the-tree/kernel/smpboot.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Common SMP CPU bringup/teardown functions
0004  */
0005 #include <linux/cpu.h>
0006 #include <linux/err.h>
0007 #include <linux/smp.h>
0008 #include <linux/delay.h>
0009 #include <linux/init.h>
0010 #include <linux/list.h>
0011 #include <linux/slab.h>
0012 #include <linux/sched.h>
0013 #include <linux/sched/task.h>
0014 #include <linux/export.h>
0015 #include <linux/percpu.h>
0016 #include <linux/kthread.h>
0017 #include <linux/smpboot.h>
0018
0019 #include "smpboot.h"
0020
0021 #ifdef CONFIG_SMP
0022
0023 #ifdef CONFIG_GENERIC_SMP_IDLE_THREAD
0024 /*
0025  * For the hotplug case we keep the task structs around and reuse
0026  * them.
0027  */
0028 static DEFINE_PER_CPU(struct task_struct *, idle_threads);
0029
0030 struct task_struct *idle_thread_get(unsigned int cpu)
0031 {
0032     struct task_struct *tsk = per_cpu(idle_threads, cpu);
0033
0034     if (!tsk)
0035         return ERR_PTR(-ENOMEM);
0036     return tsk;
0037 }
0038
0039 void __init idle_thread_set_boot_cpu(void)
0040 {
0041     per_cpu(idle_threads, smp_processor_id()) = current;
0042 }
0043
0044 /**
0045  * idle_init - Initialize the idle thread for a cpu
0046  * @cpu:    The cpu for which the idle thread should be initialized
0047  *
0048  * Creates the thread if it does not exist.
0049  */
0050 static __always_inline void idle_init(unsigned int cpu)
0051 {
0052     struct task_struct *tsk = per_cpu(idle_threads, cpu);
0053
0054     if (!tsk) {
0055         tsk = fork_idle(cpu);
0056         if (IS_ERR(tsk))
0057             pr_err("SMP: fork_idle() failed for CPU %u\n", cpu);
0058         else
0059             per_cpu(idle_threads, cpu) = tsk;
0060     }
0061 }
0062
0063 /**
0064  * idle_threads_init - Initialize idle threads for all cpus
0065  */
0066 void __init idle_threads_init(void)
0067 {
0068     unsigned int cpu, boot_cpu;
0069
0070     boot_cpu = smp_processor_id();
0071
0072     for_each_possible_cpu(cpu) {
0073         if (cpu != boot_cpu)
0074             idle_init(cpu);
0075     }
0076 }
0077 #endif
0078
0079 #endif /* #ifdef CONFIG_SMP */
0080
0081 static LIST_HEAD(hotplug_threads);
0082 static DEFINE_MUTEX(smpboot_threads_lock);
0083
0084 struct smpboot_thread_data {
0085     unsigned int            cpu;
0086     unsigned int            status;
0087     struct smp_hotplug_thread   *ht;
0088 };
0089
0090 enum {
0091     HP_THREAD_NONE = 0,
0092     HP_THREAD_ACTIVE,
0093     HP_THREAD_PARKED,
0094 };
0095
0096 /**
0097  * smpboot_thread_fn - percpu hotplug thread loop function
0098  * @data:   thread data pointer
0099  *
0100  * Checks for thread stop and park conditions. Calls the necessary
0101  * setup, cleanup, park and unpark functions for the registered
0102  * thread.
0103  *
0104  * Returns 1 when the thread should exit, 0 otherwise.
0105  */
0106 static int smpboot_thread_fn(void *data)
0107 {
0108     struct smpboot_thread_data *td = data;
0109     struct smp_hotplug_thread *ht = td->ht;
0110
0111     while (1) {
0112         set_current_state(TASK_INTERRUPTIBLE);
0113         preempt_disable();
0114         if (kthread_should_stop()) {
0115             __set_current_state(TASK_RUNNING);
0116             preempt_enable();
0117             /* cleanup must mirror setup */
0118             if (ht->cleanup && td->status != HP_THREAD_NONE)
0119                 ht->cleanup(td->cpu, cpu_online(td->cpu));
0120             kfree(td);
0121             return 0;
0122         }
0123
0124         if (kthread_should_park()) {
0125             __set_current_state(TASK_RUNNING);
0126             preempt_enable();
0127             if (ht->park && td->status == HP_THREAD_ACTIVE) {
0128                 BUG_ON(td->cpu != smp_processor_id());
0129                 ht->park(td->cpu);
0130                 td->status = HP_THREAD_PARKED;
0131             }
0132             kthread_parkme();
0133             /* We might have been woken for stop */
0134             continue;
0135         }
0136
0137         BUG_ON(td->cpu != smp_processor_id());
0138
0139         /* Check for state change setup */
0140         switch (td->status) {
0141         case HP_THREAD_NONE:
0142             __set_current_state(TASK_RUNNING);
0143             preempt_enable();
0144             if (ht->setup)
0145                 ht->setup(td->cpu);
0146             td->status = HP_THREAD_ACTIVE;
0147             continue;
0148
0149         case HP_THREAD_PARKED:
0150             __set_current_state(TASK_RUNNING);
0151             preempt_enable();
0152             if (ht->unpark)
0153                 ht->unpark(td->cpu);
0154             td->status = HP_THREAD_ACTIVE;
0155             continue;
0156         }
0157
0158         if (!ht->thread_should_run(td->cpu)) {
0159             preempt_enable_no_resched();
0160             schedule();
0161         } else {
0162             __set_current_state(TASK_RUNNING);
0163             preempt_enable();
0164             ht->thread_fn(td->cpu);
0165         }
0166     }
0167 }
0168
0169 static int
0170 __smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
0171 {
0172     struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
0173     struct smpboot_thread_data *td;
0174
0175     if (tsk)
0176         return 0;
0177
0178     td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
0179     if (!td)
0180         return -ENOMEM;
0181     td->cpu = cpu;
0182     td->ht = ht;
0183
0184     tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
0185                     ht->thread_comm);
0186     if (IS_ERR(tsk)) {
0187         kfree(td);
0188         return PTR_ERR(tsk);
0189     }
0190     kthread_set_per_cpu(tsk, cpu);
0191     /*
0192      * Park the thread so that it could start right on the CPU
0193      * when it is available.
0194      */
0195     kthread_park(tsk);
0196     get_task_struct(tsk);
0197     *per_cpu_ptr(ht->store, cpu) = tsk;
0198     if (ht->create) {
0199         /*
0200          * Make sure that the task has actually scheduled out
0201          * into park position, before calling the create
0202          * callback. At least the migration thread callback
0203          * requires that the task is off the runqueue.
0204          */
0205         if (!wait_task_inactive(tsk, TASK_PARKED))
0206             WARN_ON(1);
0207         else
0208             ht->create(cpu);
0209     }
0210     return 0;
0211 }
0212
0213 int smpboot_create_threads(unsigned int cpu)
0214 {
0215     struct smp_hotplug_thread *cur;
0216     int ret = 0;
0217
0218     mutex_lock(&smpboot_threads_lock);
0219     list_for_each_entry(cur, &hotplug_threads, list) {
0220         ret = __smpboot_create_thread(cur, cpu);
0221         if (ret)
0222             break;
0223     }
0224     mutex_unlock(&smpboot_threads_lock);
0225     return ret;
0226 }
0227
0228 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
0229 {
0230     struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
0231
0232     if (!ht->selfparking)
0233         kthread_unpark(tsk);
0234 }
0235
0236 int smpboot_unpark_threads(unsigned int cpu)
0237 {
0238     struct smp_hotplug_thread *cur;
0239
0240     mutex_lock(&smpboot_threads_lock);
0241     list_for_each_entry(cur, &hotplug_threads, list)
0242         smpboot_unpark_thread(cur, cpu);
0243     mutex_unlock(&smpboot_threads_lock);
0244     return 0;
0245 }
0246
0247 static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
0248 {
0249     struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
0250
0251     if (tsk && !ht->selfparking)
0252         kthread_park(tsk);
0253 }
0254
0255 int smpboot_park_threads(unsigned int cpu)
0256 {
0257     struct smp_hotplug_thread *cur;
0258
0259     mutex_lock(&smpboot_threads_lock);
0260     list_for_each_entry_reverse(cur, &hotplug_threads, list)
0261         smpboot_park_thread(cur, cpu);
0262     mutex_unlock(&smpboot_threads_lock);
0263     return 0;
0264 }
0265
0266 static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
0267 {
0268     unsigned int cpu;
0269
0270     /* We need to destroy also the parked threads of offline cpus */
0271     for_each_possible_cpu(cpu) {
0272         struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
0273
0274         if (tsk) {
0275             kthread_stop(tsk);
0276             put_task_struct(tsk);
0277             *per_cpu_ptr(ht->store, cpu) = NULL;
0278         }
0279     }
0280 }
0281
0282 /**
0283  * smpboot_register_percpu_thread - Register a per_cpu thread related
0284  *                      to hotplug
0285  * @plug_thread:    Hotplug thread descriptor
0286  *
0287  * Creates and starts the threads on all online cpus.
0288  */
0289 int smpboot_register_percpu_thread(struct smp_hotplug_thread *plug_thread)
0290 {
0291     unsigned int cpu;
0292     int ret = 0;
0293
0294     cpus_read_lock();
0295     mutex_lock(&smpboot_threads_lock);
0296     for_each_online_cpu(cpu) {
0297         ret = __smpboot_create_thread(plug_thread, cpu);
0298         if (ret) {
0299             smpboot_destroy_threads(plug_thread);
0300             goto out;
0301         }
0302         smpboot_unpark_thread(plug_thread, cpu);
0303     }
0304     list_add(&plug_thread->list, &hotplug_threads);
0305 out:
0306     mutex_unlock(&smpboot_threads_lock);
0307     cpus_read_unlock();
0308     return ret;
0309 }
0310 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
0311
0312 /**
0313  * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
0314  * @plug_thread:    Hotplug thread descriptor
0315  *
0316  * Stops all threads on all possible cpus.
0317  */
0318 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
0319 {
0320     cpus_read_lock();
0321     mutex_lock(&smpboot_threads_lock);
0322     list_del(&plug_thread->list);
0323     smpboot_destroy_threads(plug_thread);
0324     mutex_unlock(&smpboot_threads_lock);
0325     cpus_read_unlock();
0326 }
0327 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
0328
0329 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
0330
0331 /*
0332  * Called to poll specified CPU's state, for example, when waiting for
0333  * a CPU to come online.
0334  */
0335 int cpu_report_state(int cpu)
0336 {
0337     return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
0338 }
0339
0340 /*
0341  * If CPU has died properly, set its state to CPU_UP_PREPARE and
0342  * return success.  Otherwise, return -EBUSY if the CPU died after
0343  * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
0344  * if cpu_wait_death() timed out and the CPU still hasn't gotten around
0345  * to dying.  In the latter two cases, the CPU might not be set up
0346  * properly, but it is up to the arch-specific code to decide.
0347  * Finally, -EIO indicates an unanticipated problem.
0348  *
0349  * Note that it is permissible to omit this call entirely, as is
0350  * done in architectures that do no CPU-hotplug error checking.
0351  */
0352 int cpu_check_up_prepare(int cpu)
0353 {
0354     if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
0355         atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
0356         return 0;
0357     }
0358
0359     switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
0360
0361     case CPU_POST_DEAD:
0362
0363         /* The CPU died properly, so just start it up again. */
0364         atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
0365         return 0;
0366
0367     case CPU_DEAD_FROZEN:
0368
0369         /*
0370          * Timeout during CPU death, so let caller know.
0371          * The outgoing CPU completed its processing, but after
0372          * cpu_wait_death() timed out and reported the error. The
0373          * caller is free to proceed, in which case the state
0374          * will be reset properly by cpu_set_state_online().
0375          * Proceeding despite this -EBUSY return makes sense
0376          * for systems where the outgoing CPUs take themselves
0377          * offline, with no post-death manipulation required from
0378          * a surviving CPU.
0379          */
0380         return -EBUSY;
0381
0382     case CPU_BROKEN:
0383
0384         /*
0385          * The most likely reason we got here is that there was
0386          * a timeout during CPU death, and the outgoing CPU never
0387          * did complete its processing.  This could happen on
0388          * a virtualized system if the outgoing VCPU gets preempted
0389          * for more than five seconds, and the user attempts to
0390          * immediately online that same CPU.  Trying again later
0391          * might return -EBUSY above, hence -EAGAIN.
0392          */
0393         return -EAGAIN;
0394
0395     case CPU_UP_PREPARE:
0396         /*
0397          * Timeout while waiting for the CPU to show up. Allow to try
0398          * again later.
0399          */
0400         return 0;
0401
0402     default:
0403
0404         /* Should not happen.  Famous last words. */
0405         return -EIO;
0406     }
0407 }
0408
0409 /*
0410  * Mark the specified CPU online.
0411  *
0412  * Note that it is permissible to omit this call entirely, as is
0413  * done in architectures that do no CPU-hotplug error checking.
0414  */
0415 void cpu_set_state_online(int cpu)
0416 {
0417     (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
0418 }
0419
0420 #ifdef CONFIG_HOTPLUG_CPU
0421
0422 /*
0423  * Wait for the specified CPU to exit the idle loop and die.
0424  */
0425 bool cpu_wait_death(unsigned int cpu, int seconds)
0426 {
0427     int jf_left = seconds * HZ;
0428     int oldstate;
0429     bool ret = true;
0430     int sleep_jf = 1;
0431
0432     might_sleep();
0433
0434     /* The outgoing CPU will normally get done quite quickly. */
0435     if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
0436         goto update_state;
0437     udelay(5);
0438
0439     /* But if the outgoing CPU dawdles, wait increasingly long times. */
0440     while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
0441         schedule_timeout_uninterruptible(sleep_jf);
0442         jf_left -= sleep_jf;
0443         if (jf_left <= 0)
0444             break;
0445         sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
0446     }
0447 update_state:
0448     oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
0449     if (oldstate == CPU_DEAD) {
0450         /* Outgoing CPU died normally, update state. */
0451         smp_mb(); /* atomic_read() before update. */
0452         atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
0453     } else {
0454         /* Outgoing CPU still hasn't died, set state accordingly. */
0455         if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
0456                    oldstate, CPU_BROKEN) != oldstate)
0457             goto update_state;
0458         ret = false;
0459     }
0460     return ret;
0461 }
0462
0463 /*
0464  * Called by the outgoing CPU to report its successful death.  Return
0465  * false if this report follows the surviving CPU's timing out.
0466  *
0467  * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
0468  * timed out.  This approach allows architectures to omit calls to
0469  * cpu_check_up_prepare() and cpu_set_state_online() without defeating
0470  * the next cpu_wait_death()'s polling loop.
0471  */
0472 bool cpu_report_death(void)
0473 {
0474     int oldstate;
0475     int newstate;
0476     int cpu = smp_processor_id();
0477
0478     do {
0479         oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
0480         if (oldstate != CPU_BROKEN)
0481             newstate = CPU_DEAD;
0482         else
0483             newstate = CPU_DEAD_FROZEN;
0484     } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
0485                 oldstate, newstate) != oldstate);
0486     return newstate == CPU_DEAD;
0487 }
0488
0489 #endif /* #ifdef CONFIG_HOTPLUG_CPU */