kernel/smpboot.c

   1 /*
   2  * Common SMP CPU bringup/teardown functions
   3  */
   4 #include <linux/cpu.h>
   5 #include <linux/err.h>
   6 #include <linux/smp.h>
   7 #include <linux/delay.h>
   8 #include <linux/init.h>
   9 #include <linux/list.h>
  10 #include <linux/slab.h>
  11 #include <linux/sched.h>
  12 #include <linux/export.h>
  13 #include <linux/percpu.h>
  14 #include <linux/kthread.h>
  15 #include <linux/smpboot.h>
  16
  17 #include "smpboot.h"
  18
  19 #ifdef CONFIG_SMP
  20
  21 #ifdef CONFIG_GENERIC_SMP_IDLE_THREAD
  22 /*
  23  * For the hotplug case we keep the task structs around and reuse
  24  * them.
  25  */
  26 static DEFINE_PER_CPU(struct task_struct *, idle_threads);
  27
  28 struct task_struct *idle_thread_get(unsigned int cpu)
  29 {
  30         struct task_struct *tsk = per_cpu(idle_threads, cpu);
  31
  32         if (!tsk)
  33                 return ERR_PTR(-ENOMEM);
  34         init_idle(tsk, cpu);
  35         return tsk;
  36 }
  37
  38 void __init idle_thread_set_boot_cpu(void)
  39 {
  40         per_cpu(idle_threads, smp_processor_id()) = current;
  41 }
  42
  43 /**
  44  * idle_init - Initialize the idle thread for a cpu
  45  * @cpu:        The cpu for which the idle thread should be initialized
  46  *
  47  * Creates the thread if it does not exist.
  48  */
  49 static inline void idle_init(unsigned int cpu)
  50 {
  51         struct task_struct *tsk = per_cpu(idle_threads, cpu);
  52
  53         if (!tsk) {
  54                 tsk = fork_idle(cpu);
  55                 if (IS_ERR(tsk))
  56                         pr_err("SMP: fork_idle() failed for CPU %u\n", cpu);
  57                 else
  58                         per_cpu(idle_threads, cpu) = tsk;
  59         }
  60 }
  61
  62 /**
  63  * idle_threads_init - Initialize idle threads for all cpus
  64  */
  65 void __init idle_threads_init(void)
  66 {
  67         unsigned int cpu, boot_cpu;
  68
  69         boot_cpu = smp_processor_id();
  70
  71         for_each_possible_cpu(cpu) {
  72                 if (cpu != boot_cpu)
  73                         idle_init(cpu);
  74         }
  75 }
  76 #endif
  77
  78 #endif /* #ifdef CONFIG_SMP */
  79
  80 static LIST_HEAD(hotplug_threads);
  81 static DEFINE_MUTEX(smpboot_threads_lock);
  82
  83 struct smpboot_thread_data {
  84         unsigned int                    cpu;
  85         unsigned int                    status;
  86         struct smp_hotplug_thread       *ht;
  87 };
  88
  89 enum {
  90         HP_THREAD_NONE = 0,
  91         HP_THREAD_ACTIVE,
  92         HP_THREAD_PARKED,
  93 };
  94
  95 /**
  96  * smpboot_thread_fn - percpu hotplug thread loop function
  97  * @data:       thread data pointer
  98  *
  99  * Checks for thread stop and park conditions. Calls the necessary
 100  * setup, cleanup, park and unpark functions for the registered
 101  * thread.
 102  *
 103  * Returns 1 when the thread should exit, 0 otherwise.
 104  */
 105 static int smpboot_thread_fn(void *data)
 106 {
 107         struct smpboot_thread_data *td = data;
 108         struct smp_hotplug_thread *ht = td->ht;
 109
 110         while (1) {
 111                 set_current_state(TASK_INTERRUPTIBLE);
 112                 preempt_disable();
 113                 if (kthread_should_stop()) {
 114                         __set_current_state(TASK_RUNNING);
 115                         preempt_enable();
 116                         /* cleanup must mirror setup */
 117                         if (ht->cleanup && td->status != HP_THREAD_NONE)
 118                                 ht->cleanup(td->cpu, cpu_online(td->cpu));
 119                         kfree(td);
 120                         return 0;
 121                 }
 122
 123                 if (kthread_should_park()) {
 124                         __set_current_state(TASK_RUNNING);
 125                         preempt_enable();
 126                         if (ht->park && td->status == HP_THREAD_ACTIVE) {
 127                                 BUG_ON(td->cpu != smp_processor_id());
 128                                 ht->park(td->cpu);
 129                                 td->status = HP_THREAD_PARKED;
 130                         }
 131                         kthread_parkme();
 132                         /* We might have been woken for stop */
 133                         continue;
 134                 }
 135
 136                 BUG_ON(td->cpu != smp_processor_id());
 137
 138                 /* Check for state change setup */
 139                 switch (td->status) {
 140                 case HP_THREAD_NONE:
 141                         __set_current_state(TASK_RUNNING);
 142                         preempt_enable();
 143                         if (ht->setup)
 144                                 ht->setup(td->cpu);
 145                         td->status = HP_THREAD_ACTIVE;
 146                         continue;
 147
 148                 case HP_THREAD_PARKED:
 149                         __set_current_state(TASK_RUNNING);
 150                         preempt_enable();
 151                         if (ht->unpark)
 152                                 ht->unpark(td->cpu);
 153                         td->status = HP_THREAD_ACTIVE;
 154                         continue;
 155                 }
 156
 157                 if (!ht->thread_should_run(td->cpu)) {
 158                         preempt_enable_no_resched();
 159                         schedule();
 160                 } else {
 161                         __set_current_state(TASK_RUNNING);
 162                         preempt_enable();
 163                         ht->thread_fn(td->cpu);
 164                 }
 165         }
 166 }
 167
 168 static int
 169 __smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 170 {
 171         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 172         struct smpboot_thread_data *td;
 173
 174         if (tsk)
 175                 return 0;
 176
 177         td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
 178         if (!td)
 179                 return -ENOMEM;
 180         td->cpu = cpu;
 181         td->ht = ht;
 182
 183         tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
 184                                     ht->thread_comm);
 185         if (IS_ERR(tsk)) {
 186                 kfree(td);
 187                 return PTR_ERR(tsk);
 188         }
 189         get_task_struct(tsk);
 190         *per_cpu_ptr(ht->store, cpu) = tsk;
 191         if (ht->create) {
 192                 /*
 193                  * Make sure that the task has actually scheduled out
 194                  * into park position, before calling the create
 195                  * callback. At least the migration thread callback
 196                  * requires that the task is off the runqueue.
 197                  */
 198                 if (!wait_task_inactive(tsk, TASK_PARKED))
 199                         WARN_ON(1);
 200                 else
 201                         ht->create(cpu);
 202         }
 203         return 0;
 204 }
 205
 206 int smpboot_create_threads(unsigned int cpu)
 207 {
 208         struct smp_hotplug_thread *cur;
 209         int ret = 0;
 210
 211         mutex_lock(&smpboot_threads_lock);
 212         list_for_each_entry(cur, &hotplug_threads, list) {
 213                 ret = __smpboot_create_thread(cur, cpu);
 214                 if (ret)
 215                         break;
 216         }
 217         mutex_unlock(&smpboot_threads_lock);
 218         return ret;
 219 }
 220
 221 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 222 {
 223         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 224
 225         if (!ht->selfparking)
 226                 kthread_unpark(tsk);
 227 }
 228
 229 int smpboot_unpark_threads(unsigned int cpu)
 230 {
 231         struct smp_hotplug_thread *cur;
 232
 233         mutex_lock(&smpboot_threads_lock);
 234         list_for_each_entry(cur, &hotplug_threads, list)
 235                 if (cpumask_test_cpu(cpu, cur->cpumask))
 236                         smpboot_unpark_thread(cur, cpu);
 237         mutex_unlock(&smpboot_threads_lock);
 238         return 0;
 239 }
 240
 241 static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 242 {
 243         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 244
 245         if (tsk && !ht->selfparking)
 246                 kthread_park(tsk);
 247 }
 248
 249 int smpboot_park_threads(unsigned int cpu)
 250 {
 251         struct smp_hotplug_thread *cur;
 252
 253         mutex_lock(&smpboot_threads_lock);
 254         list_for_each_entry_reverse(cur, &hotplug_threads, list)
 255                 smpboot_park_thread(cur, cpu);
 256         mutex_unlock(&smpboot_threads_lock);
 257         return 0;
 258 }
 259
 260 static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
 261 {
 262         unsigned int cpu;
 263
 264         /* We need to destroy also the parked threads of offline cpus */
 265         for_each_possible_cpu(cpu) {
 266                 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 267
 268                 if (tsk) {
 269                         kthread_stop(tsk);
 270                         put_task_struct(tsk);
 271                         *per_cpu_ptr(ht->store, cpu) = NULL;
 272                 }
 273         }
 274 }
 275
 276 /**
 277  * smpboot_register_percpu_thread_cpumask - Register a per_cpu thread related
 278  *                                          to hotplug
 279  * @plug_thread:        Hotplug thread descriptor
 280  * @cpumask:            The cpumask where threads run
 281  *
 282  * Creates and starts the threads on all online cpus.
 283  */
 284 int smpboot_register_percpu_thread_cpumask(struct smp_hotplug_thread *plug_thread,
 285                                            const struct cpumask *cpumask)
 286 {
 287         unsigned int cpu;
 288         int ret = 0;
 289
 290         if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
 291                 return -ENOMEM;
 292         cpumask_copy(plug_thread->cpumask, cpumask);
 293
 294         get_online_cpus();
 295         mutex_lock(&smpboot_threads_lock);
 296         for_each_online_cpu(cpu) {
 297                 ret = __smpboot_create_thread(plug_thread, cpu);
 298                 if (ret) {
 299                         smpboot_destroy_threads(plug_thread);
 300                         free_cpumask_var(plug_thread->cpumask);
 301                         goto out;
 302                 }
 303                 if (cpumask_test_cpu(cpu, cpumask))
 304                         smpboot_unpark_thread(plug_thread, cpu);
 305         }
 306         list_add(&plug_thread->list, &hotplug_threads);
 307 out:
 308         mutex_unlock(&smpboot_threads_lock);
 309         put_online_cpus();
 310         return ret;
 311 }
 312 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread_cpumask);
 313
 314 /**
 315  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 316  * @plug_thread:        Hotplug thread descriptor
 317  *
 318  * Stops all threads on all possible cpus.
 319  */
 320 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 321 {
 322         get_online_cpus();
 323         mutex_lock(&smpboot_threads_lock);
 324         list_del(&plug_thread->list);
 325         smpboot_destroy_threads(plug_thread);
 326         mutex_unlock(&smpboot_threads_lock);
 327         put_online_cpus();
 328         free_cpumask_var(plug_thread->cpumask);
 329 }
 330 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 331
 332 /**
 333  * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
 334  * @plug_thread:        Hotplug thread descriptor
 335  * @new:                Revised mask to use
 336  *
 337  * The cpumask field in the smp_hotplug_thread must not be updated directly
 338  * by the client, but only by calling this function.
 339  * This function can only be called on a registered smp_hotplug_thread.
 340  */
 341 int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
 342                                          const struct cpumask *new)
 343 {
 344         struct cpumask *old = plug_thread->cpumask;
 345         cpumask_var_t tmp;
 346         unsigned int cpu;
 347
 348         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 349                 return -ENOMEM;
 350
 351         get_online_cpus();
 352         mutex_lock(&smpboot_threads_lock);
 353
 354         /* Park threads that were exclusively enabled on the old mask. */
 355         cpumask_andnot(tmp, old, new);
 356         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 357                 smpboot_park_thread(plug_thread, cpu);
 358
 359         /* Unpark threads that are exclusively enabled on the new mask. */
 360         cpumask_andnot(tmp, new, old);
 361         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 362                 smpboot_unpark_thread(plug_thread, cpu);
 363
 364         cpumask_copy(old, new);
 365
 366         mutex_unlock(&smpboot_threads_lock);
 367         put_online_cpus();
 368
 369         free_cpumask_var(tmp);
 370
 371         return 0;
 372 }
 373 EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread);
 374
 375 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
 376
 377 /*
 378  * Called to poll specified CPU's state, for example, when waiting for
 379  * a CPU to come online.
 380  */
 381 int cpu_report_state(int cpu)
 382 {
 383         return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 384 }
 385
 386 /*
 387  * If CPU has died properly, set its state to CPU_UP_PREPARE and
 388  * return success.  Otherwise, return -EBUSY if the CPU died after
 389  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
 390  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
 391  * to dying.  In the latter two cases, the CPU might not be set up
 392  * properly, but it is up to the arch-specific code to decide.
 393  * Finally, -EIO indicates an unanticipated problem.
 394  *
 395  * Note that it is permissible to omit this call entirely, as is
 396  * done in architectures that do no CPU-hotplug error checking.
 397  */
 398 int cpu_check_up_prepare(int cpu)
 399 {
 400         if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
 401                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 402                 return 0;
 403         }
 404
 405         switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
 406
 407         case CPU_POST_DEAD:
 408
 409                 /* The CPU died properly, so just start it up again. */
 410                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 411                 return 0;
 412
 413         case CPU_DEAD_FROZEN:
 414
 415                 /*
 416                  * Timeout during CPU death, so let caller know.
 417                  * The outgoing CPU completed its processing, but after
 418                  * cpu_wait_death() timed out and reported the error. The
 419                  * caller is free to proceed, in which case the state
 420                  * will be reset properly by cpu_set_state_online().
 421                  * Proceeding despite this -EBUSY return makes sense
 422                  * for systems where the outgoing CPUs take themselves
 423                  * offline, with no post-death manipulation required from
 424                  * a surviving CPU.
 425                  */
 426                 return -EBUSY;
 427
 428         case CPU_BROKEN:
 429
 430                 /*
 431                  * The most likely reason we got here is that there was
 432                  * a timeout during CPU death, and the outgoing CPU never
 433                  * did complete its processing.  This could happen on
 434                  * a virtualized system if the outgoing VCPU gets preempted
 435                  * for more than five seconds, and the user attempts to
 436                  * immediately online that same CPU.  Trying again later
 437                  * might return -EBUSY above, hence -EAGAIN.
 438                  */
 439                 return -EAGAIN;
 440
 441         default:
 442
 443                 /* Should not happen.  Famous last words. */
 444                 return -EIO;
 445         }
 446 }
 447
 448 /*
 449  * Mark the specified CPU online.
 450  *
 451  * Note that it is permissible to omit this call entirely, as is
 452  * done in architectures that do no CPU-hotplug error checking.
 453  */
 454 void cpu_set_state_online(int cpu)
 455 {
 456         (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
 457 }
 458
 459 #ifdef CONFIG_HOTPLUG_CPU
 460
 461 /*
 462  * Wait for the specified CPU to exit the idle loop and die.
 463  */
 464 bool cpu_wait_death(unsigned int cpu, int seconds)
 465 {
 466         int jf_left = seconds * HZ;
 467         int oldstate;
 468         bool ret = true;
 469         int sleep_jf = 1;
 470
 471         might_sleep();
 472
 473         /* The outgoing CPU will normally get done quite quickly. */
 474         if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
 475                 goto update_state;
 476         udelay(5);
 477
 478         /* But if the outgoing CPU dawdles, wait increasingly long times. */
 479         while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
 480                 schedule_timeout_uninterruptible(sleep_jf);
 481                 jf_left -= sleep_jf;
 482                 if (jf_left <= 0)
 483                         break;
 484                 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
 485         }
 486 update_state:
 487         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 488         if (oldstate == CPU_DEAD) {
 489                 /* Outgoing CPU died normally, update state. */
 490                 smp_mb(); /* atomic_read() before update. */
 491                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
 492         } else {
 493                 /* Outgoing CPU still hasn't died, set state accordingly. */
 494                 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 495                                    oldstate, CPU_BROKEN) != oldstate)
 496                         goto update_state;
 497                 ret = false;
 498         }
 499         return ret;
 500 }
 501
 502 /*
 503  * Called by the outgoing CPU to report its successful death.  Return
 504  * false if this report follows the surviving CPU's timing out.
 505  *
 506  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
 507  * timed out.  This approach allows architectures to omit calls to
 508  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
 509  * the next cpu_wait_death()'s polling loop.
 510  */
 511 bool cpu_report_death(void)
 512 {
 513         int oldstate;
 514         int newstate;
 515         int cpu = smp_processor_id();
 516
 517         do {
 518                 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 519                 if (oldstate != CPU_BROKEN)
 520                         newstate = CPU_DEAD;
 521                 else
 522                         newstate = CPU_DEAD_FROZEN;
 523         } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 524                                 oldstate, newstate) != oldstate);
 525         return newstate == CPU_DEAD;
 526 }
 527
 528 #endif /* #ifdef CONFIG_HOTPLUG_CPU */