1 /* 2 * Generic helpers for smp ipi calls 3 * 4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008 5 */ 6 #include <linux/rcupdate.h> 7 #include <linux/rculist.h> 8 #include <linux/kernel.h> 9 #include <linux/export.h> 10 #include <linux/percpu.h> 11 #include <linux/init.h> 12 #include <linux/gfp.h> 13 #include <linux/smp.h> 14 #include <linux/cpu.h> 15 16 #include "smpboot.h" 17 18 enum { 19 CSD_FLAG_LOCK = 0x01, 20 CSD_FLAG_WAIT = 0x02, 21 }; 22 23 struct call_function_data { 24 struct call_single_data __percpu *csd; 25 cpumask_var_t cpumask; 26 }; 27 28 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data); 29 30 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue); 31 32 static int 33 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu) 34 { 35 long cpu = (long)hcpu; 36 struct call_function_data *cfd = &per_cpu(cfd_data, cpu); 37 38 switch (action) { 39 case CPU_UP_PREPARE: 40 case CPU_UP_PREPARE_FROZEN: 41 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL, 42 cpu_to_node(cpu))) 43 return notifier_from_errno(-ENOMEM); 44 cfd->csd = alloc_percpu(struct call_single_data); 45 if (!cfd->csd) { 46 free_cpumask_var(cfd->cpumask); 47 return notifier_from_errno(-ENOMEM); 48 } 49 break; 50 51 #ifdef CONFIG_HOTPLUG_CPU 52 case CPU_UP_CANCELED: 53 case CPU_UP_CANCELED_FROZEN: 54 55 case CPU_DEAD: 56 case CPU_DEAD_FROZEN: 57 free_cpumask_var(cfd->cpumask); 58 free_percpu(cfd->csd); 59 break; 60 #endif 61 }; 62 63 return NOTIFY_OK; 64 } 65 66 static struct notifier_block hotplug_cfd_notifier = { 67 .notifier_call = hotplug_cfd, 68 }; 69 70 void __init call_function_init(void) 71 { 72 void *cpu = (void *)(long)smp_processor_id(); 73 int i; 74 75 for_each_possible_cpu(i) 76 init_llist_head(&per_cpu(call_single_queue, i)); 77 78 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu); 79 register_cpu_notifier(&hotplug_cfd_notifier); 80 } 81 82 /* 83 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources 84 * 85 * For non-synchronous ipi calls the csd can still be in use by the 86 * previous function call. For multi-cpu calls its even more interesting 87 * as we'll have to ensure no other cpu is observing our csd. 88 */ 89 static void csd_lock_wait(struct call_single_data *csd) 90 { 91 while (csd->flags & CSD_FLAG_LOCK) 92 cpu_relax(); 93 } 94 95 static void csd_lock(struct call_single_data *csd) 96 { 97 csd_lock_wait(csd); 98 csd->flags |= CSD_FLAG_LOCK; 99 100 /* 101 * prevent CPU from reordering the above assignment 102 * to ->flags with any subsequent assignments to other 103 * fields of the specified call_single_data structure: 104 */ 105 smp_mb(); 106 } 107 108 static void csd_unlock(struct call_single_data *csd) 109 { 110 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK)); 111 112 /* 113 * ensure we're all done before releasing data: 114 */ 115 smp_mb(); 116 117 csd->flags &= ~CSD_FLAG_LOCK; 118 } 119 120 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data); 121 122 /* 123 * Insert a previously allocated call_single_data element 124 * for execution on the given CPU. data must already have 125 * ->func, ->info, and ->flags set. 126 */ 127 static int generic_exec_single(int cpu, struct call_single_data *csd, 128 smp_call_func_t func, void *info, int wait) 129 { 130 struct call_single_data csd_stack = { .flags = 0 }; 131 unsigned long flags; 132 133 134 if (cpu == smp_processor_id()) { 135 local_irq_save(flags); 136 func(info); 137 local_irq_restore(flags); 138 return 0; 139 } 140 141 142 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) 143 return -ENXIO; 144 145 146 if (!csd) { 147 csd = &csd_stack; 148 if (!wait) 149 csd = &__get_cpu_var(csd_data); 150 } 151 152 csd_lock(csd); 153 154 csd->func = func; 155 csd->info = info; 156 157 if (wait) 158 csd->flags |= CSD_FLAG_WAIT; 159 160 /* 161 * The list addition should be visible before sending the IPI 162 * handler locks the list to pull the entry off it because of 163 * normal cache coherency rules implied by spinlocks. 164 * 165 * If IPIs can go out of order to the cache coherency protocol 166 * in an architecture, sufficient synchronisation should be added 167 * to arch code to make it appear to obey cache coherency WRT 168 * locking and barrier primitives. Generic code isn't really 169 * equipped to do the right thing... 170 */ 171 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu))) 172 arch_send_call_function_single_ipi(cpu); 173 174 if (wait) 175 csd_lock_wait(csd); 176 177 return 0; 178 } 179 180 /* 181 * Invoked by arch to handle an IPI for call function single. Must be 182 * called from the arch with interrupts disabled. 183 */ 184 void generic_smp_call_function_single_interrupt(void) 185 { 186 struct llist_node *entry; 187 struct call_single_data *csd, *csd_next; 188 189 /* 190 * Shouldn't receive this interrupt on a cpu that is not yet online. 191 */ 192 WARN_ON_ONCE(!cpu_online(smp_processor_id())); 193 194 entry = llist_del_all(&__get_cpu_var(call_single_queue)); 195 entry = llist_reverse_order(entry); 196 197 llist_for_each_entry_safe(csd, csd_next, entry, llist) { 198 csd->func(csd->info); 199 csd_unlock(csd); 200 } 201 } 202 203 /* 204 * smp_call_function_single - Run a function on a specific CPU 205 * @func: The function to run. This must be fast and non-blocking. 206 * @info: An arbitrary pointer to pass to the function. 207 * @wait: If true, wait until function has completed on other CPUs. 208 * 209 * Returns 0 on success, else a negative status code. 210 */ 211 int smp_call_function_single(int cpu, smp_call_func_t func, void *info, 212 int wait) 213 { 214 int this_cpu; 215 int err; 216 217 /* 218 * prevent preemption and reschedule on another processor, 219 * as well as CPU removal 220 */ 221 this_cpu = get_cpu(); 222 223 /* 224 * Can deadlock when called with interrupts disabled. 225 * We allow cpu's that are not yet online though, as no one else can 226 * send smp call function interrupt to this cpu and as such deadlocks 227 * can't happen. 228 */ 229 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() 230 && !oops_in_progress); 231 232 err = generic_exec_single(cpu, NULL, func, info, wait); 233 234 put_cpu(); 235 236 return err; 237 } 238 EXPORT_SYMBOL(smp_call_function_single); 239 240 /** 241 * smp_call_function_single_async(): Run an asynchronous function on a 242 * specific CPU. 243 * @cpu: The CPU to run on. 244 * @csd: Pre-allocated and setup data structure 245 * 246 * Like smp_call_function_single(), but the call is asynchonous and 247 * can thus be done from contexts with disabled interrupts. 248 * 249 * The caller passes his own pre-allocated data structure 250 * (ie: embedded in an object) and is responsible for synchronizing it 251 * such that the IPIs performed on the @csd are strictly serialized. 252 * 253 * NOTE: Be careful, there is unfortunately no current debugging facility to 254 * validate the correctness of this serialization. 255 */ 256 int smp_call_function_single_async(int cpu, struct call_single_data *csd) 257 { 258 int err = 0; 259 260 preempt_disable(); 261 err = generic_exec_single(cpu, csd, csd->func, csd->info, 0); 262 preempt_enable(); 263 264 return err; 265 } 266 EXPORT_SYMBOL_GPL(smp_call_function_single_async); 267 268 /* 269 * smp_call_function_any - Run a function on any of the given cpus 270 * @mask: The mask of cpus it can run on. 271 * @func: The function to run. This must be fast and non-blocking. 272 * @info: An arbitrary pointer to pass to the function. 273 * @wait: If true, wait until function has completed. 274 * 275 * Returns 0 on success, else a negative status code (if no cpus were online). 276 * 277 * Selection preference: 278 * 1) current cpu if in @mask 279 * 2) any cpu of current node if in @mask 280 * 3) any other online cpu in @mask 281 */ 282 int smp_call_function_any(const struct cpumask *mask, 283 smp_call_func_t func, void *info, int wait) 284 { 285 unsigned int cpu; 286 const struct cpumask *nodemask; 287 int ret; 288 289 /* Try for same CPU (cheapest) */ 290 cpu = get_cpu(); 291 if (cpumask_test_cpu(cpu, mask)) 292 goto call; 293 294 /* Try for same node. */ 295 nodemask = cpumask_of_node(cpu_to_node(cpu)); 296 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids; 297 cpu = cpumask_next_and(cpu, nodemask, mask)) { 298 if (cpu_online(cpu)) 299 goto call; 300 } 301 302 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */ 303 cpu = cpumask_any_and(mask, cpu_online_mask); 304 call: 305 ret = smp_call_function_single(cpu, func, info, wait); 306 put_cpu(); 307 return ret; 308 } 309 EXPORT_SYMBOL_GPL(smp_call_function_any); 310 311 /** 312 * smp_call_function_many(): Run a function on a set of other CPUs. 313 * @mask: The set of cpus to run on (only runs on online subset). 314 * @func: The function to run. This must be fast and non-blocking. 315 * @info: An arbitrary pointer to pass to the function. 316 * @wait: If true, wait (atomically) until function has completed 317 * on other CPUs. 318 * 319 * If @wait is true, then returns once @func has returned. 320 * 321 * You must not call this function with disabled interrupts or from a 322 * hardware interrupt handler or from a bottom half handler. Preemption 323 * must be disabled when calling this function. 324 */ 325 void smp_call_function_many(const struct cpumask *mask, 326 smp_call_func_t func, void *info, bool wait) 327 { 328 struct call_function_data *cfd; 329 int cpu, next_cpu, this_cpu = smp_processor_id(); 330 331 /* 332 * Can deadlock when called with interrupts disabled. 333 * We allow cpu's that are not yet online though, as no one else can 334 * send smp call function interrupt to this cpu and as such deadlocks 335 * can't happen. 336 */ 337 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() 338 && !oops_in_progress && !early_boot_irqs_disabled); 339 340 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */ 341 cpu = cpumask_first_and(mask, cpu_online_mask); 342 if (cpu == this_cpu) 343 cpu = cpumask_next_and(cpu, mask, cpu_online_mask); 344 345 /* No online cpus? We're done. */ 346 if (cpu >= nr_cpu_ids) 347 return; 348 349 /* Do we have another CPU which isn't us? */ 350 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask); 351 if (next_cpu == this_cpu) 352 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask); 353 354 /* Fastpath: do that cpu by itself. */ 355 if (next_cpu >= nr_cpu_ids) { 356 smp_call_function_single(cpu, func, info, wait); 357 return; 358 } 359 360 cfd = &__get_cpu_var(cfd_data); 361 362 cpumask_and(cfd->cpumask, mask, cpu_online_mask); 363 cpumask_clear_cpu(this_cpu, cfd->cpumask); 364 365 /* Some callers race with other cpus changing the passed mask */ 366 if (unlikely(!cpumask_weight(cfd->cpumask))) 367 return; 368 369 for_each_cpu(cpu, cfd->cpumask) { 370 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu); 371 372 csd_lock(csd); 373 csd->func = func; 374 csd->info = info; 375 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)); 376 } 377 378 /* Send a message to all CPUs in the map */ 379 arch_send_call_function_ipi_mask(cfd->cpumask); 380 381 if (wait) { 382 for_each_cpu(cpu, cfd->cpumask) { 383 struct call_single_data *csd; 384 385 csd = per_cpu_ptr(cfd->csd, cpu); 386 csd_lock_wait(csd); 387 } 388 } 389 } 390 EXPORT_SYMBOL(smp_call_function_many); 391 392 /** 393 * smp_call_function(): Run a function on all other CPUs. 394 * @func: The function to run. This must be fast and non-blocking. 395 * @info: An arbitrary pointer to pass to the function. 396 * @wait: If true, wait (atomically) until function has completed 397 * on other CPUs. 398 * 399 * Returns 0. 400 * 401 * If @wait is true, then returns once @func has returned; otherwise 402 * it returns just before the target cpu calls @func. 403 * 404 * You must not call this function with disabled interrupts or from a 405 * hardware interrupt handler or from a bottom half handler. 406 */ 407 int smp_call_function(smp_call_func_t func, void *info, int wait) 408 { 409 preempt_disable(); 410 smp_call_function_many(cpu_online_mask, func, info, wait); 411 preempt_enable(); 412 413 return 0; 414 } 415 EXPORT_SYMBOL(smp_call_function); 416 417 /* Setup configured maximum number of CPUs to activate */ 418 unsigned int setup_max_cpus = NR_CPUS; 419 EXPORT_SYMBOL(setup_max_cpus); 420 421 422 /* 423 * Setup routine for controlling SMP activation 424 * 425 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP 426 * activation entirely (the MPS table probe still happens, though). 427 * 428 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer 429 * greater than 0, limits the maximum number of CPUs activated in 430 * SMP mode to <NUM>. 431 */ 432 433 void __weak arch_disable_smp_support(void) { } 434 435 static int __init nosmp(char *str) 436 { 437 setup_max_cpus = 0; 438 arch_disable_smp_support(); 439 440 return 0; 441 } 442 443 early_param("nosmp", nosmp); 444 445 /* this is hard limit */ 446 static int __init nrcpus(char *str) 447 { 448 int nr_cpus; 449 450 get_option(&str, &nr_cpus); 451 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids) 452 nr_cpu_ids = nr_cpus; 453 454 return 0; 455 } 456 457 early_param("nr_cpus", nrcpus); 458 459 static int __init maxcpus(char *str) 460 { 461 get_option(&str, &setup_max_cpus); 462 if (setup_max_cpus == 0) 463 arch_disable_smp_support(); 464 465 return 0; 466 } 467 468 early_param("maxcpus", maxcpus); 469 470 /* Setup number of possible processor ids */ 471 int nr_cpu_ids __read_mostly = NR_CPUS; 472 EXPORT_SYMBOL(nr_cpu_ids); 473 474 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */ 475 void __init setup_nr_cpu_ids(void) 476 { 477 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1; 478 } 479 480 void __weak smp_announce(void) 481 { 482 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus()); 483 } 484 485 /* Called by boot processor to activate the rest. */ 486 void __init smp_init(void) 487 { 488 unsigned int cpu; 489 490 idle_threads_init(); 491 492 /* FIXME: This should be done in userspace --RR */ 493 for_each_present_cpu(cpu) { 494 if (num_online_cpus() >= setup_max_cpus) 495 break; 496 if (!cpu_online(cpu)) 497 cpu_up(cpu); 498 } 499 500 /* Any cleanup work */ 501 smp_announce(); 502 smp_cpus_done(setup_max_cpus); 503 } 504 505 /* 506 * Call a function on all processors. May be used during early boot while 507 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead 508 * of local_irq_disable/enable(). 509 */ 510 int on_each_cpu(void (*func) (void *info), void *info, int wait) 511 { 512 unsigned long flags; 513 int ret = 0; 514 515 preempt_disable(); 516 ret = smp_call_function(func, info, wait); 517 local_irq_save(flags); 518 func(info); 519 local_irq_restore(flags); 520 preempt_enable(); 521 return ret; 522 } 523 EXPORT_SYMBOL(on_each_cpu); 524 525 /** 526 * on_each_cpu_mask(): Run a function on processors specified by 527 * cpumask, which may include the local processor. 528 * @mask: The set of cpus to run on (only runs on online subset). 529 * @func: The function to run. This must be fast and non-blocking. 530 * @info: An arbitrary pointer to pass to the function. 531 * @wait: If true, wait (atomically) until function has completed 532 * on other CPUs. 533 * 534 * If @wait is true, then returns once @func has returned. 535 * 536 * You must not call this function with disabled interrupts or from a 537 * hardware interrupt handler or from a bottom half handler. The 538 * exception is that it may be used during early boot while 539 * early_boot_irqs_disabled is set. 540 */ 541 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func, 542 void *info, bool wait) 543 { 544 int cpu = get_cpu(); 545 546 smp_call_function_many(mask, func, info, wait); 547 if (cpumask_test_cpu(cpu, mask)) { 548 unsigned long flags; 549 local_irq_save(flags); 550 func(info); 551 local_irq_restore(flags); 552 } 553 put_cpu(); 554 } 555 EXPORT_SYMBOL(on_each_cpu_mask); 556 557 /* 558 * on_each_cpu_cond(): Call a function on each processor for which 559 * the supplied function cond_func returns true, optionally waiting 560 * for all the required CPUs to finish. This may include the local 561 * processor. 562 * @cond_func: A callback function that is passed a cpu id and 563 * the the info parameter. The function is called 564 * with preemption disabled. The function should 565 * return a blooean value indicating whether to IPI 566 * the specified CPU. 567 * @func: The function to run on all applicable CPUs. 568 * This must be fast and non-blocking. 569 * @info: An arbitrary pointer to pass to both functions. 570 * @wait: If true, wait (atomically) until function has 571 * completed on other CPUs. 572 * @gfp_flags: GFP flags to use when allocating the cpumask 573 * used internally by the function. 574 * 575 * The function might sleep if the GFP flags indicates a non 576 * atomic allocation is allowed. 577 * 578 * Preemption is disabled to protect against CPUs going offline but not online. 579 * CPUs going online during the call will not be seen or sent an IPI. 580 * 581 * You must not call this function with disabled interrupts or 582 * from a hardware interrupt handler or from a bottom half handler. 583 */ 584 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info), 585 smp_call_func_t func, void *info, bool wait, 586 gfp_t gfp_flags) 587 { 588 cpumask_var_t cpus; 589 int cpu, ret; 590 591 might_sleep_if(gfp_flags & __GFP_WAIT); 592 593 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) { 594 preempt_disable(); 595 for_each_online_cpu(cpu) 596 if (cond_func(cpu, info)) 597 cpumask_set_cpu(cpu, cpus); 598 on_each_cpu_mask(cpus, func, info, wait); 599 preempt_enable(); 600 free_cpumask_var(cpus); 601 } else { 602 /* 603 * No free cpumask, bother. No matter, we'll 604 * just have to IPI them one by one. 605 */ 606 preempt_disable(); 607 for_each_online_cpu(cpu) 608 if (cond_func(cpu, info)) { 609 ret = smp_call_function_single(cpu, func, 610 info, wait); 611 WARN_ON_ONCE(!ret); 612 } 613 preempt_enable(); 614 } 615 } 616 EXPORT_SYMBOL(on_each_cpu_cond); 617 618 static void do_nothing(void *unused) 619 { 620 } 621 622 /** 623 * kick_all_cpus_sync - Force all cpus out of idle 624 * 625 * Used to synchronize the update of pm_idle function pointer. It's 626 * called after the pointer is updated and returns after the dummy 627 * callback function has been executed on all cpus. The execution of 628 * the function can only happen on the remote cpus after they have 629 * left the idle function which had been called via pm_idle function 630 * pointer. So it's guaranteed that nothing uses the previous pointer 631 * anymore. 632 */ 633 void kick_all_cpus_sync(void) 634 { 635 /* Make sure the change is visible before we kick the cpus */ 636 smp_mb(); 637 smp_call_function(do_nothing, NULL, 1); 638 } 639 EXPORT_SYMBOL_GPL(kick_all_cpus_sync); 640