diff options
-rw-r--r-- | include/linux/stop_machine.h | 20 | ||||
-rw-r--r-- | kernel/cpu.c | 13 | ||||
-rw-r--r-- | kernel/stop_machine.c | 293 |
3 files changed, 136 insertions, 190 deletions
diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h index 18af011c13af..36c2c7284eb3 100644 --- a/include/linux/stop_machine.h +++ b/include/linux/stop_machine.h | |||
@@ -17,13 +17,12 @@ | |||
17 | * @data: the data ptr for the @fn() | 17 | * @data: the data ptr for the @fn() |
18 | * @cpu: if @cpu == n, run @fn() on cpu n | 18 | * @cpu: if @cpu == n, run @fn() on cpu n |
19 | * if @cpu == NR_CPUS, run @fn() on any cpu | 19 | * if @cpu == NR_CPUS, run @fn() on any cpu |
20 | * if @cpu == ALL_CPUS, run @fn() first on the calling cpu, and then | 20 | * if @cpu == ALL_CPUS, run @fn() on every online CPU. |
21 | * concurrently on all the other cpus | ||
22 | * | 21 | * |
23 | * Description: This causes a thread to be scheduled on every other cpu, | 22 | * Description: This causes a thread to be scheduled on every cpu, |
24 | * each of which disables interrupts, and finally interrupts are disabled | 23 | * each of which disables interrupts. The result is that noone is |
25 | * on the current CPU. The result is that noone is holding a spinlock | 24 | * holding a spinlock or inside any other preempt-disabled region when |
26 | * or inside any other preempt-disabled region when @fn() runs. | 25 | * @fn() runs. |
27 | * | 26 | * |
28 | * This can be thought of as a very heavy write lock, equivalent to | 27 | * This can be thought of as a very heavy write lock, equivalent to |
29 | * grabbing every spinlock in the kernel. */ | 28 | * grabbing every spinlock in the kernel. */ |
@@ -35,13 +34,10 @@ int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu); | |||
35 | * @data: the data ptr for the @fn | 34 | * @data: the data ptr for the @fn |
36 | * @cpu: the cpu to run @fn on (or any, if @cpu == NR_CPUS. | 35 | * @cpu: the cpu to run @fn on (or any, if @cpu == NR_CPUS. |
37 | * | 36 | * |
38 | * Description: This is a special version of the above, which returns the | 37 | * Description: This is a special version of the above, which assumes cpus |
39 | * thread which has run @fn(): kthread_stop will return the return value | 38 | * won't come or go while it's being called. Used by hotplug cpu. |
40 | * of @fn(). Used by hotplug cpu. | ||
41 | */ | 39 | */ |
42 | struct task_struct *__stop_machine_run(int (*fn)(void *), void *data, | 40 | int __stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu); |
43 | unsigned int cpu); | ||
44 | |||
45 | #else | 41 | #else |
46 | 42 | ||
47 | static inline int stop_machine_run(int (*fn)(void *), void *data, | 43 | static inline int stop_machine_run(int (*fn)(void *), void *data, |
diff --git a/kernel/cpu.c b/kernel/cpu.c index 10ba5f1004a5..cf79bb911371 100644 --- a/kernel/cpu.c +++ b/kernel/cpu.c | |||
@@ -216,7 +216,6 @@ static int __ref take_cpu_down(void *_param) | |||
216 | static int __ref _cpu_down(unsigned int cpu, int tasks_frozen) | 216 | static int __ref _cpu_down(unsigned int cpu, int tasks_frozen) |
217 | { | 217 | { |
218 | int err, nr_calls = 0; | 218 | int err, nr_calls = 0; |
219 | struct task_struct *p; | ||
220 | cpumask_t old_allowed, tmp; | 219 | cpumask_t old_allowed, tmp; |
221 | void *hcpu = (void *)(long)cpu; | 220 | void *hcpu = (void *)(long)cpu; |
222 | unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0; | 221 | unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0; |
@@ -250,19 +249,15 @@ static int __ref _cpu_down(unsigned int cpu, int tasks_frozen) | |||
250 | cpu_clear(cpu, tmp); | 249 | cpu_clear(cpu, tmp); |
251 | set_cpus_allowed_ptr(current, &tmp); | 250 | set_cpus_allowed_ptr(current, &tmp); |
252 | 251 | ||
253 | p = __stop_machine_run(take_cpu_down, &tcd_param, cpu); | 252 | err = __stop_machine_run(take_cpu_down, &tcd_param, cpu); |
254 | 253 | ||
255 | if (IS_ERR(p) || cpu_online(cpu)) { | 254 | if (err || cpu_online(cpu)) { |
256 | /* CPU didn't die: tell everyone. Can't complain. */ | 255 | /* CPU didn't die: tell everyone. Can't complain. */ |
257 | if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod, | 256 | if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod, |
258 | hcpu) == NOTIFY_BAD) | 257 | hcpu) == NOTIFY_BAD) |
259 | BUG(); | 258 | BUG(); |
260 | 259 | ||
261 | if (IS_ERR(p)) { | 260 | goto out_allowed; |
262 | err = PTR_ERR(p); | ||
263 | goto out_allowed; | ||
264 | } | ||
265 | goto out_thread; | ||
266 | } | 261 | } |
267 | 262 | ||
268 | /* Wait for it to sleep (leaving idle task). */ | 263 | /* Wait for it to sleep (leaving idle task). */ |
@@ -279,8 +274,6 @@ static int __ref _cpu_down(unsigned int cpu, int tasks_frozen) | |||
279 | 274 | ||
280 | check_for_tasks(cpu); | 275 | check_for_tasks(cpu); |
281 | 276 | ||
282 | out_thread: | ||
283 | err = kthread_stop(p); | ||
284 | out_allowed: | 277 | out_allowed: |
285 | set_cpus_allowed_ptr(current, &old_allowed); | 278 | set_cpus_allowed_ptr(current, &old_allowed); |
286 | out_release: | 279 | out_release: |
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c index a473bd0cb71b..35882dccc943 100644 --- a/kernel/stop_machine.c +++ b/kernel/stop_machine.c | |||
@@ -1,4 +1,4 @@ | |||
1 | /* Copyright 2005 Rusty Russell rusty@rustcorp.com.au IBM Corporation. | 1 | /* Copyright 2008, 2005 Rusty Russell rusty@rustcorp.com.au IBM Corporation. |
2 | * GPL v2 and any later version. | 2 | * GPL v2 and any later version. |
3 | */ | 3 | */ |
4 | #include <linux/cpu.h> | 4 | #include <linux/cpu.h> |
@@ -13,220 +13,177 @@ | |||
13 | #include <asm/atomic.h> | 13 | #include <asm/atomic.h> |
14 | #include <asm/uaccess.h> | 14 | #include <asm/uaccess.h> |
15 | 15 | ||
16 | /* Since we effect priority and affinity (both of which are visible | 16 | /* This controls the threads on each CPU. */ |
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 { | 17 | enum stopmachine_state { |
22 | STOPMACHINE_WAIT, | 18 | /* Dummy starting state for thread. */ |
19 | STOPMACHINE_NONE, | ||
20 | /* Awaiting everyone to be scheduled. */ | ||
23 | STOPMACHINE_PREPARE, | 21 | STOPMACHINE_PREPARE, |
22 | /* Disable interrupts. */ | ||
24 | STOPMACHINE_DISABLE_IRQ, | 23 | STOPMACHINE_DISABLE_IRQ, |
24 | /* Run the function */ | ||
25 | STOPMACHINE_RUN, | 25 | STOPMACHINE_RUN, |
26 | /* Exit */ | ||
26 | STOPMACHINE_EXIT, | 27 | STOPMACHINE_EXIT, |
27 | }; | 28 | }; |
29 | static enum stopmachine_state state; | ||
28 | 30 | ||
29 | struct stop_machine_data { | 31 | struct stop_machine_data { |
30 | int (*fn)(void *); | 32 | int (*fn)(void *); |
31 | void *data; | 33 | void *data; |
32 | struct completion done; | 34 | int fnret; |
33 | int run_all; | 35 | }; |
34 | } smdata; | ||
35 | 36 | ||
36 | static enum stopmachine_state stopmachine_state; | 37 | /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */ |
37 | static unsigned int stopmachine_num_threads; | 38 | static unsigned int num_threads; |
38 | static atomic_t stopmachine_thread_ack; | 39 | static atomic_t thread_ack; |
40 | static struct completion finished; | ||
41 | static DEFINE_MUTEX(lock); | ||
39 | 42 | ||
40 | static int stopmachine(void *cpu) | 43 | static void set_state(enum stopmachine_state newstate) |
41 | { | 44 | { |
42 | int irqs_disabled = 0; | 45 | /* Reset ack counter. */ |
43 | int prepared = 0; | 46 | atomic_set(&thread_ack, num_threads); |
44 | int ran = 0; | 47 | smp_wmb(); |
45 | cpumask_of_cpu_ptr(cpumask, (int)(long)cpu); | 48 | state = newstate; |
46 | |||
47 | set_cpus_allowed_ptr(current, cpumask); | ||
48 | |||
49 | /* Ack: we are alive */ | ||
50 | smp_mb(); /* Theoretically the ack = 0 might not be on this CPU yet. */ | ||
51 | atomic_inc(&stopmachine_thread_ack); | ||
52 | |||
53 | /* Simple state machine */ | ||
54 | while (stopmachine_state != STOPMACHINE_EXIT) { | ||
55 | if (stopmachine_state == STOPMACHINE_DISABLE_IRQ | ||
56 | && !irqs_disabled) { | ||
57 | local_irq_disable(); | ||
58 | hard_irq_disable(); | ||
59 | irqs_disabled = 1; | ||
60 | /* Ack: irqs disabled. */ | ||
61 | smp_mb(); /* Must read state first. */ | ||
62 | atomic_inc(&stopmachine_thread_ack); | ||
63 | } else if (stopmachine_state == STOPMACHINE_PREPARE | ||
64 | && !prepared) { | ||
65 | /* Everyone is in place, hold CPU. */ | ||
66 | preempt_disable(); | ||
67 | prepared = 1; | ||
68 | smp_mb(); /* Must read state first. */ | ||
69 | atomic_inc(&stopmachine_thread_ack); | ||
70 | } else if (stopmachine_state == STOPMACHINE_RUN && !ran) { | ||
71 | smdata.fn(smdata.data); | ||
72 | ran = 1; | ||
73 | smp_mb(); /* Must read state first. */ | ||
74 | atomic_inc(&stopmachine_thread_ack); | ||
75 | } | ||
76 | /* Yield in first stage: migration threads need to | ||
77 | * help our sisters onto their CPUs. */ | ||
78 | if (!prepared && !irqs_disabled) | ||
79 | yield(); | ||
80 | cpu_relax(); | ||
81 | } | ||
82 | |||
83 | /* Ack: we are exiting. */ | ||
84 | smp_mb(); /* Must read state first. */ | ||
85 | atomic_inc(&stopmachine_thread_ack); | ||
86 | |||
87 | if (irqs_disabled) | ||
88 | local_irq_enable(); | ||
89 | if (prepared) | ||
90 | preempt_enable(); | ||
91 | |||
92 | return 0; | ||
93 | } | 49 | } |
94 | 50 | ||
95 | /* Change the thread state */ | 51 | /* Last one to ack a state moves to the next state. */ |
96 | static void stopmachine_set_state(enum stopmachine_state state) | 52 | static void ack_state(void) |
97 | { | 53 | { |
98 | atomic_set(&stopmachine_thread_ack, 0); | 54 | if (atomic_dec_and_test(&thread_ack)) { |
99 | smp_wmb(); | 55 | /* If we're the last one to ack the EXIT, we're finished. */ |
100 | stopmachine_state = state; | 56 | if (state == STOPMACHINE_EXIT) |
101 | while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads) | 57 | complete(&finished); |
102 | cpu_relax(); | 58 | else |
59 | set_state(state + 1); | ||
60 | } | ||
103 | } | 61 | } |
104 | 62 | ||
105 | static int stop_machine(void) | 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) | ||
106 | { | 66 | { |
107 | int i, ret = 0; | 67 | enum stopmachine_state curstate = STOPMACHINE_NONE; |
108 | 68 | int uninitialized_var(ret); | |
109 | atomic_set(&stopmachine_thread_ack, 0); | ||
110 | stopmachine_num_threads = 0; | ||
111 | stopmachine_state = STOPMACHINE_WAIT; | ||
112 | 69 | ||
113 | for_each_online_cpu(i) { | 70 | /* Simple state machine */ |
114 | if (i == raw_smp_processor_id()) | 71 | do { |
115 | continue; | 72 | /* Chill out and ensure we re-read stopmachine_state. */ |
116 | ret = kernel_thread(stopmachine, (void *)(long)i,CLONE_KERNEL); | ||
117 | if (ret < 0) | ||
118 | break; | ||
119 | stopmachine_num_threads++; | ||
120 | } | ||
121 | |||
122 | /* Wait for them all to come to life. */ | ||
123 | while (atomic_read(&stopmachine_thread_ack) != stopmachine_num_threads) { | ||
124 | yield(); | ||
125 | cpu_relax(); | 73 | cpu_relax(); |
126 | } | 74 | if (state != curstate) { |
127 | 75 | curstate = state; | |
128 | /* If some failed, kill them all. */ | 76 | switch (curstate) { |
129 | if (ret < 0) { | 77 | case STOPMACHINE_DISABLE_IRQ: |
130 | stopmachine_set_state(STOPMACHINE_EXIT); | 78 | local_irq_disable(); |
131 | return ret; | 79 | hard_irq_disable(); |
132 | } | 80 | break; |
133 | 81 | case STOPMACHINE_RUN: | |
134 | /* Now they are all started, make them hold the CPUs, ready. */ | 82 | /* |= allows error detection if functions on |
135 | preempt_disable(); | 83 | * multiple CPUs. */ |
136 | stopmachine_set_state(STOPMACHINE_PREPARE); | 84 | smdata->fnret |= smdata->fn(smdata->data); |
137 | 85 | break; | |
138 | /* Make them disable irqs. */ | 86 | default: |
139 | local_irq_disable(); | 87 | break; |
140 | hard_irq_disable(); | 88 | } |
141 | stopmachine_set_state(STOPMACHINE_DISABLE_IRQ); | 89 | ack_state(); |
142 | 90 | } | |
143 | return 0; | 91 | } while (curstate != STOPMACHINE_EXIT); |
144 | } | ||
145 | 92 | ||
146 | static void restart_machine(void) | ||
147 | { | ||
148 | stopmachine_set_state(STOPMACHINE_EXIT); | ||
149 | local_irq_enable(); | 93 | local_irq_enable(); |
150 | preempt_enable_no_resched(); | 94 | do_exit(0); |
151 | } | 95 | } |
152 | 96 | ||
153 | static void run_other_cpus(void) | 97 | /* Callback for CPUs which aren't supposed to do anything. */ |
98 | static int chill(void *unused) | ||
154 | { | 99 | { |
155 | stopmachine_set_state(STOPMACHINE_RUN); | 100 | return 0; |
156 | } | 101 | } |
157 | 102 | ||
158 | static int do_stop(void *_smdata) | 103 | int __stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu) |
159 | { | 104 | { |
160 | struct stop_machine_data *smdata = _smdata; | 105 | int i, err; |
161 | int ret; | 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); | ||
162 | 129 | ||
163 | ret = stop_machine(); | 130 | for_each_online_cpu(i) { |
164 | if (ret == 0) { | 131 | struct stop_machine_data *smdata; |
165 | ret = smdata->fn(smdata->data); | 132 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; |
166 | if (smdata->run_all) | ||
167 | run_other_cpus(); | ||
168 | restart_machine(); | ||
169 | } | ||
170 | 133 | ||
171 | /* We're done: you can kthread_stop us now */ | 134 | if (cpu == ALL_CPUS || i == cpu) |
172 | complete(&smdata->done); | 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 | } | ||
173 | 146 | ||
174 | /* Wait for kthread_stop */ | 147 | /* Place it onto correct cpu. */ |
175 | set_current_state(TASK_INTERRUPTIBLE); | 148 | kthread_bind(threads[i], i); |
176 | while (!kthread_should_stop()) { | ||
177 | schedule(); | ||
178 | set_current_state(TASK_INTERRUPTIBLE); | ||
179 | } | ||
180 | __set_current_state(TASK_RUNNING); | ||
181 | return ret; | ||
182 | } | ||
183 | 149 | ||
184 | struct task_struct *__stop_machine_run(int (*fn)(void *), void *data, | 150 | /* Make it highest prio. */ |
185 | unsigned int cpu) | 151 | if (sched_setscheduler_nocheck(threads[i], SCHED_FIFO, ¶m)) |
186 | { | 152 | BUG(); |
187 | static DEFINE_MUTEX(stopmachine_mutex); | 153 | } |
188 | struct stop_machine_data smdata; | ||
189 | struct task_struct *p; | ||
190 | 154 | ||
191 | mutex_lock(&stopmachine_mutex); | 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]); | ||
192 | 160 | ||
193 | smdata.fn = fn; | 161 | /* This will release the thread on our CPU. */ |
194 | smdata.data = data; | 162 | put_cpu(); |
195 | smdata.run_all = (cpu == ALL_CPUS) ? 1 : 0; | 163 | wait_for_completion(&finished); |
196 | init_completion(&smdata.done); | 164 | mutex_unlock(&lock); |
197 | 165 | ||
198 | smp_wmb(); /* make sure other cpus see smdata updates */ | 166 | kfree(threads); |
199 | 167 | ||
200 | /* If they don't care which CPU fn runs on, bind to any online one. */ | 168 | return active.fnret; |
201 | if (cpu == NR_CPUS || cpu == ALL_CPUS) | ||
202 | cpu = raw_smp_processor_id(); | ||
203 | 169 | ||
204 | p = kthread_create(do_stop, &smdata, "kstopmachine"); | 170 | kill_threads: |
205 | if (!IS_ERR(p)) { | 171 | for_each_online_cpu(i) |
206 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; | 172 | if (threads[i]) |
173 | kthread_stop(threads[i]); | ||
174 | mutex_unlock(&lock); | ||
207 | 175 | ||
208 | /* One high-prio thread per cpu. We'll do this one. */ | 176 | kfree(threads); |
209 | sched_setscheduler_nocheck(p, SCHED_FIFO, ¶m); | 177 | return err; |
210 | kthread_bind(p, cpu); | ||
211 | wake_up_process(p); | ||
212 | wait_for_completion(&smdata.done); | ||
213 | } | ||
214 | mutex_unlock(&stopmachine_mutex); | ||
215 | return p; | ||
216 | } | 178 | } |
217 | 179 | ||
218 | int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu) | 180 | int stop_machine_run(int (*fn)(void *), void *data, unsigned int cpu) |
219 | { | 181 | { |
220 | struct task_struct *p; | ||
221 | int ret; | 182 | int ret; |
222 | 183 | ||
223 | /* No CPUs can come up or down during this. */ | 184 | /* No CPUs can come up or down during this. */ |
224 | get_online_cpus(); | 185 | get_online_cpus(); |
225 | p = __stop_machine_run(fn, data, cpu); | 186 | ret = __stop_machine_run(fn, data, cpu); |
226 | if (!IS_ERR(p)) | ||
227 | ret = kthread_stop(p); | ||
228 | else | ||
229 | ret = PTR_ERR(p); | ||
230 | put_online_cpus(); | 187 | put_online_cpus(); |
231 | 188 | ||
232 | return ret; | 189 | return ret; |