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