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         /*
 190          * Park the thread so that it could start right on the CPU
 191          * when it is available.
 192          */
 193         kthread_park(tsk);
 194         get_task_struct(tsk);
 195         *per_cpu_ptr(ht->store, cpu) = tsk;
 196         if (ht->create) {
 197                 /*
 198                  * Make sure that the task has actually scheduled out
 199                  * into park position, before calling the create
 200                  * callback. At least the migration thread callback
 201                  * requires that the task is off the runqueue.
 202                  */
 203                 if (!wait_task_inactive(tsk, TASK_PARKED))
 204                         WARN_ON(1);
 205                 else
 206                         ht->create(cpu);
 207         }
 208         return 0;
 209 }
 210
 211 int smpboot_create_threads(unsigned int cpu)
 212 {
 213         struct smp_hotplug_thread *cur;
 214         int ret = 0;
 215
 216         mutex_lock(&smpboot_threads_lock);
 217         list_for_each_entry(cur, &hotplug_threads, list) {
 218                 ret = __smpboot_create_thread(cur, cpu);
 219                 if (ret)
 220                         break;
 221         }
 222         mutex_unlock(&smpboot_threads_lock);
 223         return ret;
 224 }
 225
 226 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 227 {
 228         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 229
 230         if (!ht->selfparking)
 231                 kthread_unpark(tsk);
 232 }
 233
 234 int smpboot_unpark_threads(unsigned int cpu)
 235 {
 236         struct smp_hotplug_thread *cur;
 237
 238         mutex_lock(&smpboot_threads_lock);
 239         list_for_each_entry(cur, &hotplug_threads, list)
 240                 if (cpumask_test_cpu(cpu, cur->cpumask))
 241                         smpboot_unpark_thread(cur, cpu);
 242         mutex_unlock(&smpboot_threads_lock);
 243         return 0;
 244 }
 245
 246 static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 247 {
 248         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 249
 250         if (tsk && !ht->selfparking)
 251                 kthread_park(tsk);
 252 }
 253
 254 int smpboot_park_threads(unsigned int cpu)
 255 {
 256         struct smp_hotplug_thread *cur;
 257
 258         mutex_lock(&smpboot_threads_lock);
 259         list_for_each_entry_reverse(cur, &hotplug_threads, list)
 260                 smpboot_park_thread(cur, cpu);
 261         mutex_unlock(&smpboot_threads_lock);
 262         return 0;
 263 }
 264
 265 static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
 266 {
 267         unsigned int cpu;
 268
 269         /* We need to destroy also the parked threads of offline cpus */
 270         for_each_possible_cpu(cpu) {
 271                 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 272
 273                 if (tsk) {
 274                         kthread_stop(tsk);
 275                         put_task_struct(tsk);
 276                         *per_cpu_ptr(ht->store, cpu) = NULL;
 277                 }
 278         }
 279 }
 280
 281 /**
 282  * smpboot_register_percpu_thread_cpumask - Register a per_cpu thread related
 283  *                                          to hotplug
 284  * @plug_thread:        Hotplug thread descriptor
 285  * @cpumask:            The cpumask where threads run
 286  *
 287  * Creates and starts the threads on all online cpus.
 288  */
 289 int smpboot_register_percpu_thread_cpumask(struct smp_hotplug_thread *plug_thread,
 290                                            const struct cpumask *cpumask)
 291 {
 292         unsigned int cpu;
 293         int ret = 0;
 294
 295         if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
 296                 return -ENOMEM;
 297         cpumask_copy(plug_thread->cpumask, cpumask);
 298
 299         get_online_cpus();
 300         mutex_lock(&smpboot_threads_lock);
 301         for_each_online_cpu(cpu) {
 302                 ret = __smpboot_create_thread(plug_thread, cpu);
 303                 if (ret) {
 304                         smpboot_destroy_threads(plug_thread);
 305                         free_cpumask_var(plug_thread->cpumask);
 306                         goto out;
 307                 }
 308                 if (cpumask_test_cpu(cpu, cpumask))
 309                         smpboot_unpark_thread(plug_thread, cpu);
 310         }
 311         list_add(&plug_thread->list, &hotplug_threads);
 312 out:
 313         mutex_unlock(&smpboot_threads_lock);
 314         put_online_cpus();
 315         return ret;
 316 }
 317 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread_cpumask);
 318
 319 /**
 320  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 321  * @plug_thread:        Hotplug thread descriptor
 322  *
 323  * Stops all threads on all possible cpus.
 324  */
 325 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 326 {
 327         get_online_cpus();
 328         mutex_lock(&smpboot_threads_lock);
 329         list_del(&plug_thread->list);
 330         smpboot_destroy_threads(plug_thread);
 331         mutex_unlock(&smpboot_threads_lock);
 332         put_online_cpus();
 333         free_cpumask_var(plug_thread->cpumask);
 334 }
 335 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 336
 337 /**
 338  * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
 339  * @plug_thread:        Hotplug thread descriptor
 340  * @new:                Revised mask to use
 341  *
 342  * The cpumask field in the smp_hotplug_thread must not be updated directly
 343  * by the client, but only by calling this function.
 344  * This function can only be called on a registered smp_hotplug_thread.
 345  */
 346 int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
 347                                          const struct cpumask *new)
 348 {
 349         struct cpumask *old = plug_thread->cpumask;
 350         cpumask_var_t tmp;
 351         unsigned int cpu;
 352
 353         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 354                 return -ENOMEM;
 355
 356         get_online_cpus();
 357         mutex_lock(&smpboot_threads_lock);
 358
 359         /* Park threads that were exclusively enabled on the old mask. */
 360         cpumask_andnot(tmp, old, new);
 361         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 362                 smpboot_park_thread(plug_thread, cpu);
 363
 364         /* Unpark threads that are exclusively enabled on the new mask. */
 365         cpumask_andnot(tmp, new, old);
 366         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 367                 smpboot_unpark_thread(plug_thread, cpu);
 368
 369         cpumask_copy(old, new);
 370
 371         mutex_unlock(&smpboot_threads_lock);
 372         put_online_cpus();
 373
 374         free_cpumask_var(tmp);
 375
 376         return 0;
 377 }
 378 EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread);
 379
 380 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
 381
 382 /*
 383  * Called to poll specified CPU's state, for example, when waiting for
 384  * a CPU to come online.
 385  */
 386 int cpu_report_state(int cpu)
 387 {
 388         return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 389 }
 390
 391 /*
 392  * If CPU has died properly, set its state to CPU_UP_PREPARE and
 393  * return success.  Otherwise, return -EBUSY if the CPU died after
 394  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
 395  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
 396  * to dying.  In the latter two cases, the CPU might not be set up
 397  * properly, but it is up to the arch-specific code to decide.
 398  * Finally, -EIO indicates an unanticipated problem.
 399  *
 400  * Note that it is permissible to omit this call entirely, as is
 401  * done in architectures that do no CPU-hotplug error checking.
 402  */
 403 int cpu_check_up_prepare(int cpu)
 404 {
 405         if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
 406                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 407                 return 0;
 408         }
 409
 410         switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
 411
 412         case CPU_POST_DEAD:
 413
 414                 /* The CPU died properly, so just start it up again. */
 415                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 416                 return 0;
 417
 418         case CPU_DEAD_FROZEN:
 419
 420                 /*
 421                  * Timeout during CPU death, so let caller know.
 422                  * The outgoing CPU completed its processing, but after
 423                  * cpu_wait_death() timed out and reported the error. The
 424                  * caller is free to proceed, in which case the state
 425                  * will be reset properly by cpu_set_state_online().
 426                  * Proceeding despite this -EBUSY return makes sense
 427                  * for systems where the outgoing CPUs take themselves
 428                  * offline, with no post-death manipulation required from
 429                  * a surviving CPU.
 430                  */
 431                 return -EBUSY;
 432
 433         case CPU_BROKEN:
 434
 435                 /*
 436                  * The most likely reason we got here is that there was
 437                  * a timeout during CPU death, and the outgoing CPU never
 438                  * did complete its processing.  This could happen on
 439                  * a virtualized system if the outgoing VCPU gets preempted
 440                  * for more than five seconds, and the user attempts to
 441                  * immediately online that same CPU.  Trying again later
 442                  * might return -EBUSY above, hence -EAGAIN.
 443                  */
 444                 return -EAGAIN;
 445
 446         default:
 447
 448                 /* Should not happen.  Famous last words. */
 449                 return -EIO;
 450         }
 451 }
 452
 453 /*
 454  * Mark the specified CPU online.
 455  *
 456  * Note that it is permissible to omit this call entirely, as is
 457  * done in architectures that do no CPU-hotplug error checking.
 458  */
 459 void cpu_set_state_online(int cpu)
 460 {
 461         (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
 462 }
 463
 464 #ifdef CONFIG_HOTPLUG_CPU
 465
 466 /*
 467  * Wait for the specified CPU to exit the idle loop and die.
 468  */
 469 bool cpu_wait_death(unsigned int cpu, int seconds)
 470 {
 471         int jf_left = seconds * HZ;
 472         int oldstate;
 473         bool ret = true;
 474         int sleep_jf = 1;
 475
 476         might_sleep();
 477
 478         /* The outgoing CPU will normally get done quite quickly. */
 479         if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
 480                 goto update_state;
 481         udelay(5);
 482
 483         /* But if the outgoing CPU dawdles, wait increasingly long times. */
 484         while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
 485                 schedule_timeout_uninterruptible(sleep_jf);
 486                 jf_left -= sleep_jf;
 487                 if (jf_left <= 0)
 488                         break;
 489                 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
 490         }
 491 update_state:
 492         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 493         if (oldstate == CPU_DEAD) {
 494                 /* Outgoing CPU died normally, update state. */
 495                 smp_mb(); /* atomic_read() before update. */
 496                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
 497         } else {
 498                 /* Outgoing CPU still hasn't died, set state accordingly. */
 499                 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 500                                    oldstate, CPU_BROKEN) != oldstate)
 501                         goto update_state;
 502                 ret = false;
 503         }
 504         return ret;
 505 }
 506
 507 /*
 508  * Called by the outgoing CPU to report its successful death.  Return
 509  * false if this report follows the surviving CPU's timing out.
 510  *
 511  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
 512  * timed out.  This approach allows architectures to omit calls to
 513  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
 514  * the next cpu_wait_death()'s polling loop.
 515  */
 516 bool cpu_report_death(void)
 517 {
 518         int oldstate;
 519         int newstate;
 520         int cpu = smp_processor_id();
 521
 522         do {
 523                 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 524                 if (oldstate != CPU_BROKEN)
 525                         newstate = CPU_DEAD;
 526                 else
 527                         newstate = CPU_DEAD_FROZEN;
 528         } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 529                                 oldstate, newstate) != oldstate);
 530         return newstate == CPU_DEAD;
 531 }
 532
 533 #endif /* #ifdef CONFIG_HOTPLUG_CPU */