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