kernel/stop_machine.c

   1 /* Copyright 2005 Rusty Russell rusty@rustcorp.com.au IBM Corporation.
   2  * GPL v2 and any later version.
   3  */
   4 #include <linux/stop_machine.h>
   5 #include <linux/kthread.h>
   6 #include <linux/sched.h>
   7 #include <linux/cpu.h>
   8 #include <linux/err.h>
   9 #include <linux/syscalls.h>
  10 #include <asm/atomic.h>
  11 #include <asm/semaphore.h>
  12 #include <asm/uaccess.h>
  13
  14 /* Since we effect priority and affinity (both of which are visible
  15  * to, and settable by outside processes) we do indirection via a
  16  * kthread. */
  17
  18 /* Thread to stop each CPU in user context. */
  19 enum stopmachine_state {
  20         STOPMACHINE_WAIT,
  21         STOPMACHINE_PREPARE,
  22         STOPMACHINE_DISABLE_IRQ,
  23         STOPMACHINE_EXIT,
  24 };
  25
  26 static enum stopmachine_state stopmachine_state;
  27 static unsigned int stopmachine_num_threads;
  28 static atomic_t stopmachine_thread_ack;
  29 static DECLARE_MUTEX(stopmachine_mutex);
  30
  31 static int stopmachine(void *cpu)
  32 {
  33         int irqs_disabled = 0;
  34         int prepared = 0;
  35
  36         set_cpus_allowed(current, cpumask_of_cpu((int)(long)cpu));
  37
  38         /* Ack: we are alive */
  39         smp_mb(); /* Theoretically the ack = 0 might not be on this CPU yet. */
  40         atomic_inc(&stopmachine_thread_ack);
  41
  42         /* Simple state machine */
  43         while (stopmachine_state != STOPMACHINE_EXIT) {
  44                 if (stopmachine_state == STOPMACHINE_DISABLE_IRQ
  45                     && !irqs_disabled) {
  46                         local_irq_disable();
  47                         irqs_disabled = 1;
  48                         /* Ack: irqs disabled. */
  49                         smp_mb(); /* Must read state first. */
  50                         atomic_inc(&stopmachine_thread_ack);
  51                 } else if (stopmachine_state == STOPMACHINE_PREPARE
  52                            && !prepared) {
  53                         /* Everyone is in place, hold CPU. */
  54                         preempt_disable();
  55                         prepared = 1;
  56                         smp_mb(); /* Must read state first. */
  57                         atomic_inc(&stopmachine_thread_ack);
  58                 }
  59                 /* Yield in first stage: migration threads need to
  60                  * help our sisters onto their CPUs. */
  61                 if (!prepared && !irqs_disabled)
  62                         yield();
  63                 else
  64                         cpu_relax();
  65         }
  66
  67         /* Ack: we are exiting. */
  68         smp_mb(); /* Must read state first. */
  69         atomic_inc(&stopmachine_thread_ack);
  70
  71         if (irqs_disabled)
  72                 local_irq_enable();
  73         if (prepared)
  74                 preempt_enable();
  75
  76         return 0;
  77 }
  78
  79 /* Change the thread state */
  80 static void stopmachine_set_state(enum stopmachine_state state)
  81 {
  82         atomic_set(&stopmachine_thread_ack, 0);
  83         smp_wmb();
  84         stopmachine_state = state;
  85         while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads)
  86                 cpu_relax();
  87 }
  88
  89 static int stop_machine(void)
  90 {
  91         int i, ret = 0;
  92         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
  93
  94         /* One high-prio thread per cpu.  We'll do this one. */
  95         sched_setscheduler(current, SCHED_FIFO, &param);
  96
  97         atomic_set(&stopmachine_thread_ack, 0);
  98         stopmachine_num_threads = 0;
  99         stopmachine_state = STOPMACHINE_WAIT;
 100
 101         for_each_online_cpu(i) {
 102                 if (i == raw_smp_processor_id())
 103                         continue;
 104                 ret = kernel_thread(stopmachine, (void *)(long)i,CLONE_KERNEL);
 105                 if (ret < 0)
 106                         break;
 107                 stopmachine_num_threads++;
 108         }
 109
 110         /* Wait for them all to come to life. */
 111         while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads)
 112                 yield();
 113
 114         /* If some failed, kill them all. */
 115         if (ret < 0) {
 116                 stopmachine_set_state(STOPMACHINE_EXIT);
 117                 return ret;
 118         }
 119
 120         /* Now they are all started, make them hold the CPUs, ready. */
 121         preempt_disable();
 122         stopmachine_set_state(STOPMACHINE_PREPARE);
 123
 124         /* Make them disable irqs. */
 125         local_irq_disable();
 126         stopmachine_set_state(STOPMACHINE_DISABLE_IRQ);
 127
 128         return 0;
 129 }
 130
 131 static void restart_machine(void)
 132 {
 133         stopmachine_set_state(STOPMACHINE_EXIT);
 134         local_irq_enable();
 135         preempt_enable_no_resched();
 136 }
 137
 138 struct stop_machine_data
 139 {
 140         int (*fn)(void *);
 141         void *data;
 142         struct completion done;
 143 };
 144
 145 static int do_stop(void *_smdata)
 146 {
 147         struct stop_machine_data *smdata = _smdata;
 148         int ret;
 149
 150         ret = stop_machine();
 151         if (ret == 0) {
 152                 ret = smdata->fn(smdata->data);
 153                 restart_machine();
 154         }
 155
 156         /* We're done: you can kthread_stop us now */
 157         complete(&smdata->done);
 158
 159         /* Wait for kthread_stop */
 160         set_current_state(TASK_INTERRUPTIBLE);
 161         while (!kthread_should_stop()) {
 162                 schedule();
 163                 set_current_state(TASK_INTERRUPTIBLE);
 164         }
 165         __set_current_state(TASK_RUNNING);
 166         return ret;
 167 }
 168
 169 struct task_struct *__stop_machine_run(int (*fn)(void *), void *data,
 170                                        unsigned int cpu)
 171 {
 172         struct stop_machine_data smdata;
 173         struct task_struct *p;
 174
 175         smdata.fn = fn;
 176         smdata.data = data;
 177         init_completion(&smdata.done);
 178
 179         down(&stopmachine_mutex);
 180
 181         /* If they don't care which CPU fn runs on, bind to any online one. */
 182         if (cpu == NR_CPUS)
 183                 cpu = raw_smp_processor_id();
 184
 185         p = kthread_create(do_stop, &smdata, "kstopmachine");
 186         if (!IS_ERR(p)) {
 187                 kthread_bind(p, cpu);
 188                 wake_up_process(p);
 189                 wait_for_completion(&smdata.done);
 190         }
 191         up(&stopmachine_mutex);
 192         return p;
 193 }
 194
 195 int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu)
 196 {
 197         struct task_struct *p;
 198         int ret;
 199
 200         /* No CPUs can come up or down during this. */
 201         lock_cpu_hotplug();
 202         p = __stop_machine_run(fn, data, cpu);
 203         if (!IS_ERR(p))
 204                 ret = kthread_stop(p);
 205         else
 206                 ret = PTR_ERR(p);
 207         unlock_cpu_hotplug();
 208
 209         return ret;
 210 }