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 /** 215 * stop_two_cpus - stops two cpus 216 * @cpu1: the cpu to stop 217 * @cpu2: the other cpu to stop 218 * @fn: function to execute 219 * @arg: argument to @fn 220 * 221 * Stops both the current and specified CPU and runs @fn on one of them. 222 * 223 * returns when both are completed. 224 */ 225 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg) 226 { 227 struct cpu_stop_done done; 228 struct cpu_stop_work work1, work2; 229 struct multi_stop_data msdata; 230 231 preempt_disable(); 232 msdata = (struct multi_stop_data){ 233 .fn = fn, 234 .data = arg, 235 .num_threads = 2, 236 .active_cpus = cpumask_of(cpu1), 237 }; 238 239 work1 = work2 = (struct cpu_stop_work){ 240 .fn = multi_cpu_stop, 241 .arg = &msdata, 242 .done = &done 243 }; 244 245 cpu_stop_init_done(&done, 2); 246 set_state(&msdata, MULTI_STOP_PREPARE); 247 248 /* 249 * If we observe both CPUs active we know _cpu_down() cannot yet have 250 * queued its stop_machine works and therefore ours will get executed 251 * first. Or its not either one of our CPUs that's getting unplugged, 252 * in which case we don't care. 253 * 254 * This relies on the stopper workqueues to be FIFO. 255 */ 256 if (!cpu_active(cpu1) || !cpu_active(cpu2)) { 257 preempt_enable(); 258 return -ENOENT; 259 } 260 261 lg_double_lock(&stop_cpus_lock, cpu1, cpu2); 262 cpu_stop_queue_work(cpu1, &work1); 263 cpu_stop_queue_work(cpu2, &work2); 264 lg_double_unlock(&stop_cpus_lock, cpu1, cpu2); 265 266 preempt_enable(); 267 268 wait_for_completion(&done.completion); 269 270 return done.executed ? done.ret : -ENOENT; 271 } 272 273 /** 274 * stop_one_cpu_nowait - stop a cpu but don't wait for completion 275 * @cpu: cpu to stop 276 * @fn: function to execute 277 * @arg: argument to @fn 278 * @work_buf: pointer to cpu_stop_work structure 279 * 280 * Similar to stop_one_cpu() but doesn't wait for completion. The 281 * caller is responsible for ensuring @work_buf is currently unused 282 * and will remain untouched until stopper starts executing @fn. 283 * 284 * CONTEXT: 285 * Don't care. 286 */ 287 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg, 288 struct cpu_stop_work *work_buf) 289 { 290 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, }; 291 cpu_stop_queue_work(cpu, work_buf); 292 } 293 294 /* static data for stop_cpus */ 295 static DEFINE_MUTEX(stop_cpus_mutex); 296 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work); 297 298 static void queue_stop_cpus_work(const struct cpumask *cpumask, 299 cpu_stop_fn_t fn, void *arg, 300 struct cpu_stop_done *done) 301 { 302 struct cpu_stop_work *work; 303 unsigned int cpu; 304 305 /* initialize works and done */ 306 for_each_cpu(cpu, cpumask) { 307 work = &per_cpu(stop_cpus_work, cpu); 308 work->fn = fn; 309 work->arg = arg; 310 work->done = done; 311 } 312 313 /* 314 * Disable preemption while queueing to avoid getting 315 * preempted by a stopper which might wait for other stoppers 316 * to enter @fn which can lead to deadlock. 317 */ 318 lg_global_lock(&stop_cpus_lock); 319 for_each_cpu(cpu, cpumask) 320 cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu)); 321 lg_global_unlock(&stop_cpus_lock); 322 } 323 324 static int __stop_cpus(const struct cpumask *cpumask, 325 cpu_stop_fn_t fn, void *arg) 326 { 327 struct cpu_stop_done done; 328 329 cpu_stop_init_done(&done, cpumask_weight(cpumask)); 330 queue_stop_cpus_work(cpumask, fn, arg, &done); 331 wait_for_completion(&done.completion); 332 return done.executed ? done.ret : -ENOENT; 333 } 334 335 /** 336 * stop_cpus - stop multiple cpus 337 * @cpumask: cpus to stop 338 * @fn: function to execute 339 * @arg: argument to @fn 340 * 341 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu, 342 * @fn is run in a process context with the highest priority 343 * preempting any task on the cpu and monopolizing it. This function 344 * returns after all executions are complete. 345 * 346 * This function doesn't guarantee the cpus in @cpumask stay online 347 * till @fn completes. If some cpus go down in the middle, execution 348 * on the cpu may happen partially or fully on different cpus. @fn 349 * should either be ready for that or the caller should ensure that 350 * the cpus stay online until this function completes. 351 * 352 * All stop_cpus() calls are serialized making it safe for @fn to wait 353 * for all cpus to start executing it. 354 * 355 * CONTEXT: 356 * Might sleep. 357 * 358 * RETURNS: 359 * -ENOENT if @fn(@arg) was not executed at all because all cpus in 360 * @cpumask were offline; otherwise, 0 if all executions of @fn 361 * returned 0, any non zero return value if any returned non zero. 362 */ 363 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg) 364 { 365 int ret; 366 367 /* static works are used, process one request at a time */ 368 mutex_lock(&stop_cpus_mutex); 369 ret = __stop_cpus(cpumask, fn, arg); 370 mutex_unlock(&stop_cpus_mutex); 371 return ret; 372 } 373 374 /** 375 * try_stop_cpus - try to stop multiple cpus 376 * @cpumask: cpus to stop 377 * @fn: function to execute 378 * @arg: argument to @fn 379 * 380 * Identical to stop_cpus() except that it fails with -EAGAIN if 381 * someone else is already using the facility. 382 * 383 * CONTEXT: 384 * Might sleep. 385 * 386 * RETURNS: 387 * -EAGAIN if someone else is already stopping cpus, -ENOENT if 388 * @fn(@arg) was not executed at all because all cpus in @cpumask were 389 * offline; otherwise, 0 if all executions of @fn returned 0, any non 390 * zero return value if any returned non zero. 391 */ 392 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg) 393 { 394 int ret; 395 396 /* static works are used, process one request at a time */ 397 if (!mutex_trylock(&stop_cpus_mutex)) 398 return -EAGAIN; 399 ret = __stop_cpus(cpumask, fn, arg); 400 mutex_unlock(&stop_cpus_mutex); 401 return ret; 402 } 403 404 static int cpu_stop_should_run(unsigned int cpu) 405 { 406 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 407 unsigned long flags; 408 int run; 409 410 spin_lock_irqsave(&stopper->lock, flags); 411 run = !list_empty(&stopper->works); 412 spin_unlock_irqrestore(&stopper->lock, flags); 413 return run; 414 } 415 416 static void cpu_stopper_thread(unsigned int cpu) 417 { 418 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 419 struct cpu_stop_work *work; 420 int ret; 421 422 repeat: 423 work = NULL; 424 spin_lock_irq(&stopper->lock); 425 if (!list_empty(&stopper->works)) { 426 work = list_first_entry(&stopper->works, 427 struct cpu_stop_work, list); 428 list_del_init(&work->list); 429 } 430 spin_unlock_irq(&stopper->lock); 431 432 if (work) { 433 cpu_stop_fn_t fn = work->fn; 434 void *arg = work->arg; 435 struct cpu_stop_done *done = work->done; 436 char ksym_buf[KSYM_NAME_LEN] __maybe_unused; 437 438 /* cpu stop callbacks are not allowed to sleep */ 439 preempt_disable(); 440 441 ret = fn(arg); 442 if (ret) 443 done->ret = ret; 444 445 /* restore preemption and check it's still balanced */ 446 preempt_enable(); 447 WARN_ONCE(preempt_count(), 448 "cpu_stop: %s(%p) leaked preempt count\n", 449 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL, 450 ksym_buf), arg); 451 452 cpu_stop_signal_done(done, true); 453 goto repeat; 454 } 455 } 456 457 extern void sched_set_stop_task(int cpu, struct task_struct *stop); 458 459 static void cpu_stop_create(unsigned int cpu) 460 { 461 sched_set_stop_task(cpu, per_cpu(cpu_stopper_task, cpu)); 462 } 463 464 static void cpu_stop_park(unsigned int cpu) 465 { 466 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 467 struct cpu_stop_work *work; 468 unsigned long flags; 469 470 /* drain remaining works */ 471 spin_lock_irqsave(&stopper->lock, flags); 472 list_for_each_entry(work, &stopper->works, list) 473 cpu_stop_signal_done(work->done, false); 474 stopper->enabled = false; 475 spin_unlock_irqrestore(&stopper->lock, flags); 476 } 477 478 static void cpu_stop_unpark(unsigned int cpu) 479 { 480 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 481 482 spin_lock_irq(&stopper->lock); 483 stopper->enabled = true; 484 spin_unlock_irq(&stopper->lock); 485 } 486 487 static struct smp_hotplug_thread cpu_stop_threads = { 488 .store = &cpu_stopper_task, 489 .thread_should_run = cpu_stop_should_run, 490 .thread_fn = cpu_stopper_thread, 491 .thread_comm = "migration/%u", 492 .create = cpu_stop_create, 493 .setup = cpu_stop_unpark, 494 .park = cpu_stop_park, 495 .pre_unpark = cpu_stop_unpark, 496 .selfparking = true, 497 }; 498 499 static int __init cpu_stop_init(void) 500 { 501 unsigned int cpu; 502 503 for_each_possible_cpu(cpu) { 504 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu); 505 506 spin_lock_init(&stopper->lock); 507 INIT_LIST_HEAD(&stopper->works); 508 } 509 510 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads)); 511 stop_machine_initialized = true; 512 return 0; 513 } 514 early_initcall(cpu_stop_init); 515 516 #ifdef CONFIG_STOP_MACHINE 517 518 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus) 519 { 520 struct multi_stop_data msdata = { 521 .fn = fn, 522 .data = data, 523 .num_threads = num_online_cpus(), 524 .active_cpus = cpus, 525 }; 526 527 if (!stop_machine_initialized) { 528 /* 529 * Handle the case where stop_machine() is called 530 * early in boot before stop_machine() has been 531 * initialized. 532 */ 533 unsigned long flags; 534 int ret; 535 536 WARN_ON_ONCE(msdata.num_threads != 1); 537 538 local_irq_save(flags); 539 hard_irq_disable(); 540 ret = (*fn)(data); 541 local_irq_restore(flags); 542 543 return ret; 544 } 545 546 /* Set the initial state and stop all online cpus. */ 547 set_state(&msdata, MULTI_STOP_PREPARE); 548 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata); 549 } 550 551 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus) 552 { 553 int ret; 554 555 /* No CPUs can come up or down during this. */ 556 get_online_cpus(); 557 ret = __stop_machine(fn, data, cpus); 558 put_online_cpus(); 559 return ret; 560 } 561 EXPORT_SYMBOL_GPL(stop_machine); 562 563 /** 564 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU 565 * @fn: the function to run 566 * @data: the data ptr for the @fn() 567 * @cpus: the cpus to run the @fn() on (NULL = any online cpu) 568 * 569 * This is identical to stop_machine() but can be called from a CPU which 570 * is not active. The local CPU is in the process of hotplug (so no other 571 * CPU hotplug can start) and not marked active and doesn't have enough 572 * context to sleep. 573 * 574 * This function provides stop_machine() functionality for such state by 575 * using busy-wait for synchronization and executing @fn directly for local 576 * CPU. 577 * 578 * CONTEXT: 579 * Local CPU is inactive. Temporarily stops all active CPUs. 580 * 581 * RETURNS: 582 * 0 if all executions of @fn returned 0, any non zero return value if any 583 * returned non zero. 584 */ 585 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data, 586 const struct cpumask *cpus) 587 { 588 struct multi_stop_data msdata = { .fn = fn, .data = data, 589 .active_cpus = cpus }; 590 struct cpu_stop_done done; 591 int ret; 592 593 /* Local CPU must be inactive and CPU hotplug in progress. */ 594 BUG_ON(cpu_active(raw_smp_processor_id())); 595 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */ 596 597 /* No proper task established and can't sleep - busy wait for lock. */ 598 while (!mutex_trylock(&stop_cpus_mutex)) 599 cpu_relax(); 600 601 /* Schedule work on other CPUs and execute directly for local CPU */ 602 set_state(&msdata, MULTI_STOP_PREPARE); 603 cpu_stop_init_done(&done, num_active_cpus()); 604 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata, 605 &done); 606 ret = multi_cpu_stop(&msdata); 607 608 /* Busy wait for completion. */ 609 while (!completion_done(&done.completion)) 610 cpu_relax(); 611 612 mutex_unlock(&stop_cpus_mutex); 613 return ret ?: done.ret; 614 } 615 616 #endif /* CONFIG_STOP_MACHINE */ 617