Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
87de1cfd PM |
15 | * along with this program; if not, you can access it online at |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
1da177e4 | 17 | * |
01c1c660 | 18 | * Copyright IBM Corporation, 2001 |
1da177e4 LT |
19 | * |
20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
21 | * Manfred Spraul <manfred@colorfullife.com> | |
a71fca58 | 22 | * |
1da177e4 LT |
23 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> |
24 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. | |
25 | * Papers: | |
26 | * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf | |
27 | * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001) | |
28 | * | |
29 | * For detailed explanation of Read-Copy Update mechanism see - | |
a71fca58 | 30 | * http://lse.sourceforge.net/locking/rcupdate.html |
1da177e4 LT |
31 | * |
32 | */ | |
33 | #include <linux/types.h> | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/init.h> | |
36 | #include <linux/spinlock.h> | |
37 | #include <linux/smp.h> | |
38 | #include <linux/interrupt.h> | |
39 | #include <linux/sched.h> | |
60063497 | 40 | #include <linux/atomic.h> |
1da177e4 | 41 | #include <linux/bitops.h> |
1da177e4 LT |
42 | #include <linux/percpu.h> |
43 | #include <linux/notifier.h> | |
1da177e4 | 44 | #include <linux/cpu.h> |
9331b315 | 45 | #include <linux/mutex.h> |
9984de1a | 46 | #include <linux/export.h> |
e3818b8d | 47 | #include <linux/hardirq.h> |
e3ebfb96 | 48 | #include <linux/delay.h> |
3705b88d | 49 | #include <linux/module.h> |
8315f422 | 50 | #include <linux/kthread.h> |
4ff475ed | 51 | #include <linux/tick.h> |
1da177e4 | 52 | |
29c00b4a | 53 | #define CREATE_TRACE_POINTS |
29c00b4a PM |
54 | |
55 | #include "rcu.h" | |
56 | ||
4102adab PM |
57 | MODULE_ALIAS("rcupdate"); |
58 | #ifdef MODULE_PARAM_PREFIX | |
59 | #undef MODULE_PARAM_PREFIX | |
60 | #endif | |
61 | #define MODULE_PARAM_PREFIX "rcupdate." | |
62 | ||
3705b88d AM |
63 | module_param(rcu_expedited, int, 0); |
64 | ||
9dd8fb16 PM |
65 | #ifdef CONFIG_PREEMPT_RCU |
66 | ||
2a3fa843 PM |
67 | /* |
68 | * Preemptible RCU implementation for rcu_read_lock(). | |
69 | * Just increment ->rcu_read_lock_nesting, shared state will be updated | |
70 | * if we block. | |
71 | */ | |
72 | void __rcu_read_lock(void) | |
73 | { | |
74 | current->rcu_read_lock_nesting++; | |
75 | barrier(); /* critical section after entry code. */ | |
76 | } | |
77 | EXPORT_SYMBOL_GPL(__rcu_read_lock); | |
78 | ||
79 | /* | |
80 | * Preemptible RCU implementation for rcu_read_unlock(). | |
81 | * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost | |
82 | * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then | |
83 | * invoke rcu_read_unlock_special() to clean up after a context switch | |
84 | * in an RCU read-side critical section and other special cases. | |
85 | */ | |
86 | void __rcu_read_unlock(void) | |
87 | { | |
88 | struct task_struct *t = current; | |
89 | ||
90 | if (t->rcu_read_lock_nesting != 1) { | |
91 | --t->rcu_read_lock_nesting; | |
92 | } else { | |
93 | barrier(); /* critical section before exit code. */ | |
94 | t->rcu_read_lock_nesting = INT_MIN; | |
95 | barrier(); /* assign before ->rcu_read_unlock_special load */ | |
1d082fd0 | 96 | if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special.s))) |
2a3fa843 PM |
97 | rcu_read_unlock_special(t); |
98 | barrier(); /* ->rcu_read_unlock_special load before assign */ | |
99 | t->rcu_read_lock_nesting = 0; | |
100 | } | |
101 | #ifdef CONFIG_PROVE_LOCKING | |
102 | { | |
103 | int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting); | |
104 | ||
105 | WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2); | |
106 | } | |
107 | #endif /* #ifdef CONFIG_PROVE_LOCKING */ | |
108 | } | |
109 | EXPORT_SYMBOL_GPL(__rcu_read_unlock); | |
110 | ||
2439b696 | 111 | #endif /* #ifdef CONFIG_PREEMPT_RCU */ |
9dd8fb16 | 112 | |
162cc279 PM |
113 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
114 | static struct lock_class_key rcu_lock_key; | |
115 | struct lockdep_map rcu_lock_map = | |
116 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key); | |
117 | EXPORT_SYMBOL_GPL(rcu_lock_map); | |
632ee200 PM |
118 | |
119 | static struct lock_class_key rcu_bh_lock_key; | |
120 | struct lockdep_map rcu_bh_lock_map = | |
121 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key); | |
122 | EXPORT_SYMBOL_GPL(rcu_bh_lock_map); | |
123 | ||
124 | static struct lock_class_key rcu_sched_lock_key; | |
125 | struct lockdep_map rcu_sched_lock_map = | |
126 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key); | |
127 | EXPORT_SYMBOL_GPL(rcu_sched_lock_map); | |
e3818b8d | 128 | |
24ef659a PM |
129 | static struct lock_class_key rcu_callback_key; |
130 | struct lockdep_map rcu_callback_map = | |
131 | STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key); | |
132 | EXPORT_SYMBOL_GPL(rcu_callback_map); | |
133 | ||
a0a5a056 | 134 | int notrace debug_lockdep_rcu_enabled(void) |
bc293d62 PM |
135 | { |
136 | return rcu_scheduler_active && debug_locks && | |
137 | current->lockdep_recursion == 0; | |
138 | } | |
139 | EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled); | |
140 | ||
85b39d30 ON |
141 | /** |
142 | * rcu_read_lock_held() - might we be in RCU read-side critical section? | |
143 | * | |
144 | * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU | |
145 | * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC, | |
146 | * this assumes we are in an RCU read-side critical section unless it can | |
147 | * prove otherwise. This is useful for debug checks in functions that | |
148 | * require that they be called within an RCU read-side critical section. | |
149 | * | |
150 | * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot | |
151 | * and while lockdep is disabled. | |
152 | * | |
153 | * Note that rcu_read_lock() and the matching rcu_read_unlock() must | |
154 | * occur in the same context, for example, it is illegal to invoke | |
155 | * rcu_read_unlock() in process context if the matching rcu_read_lock() | |
156 | * was invoked from within an irq handler. | |
157 | * | |
158 | * Note that rcu_read_lock() is disallowed if the CPU is either idle or | |
159 | * offline from an RCU perspective, so check for those as well. | |
160 | */ | |
161 | int rcu_read_lock_held(void) | |
162 | { | |
163 | if (!debug_lockdep_rcu_enabled()) | |
164 | return 1; | |
165 | if (!rcu_is_watching()) | |
166 | return 0; | |
167 | if (!rcu_lockdep_current_cpu_online()) | |
168 | return 0; | |
169 | return lock_is_held(&rcu_lock_map); | |
170 | } | |
171 | EXPORT_SYMBOL_GPL(rcu_read_lock_held); | |
172 | ||
e3818b8d | 173 | /** |
ca5ecddf | 174 | * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section? |
e3818b8d PM |
175 | * |
176 | * Check for bottom half being disabled, which covers both the | |
177 | * CONFIG_PROVE_RCU and not cases. Note that if someone uses | |
178 | * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled) | |
ca5ecddf PM |
179 | * will show the situation. This is useful for debug checks in functions |
180 | * that require that they be called within an RCU read-side critical | |
181 | * section. | |
e3818b8d PM |
182 | * |
183 | * Check debug_lockdep_rcu_enabled() to prevent false positives during boot. | |
c0d6d01b PM |
184 | * |
185 | * Note that rcu_read_lock() is disallowed if the CPU is either idle or | |
186 | * offline from an RCU perspective, so check for those as well. | |
e3818b8d PM |
187 | */ |
188 | int rcu_read_lock_bh_held(void) | |
189 | { | |
190 | if (!debug_lockdep_rcu_enabled()) | |
191 | return 1; | |
5c173eb8 | 192 | if (!rcu_is_watching()) |
e6b80a3b | 193 | return 0; |
c0d6d01b PM |
194 | if (!rcu_lockdep_current_cpu_online()) |
195 | return 0; | |
773e3f93 | 196 | return in_softirq() || irqs_disabled(); |
e3818b8d PM |
197 | } |
198 | EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held); | |
199 | ||
200 | #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
201 | ||
2c42818e PM |
202 | struct rcu_synchronize { |
203 | struct rcu_head head; | |
204 | struct completion completion; | |
205 | }; | |
206 | ||
fbf6bfca PM |
207 | /* |
208 | * Awaken the corresponding synchronize_rcu() instance now that a | |
209 | * grace period has elapsed. | |
210 | */ | |
2c42818e | 211 | static void wakeme_after_rcu(struct rcu_head *head) |
21a1ea9e | 212 | { |
01c1c660 PM |
213 | struct rcu_synchronize *rcu; |
214 | ||
215 | rcu = container_of(head, struct rcu_synchronize, head); | |
216 | complete(&rcu->completion); | |
21a1ea9e | 217 | } |
ee84b824 | 218 | |
2c42818e PM |
219 | void wait_rcu_gp(call_rcu_func_t crf) |
220 | { | |
221 | struct rcu_synchronize rcu; | |
222 | ||
223 | init_rcu_head_on_stack(&rcu.head); | |
224 | init_completion(&rcu.completion); | |
225 | /* Will wake me after RCU finished. */ | |
226 | crf(&rcu.head, wakeme_after_rcu); | |
227 | /* Wait for it. */ | |
228 | wait_for_completion(&rcu.completion); | |
229 | destroy_rcu_head_on_stack(&rcu.head); | |
230 | } | |
231 | EXPORT_SYMBOL_GPL(wait_rcu_gp); | |
232 | ||
551d55a9 | 233 | #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD |
546a9d85 | 234 | void init_rcu_head(struct rcu_head *head) |
551d55a9 MD |
235 | { |
236 | debug_object_init(head, &rcuhead_debug_descr); | |
237 | } | |
238 | ||
546a9d85 | 239 | void destroy_rcu_head(struct rcu_head *head) |
551d55a9 MD |
240 | { |
241 | debug_object_free(head, &rcuhead_debug_descr); | |
242 | } | |
243 | ||
551d55a9 MD |
244 | /* |
245 | * fixup_activate is called when: | |
246 | * - an active object is activated | |
247 | * - an unknown object is activated (might be a statically initialized object) | |
248 | * Activation is performed internally by call_rcu(). | |
249 | */ | |
250 | static int rcuhead_fixup_activate(void *addr, enum debug_obj_state state) | |
251 | { | |
252 | struct rcu_head *head = addr; | |
253 | ||
254 | switch (state) { | |
255 | ||
256 | case ODEBUG_STATE_NOTAVAILABLE: | |
257 | /* | |
258 | * This is not really a fixup. We just make sure that it is | |
259 | * tracked in the object tracker. | |
260 | */ | |
261 | debug_object_init(head, &rcuhead_debug_descr); | |
262 | debug_object_activate(head, &rcuhead_debug_descr); | |
263 | return 0; | |
551d55a9 | 264 | default: |
551d55a9 | 265 | return 1; |
551d55a9 MD |
266 | } |
267 | } | |
268 | ||
269 | /** | |
270 | * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects | |
271 | * @head: pointer to rcu_head structure to be initialized | |
272 | * | |
273 | * This function informs debugobjects of a new rcu_head structure that | |
274 | * has been allocated as an auto variable on the stack. This function | |
275 | * is not required for rcu_head structures that are statically defined or | |
276 | * that are dynamically allocated on the heap. This function has no | |
277 | * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds. | |
278 | */ | |
279 | void init_rcu_head_on_stack(struct rcu_head *head) | |
280 | { | |
281 | debug_object_init_on_stack(head, &rcuhead_debug_descr); | |
282 | } | |
283 | EXPORT_SYMBOL_GPL(init_rcu_head_on_stack); | |
284 | ||
285 | /** | |
286 | * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects | |
287 | * @head: pointer to rcu_head structure to be initialized | |
288 | * | |
289 | * This function informs debugobjects that an on-stack rcu_head structure | |
290 | * is about to go out of scope. As with init_rcu_head_on_stack(), this | |
291 | * function is not required for rcu_head structures that are statically | |
292 | * defined or that are dynamically allocated on the heap. Also as with | |
293 | * init_rcu_head_on_stack(), this function has no effect for | |
294 | * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds. | |
295 | */ | |
296 | void destroy_rcu_head_on_stack(struct rcu_head *head) | |
297 | { | |
298 | debug_object_free(head, &rcuhead_debug_descr); | |
299 | } | |
300 | EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack); | |
301 | ||
302 | struct debug_obj_descr rcuhead_debug_descr = { | |
303 | .name = "rcu_head", | |
551d55a9 | 304 | .fixup_activate = rcuhead_fixup_activate, |
551d55a9 MD |
305 | }; |
306 | EXPORT_SYMBOL_GPL(rcuhead_debug_descr); | |
307 | #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */ | |
91afaf30 | 308 | |
28f6569a | 309 | #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE) |
e66c33d5 | 310 | void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp, |
52494535 PM |
311 | unsigned long secs, |
312 | unsigned long c_old, unsigned long c) | |
91afaf30 | 313 | { |
52494535 | 314 | trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c); |
91afaf30 PM |
315 | } |
316 | EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read); | |
317 | #else | |
52494535 PM |
318 | #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \ |
319 | do { } while (0) | |
91afaf30 | 320 | #endif |
6bfc09e2 PM |
321 | |
322 | #ifdef CONFIG_RCU_STALL_COMMON | |
323 | ||
324 | #ifdef CONFIG_PROVE_RCU | |
325 | #define RCU_STALL_DELAY_DELTA (5 * HZ) | |
326 | #else | |
327 | #define RCU_STALL_DELAY_DELTA 0 | |
328 | #endif | |
329 | ||
330 | int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */ | |
01896f7e | 331 | static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT; |
6bfc09e2 PM |
332 | |
333 | module_param(rcu_cpu_stall_suppress, int, 0644); | |
334 | module_param(rcu_cpu_stall_timeout, int, 0644); | |
335 | ||
336 | int rcu_jiffies_till_stall_check(void) | |
337 | { | |
338 | int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout); | |
339 | ||
340 | /* | |
341 | * Limit check must be consistent with the Kconfig limits | |
342 | * for CONFIG_RCU_CPU_STALL_TIMEOUT. | |
343 | */ | |
344 | if (till_stall_check < 3) { | |
345 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 3; | |
346 | till_stall_check = 3; | |
347 | } else if (till_stall_check > 300) { | |
348 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 300; | |
349 | till_stall_check = 300; | |
350 | } | |
351 | return till_stall_check * HZ + RCU_STALL_DELAY_DELTA; | |
352 | } | |
353 | ||
61f38db3 RR |
354 | void rcu_sysrq_start(void) |
355 | { | |
356 | if (!rcu_cpu_stall_suppress) | |
357 | rcu_cpu_stall_suppress = 2; | |
358 | } | |
359 | ||
360 | void rcu_sysrq_end(void) | |
361 | { | |
362 | if (rcu_cpu_stall_suppress == 2) | |
363 | rcu_cpu_stall_suppress = 0; | |
364 | } | |
365 | ||
6bfc09e2 PM |
366 | static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr) |
367 | { | |
368 | rcu_cpu_stall_suppress = 1; | |
369 | return NOTIFY_DONE; | |
370 | } | |
371 | ||
372 | static struct notifier_block rcu_panic_block = { | |
373 | .notifier_call = rcu_panic, | |
374 | }; | |
375 | ||
376 | static int __init check_cpu_stall_init(void) | |
377 | { | |
378 | atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block); | |
379 | return 0; | |
380 | } | |
381 | early_initcall(check_cpu_stall_init); | |
382 | ||
383 | #endif /* #ifdef CONFIG_RCU_STALL_COMMON */ | |
8315f422 PM |
384 | |
385 | #ifdef CONFIG_TASKS_RCU | |
386 | ||
387 | /* | |
388 | * Simple variant of RCU whose quiescent states are voluntary context switch, | |
389 | * user-space execution, and idle. As such, grace periods can take one good | |
390 | * long time. There are no read-side primitives similar to rcu_read_lock() | |
391 | * and rcu_read_unlock() because this implementation is intended to get | |
392 | * the system into a safe state for some of the manipulations involved in | |
393 | * tracing and the like. Finally, this implementation does not support | |
394 | * high call_rcu_tasks() rates from multiple CPUs. If this is required, | |
395 | * per-CPU callback lists will be needed. | |
396 | */ | |
397 | ||
398 | /* Global list of callbacks and associated lock. */ | |
399 | static struct rcu_head *rcu_tasks_cbs_head; | |
400 | static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head; | |
c7b24d2b | 401 | static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq); |
8315f422 PM |
402 | static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock); |
403 | ||
3f95aa81 PM |
404 | /* Track exiting tasks in order to allow them to be waited for. */ |
405 | DEFINE_SRCU(tasks_rcu_exit_srcu); | |
406 | ||
407 | /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */ | |
52db30ab | 408 | static int rcu_task_stall_timeout __read_mostly = HZ * 60 * 10; |
3f95aa81 PM |
409 | module_param(rcu_task_stall_timeout, int, 0644); |
410 | ||
84a8f446 PM |
411 | static void rcu_spawn_tasks_kthread(void); |
412 | ||
413 | /* | |
414 | * Post an RCU-tasks callback. First call must be from process context | |
415 | * after the scheduler if fully operational. | |
416 | */ | |
8315f422 PM |
417 | void call_rcu_tasks(struct rcu_head *rhp, void (*func)(struct rcu_head *rhp)) |
418 | { | |
419 | unsigned long flags; | |
c7b24d2b | 420 | bool needwake; |
8315f422 PM |
421 | |
422 | rhp->next = NULL; | |
423 | rhp->func = func; | |
424 | raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags); | |
c7b24d2b | 425 | needwake = !rcu_tasks_cbs_head; |
8315f422 PM |
426 | *rcu_tasks_cbs_tail = rhp; |
427 | rcu_tasks_cbs_tail = &rhp->next; | |
428 | raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags); | |
84a8f446 PM |
429 | if (needwake) { |
430 | rcu_spawn_tasks_kthread(); | |
c7b24d2b | 431 | wake_up(&rcu_tasks_cbs_wq); |
84a8f446 | 432 | } |
8315f422 PM |
433 | } |
434 | EXPORT_SYMBOL_GPL(call_rcu_tasks); | |
435 | ||
53c6d4ed PM |
436 | /** |
437 | * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed. | |
438 | * | |
439 | * Control will return to the caller some time after a full rcu-tasks | |
440 | * grace period has elapsed, in other words after all currently | |
441 | * executing rcu-tasks read-side critical sections have elapsed. These | |
442 | * read-side critical sections are delimited by calls to schedule(), | |
443 | * cond_resched_rcu_qs(), idle execution, userspace execution, calls | |
444 | * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched(). | |
445 | * | |
446 | * This is a very specialized primitive, intended only for a few uses in | |
447 | * tracing and other situations requiring manipulation of function | |
448 | * preambles and profiling hooks. The synchronize_rcu_tasks() function | |
449 | * is not (yet) intended for heavy use from multiple CPUs. | |
450 | * | |
451 | * Note that this guarantee implies further memory-ordering guarantees. | |
452 | * On systems with more than one CPU, when synchronize_rcu_tasks() returns, | |
453 | * each CPU is guaranteed to have executed a full memory barrier since the | |
454 | * end of its last RCU-tasks read-side critical section whose beginning | |
455 | * preceded the call to synchronize_rcu_tasks(). In addition, each CPU | |
456 | * having an RCU-tasks read-side critical section that extends beyond | |
457 | * the return from synchronize_rcu_tasks() is guaranteed to have executed | |
458 | * a full memory barrier after the beginning of synchronize_rcu_tasks() | |
459 | * and before the beginning of that RCU-tasks read-side critical section. | |
460 | * Note that these guarantees include CPUs that are offline, idle, or | |
461 | * executing in user mode, as well as CPUs that are executing in the kernel. | |
462 | * | |
463 | * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned | |
464 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
465 | * to have executed a full memory barrier during the execution of | |
466 | * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU | |
467 | * (but again only if the system has more than one CPU). | |
468 | */ | |
469 | void synchronize_rcu_tasks(void) | |
470 | { | |
471 | /* Complain if the scheduler has not started. */ | |
472 | rcu_lockdep_assert(!rcu_scheduler_active, | |
473 | "synchronize_rcu_tasks called too soon"); | |
474 | ||
475 | /* Wait for the grace period. */ | |
476 | wait_rcu_gp(call_rcu_tasks); | |
477 | } | |
06c2a923 | 478 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks); |
53c6d4ed PM |
479 | |
480 | /** | |
481 | * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks. | |
482 | * | |
483 | * Although the current implementation is guaranteed to wait, it is not | |
484 | * obligated to, for example, if there are no pending callbacks. | |
485 | */ | |
486 | void rcu_barrier_tasks(void) | |
487 | { | |
488 | /* There is only one callback queue, so this is easy. ;-) */ | |
489 | synchronize_rcu_tasks(); | |
490 | } | |
06c2a923 | 491 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks); |
53c6d4ed | 492 | |
52db30ab PM |
493 | /* See if tasks are still holding out, complain if so. */ |
494 | static void check_holdout_task(struct task_struct *t, | |
495 | bool needreport, bool *firstreport) | |
8315f422 | 496 | { |
4ff475ed PM |
497 | int cpu; |
498 | ||
8315f422 PM |
499 | if (!ACCESS_ONCE(t->rcu_tasks_holdout) || |
500 | t->rcu_tasks_nvcsw != ACCESS_ONCE(t->nvcsw) || | |
176f8f7a PM |
501 | !ACCESS_ONCE(t->on_rq) || |
502 | (IS_ENABLED(CONFIG_NO_HZ_FULL) && | |
503 | !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) { | |
8315f422 | 504 | ACCESS_ONCE(t->rcu_tasks_holdout) = false; |
8f20a5e8 | 505 | list_del_init(&t->rcu_tasks_holdout_list); |
8315f422 | 506 | put_task_struct(t); |
52db30ab | 507 | return; |
8315f422 | 508 | } |
52db30ab PM |
509 | if (!needreport) |
510 | return; | |
511 | if (*firstreport) { | |
512 | pr_err("INFO: rcu_tasks detected stalls on tasks:\n"); | |
513 | *firstreport = false; | |
514 | } | |
4ff475ed PM |
515 | cpu = task_cpu(t); |
516 | pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n", | |
517 | t, ".I"[is_idle_task(t)], | |
518 | "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)], | |
519 | t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout, | |
520 | t->rcu_tasks_idle_cpu, cpu); | |
52db30ab | 521 | sched_show_task(t); |
8315f422 PM |
522 | } |
523 | ||
524 | /* RCU-tasks kthread that detects grace periods and invokes callbacks. */ | |
525 | static int __noreturn rcu_tasks_kthread(void *arg) | |
526 | { | |
527 | unsigned long flags; | |
528 | struct task_struct *g, *t; | |
52db30ab | 529 | unsigned long lastreport; |
8315f422 PM |
530 | struct rcu_head *list; |
531 | struct rcu_head *next; | |
532 | LIST_HEAD(rcu_tasks_holdouts); | |
533 | ||
60ced495 PM |
534 | /* Run on housekeeping CPUs by default. Sysadm can move if desired. */ |
535 | housekeeping_affine(current); | |
8315f422 PM |
536 | |
537 | /* | |
538 | * Each pass through the following loop makes one check for | |
539 | * newly arrived callbacks, and, if there are some, waits for | |
540 | * one RCU-tasks grace period and then invokes the callbacks. | |
541 | * This loop is terminated by the system going down. ;-) | |
542 | */ | |
543 | for (;;) { | |
544 | ||
545 | /* Pick up any new callbacks. */ | |
546 | raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags); | |
547 | list = rcu_tasks_cbs_head; | |
548 | rcu_tasks_cbs_head = NULL; | |
549 | rcu_tasks_cbs_tail = &rcu_tasks_cbs_head; | |
550 | raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags); | |
551 | ||
552 | /* If there were none, wait a bit and start over. */ | |
553 | if (!list) { | |
c7b24d2b PM |
554 | wait_event_interruptible(rcu_tasks_cbs_wq, |
555 | rcu_tasks_cbs_head); | |
556 | if (!rcu_tasks_cbs_head) { | |
557 | WARN_ON(signal_pending(current)); | |
558 | schedule_timeout_interruptible(HZ/10); | |
559 | } | |
8315f422 PM |
560 | continue; |
561 | } | |
562 | ||
563 | /* | |
564 | * Wait for all pre-existing t->on_rq and t->nvcsw | |
565 | * transitions to complete. Invoking synchronize_sched() | |
566 | * suffices because all these transitions occur with | |
567 | * interrupts disabled. Without this synchronize_sched(), | |
568 | * a read-side critical section that started before the | |
569 | * grace period might be incorrectly seen as having started | |
570 | * after the grace period. | |
571 | * | |
572 | * This synchronize_sched() also dispenses with the | |
573 | * need for a memory barrier on the first store to | |
574 | * ->rcu_tasks_holdout, as it forces the store to happen | |
575 | * after the beginning of the grace period. | |
576 | */ | |
577 | synchronize_sched(); | |
578 | ||
579 | /* | |
580 | * There were callbacks, so we need to wait for an | |
581 | * RCU-tasks grace period. Start off by scanning | |
582 | * the task list for tasks that are not already | |
583 | * voluntarily blocked. Mark these tasks and make | |
584 | * a list of them in rcu_tasks_holdouts. | |
585 | */ | |
586 | rcu_read_lock(); | |
587 | for_each_process_thread(g, t) { | |
588 | if (t != current && ACCESS_ONCE(t->on_rq) && | |
589 | !is_idle_task(t)) { | |
590 | get_task_struct(t); | |
591 | t->rcu_tasks_nvcsw = ACCESS_ONCE(t->nvcsw); | |
592 | ACCESS_ONCE(t->rcu_tasks_holdout) = true; | |
593 | list_add(&t->rcu_tasks_holdout_list, | |
594 | &rcu_tasks_holdouts); | |
595 | } | |
596 | } | |
597 | rcu_read_unlock(); | |
598 | ||
3f95aa81 PM |
599 | /* |
600 | * Wait for tasks that are in the process of exiting. | |
601 | * This does only part of the job, ensuring that all | |
602 | * tasks that were previously exiting reach the point | |
603 | * where they have disabled preemption, allowing the | |
604 | * later synchronize_sched() to finish the job. | |
605 | */ | |
606 | synchronize_srcu(&tasks_rcu_exit_srcu); | |
607 | ||
8315f422 PM |
608 | /* |
609 | * Each pass through the following loop scans the list | |
610 | * of holdout tasks, removing any that are no longer | |
611 | * holdouts. When the list is empty, we are done. | |
612 | */ | |
52db30ab | 613 | lastreport = jiffies; |
8315f422 | 614 | while (!list_empty(&rcu_tasks_holdouts)) { |
52db30ab PM |
615 | bool firstreport; |
616 | bool needreport; | |
617 | int rtst; | |
8f20a5e8 | 618 | struct task_struct *t1; |
52db30ab | 619 | |
8315f422 | 620 | schedule_timeout_interruptible(HZ); |
52db30ab PM |
621 | rtst = ACCESS_ONCE(rcu_task_stall_timeout); |
622 | needreport = rtst > 0 && | |
623 | time_after(jiffies, lastreport + rtst); | |
624 | if (needreport) | |
625 | lastreport = jiffies; | |
626 | firstreport = true; | |
8315f422 | 627 | WARN_ON(signal_pending(current)); |
8f20a5e8 PM |
628 | list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts, |
629 | rcu_tasks_holdout_list) { | |
52db30ab | 630 | check_holdout_task(t, needreport, &firstreport); |
8f20a5e8 PM |
631 | cond_resched(); |
632 | } | |
8315f422 PM |
633 | } |
634 | ||
635 | /* | |
636 | * Because ->on_rq and ->nvcsw are not guaranteed | |
637 | * to have a full memory barriers prior to them in the | |
638 | * schedule() path, memory reordering on other CPUs could | |
639 | * cause their RCU-tasks read-side critical sections to | |
640 | * extend past the end of the grace period. However, | |
641 | * because these ->nvcsw updates are carried out with | |
642 | * interrupts disabled, we can use synchronize_sched() | |
643 | * to force the needed ordering on all such CPUs. | |
644 | * | |
645 | * This synchronize_sched() also confines all | |
646 | * ->rcu_tasks_holdout accesses to be within the grace | |
647 | * period, avoiding the need for memory barriers for | |
648 | * ->rcu_tasks_holdout accesses. | |
3f95aa81 PM |
649 | * |
650 | * In addition, this synchronize_sched() waits for exiting | |
651 | * tasks to complete their final preempt_disable() region | |
652 | * of execution, cleaning up after the synchronize_srcu() | |
653 | * above. | |
8315f422 PM |
654 | */ |
655 | synchronize_sched(); | |
656 | ||
657 | /* Invoke the callbacks. */ | |
658 | while (list) { | |
659 | next = list->next; | |
660 | local_bh_disable(); | |
661 | list->func(list); | |
662 | local_bh_enable(); | |
663 | list = next; | |
664 | cond_resched(); | |
665 | } | |
c7b24d2b | 666 | schedule_timeout_uninterruptible(HZ/10); |
8315f422 PM |
667 | } |
668 | } | |
669 | ||
84a8f446 PM |
670 | /* Spawn rcu_tasks_kthread() at first call to call_rcu_tasks(). */ |
671 | static void rcu_spawn_tasks_kthread(void) | |
8315f422 | 672 | { |
84a8f446 PM |
673 | static DEFINE_MUTEX(rcu_tasks_kthread_mutex); |
674 | static struct task_struct *rcu_tasks_kthread_ptr; | |
675 | struct task_struct *t; | |
8315f422 | 676 | |
84a8f446 PM |
677 | if (ACCESS_ONCE(rcu_tasks_kthread_ptr)) { |
678 | smp_mb(); /* Ensure caller sees full kthread. */ | |
679 | return; | |
680 | } | |
681 | mutex_lock(&rcu_tasks_kthread_mutex); | |
682 | if (rcu_tasks_kthread_ptr) { | |
683 | mutex_unlock(&rcu_tasks_kthread_mutex); | |
684 | return; | |
685 | } | |
8315f422 PM |
686 | t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread"); |
687 | BUG_ON(IS_ERR(t)); | |
84a8f446 PM |
688 | smp_mb(); /* Ensure others see full kthread. */ |
689 | ACCESS_ONCE(rcu_tasks_kthread_ptr) = t; | |
690 | mutex_unlock(&rcu_tasks_kthread_mutex); | |
8315f422 | 691 | } |
8315f422 PM |
692 | |
693 | #endif /* #ifdef CONFIG_TASKS_RCU */ | |
aa23c6fb PK |
694 | |
695 | #ifdef CONFIG_PROVE_RCU | |
696 | ||
697 | /* | |
698 | * Early boot self test parameters, one for each flavor | |
699 | */ | |
700 | static bool rcu_self_test; | |
701 | static bool rcu_self_test_bh; | |
702 | static bool rcu_self_test_sched; | |
703 | ||
704 | module_param(rcu_self_test, bool, 0444); | |
705 | module_param(rcu_self_test_bh, bool, 0444); | |
706 | module_param(rcu_self_test_sched, bool, 0444); | |
707 | ||
708 | static int rcu_self_test_counter; | |
709 | ||
710 | static void test_callback(struct rcu_head *r) | |
711 | { | |
712 | rcu_self_test_counter++; | |
713 | pr_info("RCU test callback executed %d\n", rcu_self_test_counter); | |
714 | } | |
715 | ||
716 | static void early_boot_test_call_rcu(void) | |
717 | { | |
718 | static struct rcu_head head; | |
719 | ||
720 | call_rcu(&head, test_callback); | |
721 | } | |
722 | ||
723 | static void early_boot_test_call_rcu_bh(void) | |
724 | { | |
725 | static struct rcu_head head; | |
726 | ||
727 | call_rcu_bh(&head, test_callback); | |
728 | } | |
729 | ||
730 | static void early_boot_test_call_rcu_sched(void) | |
731 | { | |
732 | static struct rcu_head head; | |
733 | ||
734 | call_rcu_sched(&head, test_callback); | |
735 | } | |
736 | ||
737 | void rcu_early_boot_tests(void) | |
738 | { | |
739 | pr_info("Running RCU self tests\n"); | |
740 | ||
741 | if (rcu_self_test) | |
742 | early_boot_test_call_rcu(); | |
743 | if (rcu_self_test_bh) | |
744 | early_boot_test_call_rcu_bh(); | |
745 | if (rcu_self_test_sched) | |
746 | early_boot_test_call_rcu_sched(); | |
747 | } | |
748 | ||
749 | static int rcu_verify_early_boot_tests(void) | |
750 | { | |
751 | int ret = 0; | |
752 | int early_boot_test_counter = 0; | |
753 | ||
754 | if (rcu_self_test) { | |
755 | early_boot_test_counter++; | |
756 | rcu_barrier(); | |
757 | } | |
758 | if (rcu_self_test_bh) { | |
759 | early_boot_test_counter++; | |
760 | rcu_barrier_bh(); | |
761 | } | |
762 | if (rcu_self_test_sched) { | |
763 | early_boot_test_counter++; | |
764 | rcu_barrier_sched(); | |
765 | } | |
766 | ||
767 | if (rcu_self_test_counter != early_boot_test_counter) { | |
768 | WARN_ON(1); | |
769 | ret = -1; | |
770 | } | |
771 | ||
772 | return ret; | |
773 | } | |
774 | late_initcall(rcu_verify_early_boot_tests); | |
775 | #else | |
776 | void rcu_early_boot_tests(void) {} | |
777 | #endif /* CONFIG_PROVE_RCU */ |