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                         if (ht->cleanup)
 117                                 ht->cleanup(td->cpu, cpu_online(td->cpu));
 118                         kfree(td);
 119                         return 0;
 120                 }
 121
 122                 if (kthread_should_park()) {
 123                         __set_current_state(TASK_RUNNING);
 124                         preempt_enable();
 125                         if (ht->park && td->status == HP_THREAD_ACTIVE) {
 126                                 BUG_ON(td->cpu != smp_processor_id());
 127                                 ht->park(td->cpu);
 128                                 td->status = HP_THREAD_PARKED;
 129                         }
 130                         kthread_parkme();
 131                         /* We might have been woken for stop */
 132                         continue;
 133                 }
 134
 135                 BUG_ON(td->cpu != smp_processor_id());
 136
 137                 /* Check for state change setup */
 138                 switch (td->status) {
 139                 case HP_THREAD_NONE:
 140                         __set_current_state(TASK_RUNNING);
 141                         preempt_enable();
 142                         if (ht->setup)
 143                                 ht->setup(td->cpu);
 144                         td->status = HP_THREAD_ACTIVE;
 145                         continue;
 146
 147                 case HP_THREAD_PARKED:
 148                         __set_current_state(TASK_RUNNING);
 149                         preempt_enable();
 150                         if (ht->unpark)
 151                                 ht->unpark(td->cpu);
 152                         td->status = HP_THREAD_ACTIVE;
 153                         continue;
 154                 }
 155
 156                 if (!ht->thread_should_run(td->cpu)) {
 157                         preempt_enable_no_resched();
 158                         schedule();
 159                 } else {
 160                         __set_current_state(TASK_RUNNING);
 161                         preempt_enable();
 162                         ht->thread_fn(td->cpu);
 163                 }
 164         }
 165 }
 166
 167 static int
 168 __smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 169 {
 170         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 171         struct smpboot_thread_data *td;
 172
 173         if (tsk)
 174                 return 0;
 175
 176         td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
 177         if (!td)
 178                 return -ENOMEM;
 179         td->cpu = cpu;
 180         td->ht = ht;
 181
 182         tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
 183                                     ht->thread_comm);
 184         if (IS_ERR(tsk)) {
 185                 kfree(td);
 186                 return PTR_ERR(tsk);
 187         }
 188         get_task_struct(tsk);
 189         *per_cpu_ptr(ht->store, cpu) = tsk;
 190         if (ht->create) {
 191                 /*
 192                  * Make sure that the task has actually scheduled out
 193                  * into park position, before calling the create
 194                  * callback. At least the migration thread callback
 195                  * requires that the task is off the runqueue.
 196                  */
 197                 if (!wait_task_inactive(tsk, TASK_PARKED))
 198                         WARN_ON(1);
 199                 else
 200                         ht->create(cpu);
 201         }
 202         return 0;
 203 }
 204
 205 int smpboot_create_threads(unsigned int cpu)
 206 {
 207         struct smp_hotplug_thread *cur;
 208         int ret = 0;
 209
 210         mutex_lock(&smpboot_threads_lock);
 211         list_for_each_entry(cur, &hotplug_threads, list) {
 212                 ret = __smpboot_create_thread(cur, cpu);
 213                 if (ret)
 214                         break;
 215         }
 216         mutex_unlock(&smpboot_threads_lock);
 217         return ret;
 218 }
 219
 220 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 221 {
 222         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 223
 224         if (ht->pre_unpark)
 225                 ht->pre_unpark(cpu);
 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         /* Unpark any threads that were voluntarily parked. */
 263         for_each_cpu_not(cpu, ht->cpumask) {
 264                 if (cpu_online(cpu)) {
 265                         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 266                         if (tsk)
 267                                 kthread_unpark(tsk);
 268                 }
 269         }
 270
 271         /* We need to destroy also the parked threads of offline cpus */
 272         for_each_possible_cpu(cpu) {
 273                 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 274
 275                 if (tsk) {
 276                         kthread_stop(tsk);
 277                         put_task_struct(tsk);
 278                         *per_cpu_ptr(ht->store, cpu) = NULL;
 279                 }
 280         }
 281 }
 282
 283 /**
 284  * smpboot_register_percpu_thread - Register a per_cpu thread related to hotplug
 285  * @plug_thread:        Hotplug thread descriptor
 286  *
 287  * Creates and starts the threads on all online cpus.
 288  */
 289 int smpboot_register_percpu_thread(struct smp_hotplug_thread *plug_thread)
 290 {
 291         unsigned int cpu;
 292         int ret = 0;
 293
 294         if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
 295                 return -ENOMEM;
 296         cpumask_copy(plug_thread->cpumask, cpu_possible_mask);
 297
 298         get_online_cpus();
 299         mutex_lock(&smpboot_threads_lock);
 300         for_each_online_cpu(cpu) {
 301                 ret = __smpboot_create_thread(plug_thread, cpu);
 302                 if (ret) {
 303                         smpboot_destroy_threads(plug_thread);
 304                         goto out;
 305                 }
 306                 smpboot_unpark_thread(plug_thread, cpu);
 307         }
 308         list_add(&plug_thread->list, &hotplug_threads);
 309 out:
 310         mutex_unlock(&smpboot_threads_lock);
 311         put_online_cpus();
 312         return ret;
 313 }
 314 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
 315
 316 /**
 317  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 318  * @plug_thread:        Hotplug thread descriptor
 319  *
 320  * Stops all threads on all possible cpus.
 321  */
 322 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 323 {
 324         get_online_cpus();
 325         mutex_lock(&smpboot_threads_lock);
 326         list_del(&plug_thread->list);
 327         smpboot_destroy_threads(plug_thread);
 328         mutex_unlock(&smpboot_threads_lock);
 329         put_online_cpus();
 330         free_cpumask_var(plug_thread->cpumask);
 331 }
 332 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 333
 334 /**
 335  * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
 336  * @plug_thread:        Hotplug thread descriptor
 337  * @new:                Revised mask to use
 338  *
 339  * The cpumask field in the smp_hotplug_thread must not be updated directly
 340  * by the client, but only by calling this function.
 341  */
 342 int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
 343                                          const struct cpumask *new)
 344 {
 345         struct cpumask *old = plug_thread->cpumask;
 346         cpumask_var_t tmp;
 347         unsigned int cpu;
 348
 349         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 350                 return -ENOMEM;
 351
 352         get_online_cpus();
 353         mutex_lock(&smpboot_threads_lock);
 354
 355         /* Park threads that were exclusively enabled on the old mask. */
 356         cpumask_andnot(tmp, old, new);
 357         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 358                 smpboot_park_thread(plug_thread, cpu);
 359
 360         /* Unpark threads that are exclusively enabled on the new mask. */
 361         cpumask_andnot(tmp, new, old);
 362         for_each_cpu_and(cpu, tmp, cpu_online_mask)
 363                 smpboot_unpark_thread(plug_thread, cpu);
 364
 365         cpumask_copy(old, new);
 366
 367         mutex_unlock(&smpboot_threads_lock);
 368         put_online_cpus();
 369
 370         free_cpumask_var(tmp);
 371
 372         return 0;
 373 }
 374 EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread);
 375
 376 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
 377
 378 /*
 379  * Called to poll specified CPU's state, for example, when waiting for
 380  * a CPU to come online.
 381  */
 382 int cpu_report_state(int cpu)
 383 {
 384         return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 385 }
 386
 387 /*
 388  * If CPU has died properly, set its state to CPU_UP_PREPARE and
 389  * return success.  Otherwise, return -EBUSY if the CPU died after
 390  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
 391  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
 392  * to dying.  In the latter two cases, the CPU might not be set up
 393  * properly, but it is up to the arch-specific code to decide.
 394  * Finally, -EIO indicates an unanticipated problem.
 395  *
 396  * Note that it is permissible to omit this call entirely, as is
 397  * done in architectures that do no CPU-hotplug error checking.
 398  */
 399 int cpu_check_up_prepare(int cpu)
 400 {
 401         if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
 402                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 403                 return 0;
 404         }
 405
 406         switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
 407
 408         case CPU_POST_DEAD:
 409
 410                 /* The CPU died properly, so just start it up again. */
 411                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 412                 return 0;
 413
 414         case CPU_DEAD_FROZEN:
 415
 416                 /*
 417                  * Timeout during CPU death, so let caller know.
 418                  * The outgoing CPU completed its processing, but after
 419                  * cpu_wait_death() timed out and reported the error. The
 420                  * caller is free to proceed, in which case the state
 421                  * will be reset properly by cpu_set_state_online().
 422                  * Proceeding despite this -EBUSY return makes sense
 423                  * for systems where the outgoing CPUs take themselves
 424                  * offline, with no post-death manipulation required from
 425                  * a surviving CPU.
 426                  */
 427                 return -EBUSY;
 428
 429         case CPU_BROKEN:
 430
 431                 /*
 432                  * The most likely reason we got here is that there was
 433                  * a timeout during CPU death, and the outgoing CPU never
 434                  * did complete its processing.  This could happen on
 435                  * a virtualized system if the outgoing VCPU gets preempted
 436                  * for more than five seconds, and the user attempts to
 437                  * immediately online that same CPU.  Trying again later
 438                  * might return -EBUSY above, hence -EAGAIN.
 439                  */
 440                 return -EAGAIN;
 441
 442         default:
 443
 444                 /* Should not happen.  Famous last words. */
 445                 return -EIO;
 446         }
 447 }
 448
 449 /*
 450  * Mark the specified CPU online.
 451  *
 452  * Note that it is permissible to omit this call entirely, as is
 453  * done in architectures that do no CPU-hotplug error checking.
 454  */
 455 void cpu_set_state_online(int cpu)
 456 {
 457         (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
 458 }
 459
 460 #ifdef CONFIG_HOTPLUG_CPU
 461
 462 /*
 463  * Wait for the specified CPU to exit the idle loop and die.
 464  */
 465 bool cpu_wait_death(unsigned int cpu, int seconds)
 466 {
 467         int jf_left = seconds * HZ;
 468         int oldstate;
 469         bool ret = true;
 470         int sleep_jf = 1;
 471
 472         might_sleep();
 473
 474         /* The outgoing CPU will normally get done quite quickly. */
 475         if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
 476                 goto update_state;
 477         udelay(5);
 478
 479         /* But if the outgoing CPU dawdles, wait increasingly long times. */
 480         while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
 481                 schedule_timeout_uninterruptible(sleep_jf);
 482                 jf_left -= sleep_jf;
 483                 if (jf_left <= 0)
 484                         break;
 485                 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
 486         }
 487 update_state:
 488         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 489         if (oldstate == CPU_DEAD) {
 490                 /* Outgoing CPU died normally, update state. */
 491                 smp_mb(); /* atomic_read() before update. */
 492                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
 493         } else {
 494                 /* Outgoing CPU still hasn't died, set state accordingly. */
 495                 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 496                                    oldstate, CPU_BROKEN) != oldstate)
 497                         goto update_state;
 498                 ret = false;
 499         }
 500         return ret;
 501 }
 502
 503 /*
 504  * Called by the outgoing CPU to report its successful death.  Return
 505  * false if this report follows the surviving CPU's timing out.
 506  *
 507  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
 508  * timed out.  This approach allows architectures to omit calls to
 509  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
 510  * the next cpu_wait_death()'s polling loop.
 511  */
 512 bool cpu_report_death(void)
 513 {
 514         int oldstate;
 515         int newstate;
 516         int cpu = smp_processor_id();
 517
 518         do {
 519                 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 520                 if (oldstate != CPU_BROKEN)
 521                         newstate = CPU_DEAD;
 522                 else
 523                         newstate = CPU_DEAD_FROZEN;
 524         } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 525                                 oldstate, newstate) != oldstate);
 526         return newstate == CPU_DEAD;
 527 }
 528
 529 #endif /* #ifdef CONFIG_HOTPLUG_CPU */