projects/linux·blob v4.3-rc1 · refs

527 lines · 13.2 KiB

1/*
2 * Common SMP CPU bringup/teardown functions
3 */
4#include <linux/cpu.h>
5#include <linux/err.h>
6#include <linux/smp.h>
7#include <linux/delay.h>
8#include <linux/init.h>
9#include <linux/list.h>
10#include <linux/slab.h>
11#include <linux/sched.h>
12#include <linux/export.h>
13#include <linux/percpu.h>
14#include <linux/kthread.h>
15#include <linux/smpboot.h>
16
17#include "smpboot.h"
18
19#ifdef CONFIG_SMP
20
21#ifdef CONFIG_GENERIC_SMP_IDLE_THREAD
22/*
23 * For the hotplug case we keep the task structs around and reuse
24 * them.
25 */
26static DEFINE_PER_CPU(struct task_struct *, idle_threads);
27
28struct task_struct *idle_thread_get(unsigned int cpu)
29{
30 struct task_struct *tsk = per_cpu(idle_threads, cpu);
31
32 if (!tsk)
33 return ERR_PTR(-ENOMEM);
34 init_idle(tsk, cpu);
35 return tsk;
36}
37
38void __init idle_thread_set_boot_cpu(void)
39{
40 per_cpu(idle_threads, smp_processor_id()) = current;
41}
42
43/**
44 * idle_init - Initialize the idle thread for a cpu
45 * @cpu: The cpu for which the idle thread should be initialized
46 *
47 * Creates the thread if it does not exist.
48 */
49static inline void idle_init(unsigned int cpu)
50{
51 struct task_struct *tsk = per_cpu(idle_threads, cpu);
52
53 if (!tsk) {
54 tsk = fork_idle(cpu);
55 if (IS_ERR(tsk))
56 pr_err("SMP: fork_idle() failed for CPU %u\n", cpu);
57 else
58 per_cpu(idle_threads, cpu) = tsk;
59 }
60}
61
62/**
63 * idle_threads_init - Initialize idle threads for all cpus
64 */
65void __init idle_threads_init(void)
66{
67 unsigned int cpu, boot_cpu;
68
69 boot_cpu = smp_processor_id();
70
71 for_each_possible_cpu(cpu) {
72 if (cpu != boot_cpu)
73 idle_init(cpu);
74 }
75}
76#endif
77
78#endif /* #ifdef CONFIG_SMP */
79
80static LIST_HEAD(hotplug_threads);
81static DEFINE_MUTEX(smpboot_threads_lock);
82
83struct smpboot_thread_data {
84 unsigned int cpu;
85 unsigned int status;
86 struct smp_hotplug_thread *ht;
87};
88
89enum {
90 HP_THREAD_NONE = 0,
91 HP_THREAD_ACTIVE,
92 HP_THREAD_PARKED,
93};
94
95/**
96 * smpboot_thread_fn - percpu hotplug thread loop function
97 * @data: thread data pointer
98 *
99 * Checks for thread stop and park conditions. Calls the necessary
100 * setup, cleanup, park and unpark functions for the registered
101 * thread.
102 *
103 * Returns 1 when the thread should exit, 0 otherwise.
104 */
105static int smpboot_thread_fn(void *data)
106{
107 struct smpboot_thread_data *td = data;
108 struct smp_hotplug_thread *ht = td->ht;
109
110 while (1) {
111 set_current_state(TASK_INTERRUPTIBLE);
112 preempt_disable();
113 if (kthread_should_stop()) {
114 __set_current_state(TASK_RUNNING);
115 preempt_enable();
116 /* cleanup must mirror setup */
117 if (ht->cleanup && td->status != HP_THREAD_NONE)
118 ht->cleanup(td->cpu, cpu_online(td->cpu));
119 kfree(td);
120 return 0;
121 }
122
123 if (kthread_should_park()) {
124 __set_current_state(TASK_RUNNING);
125 preempt_enable();
126 if (ht->park && td->status == HP_THREAD_ACTIVE) {
127 BUG_ON(td->cpu != smp_processor_id());
128 ht->park(td->cpu);
129 td->status = HP_THREAD_PARKED;
130 }
131 kthread_parkme();
132 /* We might have been woken for stop */
133 continue;
134 }
135
136 BUG_ON(td->cpu != smp_processor_id());
137
138 /* Check for state change setup */
139 switch (td->status) {
140 case HP_THREAD_NONE:
141 __set_current_state(TASK_RUNNING);
142 preempt_enable();
143 if (ht->setup)
144 ht->setup(td->cpu);
145 td->status = HP_THREAD_ACTIVE;
146 continue;
147
148 case HP_THREAD_PARKED:
149 __set_current_state(TASK_RUNNING);
150 preempt_enable();
151 if (ht->unpark)
152 ht->unpark(td->cpu);
153 td->status = HP_THREAD_ACTIVE;
154 continue;
155 }
156
157 if (!ht->thread_should_run(td->cpu)) {
158 preempt_enable_no_resched();
159 schedule();
160 } else {
161 __set_current_state(TASK_RUNNING);
162 preempt_enable();
163 ht->thread_fn(td->cpu);
164 }
165 }
166}
167
168static int
169__smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
170{
171 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
172 struct smpboot_thread_data *td;
173
174 if (tsk)
175 return 0;
176
177 td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
178 if (!td)
179 return -ENOMEM;
180 td->cpu = cpu;
181 td->ht = ht;
182
183 tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
184 ht->thread_comm);
185 if (IS_ERR(tsk)) {
186 kfree(td);
187 return PTR_ERR(tsk);
188 }
189 get_task_struct(tsk);
190 *per_cpu_ptr(ht->store, cpu) = tsk;
191 if (ht->create) {
192 /*
193 * Make sure that the task has actually scheduled out
194 * into park position, before calling the create
195 * callback. At least the migration thread callback
196 * requires that the task is off the runqueue.
197 */
198 if (!wait_task_inactive(tsk, TASK_PARKED))
199 WARN_ON(1);
200 else
201 ht->create(cpu);
202 }
203 return 0;
204}
205
206int smpboot_create_threads(unsigned int cpu)
207{
208 struct smp_hotplug_thread *cur;
209 int ret = 0;
210
211 mutex_lock(&smpboot_threads_lock);
212 list_for_each_entry(cur, &hotplug_threads, list) {
213 ret = __smpboot_create_thread(cur, cpu);
214 if (ret)
215 break;
216 }
217 mutex_unlock(&smpboot_threads_lock);
218 return ret;
219}
220
221static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
222{
223 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
224
225 if (ht->pre_unpark)
226 ht->pre_unpark(cpu);
227 kthread_unpark(tsk);
228}
229
230void smpboot_unpark_threads(unsigned int cpu)
231{
232 struct smp_hotplug_thread *cur;
233
234 mutex_lock(&smpboot_threads_lock);
235 list_for_each_entry(cur, &hotplug_threads, list)
236 if (cpumask_test_cpu(cpu, cur->cpumask))
237 smpboot_unpark_thread(cur, cpu);
238 mutex_unlock(&smpboot_threads_lock);
239}
240
241static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
242{
243 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
244
245 if (tsk && !ht->selfparking)
246 kthread_park(tsk);
247}
248
249void smpboot_park_threads(unsigned int cpu)
250{
251 struct smp_hotplug_thread *cur;
252
253 mutex_lock(&smpboot_threads_lock);
254 list_for_each_entry_reverse(cur, &hotplug_threads, list)
255 smpboot_park_thread(cur, cpu);
256 mutex_unlock(&smpboot_threads_lock);
257}
258
259static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
260{
261 unsigned int cpu;
262
263 /* We need to destroy also the parked threads of offline cpus */
264 for_each_possible_cpu(cpu) {
265 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
266
267 if (tsk) {
268 kthread_stop(tsk);
269 put_task_struct(tsk);
270 *per_cpu_ptr(ht->store, cpu) = NULL;
271 }
272 }
273}
274
275/**
276 * smpboot_register_percpu_thread_cpumask - Register a per_cpu thread related
277 * to hotplug
278 * @plug_thread: Hotplug thread descriptor
279 * @cpumask: The cpumask where threads run
280 *
281 * Creates and starts the threads on all online cpus.
282 */
283int smpboot_register_percpu_thread_cpumask(struct smp_hotplug_thread *plug_thread,
284 const struct cpumask *cpumask)
285{
286 unsigned int cpu;
287 int ret = 0;
288
289 if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
290 return -ENOMEM;
291 cpumask_copy(plug_thread->cpumask, cpumask);
292
293 get_online_cpus();
294 mutex_lock(&smpboot_threads_lock);
295 for_each_online_cpu(cpu) {
296 ret = __smpboot_create_thread(plug_thread, cpu);
297 if (ret) {
298 smpboot_destroy_threads(plug_thread);
299 free_cpumask_var(plug_thread->cpumask);
300 goto out;
301 }
302 if (cpumask_test_cpu(cpu, cpumask))
303 smpboot_unpark_thread(plug_thread, cpu);
304 }
305 list_add(&plug_thread->list, &hotplug_threads);
306out:
307 mutex_unlock(&smpboot_threads_lock);
308 put_online_cpus();
309 return ret;
310}
311EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread_cpumask);
312
313/**
314 * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
315 * @plug_thread: Hotplug thread descriptor
316 *
317 * Stops all threads on all possible cpus.
318 */
319void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
320{
321 get_online_cpus();
322 mutex_lock(&smpboot_threads_lock);
323 list_del(&plug_thread->list);
324 smpboot_destroy_threads(plug_thread);
325 mutex_unlock(&smpboot_threads_lock);
326 put_online_cpus();
327 free_cpumask_var(plug_thread->cpumask);
328}
329EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
330
331/**
332 * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
333 * @plug_thread: Hotplug thread descriptor
334 * @new: Revised mask to use
335 *
336 * The cpumask field in the smp_hotplug_thread must not be updated directly
337 * by the client, but only by calling this function.
338 * This function can only be called on a registered smp_hotplug_thread.
339 */
340int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
341 const struct cpumask *new)
342{
343 struct cpumask *old = plug_thread->cpumask;
344 cpumask_var_t tmp;
345 unsigned int cpu;
346
347 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
348 return -ENOMEM;
349
350 get_online_cpus();
351 mutex_lock(&smpboot_threads_lock);
352
353 /* Park threads that were exclusively enabled on the old mask. */
354 cpumask_andnot(tmp, old, new);
355 for_each_cpu_and(cpu, tmp, cpu_online_mask)
356 smpboot_park_thread(plug_thread, cpu);
357
358 /* Unpark threads that are exclusively enabled on the new mask. */
359 cpumask_andnot(tmp, new, old);
360 for_each_cpu_and(cpu, tmp, cpu_online_mask)
361 smpboot_unpark_thread(plug_thread, cpu);
362
363 cpumask_copy(old, new);
364
365 mutex_unlock(&smpboot_threads_lock);
366 put_online_cpus();
367
368 free_cpumask_var(tmp);
369
370 return 0;
371}
372EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread);
373
374static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
375
376/*
377 * Called to poll specified CPU's state, for example, when waiting for
378 * a CPU to come online.
379 */
380int cpu_report_state(int cpu)
381{
382 return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
383}
384
385/*
386 * If CPU has died properly, set its state to CPU_UP_PREPARE and
387 * return success. Otherwise, return -EBUSY if the CPU died after
388 * cpu_wait_death() timed out. And yet otherwise again, return -EAGAIN
389 * if cpu_wait_death() timed out and the CPU still hasn't gotten around
390 * to dying. In the latter two cases, the CPU might not be set up
391 * properly, but it is up to the arch-specific code to decide.
392 * Finally, -EIO indicates an unanticipated problem.
393 *
394 * Note that it is permissible to omit this call entirely, as is
395 * done in architectures that do no CPU-hotplug error checking.
396 */
397int cpu_check_up_prepare(int cpu)
398{
399 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
400 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
401 return 0;
402 }
403
404 switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
405
406 case CPU_POST_DEAD:
407
408 /* The CPU died properly, so just start it up again. */
409 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
410 return 0;
411
412 case CPU_DEAD_FROZEN:
413
414 /*
415 * Timeout during CPU death, so let caller know.
416 * The outgoing CPU completed its processing, but after
417 * cpu_wait_death() timed out and reported the error. The
418 * caller is free to proceed, in which case the state
419 * will be reset properly by cpu_set_state_online().
420 * Proceeding despite this -EBUSY return makes sense
421 * for systems where the outgoing CPUs take themselves
422 * offline, with no post-death manipulation required from
423 * a surviving CPU.
424 */
425 return -EBUSY;
426
427 case CPU_BROKEN:
428
429 /*
430 * The most likely reason we got here is that there was
431 * a timeout during CPU death, and the outgoing CPU never
432 * did complete its processing. This could happen on
433 * a virtualized system if the outgoing VCPU gets preempted
434 * for more than five seconds, and the user attempts to
435 * immediately online that same CPU. Trying again later
436 * might return -EBUSY above, hence -EAGAIN.
437 */
438 return -EAGAIN;
439
440 default:
441
442 /* Should not happen. Famous last words. */
443 return -EIO;
444 }
445}
446
447/*
448 * Mark the specified CPU online.
449 *
450 * Note that it is permissible to omit this call entirely, as is
451 * done in architectures that do no CPU-hotplug error checking.
452 */
453void cpu_set_state_online(int cpu)
454{
455 (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
456}
457
458#ifdef CONFIG_HOTPLUG_CPU
459
460/*
461 * Wait for the specified CPU to exit the idle loop and die.
462 */
463bool cpu_wait_death(unsigned int cpu, int seconds)
464{
465 int jf_left = seconds * HZ;
466 int oldstate;
467 bool ret = true;
468 int sleep_jf = 1;
469
470 might_sleep();
471
472 /* The outgoing CPU will normally get done quite quickly. */
473 if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
474 goto update_state;
475 udelay(5);
476
477 /* But if the outgoing CPU dawdles, wait increasingly long times. */
478 while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
479 schedule_timeout_uninterruptible(sleep_jf);
480 jf_left -= sleep_jf;
481 if (jf_left <= 0)
482 break;
483 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
484 }
485update_state:
486 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
487 if (oldstate == CPU_DEAD) {
488 /* Outgoing CPU died normally, update state. */
489 smp_mb(); /* atomic_read() before update. */
490 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
491 } else {
492 /* Outgoing CPU still hasn't died, set state accordingly. */
493 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
494 oldstate, CPU_BROKEN) != oldstate)
495 goto update_state;
496 ret = false;
497 }
498 return ret;
499}
500
501/*
502 * Called by the outgoing CPU to report its successful death. Return
503 * false if this report follows the surviving CPU's timing out.
504 *
505 * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
506 * timed out. This approach allows architectures to omit calls to
507 * cpu_check_up_prepare() and cpu_set_state_online() without defeating
508 * the next cpu_wait_death()'s polling loop.
509 */
510bool cpu_report_death(void)
511{
512 int oldstate;
513 int newstate;
514 int cpu = smp_processor_id();
515
516 do {
517 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
518 if (oldstate != CPU_BROKEN)
519 newstate = CPU_DEAD;
520 else
521 newstate = CPU_DEAD_FROZEN;
522 } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
523 oldstate, newstate) != oldstate);
524 return newstate == CPU_DEAD;
525}
526
527#endif /* #ifdef CONFIG_HOTPLUG_CPU */