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