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         kthread_set_per_cpu(tsk, cpu);
 191         /*
 192          * Park the thread so that it could start right on the CPU
 193          * when it is available.
 194          */
 195         kthread_park(tsk);
 196         get_task_struct(tsk);
 197         *per_cpu_ptr(ht->store, cpu) = tsk;
 198         if (ht->create) {
 199                 /*
 200                  * Make sure that the task has actually scheduled out
 201                  * into park position, before calling the create
 202                  * callback. At least the migration thread callback
 203                  * requires that the task is off the runqueue.
 204                  */
 205                 if (!wait_task_inactive(tsk, TASK_PARKED))
 206                         WARN_ON(1);
 207                 else
 208                         ht->create(cpu);
 209         }
 210         return 0;
 211 }
 212
 213 int smpboot_create_threads(unsigned int cpu)
 214 {
 215         struct smp_hotplug_thread *cur;
 216         int ret = 0;
 217
 218         mutex_lock(&smpboot_threads_lock);
 219         list_for_each_entry(cur, &hotplug_threads, list) {
 220                 ret = __smpboot_create_thread(cur, cpu);
 221                 if (ret)
 222                         break;
 223         }
 224         mutex_unlock(&smpboot_threads_lock);
 225         return ret;
 226 }
 227
 228 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 229 {
 230         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 231
 232         if (!ht->selfparking)
 233                 kthread_unpark(tsk);
 234 }
 235
 236 int smpboot_unpark_threads(unsigned int cpu)
 237 {
 238         struct smp_hotplug_thread *cur;
 239
 240         mutex_lock(&smpboot_threads_lock);
 241         list_for_each_entry(cur, &hotplug_threads, list)
 242                 if (cpumask_test_cpu(cpu, cur->cpumask))
 243                         smpboot_unpark_thread(cur, cpu);
 244         mutex_unlock(&smpboot_threads_lock);
 245         return 0;
 246 }
 247
 248 static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 249 {
 250         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 251
 252         if (tsk && !ht->selfparking)
 253                 kthread_park(tsk);
 254 }
 255
 256 int smpboot_park_threads(unsigned int cpu)
 257 {
 258         struct smp_hotplug_thread *cur;
 259
 260         mutex_lock(&smpboot_threads_lock);
 261         list_for_each_entry_reverse(cur, &hotplug_threads, list)
 262                 smpboot_park_thread(cur, cpu);
 263         mutex_unlock(&smpboot_threads_lock);
 264         return 0;
 265 }
 266
 267 static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
 268 {
 269         unsigned int cpu;
 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_cpumask - Register a per_cpu thread related
 285  *                                          to hotplug
 286  * @plug_thread:        Hotplug thread descriptor
 287  * @cpumask:            The cpumask where threads run
 288  *
 289  * Creates and starts the threads on all online cpus.
 290  */
 291 int smpboot_register_percpu_thread_cpumask(struct smp_hotplug_thread *plug_thread,
 292                                            const struct cpumask *cpumask)
 293 {
 294         unsigned int cpu;
 295         int ret = 0;
 296
 297         if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
 298                 return -ENOMEM;
 299         cpumask_copy(plug_thread->cpumask, cpumask);
 300
 301         get_online_cpus();
 302         mutex_lock(&smpboot_threads_lock);
 303         for_each_online_cpu(cpu) {
 304                 ret = __smpboot_create_thread(plug_thread, cpu);
 305                 if (ret) {
 306                         smpboot_destroy_threads(plug_thread);
 307                         free_cpumask_var(plug_thread->cpumask);
 308                         goto out;
 309                 }
 310                 if (cpumask_test_cpu(cpu, cpumask))
 311                         smpboot_unpark_thread(plug_thread, cpu);
 312         }
 313         list_add(&plug_thread->list, &hotplug_threads);
 314 out:
 315         mutex_unlock(&smpboot_threads_lock);
 316         put_online_cpus();
 317         return ret;
 318 }
 319 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread_cpumask);
 320
 321 /**
 322  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 323  * @plug_thread:        Hotplug thread descriptor
 324  *
 325  * Stops all threads on all possible cpus.
 326  */
 327 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 328 {
 329         get_online_cpus();
 330         mutex_lock(&smpboot_threads_lock);
 331         list_del(&plug_thread->list);
 332         smpboot_destroy_threads(plug_thread);
 333         mutex_unlock(&smpboot_threads_lock);
 334         put_online_cpus();
 335         free_cpumask_var(plug_thread->cpumask);
 336 }
 337 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 338
 339 /**
 340  * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
 341  * @plug_thread:        Hotplug thread descriptor
 342  * @new:                Revised mask to use
 343  *
 344  * The cpumask field in the smp_hotplug_thread must not be updated directly
 345  * by the client, but only by calling this function.
 346  * This function can only be called on a registered smp_hotplug_thread.
 347  */
 348 void smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
 349                                           const struct cpumask *new)
 350 {
 351         struct cpumask *old = plug_thread->cpumask;
 352         static struct cpumask tmp;
 353         unsigned int cpu;
 354
 355         lockdep_assert_cpus_held();
 356         mutex_lock(&smpboot_threads_lock);
 357
 358         /* Park threads that were exclusively enabled on the old mask. */
 359         cpumask_andnot(&tmp, old, new);
 360         for_each_cpu_and(cpu, &tmp, cpu_online_mask)
 361                 smpboot_park_thread(plug_thread, cpu);
 362
 363         /* Unpark threads that are exclusively enabled on the new mask. */
 364         cpumask_andnot(&tmp, new, old);
 365         for_each_cpu_and(cpu, &tmp, cpu_online_mask)
 366                 smpboot_unpark_thread(plug_thread, cpu);
 367
 368         cpumask_copy(old, new);
 369
 370         mutex_unlock(&smpboot_threads_lock);
 371 }
 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 */