sched: add new API sched_setscheduler_nocheck: add a flag to control access checks
[linux-2.6.git] / 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/cpu.h>
5 #include <linux/err.h>
6 #include <linux/kthread.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
9 #include <linux/stop_machine.h>
10 #include <linux/syscalls.h>
11 #include <linux/interrupt.h>
12
13 #include <asm/atomic.h>
14 #include <asm/uaccess.h>
15
16 /* Since we effect priority and affinity (both of which are visible
17  * to, and settable by outside processes) we do indirection via a
18  * kthread. */
19
20 /* Thread to stop each CPU in user context. */
21 enum stopmachine_state {
22         STOPMACHINE_WAIT,
23         STOPMACHINE_PREPARE,
24         STOPMACHINE_DISABLE_IRQ,
25         STOPMACHINE_EXIT,
26 };
27
28 static enum stopmachine_state stopmachine_state;
29 static unsigned int stopmachine_num_threads;
30 static atomic_t stopmachine_thread_ack;
31
32 static int stopmachine(void *cpu)
33 {
34         int irqs_disabled = 0;
35         int prepared = 0;
36
37         set_cpus_allowed_ptr(current, &cpumask_of_cpu((int)(long)cpu));
38
39         /* Ack: we are alive */
40         smp_mb(); /* Theoretically the ack = 0 might not be on this CPU yet. */
41         atomic_inc(&stopmachine_thread_ack);
42
43         /* Simple state machine */
44         while (stopmachine_state != STOPMACHINE_EXIT) {
45                 if (stopmachine_state == STOPMACHINE_DISABLE_IRQ 
46                     && !irqs_disabled) {
47                         local_irq_disable();
48                         hard_irq_disable();
49                         irqs_disabled = 1;
50                         /* Ack: irqs disabled. */
51                         smp_mb(); /* Must read state first. */
52                         atomic_inc(&stopmachine_thread_ack);
53                 } else if (stopmachine_state == STOPMACHINE_PREPARE
54                            && !prepared) {
55                         /* Everyone is in place, hold CPU. */
56                         preempt_disable();
57                         prepared = 1;
58                         smp_mb(); /* Must read state first. */
59                         atomic_inc(&stopmachine_thread_ack);
60                 }
61                 /* Yield in first stage: migration threads need to
62                  * help our sisters onto their CPUs. */
63                 if (!prepared && !irqs_disabled)
64                         yield();
65                 cpu_relax();
66         }
67
68         /* Ack: we are exiting. */
69         smp_mb(); /* Must read state first. */
70         atomic_inc(&stopmachine_thread_ack);
71
72         if (irqs_disabled)
73                 local_irq_enable();
74         if (prepared)
75                 preempt_enable();
76
77         return 0;
78 }
79
80 /* Change the thread state */
81 static void stopmachine_set_state(enum stopmachine_state state)
82 {
83         atomic_set(&stopmachine_thread_ack, 0);
84         smp_wmb();
85         stopmachine_state = state;
86         while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads)
87                 cpu_relax();
88 }
89
90 static int stop_machine(void)
91 {
92         int i, ret = 0;
93
94         atomic_set(&stopmachine_thread_ack, 0);
95         stopmachine_num_threads = 0;
96         stopmachine_state = STOPMACHINE_WAIT;
97
98         for_each_online_cpu(i) {
99                 if (i == raw_smp_processor_id())
100                         continue;
101                 ret = kernel_thread(stopmachine, (void *)(long)i,CLONE_KERNEL);
102                 if (ret < 0)
103                         break;
104                 stopmachine_num_threads++;
105         }
106
107         /* Wait for them all to come to life. */
108         while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads) {
109                 yield();
110                 cpu_relax();
111         }
112
113         /* If some failed, kill them all. */
114         if (ret < 0) {
115                 stopmachine_set_state(STOPMACHINE_EXIT);
116                 return ret;
117         }
118
119         /* Now they are all started, make them hold the CPUs, ready. */
120         preempt_disable();
121         stopmachine_set_state(STOPMACHINE_PREPARE);
122
123         /* Make them disable irqs. */
124         local_irq_disable();
125         hard_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         int (*fn)(void *);
140         void *data;
141         struct completion done;
142 };
143
144 static int do_stop(void *_smdata)
145 {
146         struct stop_machine_data *smdata = _smdata;
147         int ret;
148
149         ret = stop_machine();
150         if (ret == 0) {
151                 ret = smdata->fn(smdata->data);
152                 restart_machine();
153         }
154
155         /* We're done: you can kthread_stop us now */
156         complete(&smdata->done);
157
158         /* Wait for kthread_stop */
159         set_current_state(TASK_INTERRUPTIBLE);
160         while (!kthread_should_stop()) {
161                 schedule();
162                 set_current_state(TASK_INTERRUPTIBLE);
163         }
164         __set_current_state(TASK_RUNNING);
165         return ret;
166 }
167
168 struct task_struct *__stop_machine_run(int (*fn)(void *), void *data,
169                                        unsigned int cpu)
170 {
171         static DEFINE_MUTEX(stopmachine_mutex);
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         mutex_lock(&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                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
188
189                 /* One high-prio thread per cpu.  We'll do this one. */
190                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
191                 kthread_bind(p, cpu);
192                 wake_up_process(p);
193                 wait_for_completion(&smdata.done);
194         }
195         mutex_unlock(&stopmachine_mutex);
196         return p;
197 }
198
199 int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu)
200 {
201         struct task_struct *p;
202         int ret;
203
204         /* No CPUs can come up or down during this. */
205         get_online_cpus();
206         p = __stop_machine_run(fn, data, cpu);
207         if (!IS_ERR(p))
208                 ret = kthread_stop(p);
209         else
210                 ret = PTR_ERR(p);
211         put_online_cpus();
212
213         return ret;
214 }
215 EXPORT_SYMBOL_GPL(stop_machine_run);