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 0; 178 } 179 early_param("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 #define CSD_LOCK_TIMEOUT (5ULL * NSEC_PER_SEC) 187 static atomic_t csd_bug_count = ATOMIC_INIT(0); 188 static u64 cfd_seq; 189 190 #define CFD_SEQ(s, d, t, c) \ 191 (union cfd_seq_cnt){ .u.src = s, .u.dst = d, .u.type = t, .u.cnt = c } 192 193 static u64 cfd_seq_inc(unsigned int src, unsigned int dst, unsigned int type) 194 { 195 union cfd_seq_cnt new, old; 196 197 new = CFD_SEQ(src, dst, type, 0); 198 199 do { 200 old.val = READ_ONCE(cfd_seq); 201 new.u.cnt = old.u.cnt + 1; 202 } while (cmpxchg(&cfd_seq, old.val, new.val) != old.val); 203 204 return old.val; 205 } 206 207 #define cfd_seq_store(var, src, dst, type) \ 208 do { \ 209 if (static_branch_unlikely(&csdlock_debug_extended)) \ 210 var = cfd_seq_inc(src, dst, type); \ 211 } while (0) 212 213 /* Record current CSD work for current CPU, NULL to erase. */ 214 static void __csd_lock_record(call_single_data_t *csd) 215 { 216 if (!csd) { 217 smp_mb(); /* NULL cur_csd after unlock. */ 218 __this_cpu_write(cur_csd, NULL); 219 return; 220 } 221 __this_cpu_write(cur_csd_func, csd->func); 222 __this_cpu_write(cur_csd_info, csd->info); 223 smp_wmb(); /* func and info before csd. */ 224 __this_cpu_write(cur_csd, csd); 225 smp_mb(); /* Update cur_csd before function call. */ 226 /* Or before unlock, as the case may be. */ 227 } 228 229 static __always_inline void csd_lock_record(call_single_data_t *csd) 230 { 231 if (static_branch_unlikely(&csdlock_debug_enabled)) 232 __csd_lock_record(csd); 233 } 234 235 static int csd_lock_wait_getcpu(call_single_data_t *csd) 236 { 237 unsigned int csd_type; 238 239 csd_type = CSD_TYPE(csd); 240 if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC) 241 return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */ 242 return -1; 243 } 244 245 static void cfd_seq_data_add(u64 val, unsigned int src, unsigned int dst, 246 unsigned int type, union cfd_seq_cnt *data, 247 unsigned int *n_data, unsigned int now) 248 { 249 union cfd_seq_cnt new[2]; 250 unsigned int i, j, k; 251 252 new[0].val = val; 253 new[1] = CFD_SEQ(src, dst, type, new[0].u.cnt + 1); 254 255 for (i = 0; i < 2; i++) { 256 if (new[i].u.cnt <= now) 257 new[i].u.cnt |= 0x80000000U; 258 for (j = 0; j < *n_data; j++) { 259 if (new[i].u.cnt == data[j].u.cnt) { 260 /* Direct read value trumps generated one. */ 261 if (i == 0) 262 data[j].val = new[i].val; 263 break; 264 } 265 if (new[i].u.cnt < data[j].u.cnt) { 266 for (k = *n_data; k > j; k--) 267 data[k].val = data[k - 1].val; 268 data[j].val = new[i].val; 269 (*n_data)++; 270 break; 271 } 272 } 273 if (j == *n_data) { 274 data[j].val = new[i].val; 275 (*n_data)++; 276 } 277 } 278 } 279 280 static const char *csd_lock_get_type(unsigned int type) 281 { 282 return (type >= ARRAY_SIZE(seq_type)) ? "?" : seq_type[type]; 283 } 284 285 static void csd_lock_print_extended(call_single_data_t *csd, int cpu) 286 { 287 struct cfd_seq_local *seq = &per_cpu(cfd_seq_local, cpu); 288 unsigned int srccpu = csd->node.src; 289 struct call_function_data *cfd = per_cpu_ptr(&cfd_data, srccpu); 290 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 291 unsigned int now; 292 union cfd_seq_cnt data[2 * ARRAY_SIZE(seq_type)]; 293 unsigned int n_data = 0, i; 294 295 data[0].val = READ_ONCE(cfd_seq); 296 now = data[0].u.cnt; 297 298 cfd_seq_data_add(pcpu->seq_queue, srccpu, cpu, CFD_SEQ_QUEUE, data, &n_data, now); 299 cfd_seq_data_add(pcpu->seq_ipi, srccpu, cpu, CFD_SEQ_IPI, data, &n_data, now); 300 cfd_seq_data_add(pcpu->seq_noipi, srccpu, cpu, CFD_SEQ_NOIPI, data, &n_data, now); 301 302 cfd_seq_data_add(per_cpu(cfd_seq_local.ping, srccpu), srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PING, data, &n_data, now); 303 cfd_seq_data_add(per_cpu(cfd_seq_local.pinged, srccpu), srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED, data, &n_data, now); 304 305 cfd_seq_data_add(seq->idle, CFD_SEQ_NOCPU, cpu, CFD_SEQ_IDLE, data, &n_data, now); 306 cfd_seq_data_add(seq->gotipi, CFD_SEQ_NOCPU, cpu, CFD_SEQ_GOTIPI, data, &n_data, now); 307 cfd_seq_data_add(seq->handle, CFD_SEQ_NOCPU, cpu, CFD_SEQ_HANDLE, data, &n_data, now); 308 cfd_seq_data_add(seq->dequeue, CFD_SEQ_NOCPU, cpu, CFD_SEQ_DEQUEUE, data, &n_data, now); 309 cfd_seq_data_add(seq->hdlend, CFD_SEQ_NOCPU, cpu, CFD_SEQ_HDLEND, data, &n_data, now); 310 311 for (i = 0; i < n_data; i++) { 312 pr_alert("\tcsd: cnt(%07x): %04x->%04x %s\n", 313 data[i].u.cnt & ~0x80000000U, data[i].u.src, 314 data[i].u.dst, csd_lock_get_type(data[i].u.type)); 315 } 316 pr_alert("\tcsd: cnt now: %07x\n", now); 317 } 318 319 /* 320 * Complain if too much time spent waiting. Note that only 321 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU, 322 * so waiting on other types gets much less information. 323 */ 324 static bool csd_lock_wait_toolong(call_single_data_t *csd, u64 ts0, u64 *ts1, int *bug_id) 325 { 326 int cpu = -1; 327 int cpux; 328 bool firsttime; 329 u64 ts2, ts_delta; 330 call_single_data_t *cpu_cur_csd; 331 unsigned int flags = READ_ONCE(csd->node.u_flags); 332 333 if (!(flags & CSD_FLAG_LOCK)) { 334 if (!unlikely(*bug_id)) 335 return true; 336 cpu = csd_lock_wait_getcpu(csd); 337 pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n", 338 *bug_id, raw_smp_processor_id(), cpu); 339 return true; 340 } 341 342 ts2 = sched_clock(); 343 ts_delta = ts2 - *ts1; 344 if (likely(ts_delta <= CSD_LOCK_TIMEOUT)) 345 return false; 346 347 firsttime = !*bug_id; 348 if (firsttime) 349 *bug_id = atomic_inc_return(&csd_bug_count); 350 cpu = csd_lock_wait_getcpu(csd); 351 if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu)) 352 cpux = 0; 353 else 354 cpux = cpu; 355 cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */ 356 pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n", 357 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0, 358 cpu, csd->func, csd->info); 359 if (cpu_cur_csd && csd != cpu_cur_csd) { 360 pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n", 361 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)), 362 READ_ONCE(per_cpu(cur_csd_info, cpux))); 363 } else { 364 pr_alert("\tcsd: CSD lock (#%d) %s.\n", 365 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request"); 366 } 367 if (cpu >= 0) { 368 if (static_branch_unlikely(&csdlock_debug_extended)) 369 csd_lock_print_extended(csd, cpu); 370 if (!trigger_single_cpu_backtrace(cpu)) 371 dump_cpu_task(cpu); 372 if (!cpu_cur_csd) { 373 pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu); 374 arch_send_call_function_single_ipi(cpu); 375 } 376 } 377 dump_stack(); 378 *ts1 = ts2; 379 380 return false; 381 } 382 383 /* 384 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources 385 * 386 * For non-synchronous ipi calls the csd can still be in use by the 387 * previous function call. For multi-cpu calls its even more interesting 388 * as we'll have to ensure no other cpu is observing our csd. 389 */ 390 static void __csd_lock_wait(call_single_data_t *csd) 391 { 392 int bug_id = 0; 393 u64 ts0, ts1; 394 395 ts1 = ts0 = sched_clock(); 396 for (;;) { 397 if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id)) 398 break; 399 cpu_relax(); 400 } 401 smp_acquire__after_ctrl_dep(); 402 } 403 404 static __always_inline void csd_lock_wait(call_single_data_t *csd) 405 { 406 if (static_branch_unlikely(&csdlock_debug_enabled)) { 407 __csd_lock_wait(csd); 408 return; 409 } 410 411 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK)); 412 } 413 414 static void __smp_call_single_queue_debug(int cpu, struct llist_node *node) 415 { 416 unsigned int this_cpu = smp_processor_id(); 417 struct cfd_seq_local *seq = this_cpu_ptr(&cfd_seq_local); 418 struct call_function_data *cfd = this_cpu_ptr(&cfd_data); 419 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 420 421 cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE); 422 if (llist_add(node, &per_cpu(call_single_queue, cpu))) { 423 cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI); 424 cfd_seq_store(seq->ping, this_cpu, cpu, CFD_SEQ_PING); 425 send_call_function_single_ipi(cpu); 426 cfd_seq_store(seq->pinged, this_cpu, cpu, CFD_SEQ_PINGED); 427 } else { 428 cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI); 429 } 430 } 431 #else 432 #define cfd_seq_store(var, src, dst, type) 433 434 static void csd_lock_record(call_single_data_t *csd) 435 { 436 } 437 438 static __always_inline void csd_lock_wait(call_single_data_t *csd) 439 { 440 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK)); 441 } 442 #endif 443 444 static __always_inline void csd_lock(call_single_data_t *csd) 445 { 446 csd_lock_wait(csd); 447 csd->node.u_flags |= CSD_FLAG_LOCK; 448 449 /* 450 * prevent CPU from reordering the above assignment 451 * to ->flags with any subsequent assignments to other 452 * fields of the specified call_single_data_t structure: 453 */ 454 smp_wmb(); 455 } 456 457 static __always_inline void csd_unlock(call_single_data_t *csd) 458 { 459 WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK)); 460 461 /* 462 * ensure we're all done before releasing data: 463 */ 464 smp_store_release(&csd->node.u_flags, 0); 465 } 466 467 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data); 468 469 void __smp_call_single_queue(int cpu, struct llist_node *node) 470 { 471 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 472 if (static_branch_unlikely(&csdlock_debug_extended)) { 473 unsigned int type; 474 475 type = CSD_TYPE(container_of(node, call_single_data_t, 476 node.llist)); 477 if (type == CSD_TYPE_SYNC || type == CSD_TYPE_ASYNC) { 478 __smp_call_single_queue_debug(cpu, node); 479 return; 480 } 481 } 482 #endif 483 484 /* 485 * The list addition should be visible before sending the IPI 486 * handler locks the list to pull the entry off it because of 487 * normal cache coherency rules implied by spinlocks. 488 * 489 * If IPIs can go out of order to the cache coherency protocol 490 * in an architecture, sufficient synchronisation should be added 491 * to arch code to make it appear to obey cache coherency WRT 492 * locking and barrier primitives. Generic code isn't really 493 * equipped to do the right thing... 494 */ 495 if (llist_add(node, &per_cpu(call_single_queue, cpu))) 496 send_call_function_single_ipi(cpu); 497 } 498 499 /* 500 * Insert a previously allocated call_single_data_t element 501 * for execution on the given CPU. data must already have 502 * ->func, ->info, and ->flags set. 503 */ 504 static int generic_exec_single(int cpu, call_single_data_t *csd) 505 { 506 if (cpu == smp_processor_id()) { 507 smp_call_func_t func = csd->func; 508 void *info = csd->info; 509 unsigned long flags; 510 511 /* 512 * We can unlock early even for the synchronous on-stack case, 513 * since we're doing this from the same CPU.. 514 */ 515 csd_lock_record(csd); 516 csd_unlock(csd); 517 local_irq_save(flags); 518 func(info); 519 csd_lock_record(NULL); 520 local_irq_restore(flags); 521 return 0; 522 } 523 524 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) { 525 csd_unlock(csd); 526 return -ENXIO; 527 } 528 529 __smp_call_single_queue(cpu, &csd->node.llist); 530 531 return 0; 532 } 533 534 /** 535 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks 536 * 537 * Invoked by arch to handle an IPI for call function single. 538 * Must be called with interrupts disabled. 539 */ 540 void generic_smp_call_function_single_interrupt(void) 541 { 542 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->gotipi, CFD_SEQ_NOCPU, 543 smp_processor_id(), CFD_SEQ_GOTIPI); 544 flush_smp_call_function_queue(true); 545 } 546 547 /** 548 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks 549 * 550 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an 551 * offline CPU. Skip this check if set to 'false'. 552 * 553 * Flush any pending smp-call-function callbacks queued on this CPU. This is 554 * invoked by the generic IPI handler, as well as by a CPU about to go offline, 555 * to ensure that all pending IPI callbacks are run before it goes completely 556 * offline. 557 * 558 * Loop through the call_single_queue and run all the queued callbacks. 559 * Must be called with interrupts disabled. 560 */ 561 static void flush_smp_call_function_queue(bool warn_cpu_offline) 562 { 563 call_single_data_t *csd, *csd_next; 564 struct llist_node *entry, *prev; 565 struct llist_head *head; 566 static bool warned; 567 568 lockdep_assert_irqs_disabled(); 569 570 head = this_cpu_ptr(&call_single_queue); 571 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->handle, CFD_SEQ_NOCPU, 572 smp_processor_id(), CFD_SEQ_HANDLE); 573 entry = llist_del_all(head); 574 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->dequeue, 575 /* Special meaning of source cpu: 0 == queue empty */ 576 entry ? CFD_SEQ_NOCPU : 0, 577 smp_processor_id(), CFD_SEQ_DEQUEUE); 578 entry = llist_reverse_order(entry); 579 580 /* There shouldn't be any pending callbacks on an offline CPU. */ 581 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) && 582 !warned && !llist_empty(head))) { 583 warned = true; 584 WARN(1, "IPI on offline CPU %d\n", smp_processor_id()); 585 586 /* 587 * We don't have to use the _safe() variant here 588 * because we are not invoking the IPI handlers yet. 589 */ 590 llist_for_each_entry(csd, entry, node.llist) { 591 switch (CSD_TYPE(csd)) { 592 case CSD_TYPE_ASYNC: 593 case CSD_TYPE_SYNC: 594 case CSD_TYPE_IRQ_WORK: 595 pr_warn("IPI callback %pS sent to offline CPU\n", 596 csd->func); 597 break; 598 599 case CSD_TYPE_TTWU: 600 pr_warn("IPI task-wakeup sent to offline CPU\n"); 601 break; 602 603 default: 604 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n", 605 CSD_TYPE(csd)); 606 break; 607 } 608 } 609 } 610 611 /* 612 * First; run all SYNC callbacks, people are waiting for us. 613 */ 614 prev = NULL; 615 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { 616 /* Do we wait until *after* callback? */ 617 if (CSD_TYPE(csd) == CSD_TYPE_SYNC) { 618 smp_call_func_t func = csd->func; 619 void *info = csd->info; 620 621 if (prev) { 622 prev->next = &csd_next->node.llist; 623 } else { 624 entry = &csd_next->node.llist; 625 } 626 627 csd_lock_record(csd); 628 func(info); 629 csd_unlock(csd); 630 csd_lock_record(NULL); 631 } else { 632 prev = &csd->node.llist; 633 } 634 } 635 636 if (!entry) { 637 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend, 638 0, smp_processor_id(), 639 CFD_SEQ_HDLEND); 640 return; 641 } 642 643 /* 644 * Second; run all !SYNC callbacks. 645 */ 646 prev = NULL; 647 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) { 648 int type = CSD_TYPE(csd); 649 650 if (type != CSD_TYPE_TTWU) { 651 if (prev) { 652 prev->next = &csd_next->node.llist; 653 } else { 654 entry = &csd_next->node.llist; 655 } 656 657 if (type == CSD_TYPE_ASYNC) { 658 smp_call_func_t func = csd->func; 659 void *info = csd->info; 660 661 csd_lock_record(csd); 662 csd_unlock(csd); 663 func(info); 664 csd_lock_record(NULL); 665 } else if (type == CSD_TYPE_IRQ_WORK) { 666 irq_work_single(csd); 667 } 668 669 } else { 670 prev = &csd->node.llist; 671 } 672 } 673 674 /* 675 * Third; only CSD_TYPE_TTWU is left, issue those. 676 */ 677 if (entry) 678 sched_ttwu_pending(entry); 679 680 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend, CFD_SEQ_NOCPU, 681 smp_processor_id(), CFD_SEQ_HDLEND); 682 } 683 684 void flush_smp_call_function_from_idle(void) 685 { 686 unsigned long flags; 687 688 if (llist_empty(this_cpu_ptr(&call_single_queue))) 689 return; 690 691 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->idle, CFD_SEQ_NOCPU, 692 smp_processor_id(), CFD_SEQ_IDLE); 693 local_irq_save(flags); 694 flush_smp_call_function_queue(true); 695 if (local_softirq_pending()) 696 do_softirq(); 697 698 local_irq_restore(flags); 699 } 700 701 /* 702 * smp_call_function_single - Run a function on a specific CPU 703 * @func: The function to run. This must be fast and non-blocking. 704 * @info: An arbitrary pointer to pass to the function. 705 * @wait: If true, wait until function has completed on other CPUs. 706 * 707 * Returns 0 on success, else a negative status code. 708 */ 709 int smp_call_function_single(int cpu, smp_call_func_t func, void *info, 710 int wait) 711 { 712 call_single_data_t *csd; 713 call_single_data_t csd_stack = { 714 .node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, }, 715 }; 716 int this_cpu; 717 int err; 718 719 /* 720 * prevent preemption and reschedule on another processor, 721 * as well as CPU removal 722 */ 723 this_cpu = get_cpu(); 724 725 /* 726 * Can deadlock when called with interrupts disabled. 727 * We allow cpu's that are not yet online though, as no one else can 728 * send smp call function interrupt to this cpu and as such deadlocks 729 * can't happen. 730 */ 731 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() 732 && !oops_in_progress); 733 734 /* 735 * When @wait we can deadlock when we interrupt between llist_add() and 736 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to 737 * csd_lock() on because the interrupt context uses the same csd 738 * storage. 739 */ 740 WARN_ON_ONCE(!in_task()); 741 742 csd = &csd_stack; 743 if (!wait) { 744 csd = this_cpu_ptr(&csd_data); 745 csd_lock(csd); 746 } 747 748 csd->func = func; 749 csd->info = info; 750 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 751 csd->node.src = smp_processor_id(); 752 csd->node.dst = cpu; 753 #endif 754 755 err = generic_exec_single(cpu, csd); 756 757 if (wait) 758 csd_lock_wait(csd); 759 760 put_cpu(); 761 762 return err; 763 } 764 EXPORT_SYMBOL(smp_call_function_single); 765 766 /** 767 * smp_call_function_single_async(): Run an asynchronous function on a 768 * specific CPU. 769 * @cpu: The CPU to run on. 770 * @csd: Pre-allocated and setup data structure 771 * 772 * Like smp_call_function_single(), but the call is asynchonous and 773 * can thus be done from contexts with disabled interrupts. 774 * 775 * The caller passes his own pre-allocated data structure 776 * (ie: embedded in an object) and is responsible for synchronizing it 777 * such that the IPIs performed on the @csd are strictly serialized. 778 * 779 * If the function is called with one csd which has not yet been 780 * processed by previous call to smp_call_function_single_async(), the 781 * function will return immediately with -EBUSY showing that the csd 782 * object is still in progress. 783 * 784 * NOTE: Be careful, there is unfortunately no current debugging facility to 785 * validate the correctness of this serialization. 786 */ 787 int smp_call_function_single_async(int cpu, call_single_data_t *csd) 788 { 789 int err = 0; 790 791 preempt_disable(); 792 793 if (csd->node.u_flags & CSD_FLAG_LOCK) { 794 err = -EBUSY; 795 goto out; 796 } 797 798 csd->node.u_flags = CSD_FLAG_LOCK; 799 smp_wmb(); 800 801 err = generic_exec_single(cpu, csd); 802 803 out: 804 preempt_enable(); 805 806 return err; 807 } 808 EXPORT_SYMBOL_GPL(smp_call_function_single_async); 809 810 /* 811 * smp_call_function_any - Run a function on any of the given cpus 812 * @mask: The mask of cpus it can run on. 813 * @func: The function to run. This must be fast and non-blocking. 814 * @info: An arbitrary pointer to pass to the function. 815 * @wait: If true, wait until function has completed. 816 * 817 * Returns 0 on success, else a negative status code (if no cpus were online). 818 * 819 * Selection preference: 820 * 1) current cpu if in @mask 821 * 2) any cpu of current node if in @mask 822 * 3) any other online cpu in @mask 823 */ 824 int smp_call_function_any(const struct cpumask *mask, 825 smp_call_func_t func, void *info, int wait) 826 { 827 unsigned int cpu; 828 const struct cpumask *nodemask; 829 int ret; 830 831 /* Try for same CPU (cheapest) */ 832 cpu = get_cpu(); 833 if (cpumask_test_cpu(cpu, mask)) 834 goto call; 835 836 /* Try for same node. */ 837 nodemask = cpumask_of_node(cpu_to_node(cpu)); 838 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids; 839 cpu = cpumask_next_and(cpu, nodemask, mask)) { 840 if (cpu_online(cpu)) 841 goto call; 842 } 843 844 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */ 845 cpu = cpumask_any_and(mask, cpu_online_mask); 846 call: 847 ret = smp_call_function_single(cpu, func, info, wait); 848 put_cpu(); 849 return ret; 850 } 851 EXPORT_SYMBOL_GPL(smp_call_function_any); 852 853 static void smp_call_function_many_cond(const struct cpumask *mask, 854 smp_call_func_t func, void *info, 855 bool wait, smp_cond_func_t cond_func) 856 { 857 struct call_function_data *cfd; 858 int cpu, next_cpu, this_cpu = smp_processor_id(); 859 860 /* 861 * Can deadlock when called with interrupts disabled. 862 * We allow cpu's that are not yet online though, as no one else can 863 * send smp call function interrupt to this cpu and as such deadlocks 864 * can't happen. 865 */ 866 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() 867 && !oops_in_progress && !early_boot_irqs_disabled); 868 869 /* 870 * When @wait we can deadlock when we interrupt between llist_add() and 871 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to 872 * csd_lock() on because the interrupt context uses the same csd 873 * storage. 874 */ 875 WARN_ON_ONCE(!in_task()); 876 877 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */ 878 cpu = cpumask_first_and(mask, cpu_online_mask); 879 if (cpu == this_cpu) 880 cpu = cpumask_next_and(cpu, mask, cpu_online_mask); 881 882 /* No online cpus? We're done. */ 883 if (cpu >= nr_cpu_ids) 884 return; 885 886 /* Do we have another CPU which isn't us? */ 887 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask); 888 if (next_cpu == this_cpu) 889 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask); 890 891 /* Fastpath: do that cpu by itself. */ 892 if (next_cpu >= nr_cpu_ids) { 893 if (!cond_func || cond_func(cpu, info)) 894 smp_call_function_single(cpu, func, info, wait); 895 return; 896 } 897 898 cfd = this_cpu_ptr(&cfd_data); 899 900 cpumask_and(cfd->cpumask, mask, cpu_online_mask); 901 __cpumask_clear_cpu(this_cpu, cfd->cpumask); 902 903 /* Some callers race with other cpus changing the passed mask */ 904 if (unlikely(!cpumask_weight(cfd->cpumask))) 905 return; 906 907 cpumask_clear(cfd->cpumask_ipi); 908 for_each_cpu(cpu, cfd->cpumask) { 909 struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu); 910 call_single_data_t *csd = &pcpu->csd; 911 912 if (cond_func && !cond_func(cpu, info)) 913 continue; 914 915 csd_lock(csd); 916 if (wait) 917 csd->node.u_flags |= CSD_TYPE_SYNC; 918 csd->func = func; 919 csd->info = info; 920 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG 921 csd->node.src = smp_processor_id(); 922 csd->node.dst = cpu; 923 #endif 924 cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE); 925 if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) { 926 __cpumask_set_cpu(cpu, cfd->cpumask_ipi); 927 cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI); 928 } else { 929 cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI); 930 } 931 } 932 933 /* Send a message to all CPUs in the map */ 934 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->ping, this_cpu, 935 CFD_SEQ_NOCPU, CFD_SEQ_PING); 936 arch_send_call_function_ipi_mask(cfd->cpumask_ipi); 937 cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->pinged, this_cpu, 938 CFD_SEQ_NOCPU, CFD_SEQ_PINGED); 939 940 if (wait) { 941 for_each_cpu(cpu, cfd->cpumask) { 942 call_single_data_t *csd; 943 944 csd = &per_cpu_ptr(cfd->pcpu, cpu)->csd; 945 csd_lock_wait(csd); 946 } 947 } 948 } 949 950 /** 951 * smp_call_function_many(): Run a function on a set of other CPUs. 952 * @mask: The set of cpus to run on (only runs on online subset). 953 * @func: The function to run. This must be fast and non-blocking. 954 * @info: An arbitrary pointer to pass to the function. 955 * @wait: If true, wait (atomically) until function has completed 956 * on other CPUs. 957 * 958 * If @wait is true, then returns once @func has returned. 959 * 960 * You must not call this function with disabled interrupts or from a 961 * hardware interrupt handler or from a bottom half handler. Preemption 962 * must be disabled when calling this function. 963 */ 964 void smp_call_function_many(const struct cpumask *mask, 965 smp_call_func_t func, void *info, bool wait) 966 { 967 smp_call_function_many_cond(mask, func, info, wait, NULL); 968 } 969 EXPORT_SYMBOL(smp_call_function_many); 970 971 /** 972 * smp_call_function(): Run a function on all other CPUs. 973 * @func: The function to run. This must be fast and non-blocking. 974 * @info: An arbitrary pointer to pass to the function. 975 * @wait: If true, wait (atomically) until function has completed 976 * on other CPUs. 977 * 978 * Returns 0. 979 * 980 * If @wait is true, then returns once @func has returned; otherwise 981 * it returns just before the target cpu calls @func. 982 * 983 * You must not call this function with disabled interrupts or from a 984 * hardware interrupt handler or from a bottom half handler. 985 */ 986 void smp_call_function(smp_call_func_t func, void *info, int wait) 987 { 988 preempt_disable(); 989 smp_call_function_many(cpu_online_mask, func, info, wait); 990 preempt_enable(); 991 } 992 EXPORT_SYMBOL(smp_call_function); 993 994 /* Setup configured maximum number of CPUs to activate */ 995 unsigned int setup_max_cpus = NR_CPUS; 996 EXPORT_SYMBOL(setup_max_cpus); 997 998 999 /* 1000 * Setup routine for controlling SMP activation 1001 * 1002 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP 1003 * activation entirely (the MPS table probe still happens, though). 1004 * 1005 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer 1006 * greater than 0, limits the maximum number of CPUs activated in 1007 * SMP mode to <NUM>. 1008 */ 1009 1010 void __weak arch_disable_smp_support(void) { } 1011 1012 static int __init nosmp(char *str) 1013 { 1014 setup_max_cpus = 0; 1015 arch_disable_smp_support(); 1016 1017 return 0; 1018 } 1019 1020 early_param("nosmp", nosmp); 1021 1022 /* this is hard limit */ 1023 static int __init nrcpus(char *str) 1024 { 1025 int nr_cpus; 1026 1027 if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids) 1028 nr_cpu_ids = nr_cpus; 1029 1030 return 0; 1031 } 1032 1033 early_param("nr_cpus", nrcpus); 1034 1035 static int __init maxcpus(char *str) 1036 { 1037 get_option(&str, &setup_max_cpus); 1038 if (setup_max_cpus == 0) 1039 arch_disable_smp_support(); 1040 1041 return 0; 1042 } 1043 1044 early_param("maxcpus", maxcpus); 1045 1046 /* Setup number of possible processor ids */ 1047 unsigned int nr_cpu_ids __read_mostly = NR_CPUS; 1048 EXPORT_SYMBOL(nr_cpu_ids); 1049 1050 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */ 1051 void __init setup_nr_cpu_ids(void) 1052 { 1053 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1; 1054 } 1055 1056 /* Called by boot processor to activate the rest. */ 1057 void __init smp_init(void) 1058 { 1059 int num_nodes, num_cpus; 1060 1061 idle_threads_init(); 1062 cpuhp_threads_init(); 1063 1064 pr_info("Bringing up secondary CPUs ...\n"); 1065 1066 bringup_nonboot_cpus(setup_max_cpus); 1067 1068 num_nodes = num_online_nodes(); 1069 num_cpus = num_online_cpus(); 1070 pr_info("Brought up %d node%s, %d CPU%s\n", 1071 num_nodes, (num_nodes > 1 ? "s" : ""), 1072 num_cpus, (num_cpus > 1 ? "s" : "")); 1073 1074 /* Any cleanup work */ 1075 smp_cpus_done(setup_max_cpus); 1076 } 1077 1078 /* 1079 * Call a function on all processors. May be used during early boot while 1080 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead 1081 * of local_irq_disable/enable(). 1082 */ 1083 void on_each_cpu(smp_call_func_t func, void *info, int wait) 1084 { 1085 unsigned long flags; 1086 1087 preempt_disable(); 1088 smp_call_function(func, info, wait); 1089 local_irq_save(flags); 1090 func(info); 1091 local_irq_restore(flags); 1092 preempt_enable(); 1093 } 1094 EXPORT_SYMBOL(on_each_cpu); 1095 1096 /** 1097 * on_each_cpu_mask(): Run a function on processors specified by 1098 * cpumask, which may include the local processor. 1099 * @mask: The set of cpus to run on (only runs on online subset). 1100 * @func: The function to run. This must be fast and non-blocking. 1101 * @info: An arbitrary pointer to pass to the function. 1102 * @wait: If true, wait (atomically) until function has completed 1103 * on other CPUs. 1104 * 1105 * If @wait is true, then returns once @func has returned. 1106 * 1107 * You must not call this function with disabled interrupts or from a 1108 * hardware interrupt handler or from a bottom half handler. The 1109 * exception is that it may be used during early boot while 1110 * early_boot_irqs_disabled is set. 1111 */ 1112 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func, 1113 void *info, bool wait) 1114 { 1115 int cpu = get_cpu(); 1116 1117 smp_call_function_many(mask, func, info, wait); 1118 if (cpumask_test_cpu(cpu, mask)) { 1119 unsigned long flags; 1120 local_irq_save(flags); 1121 func(info); 1122 local_irq_restore(flags); 1123 } 1124 put_cpu(); 1125 } 1126 EXPORT_SYMBOL(on_each_cpu_mask); 1127 1128 /* 1129 * on_each_cpu_cond(): Call a function on each processor for which 1130 * the supplied function cond_func returns true, optionally waiting 1131 * for all the required CPUs to finish. This may include the local 1132 * processor. 1133 * @cond_func: A callback function that is passed a cpu id and 1134 * the info parameter. The function is called 1135 * with preemption disabled. The function should 1136 * return a blooean value indicating whether to IPI 1137 * the specified CPU. 1138 * @func: The function to run on all applicable CPUs. 1139 * This must be fast and non-blocking. 1140 * @info: An arbitrary pointer to pass to both functions. 1141 * @wait: If true, wait (atomically) until function has 1142 * completed on other CPUs. 1143 * 1144 * Preemption is disabled to protect against CPUs going offline but not online. 1145 * CPUs going online during the call will not be seen or sent an IPI. 1146 * 1147 * You must not call this function with disabled interrupts or 1148 * from a hardware interrupt handler or from a bottom half handler. 1149 */ 1150 void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func, 1151 void *info, bool wait, const struct cpumask *mask) 1152 { 1153 int cpu = get_cpu(); 1154 1155 smp_call_function_many_cond(mask, func, info, wait, cond_func); 1156 if (cpumask_test_cpu(cpu, mask) && cond_func(cpu, info)) { 1157 unsigned long flags; 1158 1159 local_irq_save(flags); 1160 func(info); 1161 local_irq_restore(flags); 1162 } 1163 put_cpu(); 1164 } 1165 EXPORT_SYMBOL(on_each_cpu_cond_mask); 1166 1167 void on_each_cpu_cond(smp_cond_func_t cond_func, smp_call_func_t func, 1168 void *info, bool wait) 1169 { 1170 on_each_cpu_cond_mask(cond_func, func, info, wait, cpu_online_mask); 1171 } 1172 EXPORT_SYMBOL(on_each_cpu_cond); 1173 1174 static void do_nothing(void *unused) 1175 { 1176 } 1177 1178 /** 1179 * kick_all_cpus_sync - Force all cpus out of idle 1180 * 1181 * Used to synchronize the update of pm_idle function pointer. It's 1182 * called after the pointer is updated and returns after the dummy 1183 * callback function has been executed on all cpus. The execution of 1184 * the function can only happen on the remote cpus after they have 1185 * left the idle function which had been called via pm_idle function 1186 * pointer. So it's guaranteed that nothing uses the previous pointer 1187 * anymore. 1188 */ 1189 void kick_all_cpus_sync(void) 1190 { 1191 /* Make sure the change is visible before we kick the cpus */ 1192 smp_mb(); 1193 smp_call_function(do_nothing, NULL, 1); 1194 } 1195 EXPORT_SYMBOL_GPL(kick_all_cpus_sync); 1196 1197 /** 1198 * wake_up_all_idle_cpus - break all cpus out of idle 1199 * wake_up_all_idle_cpus try to break all cpus which is in idle state even 1200 * including idle polling cpus, for non-idle cpus, we will do nothing 1201 * for them. 1202 */ 1203 void wake_up_all_idle_cpus(void) 1204 { 1205 int cpu; 1206 1207 preempt_disable(); 1208 for_each_online_cpu(cpu) { 1209 if (cpu == smp_processor_id()) 1210 continue; 1211 1212 wake_up_if_idle(cpu); 1213 } 1214 preempt_enable(); 1215 } 1216 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus); 1217 1218 /** 1219 * smp_call_on_cpu - Call a function on a specific cpu 1220 * 1221 * Used to call a function on a specific cpu and wait for it to return. 1222 * Optionally make sure the call is done on a specified physical cpu via vcpu 1223 * pinning in order to support virtualized environments. 1224 */ 1225 struct smp_call_on_cpu_struct { 1226 struct work_struct work; 1227 struct completion done; 1228 int (*func)(void *); 1229 void *data; 1230 int ret; 1231 int cpu; 1232 }; 1233 1234 static void smp_call_on_cpu_callback(struct work_struct *work) 1235 { 1236 struct smp_call_on_cpu_struct *sscs; 1237 1238 sscs = container_of(work, struct smp_call_on_cpu_struct, work); 1239 if (sscs->cpu >= 0) 1240 hypervisor_pin_vcpu(sscs->cpu); 1241 sscs->ret = sscs->func(sscs->data); 1242 if (sscs->cpu >= 0) 1243 hypervisor_pin_vcpu(-1); 1244 1245 complete(&sscs->done); 1246 } 1247 1248 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys) 1249 { 1250 struct smp_call_on_cpu_struct sscs = { 1251 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done), 1252 .func = func, 1253 .data = par, 1254 .cpu = phys ? cpu : -1, 1255 }; 1256 1257 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback); 1258 1259 if (cpu >= nr_cpu_ids || !cpu_online(cpu)) 1260 return -ENXIO; 1261 1262 queue_work_on(cpu, system_wq, &sscs.work); 1263 wait_for_completion(&sscs.done); 1264 1265 return sscs.ret; 1266 } 1267 EXPORT_SYMBOL_GPL(smp_call_on_cpu); 1268