1 /* 2 * kernel/stop_machine.c 3 * 4 * Copyright (C) 2008, 2005 IBM Corporation. 5 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au 6 * Copyright (C) 2010 SUSE Linux Products GmbH 7 * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 8 * 9 * This file is released under the GPLv2 and any later version. 10 */ 11 #include <linux/completion.h> 12 #include <linux/cpu.h> 13 #include <linux/init.h> 14 #include <linux/kthread.h> 15 #include <linux/export.h> 16 #include <linux/percpu.h> 17 #include <linux/sched.h> 18 #include <linux/stop_machine.h> 19 #include <linux/interrupt.h> 20 #include <linux/kallsyms.h> 21 #include <linux/smpboot.h> 22 #include <linux/atomic.h> 23 #include <linux/lglock.h> 24 25 /* 26 * Structure to determine completion condition and record errors. May 27 * be shared by works on different cpus. 28 */ 29 struct cpu_stop_done { 30 atomic_t nr_todo; /* nr left to execute */ 31 bool executed; /* actually executed? */ 32 int ret; /* collected return value */ 33 struct completion completion; /* fired if nr_todo reaches 0 */ 34 }; 35 36 /* the actual stopper, one per every possible cpu, enabled on online cpus */ 37 struct cpu_stopper { 38 spinlock_t lock; 39 bool enabled; /* is this stopper enabled? */ 40 struct list_head works; /* list of pending works */ 41 }; 42 43 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper); 44 static DEFINE_PER_CPU(struct task_struct *, cpu_stopper_task); 45 static bool stop_machine_initialized = false; 46 47 /* 48 * Avoids a race between stop_two_cpus and global stop_cpus, where 49 * the stoppers could get queued up in reverse order, leading to 50 * system deadlock. Using an lglock means stop_two_cpus remains 51 * relatively cheap. 52 */ 53 DEFINE_STATIC_LGLOCK(stop_cpus_lock); 54 55 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo) 56 { 57 memset(done, 0, sizeof(*done)); 58 atomic_set(&done->nr_todo, nr_todo); 59 init_completion(&done->completion); 60 } 61 62 /* signal completion unless @done is NULL */ 63 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed) 64 { 65 if (done) { 66 if (executed) 67 done->executed = true; 68 if (atomic_dec_and_test(&done->nr_todo)) 69 complete(&done->completion); 70 } 71 } 72 73 /* queue @work to @stopper. if offline, @work is completed immediately */ 74 static void cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work) 75 { 76 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 77 struct task_struct *p = per_cpu(cpu_stopper_task, cpu); 78 79 unsigned long flags; 80 81 spin_lock_irqsave(&stopper->lock, flags); 82 83 if (stopper->enabled) { 84 list_add_tail(&work->list, &stopper->works); 85 wake_up_process(p); 86 } else 87 cpu_stop_signal_done(work->done, false); 88 89 spin_unlock_irqrestore(&stopper->lock, flags); 90 } 91 92 /** 93 * stop_one_cpu - stop a cpu 94 * @cpu: cpu to stop 95 * @fn: function to execute 96 * @arg: argument to @fn 97 * 98 * Execute @fn(@arg) on @cpu. @fn is run in a process context with 99 * the highest priority preempting any task on the cpu and 100 * monopolizing it. This function returns after the execution is 101 * complete. 102 * 103 * This function doesn't guarantee @cpu stays online till @fn 104 * completes. If @cpu goes down in the middle, execution may happen 105 * partially or fully on different cpus. @fn should either be ready 106 * for that or the caller should ensure that @cpu stays online until 107 * this function completes. 108 * 109 * CONTEXT: 110 * Might sleep. 111 * 112 * RETURNS: 113 * -ENOENT if @fn(@arg) was not executed because @cpu was offline; 114 * otherwise, the return value of @fn. 115 */ 116 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg) 117 { 118 struct cpu_stop_done done; 119 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done }; 120 121 cpu_stop_init_done(&done, 1); 122 cpu_stop_queue_work(cpu, &work); 123 wait_for_completion(&done.completion); 124 return done.executed ? done.ret : -ENOENT; 125 } 126 127 /* This controls the threads on each CPU. */ 128 enum multi_stop_state { 129 /* Dummy starting state for thread. */ 130 MULTI_STOP_NONE, 131 /* Awaiting everyone to be scheduled. */ 132 MULTI_STOP_PREPARE, 133 /* Disable interrupts. */ 134 MULTI_STOP_DISABLE_IRQ, 135 /* Run the function */ 136 MULTI_STOP_RUN, 137 /* Exit */ 138 MULTI_STOP_EXIT, 139 }; 140 141 struct multi_stop_data { 142 int (*fn)(void *); 143 void *data; 144 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */ 145 unsigned int num_threads; 146 const struct cpumask *active_cpus; 147 148 enum multi_stop_state state; 149 atomic_t thread_ack; 150 }; 151 152 static void set_state(struct multi_stop_data *msdata, 153 enum multi_stop_state newstate) 154 { 155 /* Reset ack counter. */ 156 atomic_set(&msdata->thread_ack, msdata->num_threads); 157 smp_wmb(); 158 msdata->state = newstate; 159 } 160 161 /* Last one to ack a state moves to the next state. */ 162 static void ack_state(struct multi_stop_data *msdata) 163 { 164 if (atomic_dec_and_test(&msdata->thread_ack)) 165 set_state(msdata, msdata->state + 1); 166 } 167 168 /* This is the cpu_stop function which stops the CPU. */ 169 static int multi_cpu_stop(void *data) 170 { 171 struct multi_stop_data *msdata = data; 172 enum multi_stop_state curstate = MULTI_STOP_NONE; 173 int cpu = smp_processor_id(), err = 0; 174 unsigned long flags; 175 bool is_active; 176 177 /* 178 * When called from stop_machine_from_inactive_cpu(), irq might 179 * already be disabled. Save the state and restore it on exit. 180 */ 181 local_save_flags(flags); 182 183 if (!msdata->active_cpus) 184 is_active = cpu == cpumask_first(cpu_online_mask); 185 else 186 is_active = cpumask_test_cpu(cpu, msdata->active_cpus); 187 188 /* Simple state machine */ 189 do { 190 /* Chill out and ensure we re-read multi_stop_state. */ 191 cpu_relax(); 192 if (msdata->state != curstate) { 193 curstate = msdata->state; 194 switch (curstate) { 195 case MULTI_STOP_DISABLE_IRQ: 196 local_irq_disable(); 197 hard_irq_disable(); 198 break; 199 case MULTI_STOP_RUN: 200 if (is_active) 201 err = msdata->fn(msdata->data); 202 break; 203 default: 204 break; 205 } 206 ack_state(msdata); 207 } 208 } while (curstate != MULTI_STOP_EXIT); 209 210 local_irq_restore(flags); 211 return err; 212 } 213 214 struct irq_cpu_stop_queue_work_info { 215 int cpu1; 216 int cpu2; 217 struct cpu_stop_work *work1; 218 struct cpu_stop_work *work2; 219 }; 220 221 /* 222 * This function is always run with irqs and preemption disabled. 223 * This guarantees that both work1 and work2 get queued, before 224 * our local migrate thread gets the chance to preempt us. 225 */ 226 static void irq_cpu_stop_queue_work(void *arg) 227 { 228 struct irq_cpu_stop_queue_work_info *info = arg; 229 cpu_stop_queue_work(info->cpu1, info->work1); 230 cpu_stop_queue_work(info->cpu2, info->work2); 231 } 232 233 /** 234 * stop_two_cpus - stops two cpus 235 * @cpu1: the cpu to stop 236 * @cpu2: the other cpu to stop 237 * @fn: function to execute 238 * @arg: argument to @fn 239 * 240 * Stops both the current and specified CPU and runs @fn on one of them. 241 * 242 * returns when both are completed. 243 */ 244 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg) 245 { 246 struct cpu_stop_done done; 247 struct cpu_stop_work work1, work2; 248 struct irq_cpu_stop_queue_work_info call_args; 249 struct multi_stop_data msdata; 250 251 preempt_disable(); 252 msdata = (struct multi_stop_data){ 253 .fn = fn, 254 .data = arg, 255 .num_threads = 2, 256 .active_cpus = cpumask_of(cpu1), 257 }; 258 259 work1 = work2 = (struct cpu_stop_work){ 260 .fn = multi_cpu_stop, 261 .arg = &msdata, 262 .done = &done 263 }; 264 265 call_args = (struct irq_cpu_stop_queue_work_info){ 266 .cpu1 = cpu1, 267 .cpu2 = cpu2, 268 .work1 = &work1, 269 .work2 = &work2, 270 }; 271 272 cpu_stop_init_done(&done, 2); 273 set_state(&msdata, MULTI_STOP_PREPARE); 274 275 /* 276 * If we observe both CPUs active we know _cpu_down() cannot yet have 277 * queued its stop_machine works and therefore ours will get executed 278 * first. Or its not either one of our CPUs that's getting unplugged, 279 * in which case we don't care. 280 * 281 * This relies on the stopper workqueues to be FIFO. 282 */ 283 if (!cpu_active(cpu1) || !cpu_active(cpu2)) { 284 preempt_enable(); 285 return -ENOENT; 286 } 287 288 lg_local_lock(&stop_cpus_lock); 289 /* 290 * Queuing needs to be done by the lowest numbered CPU, to ensure 291 * that works are always queued in the same order on every CPU. 292 * This prevents deadlocks. 293 */ 294 smp_call_function_single(min(cpu1, cpu2), 295 &irq_cpu_stop_queue_work, 296 &call_args, 1); 297 lg_local_unlock(&stop_cpus_lock); 298 preempt_enable(); 299 300 wait_for_completion(&done.completion); 301 302 return done.executed ? done.ret : -ENOENT; 303 } 304 305 /** 306 * stop_one_cpu_nowait - stop a cpu but don't wait for completion 307 * @cpu: cpu to stop 308 * @fn: function to execute 309 * @arg: argument to @fn 310 * 311 * Similar to stop_one_cpu() but doesn't wait for completion. The 312 * caller is responsible for ensuring @work_buf is currently unused 313 * and will remain untouched until stopper starts executing @fn. 314 * 315 * CONTEXT: 316 * Don't care. 317 */ 318 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg, 319 struct cpu_stop_work *work_buf) 320 { 321 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, }; 322 cpu_stop_queue_work(cpu, work_buf); 323 } 324 325 /* static data for stop_cpus */ 326 static DEFINE_MUTEX(stop_cpus_mutex); 327 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work); 328 329 static void queue_stop_cpus_work(const struct cpumask *cpumask, 330 cpu_stop_fn_t fn, void *arg, 331 struct cpu_stop_done *done) 332 { 333 struct cpu_stop_work *work; 334 unsigned int cpu; 335 336 /* initialize works and done */ 337 for_each_cpu(cpu, cpumask) { 338 work = &per_cpu(stop_cpus_work, cpu); 339 work->fn = fn; 340 work->arg = arg; 341 work->done = done; 342 } 343 344 /* 345 * Disable preemption while queueing to avoid getting 346 * preempted by a stopper which might wait for other stoppers 347 * to enter @fn which can lead to deadlock. 348 */ 349 lg_global_lock(&stop_cpus_lock); 350 for_each_cpu(cpu, cpumask) 351 cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu)); 352 lg_global_unlock(&stop_cpus_lock); 353 } 354 355 static int __stop_cpus(const struct cpumask *cpumask, 356 cpu_stop_fn_t fn, void *arg) 357 { 358 struct cpu_stop_done done; 359 360 cpu_stop_init_done(&done, cpumask_weight(cpumask)); 361 queue_stop_cpus_work(cpumask, fn, arg, &done); 362 wait_for_completion(&done.completion); 363 return done.executed ? done.ret : -ENOENT; 364 } 365 366 /** 367 * stop_cpus - stop multiple cpus 368 * @cpumask: cpus to stop 369 * @fn: function to execute 370 * @arg: argument to @fn 371 * 372 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu, 373 * @fn is run in a process context with the highest priority 374 * preempting any task on the cpu and monopolizing it. This function 375 * returns after all executions are complete. 376 * 377 * This function doesn't guarantee the cpus in @cpumask stay online 378 * till @fn completes. If some cpus go down in the middle, execution 379 * on the cpu may happen partially or fully on different cpus. @fn 380 * should either be ready for that or the caller should ensure that 381 * the cpus stay online until this function completes. 382 * 383 * All stop_cpus() calls are serialized making it safe for @fn to wait 384 * for all cpus to start executing it. 385 * 386 * CONTEXT: 387 * Might sleep. 388 * 389 * RETURNS: 390 * -ENOENT if @fn(@arg) was not executed at all because all cpus in 391 * @cpumask were offline; otherwise, 0 if all executions of @fn 392 * returned 0, any non zero return value if any returned non zero. 393 */ 394 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg) 395 { 396 int ret; 397 398 /* static works are used, process one request at a time */ 399 mutex_lock(&stop_cpus_mutex); 400 ret = __stop_cpus(cpumask, fn, arg); 401 mutex_unlock(&stop_cpus_mutex); 402 return ret; 403 } 404 405 /** 406 * try_stop_cpus - try to stop multiple cpus 407 * @cpumask: cpus to stop 408 * @fn: function to execute 409 * @arg: argument to @fn 410 * 411 * Identical to stop_cpus() except that it fails with -EAGAIN if 412 * someone else is already using the facility. 413 * 414 * CONTEXT: 415 * Might sleep. 416 * 417 * RETURNS: 418 * -EAGAIN if someone else is already stopping cpus, -ENOENT if 419 * @fn(@arg) was not executed at all because all cpus in @cpumask were 420 * offline; otherwise, 0 if all executions of @fn returned 0, any non 421 * zero return value if any returned non zero. 422 */ 423 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg) 424 { 425 int ret; 426 427 /* static works are used, process one request at a time */ 428 if (!mutex_trylock(&stop_cpus_mutex)) 429 return -EAGAIN; 430 ret = __stop_cpus(cpumask, fn, arg); 431 mutex_unlock(&stop_cpus_mutex); 432 return ret; 433 } 434 435 static int cpu_stop_should_run(unsigned int cpu) 436 { 437 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 438 unsigned long flags; 439 int run; 440 441 spin_lock_irqsave(&stopper->lock, flags); 442 run = !list_empty(&stopper->works); 443 spin_unlock_irqrestore(&stopper->lock, flags); 444 return run; 445 } 446 447 static void cpu_stopper_thread(unsigned int cpu) 448 { 449 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 450 struct cpu_stop_work *work; 451 int ret; 452 453 repeat: 454 work = NULL; 455 spin_lock_irq(&stopper->lock); 456 if (!list_empty(&stopper->works)) { 457 work = list_first_entry(&stopper->works, 458 struct cpu_stop_work, list); 459 list_del_init(&work->list); 460 } 461 spin_unlock_irq(&stopper->lock); 462 463 if (work) { 464 cpu_stop_fn_t fn = work->fn; 465 void *arg = work->arg; 466 struct cpu_stop_done *done = work->done; 467 char ksym_buf[KSYM_NAME_LEN] __maybe_unused; 468 469 /* cpu stop callbacks are not allowed to sleep */ 470 preempt_disable(); 471 472 ret = fn(arg); 473 if (ret) 474 done->ret = ret; 475 476 /* restore preemption and check it's still balanced */ 477 preempt_enable(); 478 WARN_ONCE(preempt_count(), 479 "cpu_stop: %s(%p) leaked preempt count\n", 480 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL, 481 ksym_buf), arg); 482 483 cpu_stop_signal_done(done, true); 484 goto repeat; 485 } 486 } 487 488 extern void sched_set_stop_task(int cpu, struct task_struct *stop); 489 490 static void cpu_stop_create(unsigned int cpu) 491 { 492 sched_set_stop_task(cpu, per_cpu(cpu_stopper_task, cpu)); 493 } 494 495 static void cpu_stop_park(unsigned int cpu) 496 { 497 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 498 struct cpu_stop_work *work; 499 unsigned long flags; 500 501 /* drain remaining works */ 502 spin_lock_irqsave(&stopper->lock, flags); 503 list_for_each_entry(work, &stopper->works, list) 504 cpu_stop_signal_done(work->done, false); 505 stopper->enabled = false; 506 spin_unlock_irqrestore(&stopper->lock, flags); 507 } 508 509 static void cpu_stop_unpark(unsigned int cpu) 510 { 511 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 512 513 spin_lock_irq(&stopper->lock); 514 stopper->enabled = true; 515 spin_unlock_irq(&stopper->lock); 516 } 517 518 static struct smp_hotplug_thread cpu_stop_threads = { 519 .store = &cpu_stopper_task, 520 .thread_should_run = cpu_stop_should_run, 521 .thread_fn = cpu_stopper_thread, 522 .thread_comm = "migration/%u", 523 .create = cpu_stop_create, 524 .setup = cpu_stop_unpark, 525 .park = cpu_stop_park, 526 .pre_unpark = cpu_stop_unpark, 527 .selfparking = true, 528 }; 529 530 static int __init cpu_stop_init(void) 531 { 532 unsigned int cpu; 533 534 for_each_possible_cpu(cpu) { 535 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 536 537 spin_lock_init(&stopper->lock); 538 INIT_LIST_HEAD(&stopper->works); 539 } 540 541 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads)); 542 stop_machine_initialized = true; 543 return 0; 544 } 545 early_initcall(cpu_stop_init); 546 547 #ifdef CONFIG_STOP_MACHINE 548 549 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus) 550 { 551 struct multi_stop_data msdata = { 552 .fn = fn, 553 .data = data, 554 .num_threads = num_online_cpus(), 555 .active_cpus = cpus, 556 }; 557 558 if (!stop_machine_initialized) { 559 /* 560 * Handle the case where stop_machine() is called 561 * early in boot before stop_machine() has been 562 * initialized. 563 */ 564 unsigned long flags; 565 int ret; 566 567 WARN_ON_ONCE(msdata.num_threads != 1); 568 569 local_irq_save(flags); 570 hard_irq_disable(); 571 ret = (*fn)(data); 572 local_irq_restore(flags); 573 574 return ret; 575 } 576 577 /* Set the initial state and stop all online cpus. */ 578 set_state(&msdata, MULTI_STOP_PREPARE); 579 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata); 580 } 581 582 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus) 583 { 584 int ret; 585 586 /* No CPUs can come up or down during this. */ 587 get_online_cpus(); 588 ret = __stop_machine(fn, data, cpus); 589 put_online_cpus(); 590 return ret; 591 } 592 EXPORT_SYMBOL_GPL(stop_machine); 593 594 /** 595 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU 596 * @fn: the function to run 597 * @data: the data ptr for the @fn() 598 * @cpus: the cpus to run the @fn() on (NULL = any online cpu) 599 * 600 * This is identical to stop_machine() but can be called from a CPU which 601 * is not active. The local CPU is in the process of hotplug (so no other 602 * CPU hotplug can start) and not marked active and doesn't have enough 603 * context to sleep. 604 * 605 * This function provides stop_machine() functionality for such state by 606 * using busy-wait for synchronization and executing @fn directly for local 607 * CPU. 608 * 609 * CONTEXT: 610 * Local CPU is inactive. Temporarily stops all active CPUs. 611 * 612 * RETURNS: 613 * 0 if all executions of @fn returned 0, any non zero return value if any 614 * returned non zero. 615 */ 616 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data, 617 const struct cpumask *cpus) 618 { 619 struct multi_stop_data msdata = { .fn = fn, .data = data, 620 .active_cpus = cpus }; 621 struct cpu_stop_done done; 622 int ret; 623 624 /* Local CPU must be inactive and CPU hotplug in progress. */ 625 BUG_ON(cpu_active(raw_smp_processor_id())); 626 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */ 627 628 /* No proper task established and can't sleep - busy wait for lock. */ 629 while (!mutex_trylock(&stop_cpus_mutex)) 630 cpu_relax(); 631 632 /* Schedule work on other CPUs and execute directly for local CPU */ 633 set_state(&msdata, MULTI_STOP_PREPARE); 634 cpu_stop_init_done(&done, num_active_cpus()); 635 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata, 636 &done); 637 ret = multi_cpu_stop(&msdata); 638 639 /* Busy wait for completion. */ 640 while (!completion_done(&done.completion)) 641 cpu_relax(); 642 643 mutex_unlock(&stop_cpus_mutex); 644 return ret ?: done.ret; 645 } 646 647 #endif /* CONFIG_STOP_MACHINE */ 648