190 static u64 cfd_seq; 191 192 #define CFD_SEQ(s, d, t, c) \ 193 (union cfd_seq_cnt){ .u.src = s, .u.dst = d, .u.type = t, .u.cnt = c } 194 195 static u64 cfd_seq_inc(unsigned int src, unsigned int dst, unsigned int type) 196 { 197 union cfd_seq_cnt new, old; 198 199 new = CFD_SEQ(src, dst, type, 0); 200 201 do { 202 old.val = READ_ONCE(cfd_seq); 203 new.u.cnt = old.u.cnt + 1; 204 } while (cmpxchg(&cfd_seq, old.val, new.val) != old.val); 205 206 return old.val; 207 } 208 209 #define cfd_seq_store(var, src, dst, type) \ 210 do { \ 211 if (static_branch_unlikely(&csdlock_debug_extended)) \ 212 var = cfd_seq_inc(src, dst, type); \ 213 } while (0) 214 215 /* Record current CSD work for current CPU, NULL to erase. */ 216 static void __csd_lock_record(struct __call_single_data *csd) 217 { 218 if (!csd) { 219 smp_mb(); /* NULL cur_csd after unlock. */ 220 __this_cpu_write(cur_csd, NULL); 221 return; 222 } 223 __this_cpu_write(cur_csd_func, csd->func); 224 __this_cpu_write(cur_csd_info, csd->info); 225 smp_wmb(); /* func and info before csd. */ 226 __this_cpu_write(cur_csd, csd); 227 smp_mb(); /* Update cur_csd before function call. */ 228 /* Or before unlock, as the case may be. */ 229 } 230 231 static __always_inline void csd_lock_record(struct __call_single_data *csd) 232 { 233 if (static_branch_unlikely(&csdlock_debug_enabled)) 234 __csd_lock_record(csd); 235 } 236 237 static int csd_lock_wait_getcpu(struct __call_single_data *csd) 238 { 239 unsigned int csd_type; 240 241 csd_type = CSD_TYPE(csd); 242 if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC) 243 return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */ 244 return -1; 245 } 246 247 static void cfd_seq_data_add(u64 val, unsigned int src, unsigned int dst, 248 unsigned int type, union cfd_seq_cnt *data, 249 unsigned int *n_data, unsigned int now) 250 { 251 union cfd_seq_cnt new[2]; 252 unsigned int i, j, k; 253 254 new[0].val = val; 255 new[1] = CFD_SEQ(src, dst, type, new[0].u.cnt + 1); 256 257 for (i = 0; i < 2; i++) { 258 if (new[i].u.cnt <= now) 259 new[i].u.cnt |= 0x80000000U; 260 for (j = 0; j < *n_data; j++) { 261 if (new[i].u.cnt == data[j].u.cnt) { 262 /* Direct read value trumps generated one. */ 263 if (i == 0) 264 data[j].val = new[i].val; 265 break; 266 } 267 if (new[i].u.cnt < data[j].u.cnt) { 268 for (k = *n_data; k > j; k--) 269 data[k].val = data[k - 1].val; 270 data[j].val = new[i].val; 271 (*n_data)++; 272 break; 273 } 274 } 275 if (j == *n_data) { 276 data[j].val = new[i].val; 277 (*n_data)++; 278 } 279 } 280 } 281 282 static const char *csd_lock_get_type(unsigned int type) 283 { 284 return (type >= ARRAY_SIZE(seq_type)) ? "?" : seq_type[type]; 285 } 286 287 static void csd_lock_print_extended(struct __call_single_data *csd, int cpu) 288 { 289 struct cfd_seq_local *seq = &per_cpu(cfd_seq_local, cpu); 290 unsigned int srccpu = csd->node.src; 291 struct call_function_data *cfd = per_cpu_ptr(&cfd_data, srccpu); 292 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 293 unsigned int now; 294 union cfd_seq_cnt data[2 * ARRAY_SIZE(seq_type)]; 295 unsigned int n_data = 0, i; 296 297 data[0].val = READ_ONCE(cfd_seq); 298 now = data[0].u.cnt; 299 300 cfd_seq_data_add(pcpu->seq_queue, srccpu, cpu, CFD_SEQ_QUEUE, data, &n_data, now); 301 cfd_seq_data_add(pcpu->seq_ipi, srccpu, cpu, CFD_SEQ_IPI, data, &n_data, now); 302 cfd_seq_data_add(pcpu->seq_noipi, srccpu, cpu, CFD_SEQ_NOIPI, data, &n_data, now); 303 304 cfd_seq_data_add(per_cpu(cfd_seq_local.ping, srccpu), srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PING, data, &n_data, now); 305 cfd_seq_data_add(per_cpu(cfd_seq_local.pinged, srccpu), srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED, data, &n_data, now); 306 307 cfd_seq_data_add(seq->idle, CFD_SEQ_NOCPU, cpu, CFD_SEQ_IDLE, data, &n_data, now); 308 cfd_seq_data_add(seq->gotipi, CFD_SEQ_NOCPU, cpu, CFD_SEQ_GOTIPI, data, &n_data, now); 309 cfd_seq_data_add(seq->handle, CFD_SEQ_NOCPU, cpu, CFD_SEQ_HANDLE, data, &n_data, now); 310 cfd_seq_data_add(seq->dequeue, CFD_SEQ_NOCPU, cpu, CFD_SEQ_DEQUEUE, data, &n_data, now); 311 cfd_seq_data_add(seq->hdlend, CFD_SEQ_NOCPU, cpu, CFD_SEQ_HDLEND, data, &n_data, now); 312 313 for (i = 0; i < n_data; i++) { 314 pr_alert("\tcsd: cnt(%07x): %04x->%04x %s\n", 315 data[i].u.cnt & ~0x80000000U, data[i].u.src, 316 data[i].u.dst, csd_lock_get_type(data[i].u.type)); 317 } 318 pr_alert("\tcsd: cnt now: %07x\n", now); 319 } 320 321 /* 322 * Complain if too much time spent waiting. Note that only 323 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU, 324 * so waiting on other types gets much less information. 325 */ 326 static bool csd_lock_wait_toolong(struct __call_single_data *csd, u64 ts0, u64 *ts1, int *bug_id) 327 { 328 int cpu = -1; 329 int cpux; 330 bool firsttime; 331 u64 ts2, ts_delta; 332 call_single_data_t *cpu_cur_csd; 333 unsigned int flags = READ_ONCE(csd->node.u_flags); 334 unsigned long long csd_lock_timeout_ns = csd_lock_timeout * NSEC_PER_MSEC; 335 336 if (!(flags & CSD_FLAG_LOCK)) { 337 if (!unlikely(*bug_id)) 338 return true; 339 cpu = csd_lock_wait_getcpu(csd); 340 pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n", 341 *bug_id, raw_smp_processor_id(), cpu); 342 return true; 343 } 344 345 ts2 = sched_clock(); 346 ts_delta = ts2 - *ts1; 347 if (likely(ts_delta <= csd_lock_timeout_ns || csd_lock_timeout_ns == 0)) 348 return false; 349 350 firsttime = !*bug_id; 351 if (firsttime) 352 *bug_id = atomic_inc_return(&csd_bug_count); 353 cpu = csd_lock_wait_getcpu(csd); 354 if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu)) 355 cpux = 0; 356 else 357 cpux = cpu; 358 cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */ 359 pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n", 360 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0, 361 cpu, csd->func, csd->info); 362 if (cpu_cur_csd && csd != cpu_cur_csd) { 363 pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n", 364 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)), 365 READ_ONCE(per_cpu(cur_csd_info, cpux))); 366 } else { 367 pr_alert("\tcsd: CSD lock (#%d) %s.\n", 368 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request"); 369 } 370 if (cpu >= 0) { 371 if (static_branch_unlikely(&csdlock_debug_extended)) 372 csd_lock_print_extended(csd, cpu); 373 dump_cpu_task(cpu); 374 if (!cpu_cur_csd) { 375 pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu); 376 arch_send_call_function_single_ipi(cpu); 377 } 378 } 379 dump_stack(); 380 *ts1 = ts2; 381 382 return false; 383 } 384 385 /* 386 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources 387 * 388 * For non-synchronous ipi calls the csd can still be in use by the 389 * previous function call. For multi-cpu calls its even more interesting 390 * as we'll have to ensure no other cpu is observing our csd. 391 */ 392 static void __csd_lock_wait(struct __call_single_data *csd) 393 { 394 int bug_id = 0; 395 u64 ts0, ts1; 396 397 ts1 = ts0 = sched_clock(); 398 for (;;) { 399 if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id)) 400 break; 401 cpu_relax(); 402 } 403 smp_acquire__after_ctrl_dep(); 404 } 405 406 static __always_inline void csd_lock_wait(struct __call_single_data *csd) 407 { 408 if (static_branch_unlikely(&csdlock_debug_enabled)) { 409 __csd_lock_wait(csd); 410 return; 411 } 412 413 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK)); 414 } 415 416 static void __smp_call_single_queue_debug(int cpu, struct llist_node *node) 417 { 418 unsigned int this_cpu = smp_processor_id(); 419 struct cfd_seq_local *seq = this_cpu_ptr(&cfd_seq_local); 420 struct call_function_data *cfd = this_cpu_ptr(&cfd_data); 421 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 422 423 cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE); 424 if (llist_add(node, &per_cpu(call_single_queue, cpu))) { 425 cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI); 426 cfd_seq_store(seq->ping, this_cpu, cpu, CFD_SEQ_PING); 427 send_call_function_single_ipi(cpu); 428 cfd_seq_store(seq->pinged, this_cpu, cpu, CFD_SEQ_PINGED); 429 } else { 430 cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI); 431 } 432 } 433 #else 434 #define cfd_seq_store(var, src, dst, type) 435 436 static void csd_lock_record(struct __call_single_data *csd) 437 { 438 } 439 440 static __always_inline void csd_lock_wait(struct __call_single_data *csd) 441 { 442 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK)); 443 } 444 #endif 445 446 static __always_inline void csd_lock(struct __call_single_data *csd) 447 { 448 csd_lock_wait(csd); 449 csd->node.u_flags |= CSD_FLAG_LOCK; 450 451 /* 452 * prevent CPU from reordering the above assignment 453 * to ->flags with any subsequent assignments to other 454 * fields of the specified call_single_data_t structure: 455 */ 456 smp_wmb(); 457 } 458 459 static __always_inline void csd_unlock(struct __call_single_data *csd) 460 { 461 WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK)); 462 463 /* 464 * ensure we're all done before releasing data: 465 */ 466 smp_store_release(&csd->node.u_flags, 0); 467 } 468 469 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data); 470 471 void __smp_call_single_queue(int cpu, struct llist_node *node) 472 { 473 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 474 if (static_branch_unlikely(&csdlock_debug_extended)) { 475 unsigned int type; 476 477 type = CSD_TYPE(container_of(node, call_single_data_t, 478 node.llist)); 479 if (type == CSD_TYPE_SYNC || type == CSD_TYPE_ASYNC) { 480 __smp_call_single_queue_debug(cpu, node); 481 return; 482 } 483 } 484 #endif 485 486 /* 487 * The list addition should be visible before sending the IPI 488 * handler locks the list to pull the entry off it because of 489 * normal cache coherency rules implied by spinlocks. 490 * 491 * If IPIs can go out of order to the cache coherency protocol 492 * in an architecture, sufficient synchronisation should be added 493 * to arch code to make it appear to obey cache coherency WRT 494 * locking and barrier primitives. Generic code isn't really 495 * equipped to do the right thing... 496 */ 497 if (llist_add(node, &per_cpu(call_single_queue, cpu))) 498 send_call_function_single_ipi(cpu); 499 } 500 501 /* 502 * Insert a previously allocated call_single_data_t element 503 * for execution on the given CPU. data must already have 504 * ->func, ->info, and ->flags set. 505 */ 506 static int generic_exec_single(int cpu, struct __call_single_data *csd) 507 { 508 if (cpu == smp_processor_id()) { 509 smp_call_func_t func = csd->func; 510 void *info = csd->info; 511 unsigned long flags; 512 513 /* 514 * We can unlock early even for the synchronous on-stack case, 515 * since we're doing this from the same CPU.. 516 */ 517 csd_lock_record(csd); 518 csd_unlock(csd); 519 local_irq_save(flags); 520 func(info); 521 csd_lock_record(NULL); 522 local_irq_restore(flags); 523 return 0; 524 } 525 526 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) { 527 csd_unlock(csd); 528 return -ENXIO; 529 } 530 531 __smp_call_single_queue(cpu, &csd->node.llist); 532 533 return 0; 534 } 535 536 /** 537 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks 538 * 539 * Invoked by arch to handle an IPI for call function single. 540 * Must be called with interrupts disabled. 541 */ 542 void generic_smp_call_function_single_interrupt(void) 543 { 544 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->gotipi, CFD_SEQ_NOCPU, 545 smp_processor_id(), CFD_SEQ_GOTIPI); 546 __flush_smp_call_function_queue(true); 547 } 548 549 /** 550 * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks 551 * 552 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an 553 * offline CPU. Skip this check if set to 'false'. 554 * 555 * Flush any pending smp-call-function callbacks queued on this CPU. This is 556 * invoked by the generic IPI handler, as well as by a CPU about to go offline, 557 * to ensure that all pending IPI callbacks are run before it goes completely 558 * offline. 559 * 560 * Loop through the call_single_queue and run all the queued callbacks. 561 * Must be called with interrupts disabled. 562 */ 563 static void __flush_smp_call_function_queue(bool warn_cpu_offline) 564 { 565 call_single_data_t *csd, *csd_next; 566 struct llist_node *entry, *prev; 567 struct llist_head *head; 568 static bool warned; 569 570 lockdep_assert_irqs_disabled(); 571 572 head = this_cpu_ptr(&call_single_queue); 573 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->handle, CFD_SEQ_NOCPU, 574 smp_processor_id(), CFD_SEQ_HANDLE); 575 entry = llist_del_all(head); 576 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->dequeue, 577 /* Special meaning of source cpu: 0 == queue empty */ 578 entry ? CFD_SEQ_NOCPU : 0, 579 smp_processor_id(), CFD_SEQ_DEQUEUE); 580 entry = llist_reverse_order(entry); 581 582 /* There shouldn't be any pending callbacks on an offline CPU. */ 583 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) && 584 !warned && entry != NULL)) { 585 warned = true; 586 WARN(1, "IPI on offline CPU %d\n", smp_processor_id()); 587 588 /* 589 * We don't have to use the _safe() variant here 590 * because we are not invoking the IPI handlers yet. 591 */ 592 llist_for_each_entry(csd, entry, node.llist) { 593 switch (CSD_TYPE(csd)) { 594 case CSD_TYPE_ASYNC: 595 case CSD_TYPE_SYNC: 596 case CSD_TYPE_IRQ_WORK: 597 pr_warn("IPI callback %pS sent to offline CPU\n", 598 csd->func); 599 break; 600 601 case CSD_TYPE_TTWU: 602 pr_warn("IPI task-wakeup sent to offline CPU\n"); 603 break; 604 605 default: 606 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n", 607 CSD_TYPE(csd)); 608 break; 609 } 610 } 611 } 612 613 /* 614 * First; run all SYNC callbacks, people are waiting for us. 615 */ 616 prev = NULL; 617 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { 618 /* Do we wait until *after* callback? */ 619 if (CSD_TYPE(csd) == CSD_TYPE_SYNC) { 620 smp_call_func_t func = csd->func; 621 void *info = csd->info; 622 623 if (prev) { 624 prev->next = &csd_next->node.llist; 625 } else { 626 entry = &csd_next->node.llist; 627 } 628 629 csd_lock_record(csd); 630 func(info); 631 csd_unlock(csd); 632 csd_lock_record(NULL); 633 } else { 634 prev = &csd->node.llist; 635 } 636 } 637 638 if (!entry) { 639 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend, 640 0, smp_processor_id(), 641 CFD_SEQ_HDLEND); 642 return; 643 } 644 645 /* 646 * Second; run all !SYNC callbacks. 647 */ 648 prev = NULL; 649 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { 650 int type = CSD_TYPE(csd); 651 652 if (type != CSD_TYPE_TTWU) { 653 if (prev) { 654 prev->next = &csd_next->node.llist; 655 } else { 656 entry = &csd_next->node.llist; 657 } 658 659 if (type == CSD_TYPE_ASYNC) { 660 smp_call_func_t func = csd->func; 661 void *info = csd->info; 662 663 csd_lock_record(csd); 664 csd_unlock(csd); 665 func(info); 666 csd_lock_record(NULL); 667 } else if (type == CSD_TYPE_IRQ_WORK) { 668 irq_work_single(csd); 669 } 670 671 } else { 672 prev = &csd->node.llist; 673 } 674 } 675 676 /* 677 * Third; only CSD_TYPE_TTWU is left, issue those. 678 */ 679 if (entry) 680 sched_ttwu_pending(entry); 681 682 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend, CFD_SEQ_NOCPU, 683 smp_processor_id(), CFD_SEQ_HDLEND); 684 } 685 686 687 /** 688 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks 689 * from task context (idle, migration thread) 690 * 691 * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it 692 * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by 693 * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to 694 * handle queued SMP function calls before scheduling. 695 * 696 * The migration thread has to ensure that an eventually pending wakeup has 697 * been handled before it migrates a task. 698 */ 699 void flush_smp_call_function_queue(void) 700 { 701 unsigned int was_pending; 702 unsigned long flags; 703 704 if (llist_empty(this_cpu_ptr(&call_single_queue))) 705 return; 706 707 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->idle, CFD_SEQ_NOCPU, 708 smp_processor_id(), CFD_SEQ_IDLE); 709 local_irq_save(flags); 710 /* Get the already pending soft interrupts for RT enabled kernels */ 711 was_pending = local_softirq_pending(); 712 __flush_smp_call_function_queue(true); 713 if (local_softirq_pending()) 714 do_softirq_post_smp_call_flush(was_pending); 715 716 local_irq_restore(flags); 717 } 718 719 /* 720 * smp_call_function_single - Run a function on a specific CPU 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 until function has completed on other CPUs. 724 * 725 * Returns 0 on success, else a negative status code. 726 */ 727 int smp_call_function_single(int cpu, smp_call_func_t func, void *info, 728 int wait) 729 { 730 call_single_data_t *csd; 731 call_single_data_t csd_stack = { 732 .node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, }, 733 }; 734 int this_cpu; 735 int err; 736 737 /* 738 * prevent preemption and reschedule on another processor, 739 * as well as CPU removal 740 */ 741 this_cpu = get_cpu(); 742 743 /* 744 * Can deadlock when called with interrupts disabled. 745 * We allow cpu's that are not yet online though, as no one else can 746 * send smp call function interrupt to this cpu and as such deadlocks 747 * can't happen. 748 */ 749 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() 750 && !oops_in_progress); 751 752 /* 753 * When @wait we can deadlock when we interrupt between llist_add() and 754 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to 755 * csd_lock() on because the interrupt context uses the same csd 756 * storage. 757 */ 758 WARN_ON_ONCE(!in_task()); 759 760 csd = &csd_stack; 761 if (!wait) { 762 csd = this_cpu_ptr(&csd_data); 763 csd_lock(csd); 764 } 765 766 csd->func = func; 767 csd->info = info; 768 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 769 csd->node.src = smp_processor_id(); 770 csd->node.dst = cpu; 771 #endif 772 773 err = generic_exec_single(cpu, csd); 774 775 if (wait) 776 csd_lock_wait(csd); 777 778 put_cpu(); 779 780 return err; 781 } 782 EXPORT_SYMBOL(smp_call_function_single); 783 784 /** 785 * smp_call_function_single_async() - Run an asynchronous function on a 786 * specific CPU. 787 * @cpu: The CPU to run on. 788 * @csd: Pre-allocated and setup data structure 789 * 790 * Like smp_call_function_single(), but the call is asynchonous and 791 * can thus be done from contexts with disabled interrupts. 792 * 793 * The caller passes his own pre-allocated data structure 794 * (ie: embedded in an object) and is responsible for synchronizing it 795 * such that the IPIs performed on the @csd are strictly serialized. 796 * 797 * If the function is called with one csd which has not yet been 798 * processed by previous call to smp_call_function_single_async(), the 799 * function will return immediately with -EBUSY showing that the csd 800 * object is still in progress. 801 * 802 * NOTE: Be careful, there is unfortunately no current debugging facility to 803 * validate the correctness of this serialization. 804 * 805 * Return: %0 on success or negative errno value on error 806 */ 807 int smp_call_function_single_async(int cpu, struct __call_single_data *csd) 808 { 809 int err = 0; 810 811 preempt_disable(); 812 813 if (csd->node.u_flags & CSD_FLAG_LOCK) { 814 err = -EBUSY; 815 goto out; 816 } 817 818 csd->node.u_flags = CSD_FLAG_LOCK; 819 smp_wmb(); 820 821 err = generic_exec_single(cpu, csd); 822 823 out: 824 preempt_enable(); 825 826 return err; 827 } 828 EXPORT_SYMBOL_GPL(smp_call_function_single_async); 829 830 /* 831 * smp_call_function_any - Run a function on any of the given cpus 832 * @mask: The mask of cpus it can run on. 833 * @func: The function to run. This must be fast and non-blocking. 834 * @info: An arbitrary pointer to pass to the function. 835 * @wait: If true, wait until function has completed. 836 * 837 * Returns 0 on success, else a negative status code (if no cpus were online). 838 * 839 * Selection preference: 840 * 1) current cpu if in @mask 841 * 2) any cpu of current node if in @mask 842 * 3) any other online cpu in @mask 843 */ 844 int smp_call_function_any(const struct cpumask *mask, 845 smp_call_func_t func, void *info, int wait) 846 { 847 unsigned int cpu; 848 const struct cpumask *nodemask; 849 int ret; 850 851 /* Try for same CPU (cheapest) */ 852 cpu = get_cpu(); 853 if (cpumask_test_cpu(cpu, mask)) 854 goto call; 855 856 /* Try for same node. */ 857 nodemask = cpumask_of_node(cpu_to_node(cpu)); 858 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids; 859 cpu = cpumask_next_and(cpu, nodemask, mask)) { 860 if (cpu_online(cpu)) 861 goto call; 862 } 863 864 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */ 865 cpu = cpumask_any_and(mask, cpu_online_mask); 866 call: 867 ret = smp_call_function_single(cpu, func, info, wait); 868 put_cpu(); 869 return ret; 870 } 871 EXPORT_SYMBOL_GPL(smp_call_function_any); 872 873 /* 874 * Flags to be used as scf_flags argument of smp_call_function_many_cond(). 875 * 876 * %SCF_WAIT: Wait until function execution is completed 877 * %SCF_RUN_LOCAL: Run also locally if local cpu is set in cpumask 878 */ 879 #define SCF_WAIT (1U << 0) 880 #define SCF_RUN_LOCAL (1U << 1) 881 882 static void smp_call_function_many_cond(const struct cpumask *mask, 883 smp_call_func_t func, void *info, 884 unsigned int scf_flags, 885 smp_cond_func_t cond_func) 886 { 887 int cpu, last_cpu, this_cpu = smp_processor_id(); 888 struct call_function_data *cfd; 889 bool wait = scf_flags & SCF_WAIT; 890 bool run_remote = false; 891 bool run_local = false; 892 int nr_cpus = 0; 893 894 lockdep_assert_preemption_disabled(); 895 896 /* 897 * Can deadlock when called with interrupts disabled. 898 * We allow cpu's that are not yet online though, as no one else can 899 * send smp call function interrupt to this cpu and as such deadlocks 900 * can't happen. 901 */ 902 if (cpu_online(this_cpu) && !oops_in_progress && 903 !early_boot_irqs_disabled) 904 lockdep_assert_irqs_enabled(); 905 906 /* 907 * When @wait we can deadlock when we interrupt between llist_add() and 908 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to 909 * csd_lock() on because the interrupt context uses the same csd 910 * storage. 911 */ 912 WARN_ON_ONCE(!in_task()); 913 914 /* Check if we need local execution. */ 915 if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask)) 916 run_local = true; 917 918 /* Check if we need remote execution, i.e., any CPU excluding this one. */ 919 cpu = cpumask_first_and(mask, cpu_online_mask); 920 if (cpu == this_cpu) 921 cpu = cpumask_next_and(cpu, mask, cpu_online_mask); 922 if (cpu < nr_cpu_ids) 923 run_remote = true; 924 925 if (run_remote) { 926 cfd = this_cpu_ptr(&cfd_data); 927 cpumask_and(cfd->cpumask, mask, cpu_online_mask); 928 __cpumask_clear_cpu(this_cpu, cfd->cpumask); 929 930 cpumask_clear(cfd->cpumask_ipi); 931 for_each_cpu(cpu, cfd->cpumask) { 932 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 933 call_single_data_t *csd = &pcpu->csd; 934 935 if (cond_func && !cond_func(cpu, info)) 936 continue; 937 938 csd_lock(csd); 939 if (wait) 940 csd->node.u_flags |= CSD_TYPE_SYNC; 941 csd->func = func; 942 csd->info = info; 943 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 944 csd->node.src = smp_processor_id(); 945 csd->node.dst = cpu; 946 #endif 947 cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE); 948 if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) { 949 __cpumask_set_cpu(cpu, cfd->cpumask_ipi); 950 nr_cpus++; 951 last_cpu = cpu; 952 953 cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI); 954 } else { 955 cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI); 956 } 957 } 958 959 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->ping, this_cpu, CFD_SEQ_NOCPU, CFD_SEQ_PING); 960 961 /* 962 * Choose the most efficient way to send an IPI. Note that the 963 * number of CPUs might be zero due to concurrent changes to the 964 * provided mask. 965 */ 966 if (nr_cpus == 1) 967 send_call_function_single_ipi(last_cpu); 968 else if (likely(nr_cpus > 1)) 969 arch_send_call_function_ipi_mask(cfd->cpumask_ipi); 970 971 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->pinged, this_cpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED); 972 } 973 974 if (run_local && (!cond_func || cond_func(this_cpu, info))) { 975 unsigned long flags; 976 977 local_irq_save(flags); 978 func(info); 979 local_irq_restore(flags); 980 } 981 982 if (run_remote && wait) { 983 for_each_cpu(cpu, cfd->cpumask) { 984 call_single_data_t *csd; 985 986 csd = &per_cpu_ptr(cfd->pcpu, cpu)->csd; 987 csd_lock_wait(csd); 988 } 989 } 990 } 991 992 /** 993 * smp_call_function_many(): Run a function on a set of CPUs. 994 * @mask: The set of cpus to run on (only runs on online subset). 995 * @func: The function to run. This must be fast and non-blocking. 996 * @info: An arbitrary pointer to pass to the function. 997 * @wait: Bitmask that controls the operation. If %SCF_WAIT is set, wait 998 * (atomically) until function has completed on other CPUs. If 999 * %SCF_RUN_LOCAL is set, the function will also be run locally 1000 * if the local CPU is set in the @cpumask. 1001 * 1002 * If @wait is true, then returns once @func has returned. 1003 * 1004 * You must not call this function with disabled interrupts or from a 1005 * hardware interrupt handler or from a bottom half handler. Preemption 1006 * must be disabled when calling this function. 1007 */ 1008 void smp_call_function_many(const struct cpumask *mask, 1009 smp_call_func_t func, void *info, bool wait) 1010 { 1011 smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL); 1012 } 1013 EXPORT_SYMBOL(smp_call_function_many); 1014 1015 /** 1016 * smp_call_function(): Run a function on all other CPUs. 1017 * @func: The function to run. This must be fast and non-blocking. 1018 * @info: An arbitrary pointer to pass to the function. 1019 * @wait: If true, wait (atomically) until function has completed 1020 * on other CPUs. 1021 * 1022 * Returns 0. 1023 * 1024 * If @wait is true, then returns once @func has returned; otherwise 1025 * it returns just before the target cpu calls @func. 1026 * 1027 * You must not call this function with disabled interrupts or from a 1028 * hardware interrupt handler or from a bottom half handler. 1029 */ 1030 void smp_call_function(smp_call_func_t func, void *info, int wait) 1031 { 1032 preempt_disable(); 1033 smp_call_function_many(cpu_online_mask, func, info, wait); 1034 preempt_enable(); 1035 } 1036 EXPORT_SYMBOL(smp_call_function); 1037 1038 /* Setup configured maximum number of CPUs to activate */ 1039 unsigned int setup_max_cpus = NR_CPUS; 1040 EXPORT_SYMBOL(setup_max_cpus); 1041 1042 1043 /* 1044 * Setup routine for controlling SMP activation 1045 * 1046 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP 1047 * activation entirely (the MPS table probe still happens, though). 1048 * 1049 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer 1050 * greater than 0, limits the maximum number of CPUs activated in 1051 * SMP mode to <NUM>. 1052 */ 1053 1054 void __weak arch_disable_smp_support(void) { } 1055 1056 static int __init nosmp(char *str) 1057 { 1058 setup_max_cpus = 0; 1059 arch_disable_smp_support(); 1060 1061 return 0; 1062 } 1063 1064 early_param("nosmp", nosmp); 1065 1066 /* this is hard limit */ 1067 static int __init nrcpus(char *str) 1068 { 1069 int nr_cpus; 1070 1071 if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids) 1072 set_nr_cpu_ids(nr_cpus); 1073 1074 return 0; 1075 } 1076 1077 early_param("nr_cpus", nrcpus); 1078 1079 static int __init maxcpus(char *str) 1080 { 1081 get_option(&str, &setup_max_cpus); 1082 if (setup_max_cpus == 0) 1083 arch_disable_smp_support(); 1084 1085 return 0; 1086 } 1087 1088 early_param("maxcpus", maxcpus); 1089 1090 #if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS) 1091 /* Setup number of possible processor ids */ 1092 unsigned int nr_cpu_ids __read_mostly = NR_CPUS; 1093 EXPORT_SYMBOL(nr_cpu_ids); 1094 #endif 1095 1096 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */ 1097 void __init setup_nr_cpu_ids(void) 1098 { 1099 set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask), NR_CPUS) + 1); 1100 } 1101 1102 /* Called by boot processor to activate the rest. */ 1103 void __init smp_init(void) 1104 { 1105 int num_nodes, num_cpus; 1106 1107 idle_threads_init(); 1108 cpuhp_threads_init(); 1109 1110 pr_info("Bringing up secondary CPUs ...\n"); 1111 1112 bringup_nonboot_cpus(setup_max_cpus); 1113 1114 num_nodes = num_online_nodes(); 1115 num_cpus = num_online_cpus(); 1116 pr_info("Brought up %d node%s, %d CPU%s\n", 1117 num_nodes, (num_nodes > 1 ? "s" : ""), 1118 num_cpus, (num_cpus > 1 ? "s" : "")); 1119 1120 /* Any cleanup work */ 1121 smp_cpus_done(setup_max_cpus); 1122 } 1123 1124 /* 1125 * on_each_cpu_cond(): Call a function on each processor for which 1126 * the supplied function cond_func returns true, optionally waiting 1127 * for all the required CPUs to finish. This may include the local 1128 * processor. 1129 * @cond_func: A callback function that is passed a cpu id and 1130 * the info parameter. The function is called 1131 * with preemption disabled. The function should 1132 * return a blooean value indicating whether to IPI 1133 * the specified CPU. 1134 * @func: The function to run on all applicable CPUs. 1135 * This must be fast and non-blocking. 1136 * @info: An arbitrary pointer to pass to both functions. 1137 * @wait: If true, wait (atomically) until function has 1138 * completed on other CPUs. 1139 * 1140 * Preemption is disabled to protect against CPUs going offline but not online. 1141 * CPUs going online during the call will not be seen or sent an IPI. 1142 * 1143 * You must not call this function with disabled interrupts or 1144 * from a hardware interrupt handler or from a bottom half handler. 1145 */ 1146 void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func, 1147 void *info, bool wait, const struct cpumask *mask) 1148 { 1149 unsigned int scf_flags = SCF_RUN_LOCAL; 1150 1151 if (wait) 1152 scf_flags |= SCF_WAIT; 1153 1154 preempt_disable(); 1155 smp_call_function_many_cond(mask, func, info, scf_flags, cond_func); 1156 preempt_enable(); 1157 } 1158 EXPORT_SYMBOL(on_each_cpu_cond_mask); 1159 1160 static void do_nothing(void *unused) 1161 { 1162 } 1163 1164 /** 1165 * kick_all_cpus_sync - Force all cpus out of idle 1166 * 1167 * Used to synchronize the update of pm_idle function pointer. It's 1168 * called after the pointer is updated and returns after the dummy 1169 * callback function has been executed on all cpus. The execution of 1170 * the function can only happen on the remote cpus after they have 1171 * left the idle function which had been called via pm_idle function 1172 * pointer. So it's guaranteed that nothing uses the previous pointer 1173 * anymore. 1174 */ 1175 void kick_all_cpus_sync(void) 1176 { 1177 /* Make sure the change is visible before we kick the cpus */ 1178 smp_mb(); 1179 smp_call_function(do_nothing, NULL, 1); 1180 } 1181 EXPORT_SYMBOL_GPL(kick_all_cpus_sync); 1182 1183 /** 1184 * wake_up_all_idle_cpus - break all cpus out of idle 1185 * wake_up_all_idle_cpus try to break all cpus which is in idle state even 1186 * including idle polling cpus, for non-idle cpus, we will do nothing 1187 * for them. 1188 */ 1189 void wake_up_all_idle_cpus(void) 1190 { 1191 int cpu; 1192 1193 for_each_possible_cpu(cpu) { 1194 preempt_disable(); 1195 if (cpu != smp_processor_id() && cpu_online(cpu)) 1196 wake_up_if_idle(cpu); 1197 preempt_enable(); 1198 } 1199 } 1200 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus); 1201 1202 /** 1203 * struct smp_call_on_cpu_struct - Call a function on a specific CPU 1204 * @work: &work_struct 1205 * @done: &completion to signal 1206 * @func: function to call 1207 * @data: function's data argument 1208 * @ret: return value from @func 1209 * @cpu: target CPU (%-1 for any CPU) 1210 * 1211 * Used to call a function on a specific cpu and wait for it to return. 1212 * Optionally make sure the call is done on a specified physical cpu via vcpu 1213 * pinning in order to support virtualized environments. 1214 */ 1215 struct smp_call_on_cpu_struct { 1216 struct work_struct work; 1217 struct completion done; 1218 int (*func)(void *); 1219 void *data; 1220 int ret; 1221 int cpu; 1222 }; 1223 1224 static void smp_call_on_cpu_callback(struct work_struct *work) 1225 { 1226 struct smp_call_on_cpu_struct *sscs; 1227 1228 sscs = container_of(work, struct smp_call_on_cpu_struct, work); 1229 if (sscs->cpu >= 0) 1230 hypervisor_pin_vcpu(sscs->cpu); 1231 sscs->ret = sscs->func(sscs->data); 1232 if (sscs->cpu >= 0) 1233 hypervisor_pin_vcpu(-1); 1234 1235 complete(&sscs->done); 1236 } 1237 1238 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys) 1239 { 1240 struct smp_call_on_cpu_struct sscs = { 1241 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done), 1242 .func = func, 1243 .data = par, 1244 .cpu = phys ? cpu : -1, 1245 }; 1246 1247 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback); 1248 1249 if (cpu >= nr_cpu_ids || !cpu_online(cpu)) 1250 return -ENXIO; 1251 1252 queue_work_on(cpu, system_wq, &sscs.work); 1253 wait_for_completion(&sscs.done); 1254 1255 return sscs.ret; 1256 } 1257 EXPORT_SYMBOL_GPL(smp_call_on_cpu); 1258