]>
Commit | Line | Data |
---|---|---|
1 | /* CPU control. | |
2 | * (C) 2001, 2002, 2003, 2004 Rusty Russell | |
3 | * | |
4 | * This code is licenced under the GPL. | |
5 | */ | |
6 | #include <linux/proc_fs.h> | |
7 | #include <linux/smp.h> | |
8 | #include <linux/init.h> | |
9 | #include <linux/notifier.h> | |
10 | #include <linux/sched/signal.h> | |
11 | #include <linux/sched/hotplug.h> | |
12 | #include <linux/sched/task.h> | |
13 | #include <linux/unistd.h> | |
14 | #include <linux/cpu.h> | |
15 | #include <linux/oom.h> | |
16 | #include <linux/rcupdate.h> | |
17 | #include <linux/export.h> | |
18 | #include <linux/bug.h> | |
19 | #include <linux/kthread.h> | |
20 | #include <linux/stop_machine.h> | |
21 | #include <linux/mutex.h> | |
22 | #include <linux/gfp.h> | |
23 | #include <linux/suspend.h> | |
24 | #include <linux/lockdep.h> | |
25 | #include <linux/tick.h> | |
26 | #include <linux/irq.h> | |
27 | #include <linux/nmi.h> | |
28 | #include <linux/smpboot.h> | |
29 | #include <linux/relay.h> | |
30 | #include <linux/slab.h> | |
31 | #include <linux/percpu-rwsem.h> | |
32 | ||
33 | #include <trace/events/power.h> | |
34 | #define CREATE_TRACE_POINTS | |
35 | #include <trace/events/cpuhp.h> | |
36 | ||
37 | #include "smpboot.h" | |
38 | ||
39 | /** | |
40 | * cpuhp_cpu_state - Per cpu hotplug state storage | |
41 | * @state: The current cpu state | |
42 | * @target: The target state | |
43 | * @thread: Pointer to the hotplug thread | |
44 | * @should_run: Thread should execute | |
45 | * @rollback: Perform a rollback | |
46 | * @single: Single callback invocation | |
47 | * @bringup: Single callback bringup or teardown selector | |
48 | * @cb_state: The state for a single callback (install/uninstall) | |
49 | * @result: Result of the operation | |
50 | * @done_up: Signal completion to the issuer of the task for cpu-up | |
51 | * @done_down: Signal completion to the issuer of the task for cpu-down | |
52 | */ | |
53 | struct cpuhp_cpu_state { | |
54 | enum cpuhp_state state; | |
55 | enum cpuhp_state target; | |
56 | enum cpuhp_state fail; | |
57 | #ifdef CONFIG_SMP | |
58 | struct task_struct *thread; | |
59 | bool should_run; | |
60 | bool rollback; | |
61 | bool single; | |
62 | bool bringup; | |
63 | bool booted_once; | |
64 | struct hlist_node *node; | |
65 | struct hlist_node *last; | |
66 | enum cpuhp_state cb_state; | |
67 | int result; | |
68 | struct completion done_up; | |
69 | struct completion done_down; | |
70 | #endif | |
71 | }; | |
72 | ||
73 | static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = { | |
74 | .fail = CPUHP_INVALID, | |
75 | }; | |
76 | ||
77 | #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP) | |
78 | static struct lockdep_map cpuhp_state_up_map = | |
79 | STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map); | |
80 | static struct lockdep_map cpuhp_state_down_map = | |
81 | STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map); | |
82 | ||
83 | ||
84 | static inline void cpuhp_lock_acquire(bool bringup) | |
85 | { | |
86 | lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map); | |
87 | } | |
88 | ||
89 | static inline void cpuhp_lock_release(bool bringup) | |
90 | { | |
91 | lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map); | |
92 | } | |
93 | #else | |
94 | ||
95 | static inline void cpuhp_lock_acquire(bool bringup) { } | |
96 | static inline void cpuhp_lock_release(bool bringup) { } | |
97 | ||
98 | #endif | |
99 | ||
100 | /** | |
101 | * cpuhp_step - Hotplug state machine step | |
102 | * @name: Name of the step | |
103 | * @startup: Startup function of the step | |
104 | * @teardown: Teardown function of the step | |
105 | * @cant_stop: Bringup/teardown can't be stopped at this step | |
106 | */ | |
107 | struct cpuhp_step { | |
108 | const char *name; | |
109 | union { | |
110 | int (*single)(unsigned int cpu); | |
111 | int (*multi)(unsigned int cpu, | |
112 | struct hlist_node *node); | |
113 | } startup; | |
114 | union { | |
115 | int (*single)(unsigned int cpu); | |
116 | int (*multi)(unsigned int cpu, | |
117 | struct hlist_node *node); | |
118 | } teardown; | |
119 | struct hlist_head list; | |
120 | bool cant_stop; | |
121 | bool multi_instance; | |
122 | }; | |
123 | ||
124 | static DEFINE_MUTEX(cpuhp_state_mutex); | |
125 | static struct cpuhp_step cpuhp_hp_states[]; | |
126 | ||
127 | static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state) | |
128 | { | |
129 | return cpuhp_hp_states + state; | |
130 | } | |
131 | ||
132 | /** | |
133 | * cpuhp_invoke_callback _ Invoke the callbacks for a given state | |
134 | * @cpu: The cpu for which the callback should be invoked | |
135 | * @state: The state to do callbacks for | |
136 | * @bringup: True if the bringup callback should be invoked | |
137 | * @node: For multi-instance, do a single entry callback for install/remove | |
138 | * @lastp: For multi-instance rollback, remember how far we got | |
139 | * | |
140 | * Called from cpu hotplug and from the state register machinery. | |
141 | */ | |
142 | static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state, | |
143 | bool bringup, struct hlist_node *node, | |
144 | struct hlist_node **lastp) | |
145 | { | |
146 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
147 | struct cpuhp_step *step = cpuhp_get_step(state); | |
148 | int (*cbm)(unsigned int cpu, struct hlist_node *node); | |
149 | int (*cb)(unsigned int cpu); | |
150 | int ret, cnt; | |
151 | ||
152 | if (st->fail == state) { | |
153 | st->fail = CPUHP_INVALID; | |
154 | ||
155 | if (!(bringup ? step->startup.single : step->teardown.single)) | |
156 | return 0; | |
157 | ||
158 | return -EAGAIN; | |
159 | } | |
160 | ||
161 | if (!step->multi_instance) { | |
162 | WARN_ON_ONCE(lastp && *lastp); | |
163 | cb = bringup ? step->startup.single : step->teardown.single; | |
164 | if (!cb) | |
165 | return 0; | |
166 | trace_cpuhp_enter(cpu, st->target, state, cb); | |
167 | ret = cb(cpu); | |
168 | trace_cpuhp_exit(cpu, st->state, state, ret); | |
169 | return ret; | |
170 | } | |
171 | cbm = bringup ? step->startup.multi : step->teardown.multi; | |
172 | if (!cbm) | |
173 | return 0; | |
174 | ||
175 | /* Single invocation for instance add/remove */ | |
176 | if (node) { | |
177 | WARN_ON_ONCE(lastp && *lastp); | |
178 | trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node); | |
179 | ret = cbm(cpu, node); | |
180 | trace_cpuhp_exit(cpu, st->state, state, ret); | |
181 | return ret; | |
182 | } | |
183 | ||
184 | /* State transition. Invoke on all instances */ | |
185 | cnt = 0; | |
186 | hlist_for_each(node, &step->list) { | |
187 | if (lastp && node == *lastp) | |
188 | break; | |
189 | ||
190 | trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node); | |
191 | ret = cbm(cpu, node); | |
192 | trace_cpuhp_exit(cpu, st->state, state, ret); | |
193 | if (ret) { | |
194 | if (!lastp) | |
195 | goto err; | |
196 | ||
197 | *lastp = node; | |
198 | return ret; | |
199 | } | |
200 | cnt++; | |
201 | } | |
202 | if (lastp) | |
203 | *lastp = NULL; | |
204 | return 0; | |
205 | err: | |
206 | /* Rollback the instances if one failed */ | |
207 | cbm = !bringup ? step->startup.multi : step->teardown.multi; | |
208 | if (!cbm) | |
209 | return ret; | |
210 | ||
211 | hlist_for_each(node, &step->list) { | |
212 | if (!cnt--) | |
213 | break; | |
214 | ||
215 | trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node); | |
216 | ret = cbm(cpu, node); | |
217 | trace_cpuhp_exit(cpu, st->state, state, ret); | |
218 | /* | |
219 | * Rollback must not fail, | |
220 | */ | |
221 | WARN_ON_ONCE(ret); | |
222 | } | |
223 | return ret; | |
224 | } | |
225 | ||
226 | #ifdef CONFIG_SMP | |
227 | static bool cpuhp_is_ap_state(enum cpuhp_state state) | |
228 | { | |
229 | /* | |
230 | * The extra check for CPUHP_TEARDOWN_CPU is only for documentation | |
231 | * purposes as that state is handled explicitly in cpu_down. | |
232 | */ | |
233 | return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU; | |
234 | } | |
235 | ||
236 | static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup) | |
237 | { | |
238 | struct completion *done = bringup ? &st->done_up : &st->done_down; | |
239 | wait_for_completion(done); | |
240 | } | |
241 | ||
242 | static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup) | |
243 | { | |
244 | struct completion *done = bringup ? &st->done_up : &st->done_down; | |
245 | complete(done); | |
246 | } | |
247 | ||
248 | /* | |
249 | * The former STARTING/DYING states, ran with IRQs disabled and must not fail. | |
250 | */ | |
251 | static bool cpuhp_is_atomic_state(enum cpuhp_state state) | |
252 | { | |
253 | return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE; | |
254 | } | |
255 | ||
256 | /* Serializes the updates to cpu_online_mask, cpu_present_mask */ | |
257 | static DEFINE_MUTEX(cpu_add_remove_lock); | |
258 | bool cpuhp_tasks_frozen; | |
259 | EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen); | |
260 | ||
261 | /* | |
262 | * The following two APIs (cpu_maps_update_begin/done) must be used when | |
263 | * attempting to serialize the updates to cpu_online_mask & cpu_present_mask. | |
264 | */ | |
265 | void cpu_maps_update_begin(void) | |
266 | { | |
267 | mutex_lock(&cpu_add_remove_lock); | |
268 | } | |
269 | ||
270 | void cpu_maps_update_done(void) | |
271 | { | |
272 | mutex_unlock(&cpu_add_remove_lock); | |
273 | } | |
274 | ||
275 | /* | |
276 | * If set, cpu_up and cpu_down will return -EBUSY and do nothing. | |
277 | * Should always be manipulated under cpu_add_remove_lock | |
278 | */ | |
279 | static int cpu_hotplug_disabled; | |
280 | ||
281 | #ifdef CONFIG_HOTPLUG_CPU | |
282 | ||
283 | DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock); | |
284 | ||
285 | void cpus_read_lock(void) | |
286 | { | |
287 | percpu_down_read(&cpu_hotplug_lock); | |
288 | } | |
289 | EXPORT_SYMBOL_GPL(cpus_read_lock); | |
290 | ||
291 | int cpus_read_trylock(void) | |
292 | { | |
293 | return percpu_down_read_trylock(&cpu_hotplug_lock); | |
294 | } | |
295 | EXPORT_SYMBOL_GPL(cpus_read_trylock); | |
296 | ||
297 | void cpus_read_unlock(void) | |
298 | { | |
299 | percpu_up_read(&cpu_hotplug_lock); | |
300 | } | |
301 | EXPORT_SYMBOL_GPL(cpus_read_unlock); | |
302 | ||
303 | void cpus_write_lock(void) | |
304 | { | |
305 | percpu_down_write(&cpu_hotplug_lock); | |
306 | } | |
307 | ||
308 | void cpus_write_unlock(void) | |
309 | { | |
310 | percpu_up_write(&cpu_hotplug_lock); | |
311 | } | |
312 | ||
313 | void lockdep_assert_cpus_held(void) | |
314 | { | |
315 | percpu_rwsem_assert_held(&cpu_hotplug_lock); | |
316 | } | |
317 | ||
318 | static void lockdep_acquire_cpus_lock(void) | |
319 | { | |
320 | rwsem_acquire(&cpu_hotplug_lock.rw_sem.dep_map, 0, 0, _THIS_IP_); | |
321 | } | |
322 | ||
323 | static void lockdep_release_cpus_lock(void) | |
324 | { | |
325 | rwsem_release(&cpu_hotplug_lock.rw_sem.dep_map, 1, _THIS_IP_); | |
326 | } | |
327 | ||
328 | /* | |
329 | * Wait for currently running CPU hotplug operations to complete (if any) and | |
330 | * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects | |
331 | * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the | |
332 | * hotplug path before performing hotplug operations. So acquiring that lock | |
333 | * guarantees mutual exclusion from any currently running hotplug operations. | |
334 | */ | |
335 | void cpu_hotplug_disable(void) | |
336 | { | |
337 | cpu_maps_update_begin(); | |
338 | cpu_hotplug_disabled++; | |
339 | cpu_maps_update_done(); | |
340 | } | |
341 | EXPORT_SYMBOL_GPL(cpu_hotplug_disable); | |
342 | ||
343 | static void __cpu_hotplug_enable(void) | |
344 | { | |
345 | if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n")) | |
346 | return; | |
347 | cpu_hotplug_disabled--; | |
348 | } | |
349 | ||
350 | void cpu_hotplug_enable(void) | |
351 | { | |
352 | cpu_maps_update_begin(); | |
353 | __cpu_hotplug_enable(); | |
354 | cpu_maps_update_done(); | |
355 | } | |
356 | EXPORT_SYMBOL_GPL(cpu_hotplug_enable); | |
357 | ||
358 | #else | |
359 | ||
360 | static void lockdep_acquire_cpus_lock(void) | |
361 | { | |
362 | } | |
363 | ||
364 | static void lockdep_release_cpus_lock(void) | |
365 | { | |
366 | } | |
367 | ||
368 | #endif /* CONFIG_HOTPLUG_CPU */ | |
369 | ||
370 | #ifdef CONFIG_HOTPLUG_SMT | |
371 | enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED; | |
372 | EXPORT_SYMBOL_GPL(cpu_smt_control); | |
373 | ||
374 | static bool cpu_smt_available __read_mostly; | |
375 | ||
376 | void __init cpu_smt_disable(bool force) | |
377 | { | |
378 | if (cpu_smt_control == CPU_SMT_FORCE_DISABLED || | |
379 | cpu_smt_control == CPU_SMT_NOT_SUPPORTED) | |
380 | return; | |
381 | ||
382 | if (force) { | |
383 | pr_info("SMT: Force disabled\n"); | |
384 | cpu_smt_control = CPU_SMT_FORCE_DISABLED; | |
385 | } else { | |
386 | pr_info("SMT: disabled\n"); | |
387 | cpu_smt_control = CPU_SMT_DISABLED; | |
388 | } | |
389 | } | |
390 | ||
391 | /* | |
392 | * The decision whether SMT is supported can only be done after the full | |
393 | * CPU identification. Called from architecture code before non boot CPUs | |
394 | * are brought up. | |
395 | */ | |
396 | void __init cpu_smt_check_topology_early(void) | |
397 | { | |
398 | if (!topology_smt_supported()) | |
399 | cpu_smt_control = CPU_SMT_NOT_SUPPORTED; | |
400 | } | |
401 | ||
402 | /* | |
403 | * If SMT was disabled by BIOS, detect it here, after the CPUs have been | |
404 | * brought online. This ensures the smt/l1tf sysfs entries are consistent | |
405 | * with reality. cpu_smt_available is set to true during the bringup of non | |
406 | * boot CPUs when a SMT sibling is detected. Note, this may overwrite | |
407 | * cpu_smt_control's previous setting. | |
408 | */ | |
409 | void __init cpu_smt_check_topology(void) | |
410 | { | |
411 | if (!cpu_smt_available) | |
412 | cpu_smt_control = CPU_SMT_NOT_SUPPORTED; | |
413 | } | |
414 | ||
415 | static int __init smt_cmdline_disable(char *str) | |
416 | { | |
417 | cpu_smt_disable(str && !strcmp(str, "force")); | |
418 | return 0; | |
419 | } | |
420 | early_param("nosmt", smt_cmdline_disable); | |
421 | ||
422 | static inline bool cpu_smt_allowed(unsigned int cpu) | |
423 | { | |
424 | if (topology_is_primary_thread(cpu)) | |
425 | return true; | |
426 | ||
427 | /* | |
428 | * If the CPU is not a 'primary' thread and the booted_once bit is | |
429 | * set then the processor has SMT support. Store this information | |
430 | * for the late check of SMT support in cpu_smt_check_topology(). | |
431 | */ | |
432 | if (per_cpu(cpuhp_state, cpu).booted_once) | |
433 | cpu_smt_available = true; | |
434 | ||
435 | if (cpu_smt_control == CPU_SMT_ENABLED) | |
436 | return true; | |
437 | ||
438 | /* | |
439 | * On x86 it's required to boot all logical CPUs at least once so | |
440 | * that the init code can get a chance to set CR4.MCE on each | |
441 | * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any | |
442 | * core will shutdown the machine. | |
443 | */ | |
444 | return !per_cpu(cpuhp_state, cpu).booted_once; | |
445 | } | |
446 | #else | |
447 | static inline bool cpu_smt_allowed(unsigned int cpu) { return true; } | |
448 | #endif | |
449 | ||
450 | static inline enum cpuhp_state | |
451 | cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target) | |
452 | { | |
453 | enum cpuhp_state prev_state = st->state; | |
454 | ||
455 | st->rollback = false; | |
456 | st->last = NULL; | |
457 | ||
458 | st->target = target; | |
459 | st->single = false; | |
460 | st->bringup = st->state < target; | |
461 | ||
462 | return prev_state; | |
463 | } | |
464 | ||
465 | static inline void | |
466 | cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state) | |
467 | { | |
468 | st->rollback = true; | |
469 | ||
470 | /* | |
471 | * If we have st->last we need to undo partial multi_instance of this | |
472 | * state first. Otherwise start undo at the previous state. | |
473 | */ | |
474 | if (!st->last) { | |
475 | if (st->bringup) | |
476 | st->state--; | |
477 | else | |
478 | st->state++; | |
479 | } | |
480 | ||
481 | st->target = prev_state; | |
482 | st->bringup = !st->bringup; | |
483 | } | |
484 | ||
485 | /* Regular hotplug invocation of the AP hotplug thread */ | |
486 | static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st) | |
487 | { | |
488 | if (!st->single && st->state == st->target) | |
489 | return; | |
490 | ||
491 | st->result = 0; | |
492 | /* | |
493 | * Make sure the above stores are visible before should_run becomes | |
494 | * true. Paired with the mb() above in cpuhp_thread_fun() | |
495 | */ | |
496 | smp_mb(); | |
497 | st->should_run = true; | |
498 | wake_up_process(st->thread); | |
499 | wait_for_ap_thread(st, st->bringup); | |
500 | } | |
501 | ||
502 | static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target) | |
503 | { | |
504 | enum cpuhp_state prev_state; | |
505 | int ret; | |
506 | ||
507 | prev_state = cpuhp_set_state(st, target); | |
508 | __cpuhp_kick_ap(st); | |
509 | if ((ret = st->result)) { | |
510 | cpuhp_reset_state(st, prev_state); | |
511 | __cpuhp_kick_ap(st); | |
512 | } | |
513 | ||
514 | return ret; | |
515 | } | |
516 | ||
517 | static int bringup_wait_for_ap(unsigned int cpu) | |
518 | { | |
519 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
520 | ||
521 | /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */ | |
522 | wait_for_ap_thread(st, true); | |
523 | if (WARN_ON_ONCE((!cpu_online(cpu)))) | |
524 | return -ECANCELED; | |
525 | ||
526 | /* Unpark the stopper thread and the hotplug thread of the target cpu */ | |
527 | stop_machine_unpark(cpu); | |
528 | kthread_unpark(st->thread); | |
529 | ||
530 | /* | |
531 | * SMT soft disabling on X86 requires to bring the CPU out of the | |
532 | * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The | |
533 | * CPU marked itself as booted_once in cpu_notify_starting() so the | |
534 | * cpu_smt_allowed() check will now return false if this is not the | |
535 | * primary sibling. | |
536 | */ | |
537 | if (!cpu_smt_allowed(cpu)) | |
538 | return -ECANCELED; | |
539 | ||
540 | if (st->target <= CPUHP_AP_ONLINE_IDLE) | |
541 | return 0; | |
542 | ||
543 | return cpuhp_kick_ap(st, st->target); | |
544 | } | |
545 | ||
546 | static int bringup_cpu(unsigned int cpu) | |
547 | { | |
548 | struct task_struct *idle = idle_thread_get(cpu); | |
549 | int ret; | |
550 | ||
551 | /* | |
552 | * Some architectures have to walk the irq descriptors to | |
553 | * setup the vector space for the cpu which comes online. | |
554 | * Prevent irq alloc/free across the bringup. | |
555 | */ | |
556 | irq_lock_sparse(); | |
557 | ||
558 | /* Arch-specific enabling code. */ | |
559 | ret = __cpu_up(cpu, idle); | |
560 | irq_unlock_sparse(); | |
561 | if (ret) | |
562 | return ret; | |
563 | return bringup_wait_for_ap(cpu); | |
564 | } | |
565 | ||
566 | /* | |
567 | * Hotplug state machine related functions | |
568 | */ | |
569 | ||
570 | static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st) | |
571 | { | |
572 | for (st->state--; st->state > st->target; st->state--) | |
573 | cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL); | |
574 | } | |
575 | ||
576 | static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st, | |
577 | enum cpuhp_state target) | |
578 | { | |
579 | enum cpuhp_state prev_state = st->state; | |
580 | int ret = 0; | |
581 | ||
582 | while (st->state < target) { | |
583 | st->state++; | |
584 | ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL); | |
585 | if (ret) { | |
586 | st->target = prev_state; | |
587 | undo_cpu_up(cpu, st); | |
588 | break; | |
589 | } | |
590 | } | |
591 | return ret; | |
592 | } | |
593 | ||
594 | /* | |
595 | * The cpu hotplug threads manage the bringup and teardown of the cpus | |
596 | */ | |
597 | static void cpuhp_create(unsigned int cpu) | |
598 | { | |
599 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
600 | ||
601 | init_completion(&st->done_up); | |
602 | init_completion(&st->done_down); | |
603 | } | |
604 | ||
605 | static int cpuhp_should_run(unsigned int cpu) | |
606 | { | |
607 | struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state); | |
608 | ||
609 | return st->should_run; | |
610 | } | |
611 | ||
612 | /* | |
613 | * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke | |
614 | * callbacks when a state gets [un]installed at runtime. | |
615 | * | |
616 | * Each invocation of this function by the smpboot thread does a single AP | |
617 | * state callback. | |
618 | * | |
619 | * It has 3 modes of operation: | |
620 | * - single: runs st->cb_state | |
621 | * - up: runs ++st->state, while st->state < st->target | |
622 | * - down: runs st->state--, while st->state > st->target | |
623 | * | |
624 | * When complete or on error, should_run is cleared and the completion is fired. | |
625 | */ | |
626 | static void cpuhp_thread_fun(unsigned int cpu) | |
627 | { | |
628 | struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state); | |
629 | bool bringup = st->bringup; | |
630 | enum cpuhp_state state; | |
631 | ||
632 | if (WARN_ON_ONCE(!st->should_run)) | |
633 | return; | |
634 | ||
635 | /* | |
636 | * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures | |
637 | * that if we see ->should_run we also see the rest of the state. | |
638 | */ | |
639 | smp_mb(); | |
640 | ||
641 | /* | |
642 | * The BP holds the hotplug lock, but we're now running on the AP, | |
643 | * ensure that anybody asserting the lock is held, will actually find | |
644 | * it so. | |
645 | */ | |
646 | lockdep_acquire_cpus_lock(); | |
647 | cpuhp_lock_acquire(bringup); | |
648 | ||
649 | if (st->single) { | |
650 | state = st->cb_state; | |
651 | st->should_run = false; | |
652 | } else { | |
653 | if (bringup) { | |
654 | st->state++; | |
655 | state = st->state; | |
656 | st->should_run = (st->state < st->target); | |
657 | WARN_ON_ONCE(st->state > st->target); | |
658 | } else { | |
659 | state = st->state; | |
660 | st->state--; | |
661 | st->should_run = (st->state > st->target); | |
662 | WARN_ON_ONCE(st->state < st->target); | |
663 | } | |
664 | } | |
665 | ||
666 | WARN_ON_ONCE(!cpuhp_is_ap_state(state)); | |
667 | ||
668 | if (cpuhp_is_atomic_state(state)) { | |
669 | local_irq_disable(); | |
670 | st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last); | |
671 | local_irq_enable(); | |
672 | ||
673 | /* | |
674 | * STARTING/DYING must not fail! | |
675 | */ | |
676 | WARN_ON_ONCE(st->result); | |
677 | } else { | |
678 | st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last); | |
679 | } | |
680 | ||
681 | if (st->result) { | |
682 | /* | |
683 | * If we fail on a rollback, we're up a creek without no | |
684 | * paddle, no way forward, no way back. We loose, thanks for | |
685 | * playing. | |
686 | */ | |
687 | WARN_ON_ONCE(st->rollback); | |
688 | st->should_run = false; | |
689 | } | |
690 | ||
691 | cpuhp_lock_release(bringup); | |
692 | lockdep_release_cpus_lock(); | |
693 | ||
694 | if (!st->should_run) | |
695 | complete_ap_thread(st, bringup); | |
696 | } | |
697 | ||
698 | /* Invoke a single callback on a remote cpu */ | |
699 | static int | |
700 | cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup, | |
701 | struct hlist_node *node) | |
702 | { | |
703 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
704 | int ret; | |
705 | ||
706 | if (!cpu_online(cpu)) | |
707 | return 0; | |
708 | ||
709 | cpuhp_lock_acquire(false); | |
710 | cpuhp_lock_release(false); | |
711 | ||
712 | cpuhp_lock_acquire(true); | |
713 | cpuhp_lock_release(true); | |
714 | ||
715 | /* | |
716 | * If we are up and running, use the hotplug thread. For early calls | |
717 | * we invoke the thread function directly. | |
718 | */ | |
719 | if (!st->thread) | |
720 | return cpuhp_invoke_callback(cpu, state, bringup, node, NULL); | |
721 | ||
722 | st->rollback = false; | |
723 | st->last = NULL; | |
724 | ||
725 | st->node = node; | |
726 | st->bringup = bringup; | |
727 | st->cb_state = state; | |
728 | st->single = true; | |
729 | ||
730 | __cpuhp_kick_ap(st); | |
731 | ||
732 | /* | |
733 | * If we failed and did a partial, do a rollback. | |
734 | */ | |
735 | if ((ret = st->result) && st->last) { | |
736 | st->rollback = true; | |
737 | st->bringup = !bringup; | |
738 | ||
739 | __cpuhp_kick_ap(st); | |
740 | } | |
741 | ||
742 | /* | |
743 | * Clean up the leftovers so the next hotplug operation wont use stale | |
744 | * data. | |
745 | */ | |
746 | st->node = st->last = NULL; | |
747 | return ret; | |
748 | } | |
749 | ||
750 | static int cpuhp_kick_ap_work(unsigned int cpu) | |
751 | { | |
752 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
753 | enum cpuhp_state prev_state = st->state; | |
754 | int ret; | |
755 | ||
756 | cpuhp_lock_acquire(false); | |
757 | cpuhp_lock_release(false); | |
758 | ||
759 | cpuhp_lock_acquire(true); | |
760 | cpuhp_lock_release(true); | |
761 | ||
762 | trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work); | |
763 | ret = cpuhp_kick_ap(st, st->target); | |
764 | trace_cpuhp_exit(cpu, st->state, prev_state, ret); | |
765 | ||
766 | return ret; | |
767 | } | |
768 | ||
769 | static struct smp_hotplug_thread cpuhp_threads = { | |
770 | .store = &cpuhp_state.thread, | |
771 | .create = &cpuhp_create, | |
772 | .thread_should_run = cpuhp_should_run, | |
773 | .thread_fn = cpuhp_thread_fun, | |
774 | .thread_comm = "cpuhp/%u", | |
775 | .selfparking = true, | |
776 | }; | |
777 | ||
778 | void __init cpuhp_threads_init(void) | |
779 | { | |
780 | BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads)); | |
781 | kthread_unpark(this_cpu_read(cpuhp_state.thread)); | |
782 | } | |
783 | ||
784 | #ifdef CONFIG_HOTPLUG_CPU | |
785 | /** | |
786 | * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU | |
787 | * @cpu: a CPU id | |
788 | * | |
789 | * This function walks all processes, finds a valid mm struct for each one and | |
790 | * then clears a corresponding bit in mm's cpumask. While this all sounds | |
791 | * trivial, there are various non-obvious corner cases, which this function | |
792 | * tries to solve in a safe manner. | |
793 | * | |
794 | * Also note that the function uses a somewhat relaxed locking scheme, so it may | |
795 | * be called only for an already offlined CPU. | |
796 | */ | |
797 | void clear_tasks_mm_cpumask(int cpu) | |
798 | { | |
799 | struct task_struct *p; | |
800 | ||
801 | /* | |
802 | * This function is called after the cpu is taken down and marked | |
803 | * offline, so its not like new tasks will ever get this cpu set in | |
804 | * their mm mask. -- Peter Zijlstra | |
805 | * Thus, we may use rcu_read_lock() here, instead of grabbing | |
806 | * full-fledged tasklist_lock. | |
807 | */ | |
808 | WARN_ON(cpu_online(cpu)); | |
809 | rcu_read_lock(); | |
810 | for_each_process(p) { | |
811 | struct task_struct *t; | |
812 | ||
813 | /* | |
814 | * Main thread might exit, but other threads may still have | |
815 | * a valid mm. Find one. | |
816 | */ | |
817 | t = find_lock_task_mm(p); | |
818 | if (!t) | |
819 | continue; | |
820 | cpumask_clear_cpu(cpu, mm_cpumask(t->mm)); | |
821 | task_unlock(t); | |
822 | } | |
823 | rcu_read_unlock(); | |
824 | } | |
825 | ||
826 | /* Take this CPU down. */ | |
827 | static int take_cpu_down(void *_param) | |
828 | { | |
829 | struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state); | |
830 | enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE); | |
831 | int err, cpu = smp_processor_id(); | |
832 | int ret; | |
833 | ||
834 | /* Ensure this CPU doesn't handle any more interrupts. */ | |
835 | err = __cpu_disable(); | |
836 | if (err < 0) | |
837 | return err; | |
838 | ||
839 | /* | |
840 | * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not | |
841 | * do this step again. | |
842 | */ | |
843 | WARN_ON(st->state != CPUHP_TEARDOWN_CPU); | |
844 | st->state--; | |
845 | /* Invoke the former CPU_DYING callbacks */ | |
846 | for (; st->state > target; st->state--) { | |
847 | ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL); | |
848 | /* | |
849 | * DYING must not fail! | |
850 | */ | |
851 | WARN_ON_ONCE(ret); | |
852 | } | |
853 | ||
854 | /* Give up timekeeping duties */ | |
855 | tick_handover_do_timer(); | |
856 | /* Park the stopper thread */ | |
857 | stop_machine_park(cpu); | |
858 | return 0; | |
859 | } | |
860 | ||
861 | static int takedown_cpu(unsigned int cpu) | |
862 | { | |
863 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
864 | int err; | |
865 | ||
866 | /* Park the smpboot threads */ | |
867 | kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread); | |
868 | ||
869 | /* | |
870 | * Prevent irq alloc/free while the dying cpu reorganizes the | |
871 | * interrupt affinities. | |
872 | */ | |
873 | irq_lock_sparse(); | |
874 | ||
875 | /* | |
876 | * So now all preempt/rcu users must observe !cpu_active(). | |
877 | */ | |
878 | err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu)); | |
879 | if (err) { | |
880 | /* CPU refused to die */ | |
881 | irq_unlock_sparse(); | |
882 | /* Unpark the hotplug thread so we can rollback there */ | |
883 | kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread); | |
884 | return err; | |
885 | } | |
886 | BUG_ON(cpu_online(cpu)); | |
887 | ||
888 | /* | |
889 | * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed | |
890 | * all runnable tasks from the CPU, there's only the idle task left now | |
891 | * that the migration thread is done doing the stop_machine thing. | |
892 | * | |
893 | * Wait for the stop thread to go away. | |
894 | */ | |
895 | wait_for_ap_thread(st, false); | |
896 | BUG_ON(st->state != CPUHP_AP_IDLE_DEAD); | |
897 | ||
898 | /* Interrupts are moved away from the dying cpu, reenable alloc/free */ | |
899 | irq_unlock_sparse(); | |
900 | ||
901 | hotplug_cpu__broadcast_tick_pull(cpu); | |
902 | /* This actually kills the CPU. */ | |
903 | __cpu_die(cpu); | |
904 | ||
905 | tick_cleanup_dead_cpu(cpu); | |
906 | rcutree_migrate_callbacks(cpu); | |
907 | return 0; | |
908 | } | |
909 | ||
910 | static void cpuhp_complete_idle_dead(void *arg) | |
911 | { | |
912 | struct cpuhp_cpu_state *st = arg; | |
913 | ||
914 | complete_ap_thread(st, false); | |
915 | } | |
916 | ||
917 | void cpuhp_report_idle_dead(void) | |
918 | { | |
919 | struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state); | |
920 | ||
921 | BUG_ON(st->state != CPUHP_AP_OFFLINE); | |
922 | rcu_report_dead(smp_processor_id()); | |
923 | st->state = CPUHP_AP_IDLE_DEAD; | |
924 | /* | |
925 | * We cannot call complete after rcu_report_dead() so we delegate it | |
926 | * to an online cpu. | |
927 | */ | |
928 | smp_call_function_single(cpumask_first(cpu_online_mask), | |
929 | cpuhp_complete_idle_dead, st, 0); | |
930 | } | |
931 | ||
932 | static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st) | |
933 | { | |
934 | for (st->state++; st->state < st->target; st->state++) | |
935 | cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL); | |
936 | } | |
937 | ||
938 | static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st, | |
939 | enum cpuhp_state target) | |
940 | { | |
941 | enum cpuhp_state prev_state = st->state; | |
942 | int ret = 0; | |
943 | ||
944 | for (; st->state > target; st->state--) { | |
945 | ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL); | |
946 | if (ret) { | |
947 | st->target = prev_state; | |
948 | if (st->state < prev_state) | |
949 | undo_cpu_down(cpu, st); | |
950 | break; | |
951 | } | |
952 | } | |
953 | return ret; | |
954 | } | |
955 | ||
956 | /* Requires cpu_add_remove_lock to be held */ | |
957 | static int __ref _cpu_down(unsigned int cpu, int tasks_frozen, | |
958 | enum cpuhp_state target) | |
959 | { | |
960 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
961 | int prev_state, ret = 0; | |
962 | ||
963 | if (num_online_cpus() == 1) | |
964 | return -EBUSY; | |
965 | ||
966 | if (!cpu_present(cpu)) | |
967 | return -EINVAL; | |
968 | ||
969 | cpus_write_lock(); | |
970 | ||
971 | cpuhp_tasks_frozen = tasks_frozen; | |
972 | ||
973 | prev_state = cpuhp_set_state(st, target); | |
974 | /* | |
975 | * If the current CPU state is in the range of the AP hotplug thread, | |
976 | * then we need to kick the thread. | |
977 | */ | |
978 | if (st->state > CPUHP_TEARDOWN_CPU) { | |
979 | st->target = max((int)target, CPUHP_TEARDOWN_CPU); | |
980 | ret = cpuhp_kick_ap_work(cpu); | |
981 | /* | |
982 | * The AP side has done the error rollback already. Just | |
983 | * return the error code.. | |
984 | */ | |
985 | if (ret) | |
986 | goto out; | |
987 | ||
988 | /* | |
989 | * We might have stopped still in the range of the AP hotplug | |
990 | * thread. Nothing to do anymore. | |
991 | */ | |
992 | if (st->state > CPUHP_TEARDOWN_CPU) | |
993 | goto out; | |
994 | ||
995 | st->target = target; | |
996 | } | |
997 | /* | |
998 | * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need | |
999 | * to do the further cleanups. | |
1000 | */ | |
1001 | ret = cpuhp_down_callbacks(cpu, st, target); | |
1002 | if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) { | |
1003 | cpuhp_reset_state(st, prev_state); | |
1004 | __cpuhp_kick_ap(st); | |
1005 | } | |
1006 | ||
1007 | out: | |
1008 | cpus_write_unlock(); | |
1009 | /* | |
1010 | * Do post unplug cleanup. This is still protected against | |
1011 | * concurrent CPU hotplug via cpu_add_remove_lock. | |
1012 | */ | |
1013 | lockup_detector_cleanup(); | |
1014 | return ret; | |
1015 | } | |
1016 | ||
1017 | static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target) | |
1018 | { | |
1019 | if (cpu_hotplug_disabled) | |
1020 | return -EBUSY; | |
1021 | return _cpu_down(cpu, 0, target); | |
1022 | } | |
1023 | ||
1024 | static int do_cpu_down(unsigned int cpu, enum cpuhp_state target) | |
1025 | { | |
1026 | int err; | |
1027 | ||
1028 | cpu_maps_update_begin(); | |
1029 | err = cpu_down_maps_locked(cpu, target); | |
1030 | cpu_maps_update_done(); | |
1031 | return err; | |
1032 | } | |
1033 | ||
1034 | int cpu_down(unsigned int cpu) | |
1035 | { | |
1036 | return do_cpu_down(cpu, CPUHP_OFFLINE); | |
1037 | } | |
1038 | EXPORT_SYMBOL(cpu_down); | |
1039 | ||
1040 | #else | |
1041 | #define takedown_cpu NULL | |
1042 | #endif /*CONFIG_HOTPLUG_CPU*/ | |
1043 | ||
1044 | /** | |
1045 | * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU | |
1046 | * @cpu: cpu that just started | |
1047 | * | |
1048 | * It must be called by the arch code on the new cpu, before the new cpu | |
1049 | * enables interrupts and before the "boot" cpu returns from __cpu_up(). | |
1050 | */ | |
1051 | void notify_cpu_starting(unsigned int cpu) | |
1052 | { | |
1053 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1054 | enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE); | |
1055 | int ret; | |
1056 | ||
1057 | rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */ | |
1058 | st->booted_once = true; | |
1059 | while (st->state < target) { | |
1060 | st->state++; | |
1061 | ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL); | |
1062 | /* | |
1063 | * STARTING must not fail! | |
1064 | */ | |
1065 | WARN_ON_ONCE(ret); | |
1066 | } | |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * Called from the idle task. Wake up the controlling task which brings the | |
1071 | * stopper and the hotplug thread of the upcoming CPU up and then delegates | |
1072 | * the rest of the online bringup to the hotplug thread. | |
1073 | */ | |
1074 | void cpuhp_online_idle(enum cpuhp_state state) | |
1075 | { | |
1076 | struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state); | |
1077 | ||
1078 | /* Happens for the boot cpu */ | |
1079 | if (state != CPUHP_AP_ONLINE_IDLE) | |
1080 | return; | |
1081 | ||
1082 | st->state = CPUHP_AP_ONLINE_IDLE; | |
1083 | complete_ap_thread(st, true); | |
1084 | } | |
1085 | ||
1086 | /* Requires cpu_add_remove_lock to be held */ | |
1087 | static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target) | |
1088 | { | |
1089 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1090 | struct task_struct *idle; | |
1091 | int ret = 0; | |
1092 | ||
1093 | cpus_write_lock(); | |
1094 | ||
1095 | if (!cpu_present(cpu)) { | |
1096 | ret = -EINVAL; | |
1097 | goto out; | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * The caller of do_cpu_up might have raced with another | |
1102 | * caller. Ignore it for now. | |
1103 | */ | |
1104 | if (st->state >= target) | |
1105 | goto out; | |
1106 | ||
1107 | if (st->state == CPUHP_OFFLINE) { | |
1108 | /* Let it fail before we try to bring the cpu up */ | |
1109 | idle = idle_thread_get(cpu); | |
1110 | if (IS_ERR(idle)) { | |
1111 | ret = PTR_ERR(idle); | |
1112 | goto out; | |
1113 | } | |
1114 | } | |
1115 | ||
1116 | cpuhp_tasks_frozen = tasks_frozen; | |
1117 | ||
1118 | cpuhp_set_state(st, target); | |
1119 | /* | |
1120 | * If the current CPU state is in the range of the AP hotplug thread, | |
1121 | * then we need to kick the thread once more. | |
1122 | */ | |
1123 | if (st->state > CPUHP_BRINGUP_CPU) { | |
1124 | ret = cpuhp_kick_ap_work(cpu); | |
1125 | /* | |
1126 | * The AP side has done the error rollback already. Just | |
1127 | * return the error code.. | |
1128 | */ | |
1129 | if (ret) | |
1130 | goto out; | |
1131 | } | |
1132 | ||
1133 | /* | |
1134 | * Try to reach the target state. We max out on the BP at | |
1135 | * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is | |
1136 | * responsible for bringing it up to the target state. | |
1137 | */ | |
1138 | target = min((int)target, CPUHP_BRINGUP_CPU); | |
1139 | ret = cpuhp_up_callbacks(cpu, st, target); | |
1140 | out: | |
1141 | cpus_write_unlock(); | |
1142 | return ret; | |
1143 | } | |
1144 | ||
1145 | static int do_cpu_up(unsigned int cpu, enum cpuhp_state target) | |
1146 | { | |
1147 | int err = 0; | |
1148 | ||
1149 | if (!cpu_possible(cpu)) { | |
1150 | pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n", | |
1151 | cpu); | |
1152 | #if defined(CONFIG_IA64) | |
1153 | pr_err("please check additional_cpus= boot parameter\n"); | |
1154 | #endif | |
1155 | return -EINVAL; | |
1156 | } | |
1157 | ||
1158 | err = try_online_node(cpu_to_node(cpu)); | |
1159 | if (err) | |
1160 | return err; | |
1161 | ||
1162 | cpu_maps_update_begin(); | |
1163 | ||
1164 | if (cpu_hotplug_disabled) { | |
1165 | err = -EBUSY; | |
1166 | goto out; | |
1167 | } | |
1168 | if (!cpu_smt_allowed(cpu)) { | |
1169 | err = -EPERM; | |
1170 | goto out; | |
1171 | } | |
1172 | ||
1173 | err = _cpu_up(cpu, 0, target); | |
1174 | out: | |
1175 | cpu_maps_update_done(); | |
1176 | return err; | |
1177 | } | |
1178 | ||
1179 | int cpu_up(unsigned int cpu) | |
1180 | { | |
1181 | return do_cpu_up(cpu, CPUHP_ONLINE); | |
1182 | } | |
1183 | EXPORT_SYMBOL_GPL(cpu_up); | |
1184 | ||
1185 | #ifdef CONFIG_PM_SLEEP_SMP | |
1186 | static cpumask_var_t frozen_cpus; | |
1187 | ||
1188 | int freeze_secondary_cpus(int primary) | |
1189 | { | |
1190 | int cpu, error = 0; | |
1191 | ||
1192 | cpu_maps_update_begin(); | |
1193 | if (!cpu_online(primary)) | |
1194 | primary = cpumask_first(cpu_online_mask); | |
1195 | /* | |
1196 | * We take down all of the non-boot CPUs in one shot to avoid races | |
1197 | * with the userspace trying to use the CPU hotplug at the same time | |
1198 | */ | |
1199 | cpumask_clear(frozen_cpus); | |
1200 | ||
1201 | pr_info("Disabling non-boot CPUs ...\n"); | |
1202 | for_each_online_cpu(cpu) { | |
1203 | if (cpu == primary) | |
1204 | continue; | |
1205 | trace_suspend_resume(TPS("CPU_OFF"), cpu, true); | |
1206 | error = _cpu_down(cpu, 1, CPUHP_OFFLINE); | |
1207 | trace_suspend_resume(TPS("CPU_OFF"), cpu, false); | |
1208 | if (!error) | |
1209 | cpumask_set_cpu(cpu, frozen_cpus); | |
1210 | else { | |
1211 | pr_err("Error taking CPU%d down: %d\n", cpu, error); | |
1212 | break; | |
1213 | } | |
1214 | } | |
1215 | ||
1216 | if (!error) | |
1217 | BUG_ON(num_online_cpus() > 1); | |
1218 | else | |
1219 | pr_err("Non-boot CPUs are not disabled\n"); | |
1220 | ||
1221 | /* | |
1222 | * Make sure the CPUs won't be enabled by someone else. We need to do | |
1223 | * this even in case of failure as all disable_nonboot_cpus() users are | |
1224 | * supposed to do enable_nonboot_cpus() on the failure path. | |
1225 | */ | |
1226 | cpu_hotplug_disabled++; | |
1227 | ||
1228 | cpu_maps_update_done(); | |
1229 | return error; | |
1230 | } | |
1231 | ||
1232 | void __weak arch_enable_nonboot_cpus_begin(void) | |
1233 | { | |
1234 | } | |
1235 | ||
1236 | void __weak arch_enable_nonboot_cpus_end(void) | |
1237 | { | |
1238 | } | |
1239 | ||
1240 | void enable_nonboot_cpus(void) | |
1241 | { | |
1242 | int cpu, error; | |
1243 | ||
1244 | /* Allow everyone to use the CPU hotplug again */ | |
1245 | cpu_maps_update_begin(); | |
1246 | __cpu_hotplug_enable(); | |
1247 | if (cpumask_empty(frozen_cpus)) | |
1248 | goto out; | |
1249 | ||
1250 | pr_info("Enabling non-boot CPUs ...\n"); | |
1251 | ||
1252 | arch_enable_nonboot_cpus_begin(); | |
1253 | ||
1254 | for_each_cpu(cpu, frozen_cpus) { | |
1255 | trace_suspend_resume(TPS("CPU_ON"), cpu, true); | |
1256 | error = _cpu_up(cpu, 1, CPUHP_ONLINE); | |
1257 | trace_suspend_resume(TPS("CPU_ON"), cpu, false); | |
1258 | if (!error) { | |
1259 | pr_info("CPU%d is up\n", cpu); | |
1260 | continue; | |
1261 | } | |
1262 | pr_warn("Error taking CPU%d up: %d\n", cpu, error); | |
1263 | } | |
1264 | ||
1265 | arch_enable_nonboot_cpus_end(); | |
1266 | ||
1267 | cpumask_clear(frozen_cpus); | |
1268 | out: | |
1269 | cpu_maps_update_done(); | |
1270 | } | |
1271 | ||
1272 | static int __init alloc_frozen_cpus(void) | |
1273 | { | |
1274 | if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO)) | |
1275 | return -ENOMEM; | |
1276 | return 0; | |
1277 | } | |
1278 | core_initcall(alloc_frozen_cpus); | |
1279 | ||
1280 | /* | |
1281 | * When callbacks for CPU hotplug notifications are being executed, we must | |
1282 | * ensure that the state of the system with respect to the tasks being frozen | |
1283 | * or not, as reported by the notification, remains unchanged *throughout the | |
1284 | * duration* of the execution of the callbacks. | |
1285 | * Hence we need to prevent the freezer from racing with regular CPU hotplug. | |
1286 | * | |
1287 | * This synchronization is implemented by mutually excluding regular CPU | |
1288 | * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/ | |
1289 | * Hibernate notifications. | |
1290 | */ | |
1291 | static int | |
1292 | cpu_hotplug_pm_callback(struct notifier_block *nb, | |
1293 | unsigned long action, void *ptr) | |
1294 | { | |
1295 | switch (action) { | |
1296 | ||
1297 | case PM_SUSPEND_PREPARE: | |
1298 | case PM_HIBERNATION_PREPARE: | |
1299 | cpu_hotplug_disable(); | |
1300 | break; | |
1301 | ||
1302 | case PM_POST_SUSPEND: | |
1303 | case PM_POST_HIBERNATION: | |
1304 | cpu_hotplug_enable(); | |
1305 | break; | |
1306 | ||
1307 | default: | |
1308 | return NOTIFY_DONE; | |
1309 | } | |
1310 | ||
1311 | return NOTIFY_OK; | |
1312 | } | |
1313 | ||
1314 | ||
1315 | static int __init cpu_hotplug_pm_sync_init(void) | |
1316 | { | |
1317 | /* | |
1318 | * cpu_hotplug_pm_callback has higher priority than x86 | |
1319 | * bsp_pm_callback which depends on cpu_hotplug_pm_callback | |
1320 | * to disable cpu hotplug to avoid cpu hotplug race. | |
1321 | */ | |
1322 | pm_notifier(cpu_hotplug_pm_callback, 0); | |
1323 | return 0; | |
1324 | } | |
1325 | core_initcall(cpu_hotplug_pm_sync_init); | |
1326 | ||
1327 | #endif /* CONFIG_PM_SLEEP_SMP */ | |
1328 | ||
1329 | int __boot_cpu_id; | |
1330 | ||
1331 | #endif /* CONFIG_SMP */ | |
1332 | ||
1333 | /* Boot processor state steps */ | |
1334 | static struct cpuhp_step cpuhp_hp_states[] = { | |
1335 | [CPUHP_OFFLINE] = { | |
1336 | .name = "offline", | |
1337 | .startup.single = NULL, | |
1338 | .teardown.single = NULL, | |
1339 | }, | |
1340 | #ifdef CONFIG_SMP | |
1341 | [CPUHP_CREATE_THREADS]= { | |
1342 | .name = "threads:prepare", | |
1343 | .startup.single = smpboot_create_threads, | |
1344 | .teardown.single = NULL, | |
1345 | .cant_stop = true, | |
1346 | }, | |
1347 | [CPUHP_PERF_PREPARE] = { | |
1348 | .name = "perf:prepare", | |
1349 | .startup.single = perf_event_init_cpu, | |
1350 | .teardown.single = perf_event_exit_cpu, | |
1351 | }, | |
1352 | [CPUHP_WORKQUEUE_PREP] = { | |
1353 | .name = "workqueue:prepare", | |
1354 | .startup.single = workqueue_prepare_cpu, | |
1355 | .teardown.single = NULL, | |
1356 | }, | |
1357 | [CPUHP_HRTIMERS_PREPARE] = { | |
1358 | .name = "hrtimers:prepare", | |
1359 | .startup.single = hrtimers_prepare_cpu, | |
1360 | .teardown.single = hrtimers_dead_cpu, | |
1361 | }, | |
1362 | [CPUHP_SMPCFD_PREPARE] = { | |
1363 | .name = "smpcfd:prepare", | |
1364 | .startup.single = smpcfd_prepare_cpu, | |
1365 | .teardown.single = smpcfd_dead_cpu, | |
1366 | }, | |
1367 | [CPUHP_RELAY_PREPARE] = { | |
1368 | .name = "relay:prepare", | |
1369 | .startup.single = relay_prepare_cpu, | |
1370 | .teardown.single = NULL, | |
1371 | }, | |
1372 | [CPUHP_SLAB_PREPARE] = { | |
1373 | .name = "slab:prepare", | |
1374 | .startup.single = slab_prepare_cpu, | |
1375 | .teardown.single = slab_dead_cpu, | |
1376 | }, | |
1377 | [CPUHP_RCUTREE_PREP] = { | |
1378 | .name = "RCU/tree:prepare", | |
1379 | .startup.single = rcutree_prepare_cpu, | |
1380 | .teardown.single = rcutree_dead_cpu, | |
1381 | }, | |
1382 | /* | |
1383 | * On the tear-down path, timers_dead_cpu() must be invoked | |
1384 | * before blk_mq_queue_reinit_notify() from notify_dead(), | |
1385 | * otherwise a RCU stall occurs. | |
1386 | */ | |
1387 | [CPUHP_TIMERS_PREPARE] = { | |
1388 | .name = "timers:prepare", | |
1389 | .startup.single = timers_prepare_cpu, | |
1390 | .teardown.single = timers_dead_cpu, | |
1391 | }, | |
1392 | /* Kicks the plugged cpu into life */ | |
1393 | [CPUHP_BRINGUP_CPU] = { | |
1394 | .name = "cpu:bringup", | |
1395 | .startup.single = bringup_cpu, | |
1396 | .teardown.single = NULL, | |
1397 | .cant_stop = true, | |
1398 | }, | |
1399 | /* Final state before CPU kills itself */ | |
1400 | [CPUHP_AP_IDLE_DEAD] = { | |
1401 | .name = "idle:dead", | |
1402 | }, | |
1403 | /* | |
1404 | * Last state before CPU enters the idle loop to die. Transient state | |
1405 | * for synchronization. | |
1406 | */ | |
1407 | [CPUHP_AP_OFFLINE] = { | |
1408 | .name = "ap:offline", | |
1409 | .cant_stop = true, | |
1410 | }, | |
1411 | /* First state is scheduler control. Interrupts are disabled */ | |
1412 | [CPUHP_AP_SCHED_STARTING] = { | |
1413 | .name = "sched:starting", | |
1414 | .startup.single = sched_cpu_starting, | |
1415 | .teardown.single = sched_cpu_dying, | |
1416 | }, | |
1417 | [CPUHP_AP_RCUTREE_DYING] = { | |
1418 | .name = "RCU/tree:dying", | |
1419 | .startup.single = NULL, | |
1420 | .teardown.single = rcutree_dying_cpu, | |
1421 | }, | |
1422 | [CPUHP_AP_SMPCFD_DYING] = { | |
1423 | .name = "smpcfd:dying", | |
1424 | .startup.single = NULL, | |
1425 | .teardown.single = smpcfd_dying_cpu, | |
1426 | }, | |
1427 | /* Entry state on starting. Interrupts enabled from here on. Transient | |
1428 | * state for synchronsization */ | |
1429 | [CPUHP_AP_ONLINE] = { | |
1430 | .name = "ap:online", | |
1431 | }, | |
1432 | /* | |
1433 | * Handled on controll processor until the plugged processor manages | |
1434 | * this itself. | |
1435 | */ | |
1436 | [CPUHP_TEARDOWN_CPU] = { | |
1437 | .name = "cpu:teardown", | |
1438 | .startup.single = NULL, | |
1439 | .teardown.single = takedown_cpu, | |
1440 | .cant_stop = true, | |
1441 | }, | |
1442 | /* Handle smpboot threads park/unpark */ | |
1443 | [CPUHP_AP_SMPBOOT_THREADS] = { | |
1444 | .name = "smpboot/threads:online", | |
1445 | .startup.single = smpboot_unpark_threads, | |
1446 | .teardown.single = smpboot_park_threads, | |
1447 | }, | |
1448 | [CPUHP_AP_IRQ_AFFINITY_ONLINE] = { | |
1449 | .name = "irq/affinity:online", | |
1450 | .startup.single = irq_affinity_online_cpu, | |
1451 | .teardown.single = NULL, | |
1452 | }, | |
1453 | [CPUHP_AP_PERF_ONLINE] = { | |
1454 | .name = "perf:online", | |
1455 | .startup.single = perf_event_init_cpu, | |
1456 | .teardown.single = perf_event_exit_cpu, | |
1457 | }, | |
1458 | [CPUHP_AP_WATCHDOG_ONLINE] = { | |
1459 | .name = "lockup_detector:online", | |
1460 | .startup.single = lockup_detector_online_cpu, | |
1461 | .teardown.single = lockup_detector_offline_cpu, | |
1462 | }, | |
1463 | [CPUHP_AP_WORKQUEUE_ONLINE] = { | |
1464 | .name = "workqueue:online", | |
1465 | .startup.single = workqueue_online_cpu, | |
1466 | .teardown.single = workqueue_offline_cpu, | |
1467 | }, | |
1468 | [CPUHP_AP_RCUTREE_ONLINE] = { | |
1469 | .name = "RCU/tree:online", | |
1470 | .startup.single = rcutree_online_cpu, | |
1471 | .teardown.single = rcutree_offline_cpu, | |
1472 | }, | |
1473 | #endif | |
1474 | /* | |
1475 | * The dynamically registered state space is here | |
1476 | */ | |
1477 | ||
1478 | #ifdef CONFIG_SMP | |
1479 | /* Last state is scheduler control setting the cpu active */ | |
1480 | [CPUHP_AP_ACTIVE] = { | |
1481 | .name = "sched:active", | |
1482 | .startup.single = sched_cpu_activate, | |
1483 | .teardown.single = sched_cpu_deactivate, | |
1484 | }, | |
1485 | #endif | |
1486 | ||
1487 | /* CPU is fully up and running. */ | |
1488 | [CPUHP_ONLINE] = { | |
1489 | .name = "online", | |
1490 | .startup.single = NULL, | |
1491 | .teardown.single = NULL, | |
1492 | }, | |
1493 | }; | |
1494 | ||
1495 | /* Sanity check for callbacks */ | |
1496 | static int cpuhp_cb_check(enum cpuhp_state state) | |
1497 | { | |
1498 | if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE) | |
1499 | return -EINVAL; | |
1500 | return 0; | |
1501 | } | |
1502 | ||
1503 | /* | |
1504 | * Returns a free for dynamic slot assignment of the Online state. The states | |
1505 | * are protected by the cpuhp_slot_states mutex and an empty slot is identified | |
1506 | * by having no name assigned. | |
1507 | */ | |
1508 | static int cpuhp_reserve_state(enum cpuhp_state state) | |
1509 | { | |
1510 | enum cpuhp_state i, end; | |
1511 | struct cpuhp_step *step; | |
1512 | ||
1513 | switch (state) { | |
1514 | case CPUHP_AP_ONLINE_DYN: | |
1515 | step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN; | |
1516 | end = CPUHP_AP_ONLINE_DYN_END; | |
1517 | break; | |
1518 | case CPUHP_BP_PREPARE_DYN: | |
1519 | step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN; | |
1520 | end = CPUHP_BP_PREPARE_DYN_END; | |
1521 | break; | |
1522 | default: | |
1523 | return -EINVAL; | |
1524 | } | |
1525 | ||
1526 | for (i = state; i <= end; i++, step++) { | |
1527 | if (!step->name) | |
1528 | return i; | |
1529 | } | |
1530 | WARN(1, "No more dynamic states available for CPU hotplug\n"); | |
1531 | return -ENOSPC; | |
1532 | } | |
1533 | ||
1534 | static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name, | |
1535 | int (*startup)(unsigned int cpu), | |
1536 | int (*teardown)(unsigned int cpu), | |
1537 | bool multi_instance) | |
1538 | { | |
1539 | /* (Un)Install the callbacks for further cpu hotplug operations */ | |
1540 | struct cpuhp_step *sp; | |
1541 | int ret = 0; | |
1542 | ||
1543 | /* | |
1544 | * If name is NULL, then the state gets removed. | |
1545 | * | |
1546 | * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on | |
1547 | * the first allocation from these dynamic ranges, so the removal | |
1548 | * would trigger a new allocation and clear the wrong (already | |
1549 | * empty) state, leaving the callbacks of the to be cleared state | |
1550 | * dangling, which causes wreckage on the next hotplug operation. | |
1551 | */ | |
1552 | if (name && (state == CPUHP_AP_ONLINE_DYN || | |
1553 | state == CPUHP_BP_PREPARE_DYN)) { | |
1554 | ret = cpuhp_reserve_state(state); | |
1555 | if (ret < 0) | |
1556 | return ret; | |
1557 | state = ret; | |
1558 | } | |
1559 | sp = cpuhp_get_step(state); | |
1560 | if (name && sp->name) | |
1561 | return -EBUSY; | |
1562 | ||
1563 | sp->startup.single = startup; | |
1564 | sp->teardown.single = teardown; | |
1565 | sp->name = name; | |
1566 | sp->multi_instance = multi_instance; | |
1567 | INIT_HLIST_HEAD(&sp->list); | |
1568 | return ret; | |
1569 | } | |
1570 | ||
1571 | static void *cpuhp_get_teardown_cb(enum cpuhp_state state) | |
1572 | { | |
1573 | return cpuhp_get_step(state)->teardown.single; | |
1574 | } | |
1575 | ||
1576 | /* | |
1577 | * Call the startup/teardown function for a step either on the AP or | |
1578 | * on the current CPU. | |
1579 | */ | |
1580 | static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup, | |
1581 | struct hlist_node *node) | |
1582 | { | |
1583 | struct cpuhp_step *sp = cpuhp_get_step(state); | |
1584 | int ret; | |
1585 | ||
1586 | /* | |
1587 | * If there's nothing to do, we done. | |
1588 | * Relies on the union for multi_instance. | |
1589 | */ | |
1590 | if ((bringup && !sp->startup.single) || | |
1591 | (!bringup && !sp->teardown.single)) | |
1592 | return 0; | |
1593 | /* | |
1594 | * The non AP bound callbacks can fail on bringup. On teardown | |
1595 | * e.g. module removal we crash for now. | |
1596 | */ | |
1597 | #ifdef CONFIG_SMP | |
1598 | if (cpuhp_is_ap_state(state)) | |
1599 | ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node); | |
1600 | else | |
1601 | ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL); | |
1602 | #else | |
1603 | ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL); | |
1604 | #endif | |
1605 | BUG_ON(ret && !bringup); | |
1606 | return ret; | |
1607 | } | |
1608 | ||
1609 | /* | |
1610 | * Called from __cpuhp_setup_state on a recoverable failure. | |
1611 | * | |
1612 | * Note: The teardown callbacks for rollback are not allowed to fail! | |
1613 | */ | |
1614 | static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state, | |
1615 | struct hlist_node *node) | |
1616 | { | |
1617 | int cpu; | |
1618 | ||
1619 | /* Roll back the already executed steps on the other cpus */ | |
1620 | for_each_present_cpu(cpu) { | |
1621 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1622 | int cpustate = st->state; | |
1623 | ||
1624 | if (cpu >= failedcpu) | |
1625 | break; | |
1626 | ||
1627 | /* Did we invoke the startup call on that cpu ? */ | |
1628 | if (cpustate >= state) | |
1629 | cpuhp_issue_call(cpu, state, false, node); | |
1630 | } | |
1631 | } | |
1632 | ||
1633 | int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state, | |
1634 | struct hlist_node *node, | |
1635 | bool invoke) | |
1636 | { | |
1637 | struct cpuhp_step *sp; | |
1638 | int cpu; | |
1639 | int ret; | |
1640 | ||
1641 | lockdep_assert_cpus_held(); | |
1642 | ||
1643 | sp = cpuhp_get_step(state); | |
1644 | if (sp->multi_instance == false) | |
1645 | return -EINVAL; | |
1646 | ||
1647 | mutex_lock(&cpuhp_state_mutex); | |
1648 | ||
1649 | if (!invoke || !sp->startup.multi) | |
1650 | goto add_node; | |
1651 | ||
1652 | /* | |
1653 | * Try to call the startup callback for each present cpu | |
1654 | * depending on the hotplug state of the cpu. | |
1655 | */ | |
1656 | for_each_present_cpu(cpu) { | |
1657 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1658 | int cpustate = st->state; | |
1659 | ||
1660 | if (cpustate < state) | |
1661 | continue; | |
1662 | ||
1663 | ret = cpuhp_issue_call(cpu, state, true, node); | |
1664 | if (ret) { | |
1665 | if (sp->teardown.multi) | |
1666 | cpuhp_rollback_install(cpu, state, node); | |
1667 | goto unlock; | |
1668 | } | |
1669 | } | |
1670 | add_node: | |
1671 | ret = 0; | |
1672 | hlist_add_head(node, &sp->list); | |
1673 | unlock: | |
1674 | mutex_unlock(&cpuhp_state_mutex); | |
1675 | return ret; | |
1676 | } | |
1677 | ||
1678 | int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node, | |
1679 | bool invoke) | |
1680 | { | |
1681 | int ret; | |
1682 | ||
1683 | cpus_read_lock(); | |
1684 | ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke); | |
1685 | cpus_read_unlock(); | |
1686 | return ret; | |
1687 | } | |
1688 | EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance); | |
1689 | ||
1690 | /** | |
1691 | * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state | |
1692 | * @state: The state to setup | |
1693 | * @invoke: If true, the startup function is invoked for cpus where | |
1694 | * cpu state >= @state | |
1695 | * @startup: startup callback function | |
1696 | * @teardown: teardown callback function | |
1697 | * @multi_instance: State is set up for multiple instances which get | |
1698 | * added afterwards. | |
1699 | * | |
1700 | * The caller needs to hold cpus read locked while calling this function. | |
1701 | * Returns: | |
1702 | * On success: | |
1703 | * Positive state number if @state is CPUHP_AP_ONLINE_DYN | |
1704 | * 0 for all other states | |
1705 | * On failure: proper (negative) error code | |
1706 | */ | |
1707 | int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state, | |
1708 | const char *name, bool invoke, | |
1709 | int (*startup)(unsigned int cpu), | |
1710 | int (*teardown)(unsigned int cpu), | |
1711 | bool multi_instance) | |
1712 | { | |
1713 | int cpu, ret = 0; | |
1714 | bool dynstate; | |
1715 | ||
1716 | lockdep_assert_cpus_held(); | |
1717 | ||
1718 | if (cpuhp_cb_check(state) || !name) | |
1719 | return -EINVAL; | |
1720 | ||
1721 | mutex_lock(&cpuhp_state_mutex); | |
1722 | ||
1723 | ret = cpuhp_store_callbacks(state, name, startup, teardown, | |
1724 | multi_instance); | |
1725 | ||
1726 | dynstate = state == CPUHP_AP_ONLINE_DYN; | |
1727 | if (ret > 0 && dynstate) { | |
1728 | state = ret; | |
1729 | ret = 0; | |
1730 | } | |
1731 | ||
1732 | if (ret || !invoke || !startup) | |
1733 | goto out; | |
1734 | ||
1735 | /* | |
1736 | * Try to call the startup callback for each present cpu | |
1737 | * depending on the hotplug state of the cpu. | |
1738 | */ | |
1739 | for_each_present_cpu(cpu) { | |
1740 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1741 | int cpustate = st->state; | |
1742 | ||
1743 | if (cpustate < state) | |
1744 | continue; | |
1745 | ||
1746 | ret = cpuhp_issue_call(cpu, state, true, NULL); | |
1747 | if (ret) { | |
1748 | if (teardown) | |
1749 | cpuhp_rollback_install(cpu, state, NULL); | |
1750 | cpuhp_store_callbacks(state, NULL, NULL, NULL, false); | |
1751 | goto out; | |
1752 | } | |
1753 | } | |
1754 | out: | |
1755 | mutex_unlock(&cpuhp_state_mutex); | |
1756 | /* | |
1757 | * If the requested state is CPUHP_AP_ONLINE_DYN, return the | |
1758 | * dynamically allocated state in case of success. | |
1759 | */ | |
1760 | if (!ret && dynstate) | |
1761 | return state; | |
1762 | return ret; | |
1763 | } | |
1764 | EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked); | |
1765 | ||
1766 | int __cpuhp_setup_state(enum cpuhp_state state, | |
1767 | const char *name, bool invoke, | |
1768 | int (*startup)(unsigned int cpu), | |
1769 | int (*teardown)(unsigned int cpu), | |
1770 | bool multi_instance) | |
1771 | { | |
1772 | int ret; | |
1773 | ||
1774 | cpus_read_lock(); | |
1775 | ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup, | |
1776 | teardown, multi_instance); | |
1777 | cpus_read_unlock(); | |
1778 | return ret; | |
1779 | } | |
1780 | EXPORT_SYMBOL(__cpuhp_setup_state); | |
1781 | ||
1782 | int __cpuhp_state_remove_instance(enum cpuhp_state state, | |
1783 | struct hlist_node *node, bool invoke) | |
1784 | { | |
1785 | struct cpuhp_step *sp = cpuhp_get_step(state); | |
1786 | int cpu; | |
1787 | ||
1788 | BUG_ON(cpuhp_cb_check(state)); | |
1789 | ||
1790 | if (!sp->multi_instance) | |
1791 | return -EINVAL; | |
1792 | ||
1793 | cpus_read_lock(); | |
1794 | mutex_lock(&cpuhp_state_mutex); | |
1795 | ||
1796 | if (!invoke || !cpuhp_get_teardown_cb(state)) | |
1797 | goto remove; | |
1798 | /* | |
1799 | * Call the teardown callback for each present cpu depending | |
1800 | * on the hotplug state of the cpu. This function is not | |
1801 | * allowed to fail currently! | |
1802 | */ | |
1803 | for_each_present_cpu(cpu) { | |
1804 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1805 | int cpustate = st->state; | |
1806 | ||
1807 | if (cpustate >= state) | |
1808 | cpuhp_issue_call(cpu, state, false, node); | |
1809 | } | |
1810 | ||
1811 | remove: | |
1812 | hlist_del(node); | |
1813 | mutex_unlock(&cpuhp_state_mutex); | |
1814 | cpus_read_unlock(); | |
1815 | ||
1816 | return 0; | |
1817 | } | |
1818 | EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance); | |
1819 | ||
1820 | /** | |
1821 | * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state | |
1822 | * @state: The state to remove | |
1823 | * @invoke: If true, the teardown function is invoked for cpus where | |
1824 | * cpu state >= @state | |
1825 | * | |
1826 | * The caller needs to hold cpus read locked while calling this function. | |
1827 | * The teardown callback is currently not allowed to fail. Think | |
1828 | * about module removal! | |
1829 | */ | |
1830 | void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke) | |
1831 | { | |
1832 | struct cpuhp_step *sp = cpuhp_get_step(state); | |
1833 | int cpu; | |
1834 | ||
1835 | BUG_ON(cpuhp_cb_check(state)); | |
1836 | ||
1837 | lockdep_assert_cpus_held(); | |
1838 | ||
1839 | mutex_lock(&cpuhp_state_mutex); | |
1840 | if (sp->multi_instance) { | |
1841 | WARN(!hlist_empty(&sp->list), | |
1842 | "Error: Removing state %d which has instances left.\n", | |
1843 | state); | |
1844 | goto remove; | |
1845 | } | |
1846 | ||
1847 | if (!invoke || !cpuhp_get_teardown_cb(state)) | |
1848 | goto remove; | |
1849 | ||
1850 | /* | |
1851 | * Call the teardown callback for each present cpu depending | |
1852 | * on the hotplug state of the cpu. This function is not | |
1853 | * allowed to fail currently! | |
1854 | */ | |
1855 | for_each_present_cpu(cpu) { | |
1856 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu); | |
1857 | int cpustate = st->state; | |
1858 | ||
1859 | if (cpustate >= state) | |
1860 | cpuhp_issue_call(cpu, state, false, NULL); | |
1861 | } | |
1862 | remove: | |
1863 | cpuhp_store_callbacks(state, NULL, NULL, NULL, false); | |
1864 | mutex_unlock(&cpuhp_state_mutex); | |
1865 | } | |
1866 | EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked); | |
1867 | ||
1868 | void __cpuhp_remove_state(enum cpuhp_state state, bool invoke) | |
1869 | { | |
1870 | cpus_read_lock(); | |
1871 | __cpuhp_remove_state_cpuslocked(state, invoke); | |
1872 | cpus_read_unlock(); | |
1873 | } | |
1874 | EXPORT_SYMBOL(__cpuhp_remove_state); | |
1875 | ||
1876 | #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU) | |
1877 | static ssize_t show_cpuhp_state(struct device *dev, | |
1878 | struct device_attribute *attr, char *buf) | |
1879 | { | |
1880 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id); | |
1881 | ||
1882 | return sprintf(buf, "%d\n", st->state); | |
1883 | } | |
1884 | static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL); | |
1885 | ||
1886 | static ssize_t write_cpuhp_target(struct device *dev, | |
1887 | struct device_attribute *attr, | |
1888 | const char *buf, size_t count) | |
1889 | { | |
1890 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id); | |
1891 | struct cpuhp_step *sp; | |
1892 | int target, ret; | |
1893 | ||
1894 | ret = kstrtoint(buf, 10, &target); | |
1895 | if (ret) | |
1896 | return ret; | |
1897 | ||
1898 | #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL | |
1899 | if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE) | |
1900 | return -EINVAL; | |
1901 | #else | |
1902 | if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE) | |
1903 | return -EINVAL; | |
1904 | #endif | |
1905 | ||
1906 | ret = lock_device_hotplug_sysfs(); | |
1907 | if (ret) | |
1908 | return ret; | |
1909 | ||
1910 | mutex_lock(&cpuhp_state_mutex); | |
1911 | sp = cpuhp_get_step(target); | |
1912 | ret = !sp->name || sp->cant_stop ? -EINVAL : 0; | |
1913 | mutex_unlock(&cpuhp_state_mutex); | |
1914 | if (ret) | |
1915 | goto out; | |
1916 | ||
1917 | if (st->state < target) | |
1918 | ret = do_cpu_up(dev->id, target); | |
1919 | else | |
1920 | ret = do_cpu_down(dev->id, target); | |
1921 | out: | |
1922 | unlock_device_hotplug(); | |
1923 | return ret ? ret : count; | |
1924 | } | |
1925 | ||
1926 | static ssize_t show_cpuhp_target(struct device *dev, | |
1927 | struct device_attribute *attr, char *buf) | |
1928 | { | |
1929 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id); | |
1930 | ||
1931 | return sprintf(buf, "%d\n", st->target); | |
1932 | } | |
1933 | static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target); | |
1934 | ||
1935 | ||
1936 | static ssize_t write_cpuhp_fail(struct device *dev, | |
1937 | struct device_attribute *attr, | |
1938 | const char *buf, size_t count) | |
1939 | { | |
1940 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id); | |
1941 | struct cpuhp_step *sp; | |
1942 | int fail, ret; | |
1943 | ||
1944 | ret = kstrtoint(buf, 10, &fail); | |
1945 | if (ret) | |
1946 | return ret; | |
1947 | ||
1948 | /* | |
1949 | * Cannot fail STARTING/DYING callbacks. | |
1950 | */ | |
1951 | if (cpuhp_is_atomic_state(fail)) | |
1952 | return -EINVAL; | |
1953 | ||
1954 | /* | |
1955 | * Cannot fail anything that doesn't have callbacks. | |
1956 | */ | |
1957 | mutex_lock(&cpuhp_state_mutex); | |
1958 | sp = cpuhp_get_step(fail); | |
1959 | if (!sp->startup.single && !sp->teardown.single) | |
1960 | ret = -EINVAL; | |
1961 | mutex_unlock(&cpuhp_state_mutex); | |
1962 | if (ret) | |
1963 | return ret; | |
1964 | ||
1965 | st->fail = fail; | |
1966 | ||
1967 | return count; | |
1968 | } | |
1969 | ||
1970 | static ssize_t show_cpuhp_fail(struct device *dev, | |
1971 | struct device_attribute *attr, char *buf) | |
1972 | { | |
1973 | struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id); | |
1974 | ||
1975 | return sprintf(buf, "%d\n", st->fail); | |
1976 | } | |
1977 | ||
1978 | static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail); | |
1979 | ||
1980 | static struct attribute *cpuhp_cpu_attrs[] = { | |
1981 | &dev_attr_state.attr, | |
1982 | &dev_attr_target.attr, | |
1983 | &dev_attr_fail.attr, | |
1984 | NULL | |
1985 | }; | |
1986 | ||
1987 | static const struct attribute_group cpuhp_cpu_attr_group = { | |
1988 | .attrs = cpuhp_cpu_attrs, | |
1989 | .name = "hotplug", | |
1990 | NULL | |
1991 | }; | |
1992 | ||
1993 | static ssize_t show_cpuhp_states(struct device *dev, | |
1994 | struct device_attribute *attr, char *buf) | |
1995 | { | |
1996 | ssize_t cur, res = 0; | |
1997 | int i; | |
1998 | ||
1999 | mutex_lock(&cpuhp_state_mutex); | |
2000 | for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) { | |
2001 | struct cpuhp_step *sp = cpuhp_get_step(i); | |
2002 | ||
2003 | if (sp->name) { | |
2004 | cur = sprintf(buf, "%3d: %s\n", i, sp->name); | |
2005 | buf += cur; | |
2006 | res += cur; | |
2007 | } | |
2008 | } | |
2009 | mutex_unlock(&cpuhp_state_mutex); | |
2010 | return res; | |
2011 | } | |
2012 | static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL); | |
2013 | ||
2014 | static struct attribute *cpuhp_cpu_root_attrs[] = { | |
2015 | &dev_attr_states.attr, | |
2016 | NULL | |
2017 | }; | |
2018 | ||
2019 | static const struct attribute_group cpuhp_cpu_root_attr_group = { | |
2020 | .attrs = cpuhp_cpu_root_attrs, | |
2021 | .name = "hotplug", | |
2022 | NULL | |
2023 | }; | |
2024 | ||
2025 | #ifdef CONFIG_HOTPLUG_SMT | |
2026 | ||
2027 | static const char *smt_states[] = { | |
2028 | [CPU_SMT_ENABLED] = "on", | |
2029 | [CPU_SMT_DISABLED] = "off", | |
2030 | [CPU_SMT_FORCE_DISABLED] = "forceoff", | |
2031 | [CPU_SMT_NOT_SUPPORTED] = "notsupported", | |
2032 | }; | |
2033 | ||
2034 | static ssize_t | |
2035 | show_smt_control(struct device *dev, struct device_attribute *attr, char *buf) | |
2036 | { | |
2037 | return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]); | |
2038 | } | |
2039 | ||
2040 | static void cpuhp_offline_cpu_device(unsigned int cpu) | |
2041 | { | |
2042 | struct device *dev = get_cpu_device(cpu); | |
2043 | ||
2044 | dev->offline = true; | |
2045 | /* Tell user space about the state change */ | |
2046 | kobject_uevent(&dev->kobj, KOBJ_OFFLINE); | |
2047 | } | |
2048 | ||
2049 | static void cpuhp_online_cpu_device(unsigned int cpu) | |
2050 | { | |
2051 | struct device *dev = get_cpu_device(cpu); | |
2052 | ||
2053 | dev->offline = false; | |
2054 | /* Tell user space about the state change */ | |
2055 | kobject_uevent(&dev->kobj, KOBJ_ONLINE); | |
2056 | } | |
2057 | ||
2058 | /* | |
2059 | * Architectures that need SMT-specific errata handling during SMT hotplug | |
2060 | * should override this. | |
2061 | */ | |
2062 | void __weak arch_smt_update(void) { }; | |
2063 | ||
2064 | static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval) | |
2065 | { | |
2066 | int cpu, ret = 0; | |
2067 | ||
2068 | cpu_maps_update_begin(); | |
2069 | for_each_online_cpu(cpu) { | |
2070 | if (topology_is_primary_thread(cpu)) | |
2071 | continue; | |
2072 | ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE); | |
2073 | if (ret) | |
2074 | break; | |
2075 | /* | |
2076 | * As this needs to hold the cpu maps lock it's impossible | |
2077 | * to call device_offline() because that ends up calling | |
2078 | * cpu_down() which takes cpu maps lock. cpu maps lock | |
2079 | * needs to be held as this might race against in kernel | |
2080 | * abusers of the hotplug machinery (thermal management). | |
2081 | * | |
2082 | * So nothing would update device:offline state. That would | |
2083 | * leave the sysfs entry stale and prevent onlining after | |
2084 | * smt control has been changed to 'off' again. This is | |
2085 | * called under the sysfs hotplug lock, so it is properly | |
2086 | * serialized against the regular offline usage. | |
2087 | */ | |
2088 | cpuhp_offline_cpu_device(cpu); | |
2089 | } | |
2090 | if (!ret) { | |
2091 | cpu_smt_control = ctrlval; | |
2092 | arch_smt_update(); | |
2093 | } | |
2094 | cpu_maps_update_done(); | |
2095 | return ret; | |
2096 | } | |
2097 | ||
2098 | static int cpuhp_smt_enable(void) | |
2099 | { | |
2100 | int cpu, ret = 0; | |
2101 | ||
2102 | cpu_maps_update_begin(); | |
2103 | cpu_smt_control = CPU_SMT_ENABLED; | |
2104 | arch_smt_update(); | |
2105 | for_each_present_cpu(cpu) { | |
2106 | /* Skip online CPUs and CPUs on offline nodes */ | |
2107 | if (cpu_online(cpu) || !node_online(cpu_to_node(cpu))) | |
2108 | continue; | |
2109 | ret = _cpu_up(cpu, 0, CPUHP_ONLINE); | |
2110 | if (ret) | |
2111 | break; | |
2112 | /* See comment in cpuhp_smt_disable() */ | |
2113 | cpuhp_online_cpu_device(cpu); | |
2114 | } | |
2115 | cpu_maps_update_done(); | |
2116 | return ret; | |
2117 | } | |
2118 | ||
2119 | static ssize_t | |
2120 | store_smt_control(struct device *dev, struct device_attribute *attr, | |
2121 | const char *buf, size_t count) | |
2122 | { | |
2123 | int ctrlval, ret; | |
2124 | ||
2125 | if (sysfs_streq(buf, "on")) | |
2126 | ctrlval = CPU_SMT_ENABLED; | |
2127 | else if (sysfs_streq(buf, "off")) | |
2128 | ctrlval = CPU_SMT_DISABLED; | |
2129 | else if (sysfs_streq(buf, "forceoff")) | |
2130 | ctrlval = CPU_SMT_FORCE_DISABLED; | |
2131 | else | |
2132 | return -EINVAL; | |
2133 | ||
2134 | if (cpu_smt_control == CPU_SMT_FORCE_DISABLED) | |
2135 | return -EPERM; | |
2136 | ||
2137 | if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED) | |
2138 | return -ENODEV; | |
2139 | ||
2140 | ret = lock_device_hotplug_sysfs(); | |
2141 | if (ret) | |
2142 | return ret; | |
2143 | ||
2144 | if (ctrlval != cpu_smt_control) { | |
2145 | switch (ctrlval) { | |
2146 | case CPU_SMT_ENABLED: | |
2147 | ret = cpuhp_smt_enable(); | |
2148 | break; | |
2149 | case CPU_SMT_DISABLED: | |
2150 | case CPU_SMT_FORCE_DISABLED: | |
2151 | ret = cpuhp_smt_disable(ctrlval); | |
2152 | break; | |
2153 | } | |
2154 | } | |
2155 | ||
2156 | unlock_device_hotplug(); | |
2157 | return ret ? ret : count; | |
2158 | } | |
2159 | static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control); | |
2160 | ||
2161 | static ssize_t | |
2162 | show_smt_active(struct device *dev, struct device_attribute *attr, char *buf) | |
2163 | { | |
2164 | bool active = topology_max_smt_threads() > 1; | |
2165 | ||
2166 | return snprintf(buf, PAGE_SIZE - 2, "%d\n", active); | |
2167 | } | |
2168 | static DEVICE_ATTR(active, 0444, show_smt_active, NULL); | |
2169 | ||
2170 | static struct attribute *cpuhp_smt_attrs[] = { | |
2171 | &dev_attr_control.attr, | |
2172 | &dev_attr_active.attr, | |
2173 | NULL | |
2174 | }; | |
2175 | ||
2176 | static const struct attribute_group cpuhp_smt_attr_group = { | |
2177 | .attrs = cpuhp_smt_attrs, | |
2178 | .name = "smt", | |
2179 | NULL | |
2180 | }; | |
2181 | ||
2182 | static int __init cpu_smt_state_init(void) | |
2183 | { | |
2184 | return sysfs_create_group(&cpu_subsys.dev_root->kobj, | |
2185 | &cpuhp_smt_attr_group); | |
2186 | } | |
2187 | ||
2188 | #else | |
2189 | static inline int cpu_smt_state_init(void) { return 0; } | |
2190 | #endif | |
2191 | ||
2192 | static int __init cpuhp_sysfs_init(void) | |
2193 | { | |
2194 | int cpu, ret; | |
2195 | ||
2196 | ret = cpu_smt_state_init(); | |
2197 | if (ret) | |
2198 | return ret; | |
2199 | ||
2200 | ret = sysfs_create_group(&cpu_subsys.dev_root->kobj, | |
2201 | &cpuhp_cpu_root_attr_group); | |
2202 | if (ret) | |
2203 | return ret; | |
2204 | ||
2205 | for_each_possible_cpu(cpu) { | |
2206 | struct device *dev = get_cpu_device(cpu); | |
2207 | ||
2208 | if (!dev) | |
2209 | continue; | |
2210 | ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group); | |
2211 | if (ret) | |
2212 | return ret; | |
2213 | } | |
2214 | return 0; | |
2215 | } | |
2216 | device_initcall(cpuhp_sysfs_init); | |
2217 | #endif | |
2218 | ||
2219 | /* | |
2220 | * cpu_bit_bitmap[] is a special, "compressed" data structure that | |
2221 | * represents all NR_CPUS bits binary values of 1<<nr. | |
2222 | * | |
2223 | * It is used by cpumask_of() to get a constant address to a CPU | |
2224 | * mask value that has a single bit set only. | |
2225 | */ | |
2226 | ||
2227 | /* cpu_bit_bitmap[0] is empty - so we can back into it */ | |
2228 | #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x)) | |
2229 | #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1) | |
2230 | #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2) | |
2231 | #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4) | |
2232 | ||
2233 | const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = { | |
2234 | ||
2235 | MASK_DECLARE_8(0), MASK_DECLARE_8(8), | |
2236 | MASK_DECLARE_8(16), MASK_DECLARE_8(24), | |
2237 | #if BITS_PER_LONG > 32 | |
2238 | MASK_DECLARE_8(32), MASK_DECLARE_8(40), | |
2239 | MASK_DECLARE_8(48), MASK_DECLARE_8(56), | |
2240 | #endif | |
2241 | }; | |
2242 | EXPORT_SYMBOL_GPL(cpu_bit_bitmap); | |
2243 | ||
2244 | const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL; | |
2245 | EXPORT_SYMBOL(cpu_all_bits); | |
2246 | ||
2247 | #ifdef CONFIG_INIT_ALL_POSSIBLE | |
2248 | struct cpumask __cpu_possible_mask __read_mostly | |
2249 | = {CPU_BITS_ALL}; | |
2250 | #else | |
2251 | struct cpumask __cpu_possible_mask __read_mostly; | |
2252 | #endif | |
2253 | EXPORT_SYMBOL(__cpu_possible_mask); | |
2254 | ||
2255 | struct cpumask __cpu_online_mask __read_mostly; | |
2256 | EXPORT_SYMBOL(__cpu_online_mask); | |
2257 | ||
2258 | struct cpumask __cpu_present_mask __read_mostly; | |
2259 | EXPORT_SYMBOL(__cpu_present_mask); | |
2260 | ||
2261 | struct cpumask __cpu_active_mask __read_mostly; | |
2262 | EXPORT_SYMBOL(__cpu_active_mask); | |
2263 | ||
2264 | void init_cpu_present(const struct cpumask *src) | |
2265 | { | |
2266 | cpumask_copy(&__cpu_present_mask, src); | |
2267 | } | |
2268 | ||
2269 | void init_cpu_possible(const struct cpumask *src) | |
2270 | { | |
2271 | cpumask_copy(&__cpu_possible_mask, src); | |
2272 | } | |
2273 | ||
2274 | void init_cpu_online(const struct cpumask *src) | |
2275 | { | |
2276 | cpumask_copy(&__cpu_online_mask, src); | |
2277 | } | |
2278 | ||
2279 | /* | |
2280 | * Activate the first processor. | |
2281 | */ | |
2282 | void __init boot_cpu_init(void) | |
2283 | { | |
2284 | int cpu = smp_processor_id(); | |
2285 | ||
2286 | /* Mark the boot cpu "present", "online" etc for SMP and UP case */ | |
2287 | set_cpu_online(cpu, true); | |
2288 | set_cpu_active(cpu, true); | |
2289 | set_cpu_present(cpu, true); | |
2290 | set_cpu_possible(cpu, true); | |
2291 | ||
2292 | #ifdef CONFIG_SMP | |
2293 | __boot_cpu_id = cpu; | |
2294 | #endif | |
2295 | } | |
2296 | ||
2297 | /* | |
2298 | * Must be called _AFTER_ setting up the per_cpu areas | |
2299 | */ | |
2300 | void __init boot_cpu_hotplug_init(void) | |
2301 | { | |
2302 | #ifdef CONFIG_SMP | |
2303 | this_cpu_write(cpuhp_state.booted_once, true); | |
2304 | #endif | |
2305 | this_cpu_write(cpuhp_state.state, CPUHP_ONLINE); | |
2306 | } |