xref: /linux/kernel/stop_machine.c (revision 776cfebb430c7b22c208b1b17add97f354d97cab)
1 #include <linux/stop_machine.h>
2 #include <linux/kthread.h>
3 #include <linux/sched.h>
4 #include <linux/cpu.h>
5 #include <linux/err.h>
6 #include <linux/syscalls.h>
7 #include <asm/atomic.h>
8 #include <asm/semaphore.h>
9 #include <asm/uaccess.h>
10 
11 /* Since we effect priority and affinity (both of which are visible
12  * to, and settable by outside processes) we do indirection via a
13  * kthread. */
14 
15 /* Thread to stop each CPU in user context. */
16 enum stopmachine_state {
17 	STOPMACHINE_WAIT,
18 	STOPMACHINE_PREPARE,
19 	STOPMACHINE_DISABLE_IRQ,
20 	STOPMACHINE_EXIT,
21 };
22 
23 static enum stopmachine_state stopmachine_state;
24 static unsigned int stopmachine_num_threads;
25 static atomic_t stopmachine_thread_ack;
26 static DECLARE_MUTEX(stopmachine_mutex);
27 
28 static int stopmachine(void *cpu)
29 {
30 	int irqs_disabled = 0;
31 	int prepared = 0;
32 
33 	set_cpus_allowed(current, cpumask_of_cpu((int)(long)cpu));
34 
35 	/* Ack: we are alive */
36 	smp_mb(); /* Theoretically the ack = 0 might not be on this CPU yet. */
37 	atomic_inc(&stopmachine_thread_ack);
38 
39 	/* Simple state machine */
40 	while (stopmachine_state != STOPMACHINE_EXIT) {
41 		if (stopmachine_state == STOPMACHINE_DISABLE_IRQ
42 		    && !irqs_disabled) {
43 			local_irq_disable();
44 			irqs_disabled = 1;
45 			/* Ack: irqs disabled. */
46 			smp_mb(); /* Must read state first. */
47 			atomic_inc(&stopmachine_thread_ack);
48 		} else if (stopmachine_state == STOPMACHINE_PREPARE
49 			   && !prepared) {
50 			/* Everyone is in place, hold CPU. */
51 			preempt_disable();
52 			prepared = 1;
53 			smp_mb(); /* Must read state first. */
54 			atomic_inc(&stopmachine_thread_ack);
55 		}
56 		/* Yield in first stage: migration threads need to
57 		 * help our sisters onto their CPUs. */
58 		if (!prepared && !irqs_disabled)
59 			yield();
60 		else
61 			cpu_relax();
62 	}
63 
64 	/* Ack: we are exiting. */
65 	smp_mb(); /* Must read state first. */
66 	atomic_inc(&stopmachine_thread_ack);
67 
68 	if (irqs_disabled)
69 		local_irq_enable();
70 	if (prepared)
71 		preempt_enable();
72 
73 	return 0;
74 }
75 
76 /* Change the thread state */
77 static void stopmachine_set_state(enum stopmachine_state state)
78 {
79 	atomic_set(&stopmachine_thread_ack, 0);
80 	smp_wmb();
81 	stopmachine_state = state;
82 	while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads)
83 		cpu_relax();
84 }
85 
86 static int stop_machine(void)
87 {
88 	int i, ret = 0;
89 	struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
90 	mm_segment_t old_fs = get_fs();
91 
92 	/* One high-prio thread per cpu.  We'll do this one. */
93 	set_fs(KERNEL_DS);
94 	sys_sched_setscheduler(current->pid, SCHED_FIFO,
95 				(struct sched_param __user *)&param);
96 	set_fs(old_fs);
97 
98 	atomic_set(&stopmachine_thread_ack, 0);
99 	stopmachine_num_threads = 0;
100 	stopmachine_state = STOPMACHINE_WAIT;
101 
102 	for_each_online_cpu(i) {
103 		if (i == _smp_processor_id())
104 			continue;
105 		ret = kernel_thread(stopmachine, (void *)(long)i,CLONE_KERNEL);
106 		if (ret < 0)
107 			break;
108 		stopmachine_num_threads++;
109 	}
110 
111 	/* Wait for them all to come to life. */
112 	while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads)
113 		yield();
114 
115 	/* If some failed, kill them all. */
116 	if (ret < 0) {
117 		stopmachine_set_state(STOPMACHINE_EXIT);
118 		up(&stopmachine_mutex);
119 		return ret;
120 	}
121 
122 	/* Don't schedule us away at this point, please. */
123 	local_irq_disable();
124 
125 	/* Now they are all started, make them hold the CPUs, ready. */
126 	stopmachine_set_state(STOPMACHINE_PREPARE);
127 
128 	/* Make them disable irqs. */
129 	stopmachine_set_state(STOPMACHINE_DISABLE_IRQ);
130 
131 	return 0;
132 }
133 
134 static void restart_machine(void)
135 {
136 	stopmachine_set_state(STOPMACHINE_EXIT);
137 	local_irq_enable();
138 }
139 
140 struct stop_machine_data
141 {
142 	int (*fn)(void *);
143 	void *data;
144 	struct completion done;
145 };
146 
147 static int do_stop(void *_smdata)
148 {
149 	struct stop_machine_data *smdata = _smdata;
150 	int ret;
151 
152 	ret = stop_machine();
153 	if (ret == 0) {
154 		ret = smdata->fn(smdata->data);
155 		restart_machine();
156 	}
157 
158 	/* We're done: you can kthread_stop us now */
159 	complete(&smdata->done);
160 
161 	/* Wait for kthread_stop */
162 	set_current_state(TASK_INTERRUPTIBLE);
163 	while (!kthread_should_stop()) {
164 		schedule();
165 		set_current_state(TASK_INTERRUPTIBLE);
166 	}
167 	__set_current_state(TASK_RUNNING);
168 	return ret;
169 }
170 
171 struct task_struct *__stop_machine_run(int (*fn)(void *), void *data,
172 				       unsigned int cpu)
173 {
174 	struct stop_machine_data smdata;
175 	struct task_struct *p;
176 
177 	smdata.fn = fn;
178 	smdata.data = data;
179 	init_completion(&smdata.done);
180 
181 	down(&stopmachine_mutex);
182 
183 	/* If they don't care which CPU fn runs on, bind to any online one. */
184 	if (cpu == NR_CPUS)
185 		cpu = _smp_processor_id();
186 
187 	p = kthread_create(do_stop, &smdata, "kstopmachine");
188 	if (!IS_ERR(p)) {
189 		kthread_bind(p, cpu);
190 		wake_up_process(p);
191 		wait_for_completion(&smdata.done);
192 	}
193 	up(&stopmachine_mutex);
194 	return p;
195 }
196 
197 int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu)
198 {
199 	struct task_struct *p;
200 	int ret;
201 
202 	/* No CPUs can come up or down during this. */
203 	lock_cpu_hotplug();
204 	p = __stop_machine_run(fn, data, cpu);
205 	if (!IS_ERR(p))
206 		ret = kthread_stop(p);
207 	else
208 		ret = PTR_ERR(p);
209 	unlock_cpu_hotplug();
210 
211 	return ret;
212 }
213