kernel/smpboot.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Common SMP CPU bringup/teardown functions
   4  */
   5 #include <linux/cpu.h>
   6 #include <linux/err.h>
   7 #include <linux/smp.h>
   8 #include <linux/delay.h>
   9 #include <linux/init.h>
  10 #include <linux/list.h>
  11 #include <linux/slab.h>
  12 #include <linux/sched.h>
  13 #include <linux/sched/task.h>
  14 #include <linux/export.h>
  15 #include <linux/percpu.h>
  16 #include <linux/kthread.h>
  17 #include <linux/smpboot.h>
  18
  19 #include "smpboot.h"
  20
  21 #ifdef CONFIG_SMP
  22
  23 #ifdef CONFIG_GENERIC_SMP_IDLE_THREAD
  24 /*
  25  * For the hotplug case we keep the task structs around and reuse
  26  * them.
  27  */
  28 static DEFINE_PER_CPU(struct task_struct *, idle_threads);
  29
  30 struct task_struct *idle_thread_get(unsigned int cpu)
  31 {
  32         struct task_struct *tsk = per_cpu(idle_threads, cpu);
  33
  34         if (!tsk)
  35                 return ERR_PTR(-ENOMEM);
  36         init_idle(tsk, cpu);
  37         return tsk;
  38 }
  39
  40 void __init idle_thread_set_boot_cpu(void)
  41 {
  42         per_cpu(idle_threads, smp_processor_id()) = current;
  43 }
  44
  45 /**
  46  * idle_init - Initialize the idle thread for a cpu
  47  * @cpu:        The cpu for which the idle thread should be initialized
  48  *
  49  * Creates the thread if it does not exist.
  50  */
  51 static inline void idle_init(unsigned int cpu)
  52 {
  53         struct task_struct *tsk = per_cpu(idle_threads, cpu);
  54
  55         if (!tsk) {
  56                 tsk = fork_idle(cpu);
  57                 if (IS_ERR(tsk))
  58                         pr_err("SMP: fork_idle() failed for CPU %u\n", cpu);
  59                 else
  60                         per_cpu(idle_threads, cpu) = tsk;
  61         }
  62 }
  63
  64 /**
  65  * idle_threads_init - Initialize idle threads for all cpus
  66  */
  67 void __init idle_threads_init(void)
  68 {
  69         unsigned int cpu, boot_cpu;
  70
  71         boot_cpu = smp_processor_id();
  72
  73         for_each_possible_cpu(cpu) {
  74                 if (cpu != boot_cpu)
  75                         idle_init(cpu);
  76         }
  77 }
  78 #endif
  79
  80 #endif /* #ifdef CONFIG_SMP */
  81
  82 static LIST_HEAD(hotplug_threads);
  83 static DEFINE_MUTEX(smpboot_threads_lock);
  84
  85 struct smpboot_thread_data {
  86         unsigned int                    cpu;
  87         unsigned int                    status;
  88         struct smp_hotplug_thread       *ht;
  89 };
  90
  91 enum {
  92         HP_THREAD_NONE = 0,
  93         HP_THREAD_ACTIVE,
  94         HP_THREAD_PARKED,
  95 };
  96
  97 /**
  98  * smpboot_thread_fn - percpu hotplug thread loop function
  99  * @data:       thread data pointer
 100  *
 101  * Checks for thread stop and park conditions. Calls the necessary
 102  * setup, cleanup, park and unpark functions for the registered
 103  * thread.
 104  *
 105  * Returns 1 when the thread should exit, 0 otherwise.
 106  */
 107 static int smpboot_thread_fn(void *data)
 108 {
 109         struct smpboot_thread_data *td = data;
 110         struct smp_hotplug_thread *ht = td->ht;
 111
 112         while (1) {
 113                 set_current_state(TASK_INTERRUPTIBLE);
 114                 preempt_disable();
 115                 if (kthread_should_stop()) {
 116                         __set_current_state(TASK_RUNNING);
 117                         preempt_enable();
 118                         /* cleanup must mirror setup */
 119                         if (ht->cleanup && td->status != HP_THREAD_NONE)
 120                                 ht->cleanup(td->cpu, cpu_online(td->cpu));
 121                         kfree(td);
 122                         return 0;
 123                 }
 124
 125                 if (kthread_should_park()) {
 126                         __set_current_state(TASK_RUNNING);
 127                         preempt_enable();
 128                         if (ht->park && td->status == HP_THREAD_ACTIVE) {
 129                                 BUG_ON(td->cpu != smp_processor_id());
 130                                 ht->park(td->cpu);
 131                                 td->status = HP_THREAD_PARKED;
 132                         }
 133                         kthread_parkme();
 134                         /* We might have been woken for stop */
 135                         continue;
 136                 }
 137
 138                 BUG_ON(td->cpu != smp_processor_id());
 139
 140                 /* Check for state change setup */
 141                 switch (td->status) {
 142                 case HP_THREAD_NONE:
 143                         __set_current_state(TASK_RUNNING);
 144                         preempt_enable();
 145                         if (ht->setup)
 146                                 ht->setup(td->cpu);
 147                         td->status = HP_THREAD_ACTIVE;
 148                         continue;
 149
 150                 case HP_THREAD_PARKED:
 151                         __set_current_state(TASK_RUNNING);
 152                         preempt_enable();
 153                         if (ht->unpark)
 154                                 ht->unpark(td->cpu);
 155                         td->status = HP_THREAD_ACTIVE;
 156                         continue;
 157                 }
 158
 159                 if (!ht->thread_should_run(td->cpu)) {
 160                         preempt_enable_no_resched();
 161                         schedule();
 162                 } else {
 163                         __set_current_state(TASK_RUNNING);
 164                         preempt_enable();
 165                         ht->thread_fn(td->cpu);
 166                 }
 167         }
 168 }
 169
 170 static int
 171 __smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 172 {
 173         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 174         struct smpboot_thread_data *td;
 175
 176         if (tsk)
 177                 return 0;
 178
 179         td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
 180         if (!td)
 181                 return -ENOMEM;
 182         td->cpu = cpu;
 183         td->ht = ht;
 184
 185         tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
 186                                     ht->thread_comm);
 187         if (IS_ERR(tsk)) {
 188                 kfree(td);
 189                 return PTR_ERR(tsk);
 190         }
 191         kthread_set_per_cpu(tsk, cpu);
 192         /*
 193          * Park the thread so that it could start right on the CPU
 194          * when it is available.
 195          */
 196         kthread_park(tsk);
 197         get_task_struct(tsk);
 198         *per_cpu_ptr(ht->store, cpu) = tsk;
 199         if (ht->create) {
 200                 /*
 201                  * Make sure that the task has actually scheduled out
 202                  * into park position, before calling the create
 203                  * callback. At least the migration thread callback
 204                  * requires that the task is off the runqueue.
 205                  */
 206                 if (!wait_task_inactive(tsk, TASK_PARKED))
 207                         WARN_ON(1);
 208                 else
 209                         ht->create(cpu);
 210         }
 211         return 0;
 212 }
 213
 214 int smpboot_create_threads(unsigned int cpu)
 215 {
 216         struct smp_hotplug_thread *cur;
 217         int ret = 0;
 218
 219         mutex_lock(&smpboot_threads_lock);
 220         list_for_each_entry(cur, &hotplug_threads, list) {
 221                 ret = __smpboot_create_thread(cur, cpu);
 222                 if (ret)
 223                         break;
 224         }
 225         mutex_unlock(&smpboot_threads_lock);
 226         return ret;
 227 }
 228
 229 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
 230 {
 231         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
 232
 233         if (!ht->selfparking)
 234                 kthread_unpark(tsk);
 235 }
 236
 237 int smpboot_unpark_threads(unsigned int cpu)
 238 {
 239         struct smp_hotplug_thread *cur;
 240
 241         mutex_lock(&smpboot_threads_lock);
 242         list_for_each_entry(cur, &hotplug_threads, list)
 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 - Register a per_cpu thread related
 285  *                                          to hotplug
 286  * @plug_thread:        Hotplug thread descriptor
 287  *
 288  * Creates and starts the threads on all online cpus.
 289  */
 290 int smpboot_register_percpu_thread(struct smp_hotplug_thread *plug_thread)
 291 {
 292         unsigned int cpu;
 293         int ret = 0;
 294
 295         get_online_cpus();
 296         mutex_lock(&smpboot_threads_lock);
 297         for_each_online_cpu(cpu) {
 298                 ret = __smpboot_create_thread(plug_thread, cpu);
 299                 if (ret) {
 300                         smpboot_destroy_threads(plug_thread);
 301                         goto out;
 302                 }
 303                 smpboot_unpark_thread(plug_thread, cpu);
 304         }
 305         list_add(&plug_thread->list, &hotplug_threads);
 306 out:
 307         mutex_unlock(&smpboot_threads_lock);
 308         put_online_cpus();
 309         return ret;
 310 }
 311 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
 312
 313 /**
 314  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 315  * @plug_thread:        Hotplug thread descriptor
 316  *
 317  * Stops all threads on all possible cpus.
 318  */
 319 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 320 {
 321         get_online_cpus();
 322         mutex_lock(&smpboot_threads_lock);
 323         list_del(&plug_thread->list);
 324         smpboot_destroy_threads(plug_thread);
 325         mutex_unlock(&smpboot_threads_lock);
 326         put_online_cpus();
 327 }
 328 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 329
 330 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
 331
 332 /*
 333  * Called to poll specified CPU's state, for example, when waiting for
 334  * a CPU to come online.
 335  */
 336 int cpu_report_state(int cpu)
 337 {
 338         return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 339 }
 340
 341 /*
 342  * If CPU has died properly, set its state to CPU_UP_PREPARE and
 343  * return success.  Otherwise, return -EBUSY if the CPU died after
 344  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
 345  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
 346  * to dying.  In the latter two cases, the CPU might not be set up
 347  * properly, but it is up to the arch-specific code to decide.
 348  * Finally, -EIO indicates an unanticipated problem.
 349  *
 350  * Note that it is permissible to omit this call entirely, as is
 351  * done in architectures that do no CPU-hotplug error checking.
 352  */
 353 int cpu_check_up_prepare(int cpu)
 354 {
 355         if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
 356                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 357                 return 0;
 358         }
 359
 360         switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
 361
 362         case CPU_POST_DEAD:
 363
 364                 /* The CPU died properly, so just start it up again. */
 365                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 366                 return 0;
 367
 368         case CPU_DEAD_FROZEN:
 369
 370                 /*
 371                  * Timeout during CPU death, so let caller know.
 372                  * The outgoing CPU completed its processing, but after
 373                  * cpu_wait_death() timed out and reported the error. The
 374                  * caller is free to proceed, in which case the state
 375                  * will be reset properly by cpu_set_state_online().
 376                  * Proceeding despite this -EBUSY return makes sense
 377                  * for systems where the outgoing CPUs take themselves
 378                  * offline, with no post-death manipulation required from
 379                  * a surviving CPU.
 380                  */
 381                 return -EBUSY;
 382
 383         case CPU_BROKEN:
 384
 385                 /*
 386                  * The most likely reason we got here is that there was
 387                  * a timeout during CPU death, and the outgoing CPU never
 388                  * did complete its processing.  This could happen on
 389                  * a virtualized system if the outgoing VCPU gets preempted
 390                  * for more than five seconds, and the user attempts to
 391                  * immediately online that same CPU.  Trying again later
 392                  * might return -EBUSY above, hence -EAGAIN.
 393                  */
 394                 return -EAGAIN;
 395
 396         default:
 397
 398                 /* Should not happen.  Famous last words. */
 399                 return -EIO;
 400         }
 401 }
 402
 403 /*
 404  * Mark the specified CPU online.
 405  *
 406  * Note that it is permissible to omit this call entirely, as is
 407  * done in architectures that do no CPU-hotplug error checking.
 408  */
 409 void cpu_set_state_online(int cpu)
 410 {
 411         (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
 412 }
 413
 414 #ifdef CONFIG_HOTPLUG_CPU
 415
 416 /*
 417  * Wait for the specified CPU to exit the idle loop and die.
 418  */
 419 bool cpu_wait_death(unsigned int cpu, int seconds)
 420 {
 421         int jf_left = seconds * HZ;
 422         int oldstate;
 423         bool ret = true;
 424         int sleep_jf = 1;
 425
 426         might_sleep();
 427
 428         /* The outgoing CPU will normally get done quite quickly. */
 429         if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
 430                 goto update_state;
 431         udelay(5);
 432
 433         /* But if the outgoing CPU dawdles, wait increasingly long times. */
 434         while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
 435                 schedule_timeout_uninterruptible(sleep_jf);
 436                 jf_left -= sleep_jf;
 437                 if (jf_left <= 0)
 438                         break;
 439                 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
 440         }
 441 update_state:
 442         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 443         if (oldstate == CPU_DEAD) {
 444                 /* Outgoing CPU died normally, update state. */
 445                 smp_mb(); /* atomic_read() before update. */
 446                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
 447         } else {
 448                 /* Outgoing CPU still hasn't died, set state accordingly. */
 449                 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 450                                    oldstate, CPU_BROKEN) != oldstate)
 451                         goto update_state;
 452                 ret = false;
 453         }
 454         return ret;
 455 }
 456
 457 /*
 458  * Called by the outgoing CPU to report its successful death.  Return
 459  * false if this report follows the surviving CPU's timing out.
 460  *
 461  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
 462  * timed out.  This approach allows architectures to omit calls to
 463  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
 464  * the next cpu_wait_death()'s polling loop.
 465  */
 466 bool cpu_report_death(void)
 467 {
 468         int oldstate;
 469         int newstate;
 470         int cpu = smp_processor_id();
 471
 472         do {
 473                 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 474                 if (oldstate != CPU_BROKEN)
 475                         newstate = CPU_DEAD;
 476                 else
 477                         newstate = CPU_DEAD_FROZEN;
 478         } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 479                                 oldstate, newstate) != oldstate);
 480         return newstate == CPU_DEAD;
 481 }
 482
 483 #endif /* #ifdef CONFIG_HOTPLUG_CPU */