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         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 __always_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                 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 - Register a per_cpu thread related
 284  *                                          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         cpus_read_lock();
 295         mutex_lock(&smpboot_threads_lock);
 296         for_each_online_cpu(cpu) {
 297                 ret = __smpboot_create_thread(plug_thread, cpu);
 298                 if (ret) {
 299                         smpboot_destroy_threads(plug_thread);
 300                         goto out;
 301                 }
 302                 smpboot_unpark_thread(plug_thread, cpu);
 303         }
 304         list_add(&plug_thread->list, &hotplug_threads);
 305 out:
 306         mutex_unlock(&smpboot_threads_lock);
 307         cpus_read_unlock();
 308         return ret;
 309 }
 310 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
 311
 312 /**
 313  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
 314  * @plug_thread:        Hotplug thread descriptor
 315  *
 316  * Stops all threads on all possible cpus.
 317  */
 318 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
 319 {
 320         cpus_read_lock();
 321         mutex_lock(&smpboot_threads_lock);
 322         list_del(&plug_thread->list);
 323         smpboot_destroy_threads(plug_thread);
 324         mutex_unlock(&smpboot_threads_lock);
 325         cpus_read_unlock();
 326 }
 327 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
 328
 329 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
 330
 331 /*
 332  * Called to poll specified CPU's state, for example, when waiting for
 333  * a CPU to come online.
 334  */
 335 int cpu_report_state(int cpu)
 336 {
 337         return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 338 }
 339
 340 /*
 341  * If CPU has died properly, set its state to CPU_UP_PREPARE and
 342  * return success.  Otherwise, return -EBUSY if the CPU died after
 343  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
 344  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
 345  * to dying.  In the latter two cases, the CPU might not be set up
 346  * properly, but it is up to the arch-specific code to decide.
 347  * Finally, -EIO indicates an unanticipated problem.
 348  *
 349  * Note that it is permissible to omit this call entirely, as is
 350  * done in architectures that do no CPU-hotplug error checking.
 351  */
 352 int cpu_check_up_prepare(int cpu)
 353 {
 354         if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
 355                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 356                 return 0;
 357         }
 358
 359         switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
 360
 361         case CPU_POST_DEAD:
 362
 363                 /* The CPU died properly, so just start it up again. */
 364                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
 365                 return 0;
 366
 367         case CPU_DEAD_FROZEN:
 368
 369                 /*
 370                  * Timeout during CPU death, so let caller know.
 371                  * The outgoing CPU completed its processing, but after
 372                  * cpu_wait_death() timed out and reported the error. The
 373                  * caller is free to proceed, in which case the state
 374                  * will be reset properly by cpu_set_state_online().
 375                  * Proceeding despite this -EBUSY return makes sense
 376                  * for systems where the outgoing CPUs take themselves
 377                  * offline, with no post-death manipulation required from
 378                  * a surviving CPU.
 379                  */
 380                 return -EBUSY;
 381
 382         case CPU_BROKEN:
 383
 384                 /*
 385                  * The most likely reason we got here is that there was
 386                  * a timeout during CPU death, and the outgoing CPU never
 387                  * did complete its processing.  This could happen on
 388                  * a virtualized system if the outgoing VCPU gets preempted
 389                  * for more than five seconds, and the user attempts to
 390                  * immediately online that same CPU.  Trying again later
 391                  * might return -EBUSY above, hence -EAGAIN.
 392                  */
 393                 return -EAGAIN;
 394
 395         case CPU_UP_PREPARE:
 396                 /*
 397                  * Timeout while waiting for the CPU to show up. Allow to try
 398                  * again later.
 399                  */
 400                 return 0;
 401
 402         default:
 403
 404                 /* Should not happen.  Famous last words. */
 405                 return -EIO;
 406         }
 407 }
 408
 409 /*
 410  * Mark the specified CPU online.
 411  *
 412  * Note that it is permissible to omit this call entirely, as is
 413  * done in architectures that do no CPU-hotplug error checking.
 414  */
 415 void cpu_set_state_online(int cpu)
 416 {
 417         (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
 418 }
 419
 420 #ifdef CONFIG_HOTPLUG_CPU
 421
 422 /*
 423  * Wait for the specified CPU to exit the idle loop and die.
 424  */
 425 bool cpu_wait_death(unsigned int cpu, int seconds)
 426 {
 427         int jf_left = seconds * HZ;
 428         int oldstate;
 429         bool ret = true;
 430         int sleep_jf = 1;
 431
 432         might_sleep();
 433
 434         /* The outgoing CPU will normally get done quite quickly. */
 435         if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
 436                 goto update_state_early;
 437         udelay(5);
 438
 439         /* But if the outgoing CPU dawdles, wait increasingly long times. */
 440         while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
 441                 schedule_timeout_uninterruptible(sleep_jf);
 442                 jf_left -= sleep_jf;
 443                 if (jf_left <= 0)
 444                         break;
 445                 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
 446         }
 447 update_state_early:
 448         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 449 update_state:
 450         if (oldstate == CPU_DEAD) {
 451                 /* Outgoing CPU died normally, update state. */
 452                 smp_mb(); /* atomic_read() before update. */
 453                 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
 454         } else {
 455                 /* Outgoing CPU still hasn't died, set state accordingly. */
 456                 if (!atomic_try_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 457                                         &oldstate, CPU_BROKEN))
 458                         goto update_state;
 459                 ret = false;
 460         }
 461         return ret;
 462 }
 463
 464 /*
 465  * Called by the outgoing CPU to report its successful death.  Return
 466  * false if this report follows the surviving CPU's timing out.
 467  *
 468  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
 469  * timed out.  This approach allows architectures to omit calls to
 470  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
 471  * the next cpu_wait_death()'s polling loop.
 472  */
 473 bool cpu_report_death(void)
 474 {
 475         int oldstate;
 476         int newstate;
 477         int cpu = smp_processor_id();
 478
 479         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
 480         do {
 481                 if (oldstate != CPU_BROKEN)
 482                         newstate = CPU_DEAD;
 483                 else
 484                         newstate = CPU_DEAD_FROZEN;
 485         } while (!atomic_try_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
 486                                      &oldstate, newstate));
 487         return newstate == CPU_DEAD;
 488 }
 489
 490 #endif /* #ifdef CONFIG_HOTPLUG_CPU */