]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - kernel/rcu/update.c
Merge branch 'work.sys_wait' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[mirror_ubuntu-artful-kernel.git] / kernel / rcu / update.c
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
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
17 *
18 * Copyright IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 *
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 -
30 * http://lse.sourceforge.net/locking/rcupdate.html
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/signal.h>
40 #include <linux/sched/debug.h>
41 #include <linux/atomic.h>
42 #include <linux/bitops.h>
43 #include <linux/percpu.h>
44 #include <linux/notifier.h>
45 #include <linux/cpu.h>
46 #include <linux/mutex.h>
47 #include <linux/export.h>
48 #include <linux/hardirq.h>
49 #include <linux/delay.h>
50 #include <linux/moduleparam.h>
51 #include <linux/kthread.h>
52 #include <linux/tick.h>
53 #include <linux/rcupdate_wait.h>
54
55 #define CREATE_TRACE_POINTS
56
57 #include "rcu.h"
58
59 #ifdef MODULE_PARAM_PREFIX
60 #undef MODULE_PARAM_PREFIX
61 #endif
62 #define MODULE_PARAM_PREFIX "rcupdate."
63
64 #ifndef CONFIG_TINY_RCU
65 extern int rcu_expedited; /* from sysctl */
66 module_param(rcu_expedited, int, 0);
67 extern int rcu_normal; /* from sysctl */
68 module_param(rcu_normal, int, 0);
69 static int rcu_normal_after_boot;
70 module_param(rcu_normal_after_boot, int, 0);
71 #endif /* #ifndef CONFIG_TINY_RCU */
72
73 #ifdef CONFIG_DEBUG_LOCK_ALLOC
74 /**
75 * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
76 *
77 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
78 * RCU-sched read-side critical section. In absence of
79 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
80 * critical section unless it can prove otherwise. Note that disabling
81 * of preemption (including disabling irqs) counts as an RCU-sched
82 * read-side critical section. This is useful for debug checks in functions
83 * that required that they be called within an RCU-sched read-side
84 * critical section.
85 *
86 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
87 * and while lockdep is disabled.
88 *
89 * Note that if the CPU is in the idle loop from an RCU point of
90 * view (ie: that we are in the section between rcu_idle_enter() and
91 * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
92 * did an rcu_read_lock(). The reason for this is that RCU ignores CPUs
93 * that are in such a section, considering these as in extended quiescent
94 * state, so such a CPU is effectively never in an RCU read-side critical
95 * section regardless of what RCU primitives it invokes. This state of
96 * affairs is required --- we need to keep an RCU-free window in idle
97 * where the CPU may possibly enter into low power mode. This way we can
98 * notice an extended quiescent state to other CPUs that started a grace
99 * period. Otherwise we would delay any grace period as long as we run in
100 * the idle task.
101 *
102 * Similarly, we avoid claiming an SRCU read lock held if the current
103 * CPU is offline.
104 */
105 int rcu_read_lock_sched_held(void)
106 {
107 int lockdep_opinion = 0;
108
109 if (!debug_lockdep_rcu_enabled())
110 return 1;
111 if (!rcu_is_watching())
112 return 0;
113 if (!rcu_lockdep_current_cpu_online())
114 return 0;
115 if (debug_locks)
116 lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
117 return lockdep_opinion || !preemptible();
118 }
119 EXPORT_SYMBOL(rcu_read_lock_sched_held);
120 #endif
121
122 #ifndef CONFIG_TINY_RCU
123
124 /*
125 * Should expedited grace-period primitives always fall back to their
126 * non-expedited counterparts? Intended for use within RCU. Note
127 * that if the user specifies both rcu_expedited and rcu_normal, then
128 * rcu_normal wins. (Except during the time period during boot from
129 * when the first task is spawned until the rcu_set_runtime_mode()
130 * core_initcall() is invoked, at which point everything is expedited.)
131 */
132 bool rcu_gp_is_normal(void)
133 {
134 return READ_ONCE(rcu_normal) &&
135 rcu_scheduler_active != RCU_SCHEDULER_INIT;
136 }
137 EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
138
139 static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
140
141 /*
142 * Should normal grace-period primitives be expedited? Intended for
143 * use within RCU. Note that this function takes the rcu_expedited
144 * sysfs/boot variable and rcu_scheduler_active into account as well
145 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
146 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
147 */
148 bool rcu_gp_is_expedited(void)
149 {
150 return rcu_expedited || atomic_read(&rcu_expedited_nesting) ||
151 rcu_scheduler_active == RCU_SCHEDULER_INIT;
152 }
153 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
154
155 /**
156 * rcu_expedite_gp - Expedite future RCU grace periods
157 *
158 * After a call to this function, future calls to synchronize_rcu() and
159 * friends act as the corresponding synchronize_rcu_expedited() function
160 * had instead been called.
161 */
162 void rcu_expedite_gp(void)
163 {
164 atomic_inc(&rcu_expedited_nesting);
165 }
166 EXPORT_SYMBOL_GPL(rcu_expedite_gp);
167
168 /**
169 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
170 *
171 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
172 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
173 * and if the rcu_expedited sysfs/boot parameter is not set, then all
174 * subsequent calls to synchronize_rcu() and friends will return to
175 * their normal non-expedited behavior.
176 */
177 void rcu_unexpedite_gp(void)
178 {
179 atomic_dec(&rcu_expedited_nesting);
180 }
181 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
182
183 /*
184 * Inform RCU of the end of the in-kernel boot sequence.
185 */
186 void rcu_end_inkernel_boot(void)
187 {
188 rcu_unexpedite_gp();
189 if (rcu_normal_after_boot)
190 WRITE_ONCE(rcu_normal, 1);
191 }
192
193 #endif /* #ifndef CONFIG_TINY_RCU */
194
195 /*
196 * Test each non-SRCU synchronous grace-period wait API. This is
197 * useful just after a change in mode for these primitives, and
198 * during early boot.
199 */
200 void rcu_test_sync_prims(void)
201 {
202 if (!IS_ENABLED(CONFIG_PROVE_RCU))
203 return;
204 synchronize_rcu();
205 synchronize_rcu_bh();
206 synchronize_sched();
207 synchronize_rcu_expedited();
208 synchronize_rcu_bh_expedited();
209 synchronize_sched_expedited();
210 }
211
212 #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
213
214 /*
215 * Switch to run-time mode once RCU has fully initialized.
216 */
217 static int __init rcu_set_runtime_mode(void)
218 {
219 rcu_test_sync_prims();
220 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
221 rcu_test_sync_prims();
222 return 0;
223 }
224 core_initcall(rcu_set_runtime_mode);
225
226 #endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
227
228 #ifdef CONFIG_PREEMPT_RCU
229
230 /*
231 * Preemptible RCU implementation for rcu_read_lock().
232 * Just increment ->rcu_read_lock_nesting, shared state will be updated
233 * if we block.
234 */
235 void __rcu_read_lock(void)
236 {
237 current->rcu_read_lock_nesting++;
238 barrier(); /* critical section after entry code. */
239 }
240 EXPORT_SYMBOL_GPL(__rcu_read_lock);
241
242 /*
243 * Preemptible RCU implementation for rcu_read_unlock().
244 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
245 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
246 * invoke rcu_read_unlock_special() to clean up after a context switch
247 * in an RCU read-side critical section and other special cases.
248 */
249 void __rcu_read_unlock(void)
250 {
251 struct task_struct *t = current;
252
253 if (t->rcu_read_lock_nesting != 1) {
254 --t->rcu_read_lock_nesting;
255 } else {
256 barrier(); /* critical section before exit code. */
257 t->rcu_read_lock_nesting = INT_MIN;
258 barrier(); /* assign before ->rcu_read_unlock_special load */
259 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
260 rcu_read_unlock_special(t);
261 barrier(); /* ->rcu_read_unlock_special load before assign */
262 t->rcu_read_lock_nesting = 0;
263 }
264 #ifdef CONFIG_PROVE_LOCKING
265 {
266 int rrln = READ_ONCE(t->rcu_read_lock_nesting);
267
268 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
269 }
270 #endif /* #ifdef CONFIG_PROVE_LOCKING */
271 }
272 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
273
274 #endif /* #ifdef CONFIG_PREEMPT_RCU */
275
276 #ifdef CONFIG_DEBUG_LOCK_ALLOC
277 static struct lock_class_key rcu_lock_key;
278 struct lockdep_map rcu_lock_map =
279 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
280 EXPORT_SYMBOL_GPL(rcu_lock_map);
281
282 static struct lock_class_key rcu_bh_lock_key;
283 struct lockdep_map rcu_bh_lock_map =
284 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
285 EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
286
287 static struct lock_class_key rcu_sched_lock_key;
288 struct lockdep_map rcu_sched_lock_map =
289 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
290 EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
291
292 static struct lock_class_key rcu_callback_key;
293 struct lockdep_map rcu_callback_map =
294 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
295 EXPORT_SYMBOL_GPL(rcu_callback_map);
296
297 int notrace debug_lockdep_rcu_enabled(void)
298 {
299 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
300 current->lockdep_recursion == 0;
301 }
302 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
303
304 /**
305 * rcu_read_lock_held() - might we be in RCU read-side critical section?
306 *
307 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
308 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
309 * this assumes we are in an RCU read-side critical section unless it can
310 * prove otherwise. This is useful for debug checks in functions that
311 * require that they be called within an RCU read-side critical section.
312 *
313 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
314 * and while lockdep is disabled.
315 *
316 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
317 * occur in the same context, for example, it is illegal to invoke
318 * rcu_read_unlock() in process context if the matching rcu_read_lock()
319 * was invoked from within an irq handler.
320 *
321 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
322 * offline from an RCU perspective, so check for those as well.
323 */
324 int rcu_read_lock_held(void)
325 {
326 if (!debug_lockdep_rcu_enabled())
327 return 1;
328 if (!rcu_is_watching())
329 return 0;
330 if (!rcu_lockdep_current_cpu_online())
331 return 0;
332 return lock_is_held(&rcu_lock_map);
333 }
334 EXPORT_SYMBOL_GPL(rcu_read_lock_held);
335
336 /**
337 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
338 *
339 * Check for bottom half being disabled, which covers both the
340 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
341 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
342 * will show the situation. This is useful for debug checks in functions
343 * that require that they be called within an RCU read-side critical
344 * section.
345 *
346 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
347 *
348 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
349 * offline from an RCU perspective, so check for those as well.
350 */
351 int rcu_read_lock_bh_held(void)
352 {
353 if (!debug_lockdep_rcu_enabled())
354 return 1;
355 if (!rcu_is_watching())
356 return 0;
357 if (!rcu_lockdep_current_cpu_online())
358 return 0;
359 return in_softirq() || irqs_disabled();
360 }
361 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
362
363 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
364
365 /**
366 * wakeme_after_rcu() - Callback function to awaken a task after grace period
367 * @head: Pointer to rcu_head member within rcu_synchronize structure
368 *
369 * Awaken the corresponding task now that a grace period has elapsed.
370 */
371 void wakeme_after_rcu(struct rcu_head *head)
372 {
373 struct rcu_synchronize *rcu;
374
375 rcu = container_of(head, struct rcu_synchronize, head);
376 complete(&rcu->completion);
377 }
378 EXPORT_SYMBOL_GPL(wakeme_after_rcu);
379
380 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
381 struct rcu_synchronize *rs_array)
382 {
383 int i;
384 int j;
385
386 /* Initialize and register callbacks for each flavor specified. */
387 for (i = 0; i < n; i++) {
388 if (checktiny &&
389 (crcu_array[i] == call_rcu ||
390 crcu_array[i] == call_rcu_bh)) {
391 might_sleep();
392 continue;
393 }
394 init_rcu_head_on_stack(&rs_array[i].head);
395 init_completion(&rs_array[i].completion);
396 for (j = 0; j < i; j++)
397 if (crcu_array[j] == crcu_array[i])
398 break;
399 if (j == i)
400 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
401 }
402
403 /* Wait for all callbacks to be invoked. */
404 for (i = 0; i < n; i++) {
405 if (checktiny &&
406 (crcu_array[i] == call_rcu ||
407 crcu_array[i] == call_rcu_bh))
408 continue;
409 for (j = 0; j < i; j++)
410 if (crcu_array[j] == crcu_array[i])
411 break;
412 if (j == i)
413 wait_for_completion(&rs_array[i].completion);
414 destroy_rcu_head_on_stack(&rs_array[i].head);
415 }
416 }
417 EXPORT_SYMBOL_GPL(__wait_rcu_gp);
418
419 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
420 void init_rcu_head(struct rcu_head *head)
421 {
422 debug_object_init(head, &rcuhead_debug_descr);
423 }
424
425 void destroy_rcu_head(struct rcu_head *head)
426 {
427 debug_object_free(head, &rcuhead_debug_descr);
428 }
429
430 static bool rcuhead_is_static_object(void *addr)
431 {
432 return true;
433 }
434
435 /**
436 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
437 * @head: pointer to rcu_head structure to be initialized
438 *
439 * This function informs debugobjects of a new rcu_head structure that
440 * has been allocated as an auto variable on the stack. This function
441 * is not required for rcu_head structures that are statically defined or
442 * that are dynamically allocated on the heap. This function has no
443 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
444 */
445 void init_rcu_head_on_stack(struct rcu_head *head)
446 {
447 debug_object_init_on_stack(head, &rcuhead_debug_descr);
448 }
449 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
450
451 /**
452 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
453 * @head: pointer to rcu_head structure to be initialized
454 *
455 * This function informs debugobjects that an on-stack rcu_head structure
456 * is about to go out of scope. As with init_rcu_head_on_stack(), this
457 * function is not required for rcu_head structures that are statically
458 * defined or that are dynamically allocated on the heap. Also as with
459 * init_rcu_head_on_stack(), this function has no effect for
460 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
461 */
462 void destroy_rcu_head_on_stack(struct rcu_head *head)
463 {
464 debug_object_free(head, &rcuhead_debug_descr);
465 }
466 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
467
468 struct debug_obj_descr rcuhead_debug_descr = {
469 .name = "rcu_head",
470 .is_static_object = rcuhead_is_static_object,
471 };
472 EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
473 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
474
475 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
476 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
477 unsigned long secs,
478 unsigned long c_old, unsigned long c)
479 {
480 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
481 }
482 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
483 #else
484 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
485 do { } while (0)
486 #endif
487
488 #ifdef CONFIG_RCU_STALL_COMMON
489
490 #ifdef CONFIG_PROVE_RCU
491 #define RCU_STALL_DELAY_DELTA (5 * HZ)
492 #else
493 #define RCU_STALL_DELAY_DELTA 0
494 #endif
495
496 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
497 static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
498
499 module_param(rcu_cpu_stall_suppress, int, 0644);
500 module_param(rcu_cpu_stall_timeout, int, 0644);
501
502 int rcu_jiffies_till_stall_check(void)
503 {
504 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
505
506 /*
507 * Limit check must be consistent with the Kconfig limits
508 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
509 */
510 if (till_stall_check < 3) {
511 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
512 till_stall_check = 3;
513 } else if (till_stall_check > 300) {
514 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
515 till_stall_check = 300;
516 }
517 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
518 }
519
520 void rcu_sysrq_start(void)
521 {
522 if (!rcu_cpu_stall_suppress)
523 rcu_cpu_stall_suppress = 2;
524 }
525
526 void rcu_sysrq_end(void)
527 {
528 if (rcu_cpu_stall_suppress == 2)
529 rcu_cpu_stall_suppress = 0;
530 }
531
532 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
533 {
534 rcu_cpu_stall_suppress = 1;
535 return NOTIFY_DONE;
536 }
537
538 static struct notifier_block rcu_panic_block = {
539 .notifier_call = rcu_panic,
540 };
541
542 static int __init check_cpu_stall_init(void)
543 {
544 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
545 return 0;
546 }
547 early_initcall(check_cpu_stall_init);
548
549 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */
550
551 #ifdef CONFIG_TASKS_RCU
552
553 /*
554 * Simple variant of RCU whose quiescent states are voluntary context switch,
555 * user-space execution, and idle. As such, grace periods can take one good
556 * long time. There are no read-side primitives similar to rcu_read_lock()
557 * and rcu_read_unlock() because this implementation is intended to get
558 * the system into a safe state for some of the manipulations involved in
559 * tracing and the like. Finally, this implementation does not support
560 * high call_rcu_tasks() rates from multiple CPUs. If this is required,
561 * per-CPU callback lists will be needed.
562 */
563
564 /* Global list of callbacks and associated lock. */
565 static struct rcu_head *rcu_tasks_cbs_head;
566 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
567 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
568 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
569
570 /* Track exiting tasks in order to allow them to be waited for. */
571 DEFINE_SRCU(tasks_rcu_exit_srcu);
572
573 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
574 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
575 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
576 module_param(rcu_task_stall_timeout, int, 0644);
577
578 static void rcu_spawn_tasks_kthread(void);
579 static struct task_struct *rcu_tasks_kthread_ptr;
580
581 /**
582 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
583 * @rhp: structure to be used for queueing the RCU updates.
584 * @func: actual callback function to be invoked after the grace period
585 *
586 * The callback function will be invoked some time after a full grace
587 * period elapses, in other words after all currently executing RCU
588 * read-side critical sections have completed. call_rcu_tasks() assumes
589 * that the read-side critical sections end at a voluntary context
590 * switch (not a preemption!), entry into idle, or transition to usermode
591 * execution. As such, there are no read-side primitives analogous to
592 * rcu_read_lock() and rcu_read_unlock() because this primitive is intended
593 * to determine that all tasks have passed through a safe state, not so
594 * much for data-strcuture synchronization.
595 *
596 * See the description of call_rcu() for more detailed information on
597 * memory ordering guarantees.
598 */
599 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
600 {
601 unsigned long flags;
602 bool needwake;
603 bool havetask = READ_ONCE(rcu_tasks_kthread_ptr);
604
605 rhp->next = NULL;
606 rhp->func = func;
607 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
608 needwake = !rcu_tasks_cbs_head;
609 *rcu_tasks_cbs_tail = rhp;
610 rcu_tasks_cbs_tail = &rhp->next;
611 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
612 /* We can't create the thread unless interrupts are enabled. */
613 if ((needwake && havetask) ||
614 (!havetask && !irqs_disabled_flags(flags))) {
615 rcu_spawn_tasks_kthread();
616 wake_up(&rcu_tasks_cbs_wq);
617 }
618 }
619 EXPORT_SYMBOL_GPL(call_rcu_tasks);
620
621 /**
622 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
623 *
624 * Control will return to the caller some time after a full rcu-tasks
625 * grace period has elapsed, in other words after all currently
626 * executing rcu-tasks read-side critical sections have elapsed. These
627 * read-side critical sections are delimited by calls to schedule(),
628 * cond_resched_rcu_qs(), idle execution, userspace execution, calls
629 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
630 *
631 * This is a very specialized primitive, intended only for a few uses in
632 * tracing and other situations requiring manipulation of function
633 * preambles and profiling hooks. The synchronize_rcu_tasks() function
634 * is not (yet) intended for heavy use from multiple CPUs.
635 *
636 * Note that this guarantee implies further memory-ordering guarantees.
637 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
638 * each CPU is guaranteed to have executed a full memory barrier since the
639 * end of its last RCU-tasks read-side critical section whose beginning
640 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
641 * having an RCU-tasks read-side critical section that extends beyond
642 * the return from synchronize_rcu_tasks() is guaranteed to have executed
643 * a full memory barrier after the beginning of synchronize_rcu_tasks()
644 * and before the beginning of that RCU-tasks read-side critical section.
645 * Note that these guarantees include CPUs that are offline, idle, or
646 * executing in user mode, as well as CPUs that are executing in the kernel.
647 *
648 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
649 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
650 * to have executed a full memory barrier during the execution of
651 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
652 * (but again only if the system has more than one CPU).
653 */
654 void synchronize_rcu_tasks(void)
655 {
656 /* Complain if the scheduler has not started. */
657 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
658 "synchronize_rcu_tasks called too soon");
659
660 /* Wait for the grace period. */
661 wait_rcu_gp(call_rcu_tasks);
662 }
663 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
664
665 /**
666 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
667 *
668 * Although the current implementation is guaranteed to wait, it is not
669 * obligated to, for example, if there are no pending callbacks.
670 */
671 void rcu_barrier_tasks(void)
672 {
673 /* There is only one callback queue, so this is easy. ;-) */
674 synchronize_rcu_tasks();
675 }
676 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
677
678 /* See if tasks are still holding out, complain if so. */
679 static void check_holdout_task(struct task_struct *t,
680 bool needreport, bool *firstreport)
681 {
682 int cpu;
683
684 if (!READ_ONCE(t->rcu_tasks_holdout) ||
685 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
686 !READ_ONCE(t->on_rq) ||
687 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
688 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
689 WRITE_ONCE(t->rcu_tasks_holdout, false);
690 list_del_init(&t->rcu_tasks_holdout_list);
691 put_task_struct(t);
692 return;
693 }
694 rcu_request_urgent_qs_task(t);
695 if (!needreport)
696 return;
697 if (*firstreport) {
698 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
699 *firstreport = false;
700 }
701 cpu = task_cpu(t);
702 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
703 t, ".I"[is_idle_task(t)],
704 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
705 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
706 t->rcu_tasks_idle_cpu, cpu);
707 sched_show_task(t);
708 }
709
710 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
711 static int __noreturn rcu_tasks_kthread(void *arg)
712 {
713 unsigned long flags;
714 struct task_struct *g, *t;
715 unsigned long lastreport;
716 struct rcu_head *list;
717 struct rcu_head *next;
718 LIST_HEAD(rcu_tasks_holdouts);
719
720 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
721 housekeeping_affine(current);
722
723 /*
724 * Each pass through the following loop makes one check for
725 * newly arrived callbacks, and, if there are some, waits for
726 * one RCU-tasks grace period and then invokes the callbacks.
727 * This loop is terminated by the system going down. ;-)
728 */
729 for (;;) {
730
731 /* Pick up any new callbacks. */
732 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
733 list = rcu_tasks_cbs_head;
734 rcu_tasks_cbs_head = NULL;
735 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
736 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
737
738 /* If there were none, wait a bit and start over. */
739 if (!list) {
740 wait_event_interruptible(rcu_tasks_cbs_wq,
741 rcu_tasks_cbs_head);
742 if (!rcu_tasks_cbs_head) {
743 WARN_ON(signal_pending(current));
744 schedule_timeout_interruptible(HZ/10);
745 }
746 continue;
747 }
748
749 /*
750 * Wait for all pre-existing t->on_rq and t->nvcsw
751 * transitions to complete. Invoking synchronize_sched()
752 * suffices because all these transitions occur with
753 * interrupts disabled. Without this synchronize_sched(),
754 * a read-side critical section that started before the
755 * grace period might be incorrectly seen as having started
756 * after the grace period.
757 *
758 * This synchronize_sched() also dispenses with the
759 * need for a memory barrier on the first store to
760 * ->rcu_tasks_holdout, as it forces the store to happen
761 * after the beginning of the grace period.
762 */
763 synchronize_sched();
764
765 /*
766 * There were callbacks, so we need to wait for an
767 * RCU-tasks grace period. Start off by scanning
768 * the task list for tasks that are not already
769 * voluntarily blocked. Mark these tasks and make
770 * a list of them in rcu_tasks_holdouts.
771 */
772 rcu_read_lock();
773 for_each_process_thread(g, t) {
774 if (t != current && READ_ONCE(t->on_rq) &&
775 !is_idle_task(t)) {
776 get_task_struct(t);
777 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
778 WRITE_ONCE(t->rcu_tasks_holdout, true);
779 list_add(&t->rcu_tasks_holdout_list,
780 &rcu_tasks_holdouts);
781 }
782 }
783 rcu_read_unlock();
784
785 /*
786 * Wait for tasks that are in the process of exiting.
787 * This does only part of the job, ensuring that all
788 * tasks that were previously exiting reach the point
789 * where they have disabled preemption, allowing the
790 * later synchronize_sched() to finish the job.
791 */
792 synchronize_srcu(&tasks_rcu_exit_srcu);
793
794 /*
795 * Each pass through the following loop scans the list
796 * of holdout tasks, removing any that are no longer
797 * holdouts. When the list is empty, we are done.
798 */
799 lastreport = jiffies;
800 while (!list_empty(&rcu_tasks_holdouts)) {
801 bool firstreport;
802 bool needreport;
803 int rtst;
804 struct task_struct *t1;
805
806 schedule_timeout_interruptible(HZ);
807 rtst = READ_ONCE(rcu_task_stall_timeout);
808 needreport = rtst > 0 &&
809 time_after(jiffies, lastreport + rtst);
810 if (needreport)
811 lastreport = jiffies;
812 firstreport = true;
813 WARN_ON(signal_pending(current));
814 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
815 rcu_tasks_holdout_list) {
816 check_holdout_task(t, needreport, &firstreport);
817 cond_resched();
818 }
819 }
820
821 /*
822 * Because ->on_rq and ->nvcsw are not guaranteed
823 * to have a full memory barriers prior to them in the
824 * schedule() path, memory reordering on other CPUs could
825 * cause their RCU-tasks read-side critical sections to
826 * extend past the end of the grace period. However,
827 * because these ->nvcsw updates are carried out with
828 * interrupts disabled, we can use synchronize_sched()
829 * to force the needed ordering on all such CPUs.
830 *
831 * This synchronize_sched() also confines all
832 * ->rcu_tasks_holdout accesses to be within the grace
833 * period, avoiding the need for memory barriers for
834 * ->rcu_tasks_holdout accesses.
835 *
836 * In addition, this synchronize_sched() waits for exiting
837 * tasks to complete their final preempt_disable() region
838 * of execution, cleaning up after the synchronize_srcu()
839 * above.
840 */
841 synchronize_sched();
842
843 /* Invoke the callbacks. */
844 while (list) {
845 next = list->next;
846 local_bh_disable();
847 list->func(list);
848 local_bh_enable();
849 list = next;
850 cond_resched();
851 }
852 schedule_timeout_uninterruptible(HZ/10);
853 }
854 }
855
856 /* Spawn rcu_tasks_kthread() at first call to call_rcu_tasks(). */
857 static void rcu_spawn_tasks_kthread(void)
858 {
859 static DEFINE_MUTEX(rcu_tasks_kthread_mutex);
860 struct task_struct *t;
861
862 if (READ_ONCE(rcu_tasks_kthread_ptr)) {
863 smp_mb(); /* Ensure caller sees full kthread. */
864 return;
865 }
866 mutex_lock(&rcu_tasks_kthread_mutex);
867 if (rcu_tasks_kthread_ptr) {
868 mutex_unlock(&rcu_tasks_kthread_mutex);
869 return;
870 }
871 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
872 BUG_ON(IS_ERR(t));
873 smp_mb(); /* Ensure others see full kthread. */
874 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
875 mutex_unlock(&rcu_tasks_kthread_mutex);
876 }
877
878 #endif /* #ifdef CONFIG_TASKS_RCU */
879
880 #ifndef CONFIG_TINY_RCU
881
882 /*
883 * Print any non-default Tasks RCU settings.
884 */
885 static void __init rcu_tasks_bootup_oddness(void)
886 {
887 #ifdef CONFIG_TASKS_RCU
888 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
889 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
890 else
891 pr_info("\tTasks RCU enabled.\n");
892 #endif /* #ifdef CONFIG_TASKS_RCU */
893 }
894
895 #endif /* #ifndef CONFIG_TINY_RCU */
896
897 #ifdef CONFIG_PROVE_RCU
898
899 /*
900 * Early boot self test parameters, one for each flavor
901 */
902 static bool rcu_self_test;
903 static bool rcu_self_test_bh;
904 static bool rcu_self_test_sched;
905
906 module_param(rcu_self_test, bool, 0444);
907 module_param(rcu_self_test_bh, bool, 0444);
908 module_param(rcu_self_test_sched, bool, 0444);
909
910 static int rcu_self_test_counter;
911
912 static void test_callback(struct rcu_head *r)
913 {
914 rcu_self_test_counter++;
915 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
916 }
917
918 static void early_boot_test_call_rcu(void)
919 {
920 static struct rcu_head head;
921
922 call_rcu(&head, test_callback);
923 }
924
925 static void early_boot_test_call_rcu_bh(void)
926 {
927 static struct rcu_head head;
928
929 call_rcu_bh(&head, test_callback);
930 }
931
932 static void early_boot_test_call_rcu_sched(void)
933 {
934 static struct rcu_head head;
935
936 call_rcu_sched(&head, test_callback);
937 }
938
939 void rcu_early_boot_tests(void)
940 {
941 pr_info("Running RCU self tests\n");
942
943 if (rcu_self_test)
944 early_boot_test_call_rcu();
945 if (rcu_self_test_bh)
946 early_boot_test_call_rcu_bh();
947 if (rcu_self_test_sched)
948 early_boot_test_call_rcu_sched();
949 rcu_test_sync_prims();
950 }
951
952 static int rcu_verify_early_boot_tests(void)
953 {
954 int ret = 0;
955 int early_boot_test_counter = 0;
956
957 if (rcu_self_test) {
958 early_boot_test_counter++;
959 rcu_barrier();
960 }
961 if (rcu_self_test_bh) {
962 early_boot_test_counter++;
963 rcu_barrier_bh();
964 }
965 if (rcu_self_test_sched) {
966 early_boot_test_counter++;
967 rcu_barrier_sched();
968 }
969
970 if (rcu_self_test_counter != early_boot_test_counter) {
971 WARN_ON(1);
972 ret = -1;
973 }
974
975 return ret;
976 }
977 late_initcall(rcu_verify_early_boot_tests);
978 #else
979 void rcu_early_boot_tests(void) {}
980 #endif /* CONFIG_PROVE_RCU */
981
982 #ifndef CONFIG_TINY_RCU
983
984 /*
985 * Print any significant non-default boot-time settings.
986 */
987 void __init rcupdate_announce_bootup_oddness(void)
988 {
989 if (rcu_normal)
990 pr_info("\tNo expedited grace period (rcu_normal).\n");
991 else if (rcu_normal_after_boot)
992 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
993 else if (rcu_expedited)
994 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
995 if (rcu_cpu_stall_suppress)
996 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
997 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
998 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
999 rcu_tasks_bootup_oddness();
1000 }
1001
1002 #endif /* #ifndef CONFIG_TINY_RCU */