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			if (ht->cleanup)
117				ht->cleanup(td->cpu, cpu_online(td->cpu));
118			kfree(td);
119			return 0;
120		}
121
122		if (kthread_should_park()) {
123			__set_current_state(TASK_RUNNING);
124			preempt_enable();
125			if (ht->park && td->status == HP_THREAD_ACTIVE) {
126				BUG_ON(td->cpu != smp_processor_id());
127				ht->park(td->cpu);
128				td->status = HP_THREAD_PARKED;
129			}
130			kthread_parkme();
131			/* We might have been woken for stop */
132			continue;
133		}
134
135		BUG_ON(td->cpu != smp_processor_id());
136
137		/* Check for state change setup */
138		switch (td->status) {
139		case HP_THREAD_NONE:
140			__set_current_state(TASK_RUNNING);
141			preempt_enable();
142			if (ht->setup)
143				ht->setup(td->cpu);
144			td->status = HP_THREAD_ACTIVE;
145			continue;
146
147		case HP_THREAD_PARKED:
148			__set_current_state(TASK_RUNNING);
149			preempt_enable();
150			if (ht->unpark)
151				ht->unpark(td->cpu);
152			td->status = HP_THREAD_ACTIVE;
153			continue;
154		}
155
156		if (!ht->thread_should_run(td->cpu)) {
157			preempt_enable_no_resched();
158			schedule();
159		} else {
160			__set_current_state(TASK_RUNNING);
161			preempt_enable();
162			ht->thread_fn(td->cpu);
163		}
164	}
165}
166
167static int
168__smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
169{
170	struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
171	struct smpboot_thread_data *td;
172
173	if (tsk)
174		return 0;
175
176	td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
177	if (!td)
178		return -ENOMEM;
179	td->cpu = cpu;
180	td->ht = ht;
181
182	tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
183				    ht->thread_comm);
184	if (IS_ERR(tsk)) {
185		kfree(td);
186		return PTR_ERR(tsk);
187	}
188	get_task_struct(tsk);
189	*per_cpu_ptr(ht->store, cpu) = tsk;
190	if (ht->create) {
191		/*
192		 * Make sure that the task has actually scheduled out
193		 * into park position, before calling the create
194		 * callback. At least the migration thread callback
195		 * requires that the task is off the runqueue.
196		 */
197		if (!wait_task_inactive(tsk, TASK_PARKED))
198			WARN_ON(1);
199		else
200			ht->create(cpu);
201	}
202	return 0;
203}
204
205int smpboot_create_threads(unsigned int cpu)
206{
207	struct smp_hotplug_thread *cur;
208	int ret = 0;
209
210	mutex_lock(&smpboot_threads_lock);
211	list_for_each_entry(cur, &hotplug_threads, list) {
212		ret = __smpboot_create_thread(cur, cpu);
213		if (ret)
214			break;
215	}
216	mutex_unlock(&smpboot_threads_lock);
217	return ret;
218}
219
220static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
221{
222	struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
223
224	if (ht->pre_unpark)
225		ht->pre_unpark(cpu);
226	kthread_unpark(tsk);
227}
228
229void smpboot_unpark_threads(unsigned int cpu)
230{
231	struct smp_hotplug_thread *cur;
232
233	mutex_lock(&smpboot_threads_lock);
234	list_for_each_entry(cur, &hotplug_threads, list)
235		smpboot_unpark_thread(cur, cpu);
236	mutex_unlock(&smpboot_threads_lock);
237}
238
239static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
240{
241	struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
242
243	if (tsk && !ht->selfparking)
244		kthread_park(tsk);
245}
246
247void smpboot_park_threads(unsigned int cpu)
248{
249	struct smp_hotplug_thread *cur;
250
251	mutex_lock(&smpboot_threads_lock);
252	list_for_each_entry_reverse(cur, &hotplug_threads, list)
253		smpboot_park_thread(cur, cpu);
254	mutex_unlock(&smpboot_threads_lock);
255}
256
257static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
258{
259	unsigned int cpu;
260
261	/* We need to destroy also the parked threads of offline cpus */
262	for_each_possible_cpu(cpu) {
263		struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
264
265		if (tsk) {
266			kthread_stop(tsk);
267			put_task_struct(tsk);
268			*per_cpu_ptr(ht->store, cpu) = NULL;
269		}
270	}
271}
272
273/**
274 * smpboot_register_percpu_thread - Register a per_cpu thread related to hotplug
275 * @plug_thread:	Hotplug thread descriptor
276 *
277 * Creates and starts the threads on all online cpus.
278 */
279int smpboot_register_percpu_thread(struct smp_hotplug_thread *plug_thread)
280{
281	unsigned int cpu;
282	int ret = 0;
283
284	get_online_cpus();
285	mutex_lock(&smpboot_threads_lock);
286	for_each_online_cpu(cpu) {
287		ret = __smpboot_create_thread(plug_thread, cpu);
288		if (ret) {
289			smpboot_destroy_threads(plug_thread);
290			goto out;
291		}
292		smpboot_unpark_thread(plug_thread, cpu);
293	}
294	list_add(&plug_thread->list, &hotplug_threads);
295out:
296	mutex_unlock(&smpboot_threads_lock);
297	put_online_cpus();
298	return ret;
299}
300EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
301
302/**
303 * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
304 * @plug_thread:	Hotplug thread descriptor
305 *
306 * Stops all threads on all possible cpus.
307 */
308void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
309{
310	get_online_cpus();
311	mutex_lock(&smpboot_threads_lock);
312	list_del(&plug_thread->list);
313	smpboot_destroy_threads(plug_thread);
314	mutex_unlock(&smpboot_threads_lock);
315	put_online_cpus();
316}
317EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
318
319static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
320
321/*
322 * Called to poll specified CPU's state, for example, when waiting for
323 * a CPU to come online.
324 */
325int cpu_report_state(int cpu)
326{
327	return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
328}
329
330/*
331 * If CPU has died properly, set its state to CPU_UP_PREPARE and
332 * return success.  Otherwise, return -EBUSY if the CPU died after
333 * cpu_wait_death() timed out.  And yet otherwise again, return -EAGAIN
334 * if cpu_wait_death() timed out and the CPU still hasn't gotten around
335 * to dying.  In the latter two cases, the CPU might not be set up
336 * properly, but it is up to the arch-specific code to decide.
337 * Finally, -EIO indicates an unanticipated problem.
338 *
339 * Note that it is permissible to omit this call entirely, as is
340 * done in architectures that do no CPU-hotplug error checking.
341 */
342int cpu_check_up_prepare(int cpu)
343{
344	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
345		atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
346		return 0;
347	}
348
349	switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
350
351	case CPU_POST_DEAD:
352
353		/* The CPU died properly, so just start it up again. */
354		atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
355		return 0;
356
357	case CPU_DEAD_FROZEN:
358
359		/*
360		 * Timeout during CPU death, so let caller know.
361		 * The outgoing CPU completed its processing, but after
362		 * cpu_wait_death() timed out and reported the error. The
363		 * caller is free to proceed, in which case the state
364		 * will be reset properly by cpu_set_state_online().
365		 * Proceeding despite this -EBUSY return makes sense
366		 * for systems where the outgoing CPUs take themselves
367		 * offline, with no post-death manipulation required from
368		 * a surviving CPU.
369		 */
370		return -EBUSY;
371
372	case CPU_BROKEN:
373
374		/*
375		 * The most likely reason we got here is that there was
376		 * a timeout during CPU death, and the outgoing CPU never
377		 * did complete its processing.  This could happen on
378		 * a virtualized system if the outgoing VCPU gets preempted
379		 * for more than five seconds, and the user attempts to
380		 * immediately online that same CPU.  Trying again later
381		 * might return -EBUSY above, hence -EAGAIN.
382		 */
383		return -EAGAIN;
384
385	default:
386
387		/* Should not happen.  Famous last words. */
388		return -EIO;
389	}
390}
391
392/*
393 * Mark the specified CPU online.
394 *
395 * Note that it is permissible to omit this call entirely, as is
396 * done in architectures that do no CPU-hotplug error checking.
397 */
398void cpu_set_state_online(int cpu)
399{
400	(void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
401}
402
403#ifdef CONFIG_HOTPLUG_CPU
404
405/*
406 * Wait for the specified CPU to exit the idle loop and die.
407 */
408bool cpu_wait_death(unsigned int cpu, int seconds)
409{
410	int jf_left = seconds * HZ;
411	int oldstate;
412	bool ret = true;
413	int sleep_jf = 1;
414
415	might_sleep();
416
417	/* The outgoing CPU will normally get done quite quickly. */
418	if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
419		goto update_state;
420	udelay(5);
421
422	/* But if the outgoing CPU dawdles, wait increasingly long times. */
423	while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
424		schedule_timeout_uninterruptible(sleep_jf);
425		jf_left -= sleep_jf;
426		if (jf_left <= 0)
427			break;
428		sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
429	}
430update_state:
431	oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
432	if (oldstate == CPU_DEAD) {
433		/* Outgoing CPU died normally, update state. */
434		smp_mb(); /* atomic_read() before update. */
435		atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
436	} else {
437		/* Outgoing CPU still hasn't died, set state accordingly. */
438		if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
439				   oldstate, CPU_BROKEN) != oldstate)
440			goto update_state;
441		ret = false;
442	}
443	return ret;
444}
445
446/*
447 * Called by the outgoing CPU to report its successful death.  Return
448 * false if this report follows the surviving CPU's timing out.
449 *
450 * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
451 * timed out.  This approach allows architectures to omit calls to
452 * cpu_check_up_prepare() and cpu_set_state_online() without defeating
453 * the next cpu_wait_death()'s polling loop.
454 */
455bool cpu_report_death(void)
456{
457	int oldstate;
458	int newstate;
459	int cpu = smp_processor_id();
460
461	do {
462		oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
463		if (oldstate != CPU_BROKEN)
464			newstate = CPU_DEAD;
465		else
466			newstate = CPU_DEAD_FROZEN;
467	} while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
468				oldstate, newstate) != oldstate);
469	return newstate == CPU_DEAD;
470}
471
472#endif /* #ifdef CONFIG_HOTPLUG_CPU */
473