Simplify stop_machine
[linux-2.6.git] / kernel / stop_machine.c
1 /* Copyright 2008, 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 /* This controls the threads on each CPU. */
17 enum stopmachine_state {
18         /* Dummy starting state for thread. */
19         STOPMACHINE_NONE,
20         /* Awaiting everyone to be scheduled. */
21         STOPMACHINE_PREPARE,
22         /* Disable interrupts. */
23         STOPMACHINE_DISABLE_IRQ,
24         /* Run the function */
25         STOPMACHINE_RUN,
26         /* Exit */
27         STOPMACHINE_EXIT,
28 };
29 static enum stopmachine_state state;
30
31 struct stop_machine_data {
32         int (*fn)(void *);
33         void *data;
34         int fnret;
35 };
36
37 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
38 static unsigned int num_threads;
39 static atomic_t thread_ack;
40 static struct completion finished;
41 static DEFINE_MUTEX(lock);
42
43 static void set_state(enum stopmachine_state newstate)
44 {
45         /* Reset ack counter. */
46         atomic_set(&thread_ack, num_threads);
47         smp_wmb();
48         state = newstate;
49 }
50
51 /* Last one to ack a state moves to the next state. */
52 static void ack_state(void)
53 {
54         if (atomic_dec_and_test(&thread_ack)) {
55                 /* If we're the last one to ack the EXIT, we're finished. */
56                 if (state == STOPMACHINE_EXIT)
57                         complete(&finished);
58                 else
59                         set_state(state + 1);
60         }
61 }
62
63 /* This is the actual thread which stops the CPU.  It exits by itself rather
64  * than waiting for kthread_stop(), because it's easier for hotplug CPU. */
65 static int stop_cpu(struct stop_machine_data *smdata)
66 {
67         enum stopmachine_state curstate = STOPMACHINE_NONE;
68         int uninitialized_var(ret);
69
70         /* Simple state machine */
71         do {
72                 /* Chill out and ensure we re-read stopmachine_state. */
73                 cpu_relax();
74                 if (state != curstate) {
75                         curstate = state;
76                         switch (curstate) {
77                         case STOPMACHINE_DISABLE_IRQ:
78                                 local_irq_disable();
79                                 hard_irq_disable();
80                                 break;
81                         case STOPMACHINE_RUN:
82                                 /* |= allows error detection if functions on
83                                  * multiple CPUs. */
84                                 smdata->fnret |= smdata->fn(smdata->data);
85                                 break;
86                         default:
87                                 break;
88                         }
89                         ack_state();
90                 }
91         } while (curstate != STOPMACHINE_EXIT);
92
93         local_irq_enable();
94         do_exit(0);
95 }
96
97 /* Callback for CPUs which aren't supposed to do anything. */
98 static int chill(void *unused)
99 {
100         return 0;
101 }
102
103 int __stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu)
104 {
105         int i, err;
106         struct stop_machine_data active, idle;
107         struct task_struct **threads;
108
109         active.fn = fn;
110         active.data = data;
111         active.fnret = 0;
112         idle.fn = chill;
113         idle.data = NULL;
114
115         /* If they don't care which cpu fn runs on, just pick one. */
116         if (cpu == NR_CPUS)
117                 cpu = any_online_cpu(cpu_online_map);
118
119         /* This could be too big for stack on large machines. */
120         threads = kcalloc(NR_CPUS, sizeof(threads[0]), GFP_KERNEL);
121         if (!threads)
122                 return -ENOMEM;
123
124         /* Set up initial state. */
125         mutex_lock(&lock);
126         init_completion(&finished);
127         num_threads = num_online_cpus();
128         set_state(STOPMACHINE_PREPARE);
129
130         for_each_online_cpu(i) {
131                 struct stop_machine_data *smdata;
132                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
133
134                 if (cpu == ALL_CPUS || i == cpu)
135                         smdata = &active;
136                 else
137                         smdata = &idle;
138
139                 threads[i] = kthread_create((void *)stop_cpu, smdata, "kstop%u",
140                                             i);
141                 if (IS_ERR(threads[i])) {
142                         err = PTR_ERR(threads[i]);
143                         threads[i] = NULL;
144                         goto kill_threads;
145                 }
146
147                 /* Place it onto correct cpu. */
148                 kthread_bind(threads[i], i);
149
150                 /* Make it highest prio. */
151                 if (sched_setscheduler_nocheck(threads[i], SCHED_FIFO, &param))
152                         BUG();
153         }
154
155         /* We've created all the threads.  Wake them all: hold this CPU so one
156          * doesn't hit this CPU until we're ready. */
157         cpu = get_cpu();
158         for_each_online_cpu(i)
159                 wake_up_process(threads[i]);
160
161         /* This will release the thread on our CPU. */
162         put_cpu();
163         wait_for_completion(&finished);
164         mutex_unlock(&lock);
165
166         kfree(threads);
167
168         return active.fnret;
169
170 kill_threads:
171         for_each_online_cpu(i)
172                 if (threads[i])
173                         kthread_stop(threads[i]);
174         mutex_unlock(&lock);
175
176         kfree(threads);
177         return err;
178 }
179
180 int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu)
181 {
182         int ret;
183
184         /* No CPUs can come up or down during this. */
185         get_online_cpus();
186         ret = __stop_machine_run(fn, data, cpu);
187         put_online_cpus();
188
189         return ret;
190 }
191 EXPORT_SYMBOL_GPL(stop_machine_run);