1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * Task-based RCU implementations.
5 * Copyright (C) 2020 Paul E. McKenney
8 #ifdef CONFIG_TASKS_RCU_GENERIC
10 ////////////////////////////////////////////////////////////////////////
12 // Generic data structures.
15 typedef void (*rcu_tasks_gp_func_t
)(struct rcu_tasks
*rtp
);
16 typedef void (*pregp_func_t
)(void);
17 typedef void (*pertask_func_t
)(struct task_struct
*t
, struct list_head
*hop
);
18 typedef void (*postscan_func_t
)(struct list_head
*hop
);
19 typedef void (*holdouts_func_t
)(struct list_head
*hop
, bool ndrpt
, bool *frptp
);
20 typedef void (*postgp_func_t
)(struct rcu_tasks
*rtp
);
23 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
24 * @cbs_head: Head of callback list.
25 * @cbs_tail: Tail pointer for callback list.
26 * @cbs_wq: Wait queue allowing new callback to get kthread's attention.
27 * @cbs_lock: Lock protecting callback list.
28 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
29 * @gp_func: This flavor's grace-period-wait function.
30 * @gp_state: Grace period's most recent state transition (debugging).
31 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
32 * @init_fract: Initial backoff sleep interval.
33 * @gp_jiffies: Time of last @gp_state transition.
34 * @gp_start: Most recent grace-period start in jiffies.
35 * @n_gps: Number of grace periods completed since boot.
36 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
37 * @n_ipis_fails: Number of IPI-send failures.
38 * @pregp_func: This flavor's pre-grace-period function (optional).
39 * @pertask_func: This flavor's per-task scan function (optional).
40 * @postscan_func: This flavor's post-task scan function (optional).
41 * @holdouts_func: This flavor's holdout-list scan function (optional).
42 * @postgp_func: This flavor's post-grace-period function (optional).
43 * @call_func: This flavor's call_rcu()-equivalent function.
44 * @name: This flavor's textual name.
45 * @kname: This flavor's kthread name.
48 struct rcu_head
*cbs_head
;
49 struct rcu_head
**cbs_tail
;
50 struct wait_queue_head cbs_wq
;
51 raw_spinlock_t cbs_lock
;
55 unsigned long gp_jiffies
;
56 unsigned long gp_start
;
59 unsigned long n_ipis_fails
;
60 struct task_struct
*kthread_ptr
;
61 rcu_tasks_gp_func_t gp_func
;
62 pregp_func_t pregp_func
;
63 pertask_func_t pertask_func
;
64 postscan_func_t postscan_func
;
65 holdouts_func_t holdouts_func
;
66 postgp_func_t postgp_func
;
67 call_rcu_func_t call_func
;
72 #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
73 static struct rcu_tasks rt_name = \
75 .cbs_tail = &rt_name.cbs_head, \
76 .cbs_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rt_name.cbs_wq), \
77 .cbs_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_lock), \
84 /* Track exiting tasks in order to allow them to be waited for. */
85 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu
);
87 /* Avoid IPIing CPUs early in the grace period. */
88 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
89 static int rcu_task_ipi_delay __read_mostly
= RCU_TASK_IPI_DELAY
;
90 module_param(rcu_task_ipi_delay
, int, 0644);
92 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
93 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
94 static int rcu_task_stall_timeout __read_mostly
= RCU_TASK_STALL_TIMEOUT
;
95 module_param(rcu_task_stall_timeout
, int, 0644);
97 /* RCU tasks grace-period state for debugging. */
99 #define RTGS_WAIT_WAIT_CBS 1
100 #define RTGS_WAIT_GP 2
101 #define RTGS_PRE_WAIT_GP 3
102 #define RTGS_SCAN_TASKLIST 4
103 #define RTGS_POST_SCAN_TASKLIST 5
104 #define RTGS_WAIT_SCAN_HOLDOUTS 6
105 #define RTGS_SCAN_HOLDOUTS 7
106 #define RTGS_POST_GP 8
107 #define RTGS_WAIT_READERS 9
108 #define RTGS_INVOKE_CBS 10
109 #define RTGS_WAIT_CBS 11
110 #ifndef CONFIG_TINY_RCU
111 static const char * const rcu_tasks_gp_state_names
[] = {
113 "RTGS_WAIT_WAIT_CBS",
116 "RTGS_SCAN_TASKLIST",
117 "RTGS_POST_SCAN_TASKLIST",
118 "RTGS_WAIT_SCAN_HOLDOUTS",
119 "RTGS_SCAN_HOLDOUTS",
125 #endif /* #ifndef CONFIG_TINY_RCU */
127 ////////////////////////////////////////////////////////////////////////
131 /* Record grace-period phase and time. */
132 static void set_tasks_gp_state(struct rcu_tasks
*rtp
, int newstate
)
134 rtp
->gp_state
= newstate
;
135 rtp
->gp_jiffies
= jiffies
;
138 #ifndef CONFIG_TINY_RCU
139 /* Return state name. */
140 static const char *tasks_gp_state_getname(struct rcu_tasks
*rtp
)
142 int i
= data_race(rtp
->gp_state
); // Let KCSAN detect update races
143 int j
= READ_ONCE(i
); // Prevent the compiler from reading twice
145 if (j
>= ARRAY_SIZE(rcu_tasks_gp_state_names
))
147 return rcu_tasks_gp_state_names
[j
];
149 #endif /* #ifndef CONFIG_TINY_RCU */
151 // Enqueue a callback for the specified flavor of Tasks RCU.
152 static void call_rcu_tasks_generic(struct rcu_head
*rhp
, rcu_callback_t func
,
153 struct rcu_tasks
*rtp
)
160 raw_spin_lock_irqsave(&rtp
->cbs_lock
, flags
);
161 needwake
= !rtp
->cbs_head
;
162 WRITE_ONCE(*rtp
->cbs_tail
, rhp
);
163 rtp
->cbs_tail
= &rhp
->next
;
164 raw_spin_unlock_irqrestore(&rtp
->cbs_lock
, flags
);
165 /* We can't create the thread unless interrupts are enabled. */
166 if (needwake
&& READ_ONCE(rtp
->kthread_ptr
))
167 wake_up(&rtp
->cbs_wq
);
170 // Wait for a grace period for the specified flavor of Tasks RCU.
171 static void synchronize_rcu_tasks_generic(struct rcu_tasks
*rtp
)
173 /* Complain if the scheduler has not started. */
174 RCU_LOCKDEP_WARN(rcu_scheduler_active
== RCU_SCHEDULER_INACTIVE
,
175 "synchronize_rcu_tasks called too soon");
177 /* Wait for the grace period. */
178 wait_rcu_gp(rtp
->call_func
);
181 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
182 static int __noreturn
rcu_tasks_kthread(void *arg
)
185 struct rcu_head
*list
;
186 struct rcu_head
*next
;
187 struct rcu_tasks
*rtp
= arg
;
189 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
190 housekeeping_affine(current
, HK_FLAG_RCU
);
191 WRITE_ONCE(rtp
->kthread_ptr
, current
); // Let GPs start!
194 * Each pass through the following loop makes one check for
195 * newly arrived callbacks, and, if there are some, waits for
196 * one RCU-tasks grace period and then invokes the callbacks.
197 * This loop is terminated by the system going down. ;-)
200 set_tasks_gp_state(rtp
, RTGS_WAIT_CBS
);
202 /* Pick up any new callbacks. */
203 raw_spin_lock_irqsave(&rtp
->cbs_lock
, flags
);
204 smp_mb__after_spinlock(); // Order updates vs. GP.
205 list
= rtp
->cbs_head
;
206 rtp
->cbs_head
= NULL
;
207 rtp
->cbs_tail
= &rtp
->cbs_head
;
208 raw_spin_unlock_irqrestore(&rtp
->cbs_lock
, flags
);
210 /* If there were none, wait a bit and start over. */
212 wait_event_interruptible(rtp
->cbs_wq
,
213 READ_ONCE(rtp
->cbs_head
));
214 if (!rtp
->cbs_head
) {
215 WARN_ON(signal_pending(current
));
216 set_tasks_gp_state(rtp
, RTGS_WAIT_WAIT_CBS
);
217 schedule_timeout_idle(HZ
/10);
222 // Wait for one grace period.
223 set_tasks_gp_state(rtp
, RTGS_WAIT_GP
);
224 rtp
->gp_start
= jiffies
;
228 /* Invoke the callbacks. */
229 set_tasks_gp_state(rtp
, RTGS_INVOKE_CBS
);
238 /* Paranoid sleep to keep this from entering a tight loop */
239 schedule_timeout_idle(rtp
->gp_sleep
);
243 /* Spawn RCU-tasks grace-period kthread. */
244 static void __init
rcu_spawn_tasks_kthread_generic(struct rcu_tasks
*rtp
)
246 struct task_struct
*t
;
248 t
= kthread_run(rcu_tasks_kthread
, rtp
, "%s_kthread", rtp
->kname
);
249 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__
, rtp
->name
))
251 smp_mb(); /* Ensure others see full kthread. */
254 #ifndef CONFIG_TINY_RCU
257 * Print any non-default Tasks RCU settings.
259 static void __init
rcu_tasks_bootup_oddness(void)
261 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
262 if (rcu_task_stall_timeout
!= RCU_TASK_STALL_TIMEOUT
)
263 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout
);
264 #endif /* #ifdef CONFIG_TASKS_RCU */
265 #ifdef CONFIG_TASKS_RCU
266 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
267 #endif /* #ifdef CONFIG_TASKS_RCU */
268 #ifdef CONFIG_TASKS_RUDE_RCU
269 pr_info("\tRude variant of Tasks RCU enabled.\n");
270 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
271 #ifdef CONFIG_TASKS_TRACE_RCU
272 pr_info("\tTracing variant of Tasks RCU enabled.\n");
273 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
276 #endif /* #ifndef CONFIG_TINY_RCU */
278 #ifndef CONFIG_TINY_RCU
279 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
280 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks
*rtp
, char *s
)
282 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
284 tasks_gp_state_getname(rtp
), data_race(rtp
->gp_state
),
285 jiffies
- data_race(rtp
->gp_jiffies
),
286 data_race(rtp
->n_gps
),
287 data_race(rtp
->n_ipis_fails
), data_race(rtp
->n_ipis
),
288 ".k"[!!data_race(rtp
->kthread_ptr
)],
289 ".C"[!!data_race(rtp
->cbs_head
)],
292 #endif // #ifndef CONFIG_TINY_RCU
294 static void exit_tasks_rcu_finish_trace(struct task_struct
*t
);
296 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
298 ////////////////////////////////////////////////////////////////////////
300 // Shared code between task-list-scanning variants of Tasks RCU.
302 /* Wait for one RCU-tasks grace period. */
303 static void rcu_tasks_wait_gp(struct rcu_tasks
*rtp
)
305 struct task_struct
*g
, *t
;
306 unsigned long lastreport
;
310 set_tasks_gp_state(rtp
, RTGS_PRE_WAIT_GP
);
314 * There were callbacks, so we need to wait for an RCU-tasks
315 * grace period. Start off by scanning the task list for tasks
316 * that are not already voluntarily blocked. Mark these tasks
317 * and make a list of them in holdouts.
319 set_tasks_gp_state(rtp
, RTGS_SCAN_TASKLIST
);
321 for_each_process_thread(g
, t
)
322 rtp
->pertask_func(t
, &holdouts
);
325 set_tasks_gp_state(rtp
, RTGS_POST_SCAN_TASKLIST
);
326 rtp
->postscan_func(&holdouts
);
329 * Each pass through the following loop scans the list of holdout
330 * tasks, removing any that are no longer holdouts. When the list
331 * is empty, we are done.
333 lastreport
= jiffies
;
335 // Start off with initial wait and slowly back off to 1 HZ wait.
336 fract
= rtp
->init_fract
;
338 while (!list_empty(&holdouts
)) {
343 /* Slowly back off waiting for holdouts */
344 set_tasks_gp_state(rtp
, RTGS_WAIT_SCAN_HOLDOUTS
);
345 schedule_timeout_idle(fract
);
350 rtst
= READ_ONCE(rcu_task_stall_timeout
);
351 needreport
= rtst
> 0 && time_after(jiffies
, lastreport
+ rtst
);
353 lastreport
= jiffies
;
355 WARN_ON(signal_pending(current
));
356 set_tasks_gp_state(rtp
, RTGS_SCAN_HOLDOUTS
);
357 rtp
->holdouts_func(&holdouts
, needreport
, &firstreport
);
360 set_tasks_gp_state(rtp
, RTGS_POST_GP
);
361 rtp
->postgp_func(rtp
);
364 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
366 #ifdef CONFIG_TASKS_RCU
368 ////////////////////////////////////////////////////////////////////////
370 // Simple variant of RCU whose quiescent states are voluntary context
371 // switch, cond_resched_rcu_qs(), user-space execution, and idle.
372 // As such, grace periods can take one good long time. There are no
373 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
374 // because this implementation is intended to get the system into a safe
375 // state for some of the manipulations involved in tracing and the like.
376 // Finally, this implementation does not support high call_rcu_tasks()
377 // rates from multiple CPUs. If this is required, per-CPU callback lists
380 // The implementation uses rcu_tasks_wait_gp(), which relies on function
381 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
382 // function sets these function pointers up so that rcu_tasks_wait_gp()
383 // invokes these functions in this order:
385 // rcu_tasks_pregp_step():
386 // Invokes synchronize_rcu() in order to wait for all in-flight
387 // t->on_rq and t->nvcsw transitions to complete. This works because
388 // all such transitions are carried out with interrupts disabled.
389 // rcu_tasks_pertask(), invoked on every non-idle task:
390 // For every runnable non-idle task other than the current one, use
391 // get_task_struct() to pin down that task, snapshot that task's
392 // number of voluntary context switches, and add that task to the
394 // rcu_tasks_postscan():
395 // Invoke synchronize_srcu() to ensure that all tasks that were
396 // in the process of exiting (and which thus might not know to
397 // synchronize with this RCU Tasks grace period) have completed
399 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
400 // Scans the holdout list, attempting to identify a quiescent state
401 // for each task on the list. If there is a quiescent state, the
402 // corresponding task is removed from the holdout list.
403 // rcu_tasks_postgp():
404 // Invokes synchronize_rcu() in order to ensure that all prior
405 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
406 // to have happened before the end of this RCU Tasks grace period.
407 // Again, this works because all such transitions are carried out
408 // with interrupts disabled.
410 // For each exiting task, the exit_tasks_rcu_start() and
411 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
412 // read-side critical sections waited for by rcu_tasks_postscan().
414 // Pre-grace-period update-side code is ordered before the grace via the
415 // ->cbs_lock and the smp_mb__after_spinlock(). Pre-grace-period read-side
416 // code is ordered before the grace period via synchronize_rcu() call
417 // in rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
420 /* Pre-grace-period preparation. */
421 static void rcu_tasks_pregp_step(void)
424 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
425 * to complete. Invoking synchronize_rcu() suffices because all
426 * these transitions occur with interrupts disabled. Without this
427 * synchronize_rcu(), a read-side critical section that started
428 * before the grace period might be incorrectly seen as having
429 * started after the grace period.
431 * This synchronize_rcu() also dispenses with the need for a
432 * memory barrier on the first store to t->rcu_tasks_holdout,
433 * as it forces the store to happen after the beginning of the
439 /* Per-task initial processing. */
440 static void rcu_tasks_pertask(struct task_struct
*t
, struct list_head
*hop
)
442 if (t
!= current
&& READ_ONCE(t
->on_rq
) && !is_idle_task(t
)) {
444 t
->rcu_tasks_nvcsw
= READ_ONCE(t
->nvcsw
);
445 WRITE_ONCE(t
->rcu_tasks_holdout
, true);
446 list_add(&t
->rcu_tasks_holdout_list
, hop
);
450 /* Processing between scanning taskslist and draining the holdout list. */
451 static void rcu_tasks_postscan(struct list_head
*hop
)
454 * Wait for tasks that are in the process of exiting. This
455 * does only part of the job, ensuring that all tasks that were
456 * previously exiting reach the point where they have disabled
457 * preemption, allowing the later synchronize_rcu() to finish
460 synchronize_srcu(&tasks_rcu_exit_srcu
);
463 /* See if tasks are still holding out, complain if so. */
464 static void check_holdout_task(struct task_struct
*t
,
465 bool needreport
, bool *firstreport
)
469 if (!READ_ONCE(t
->rcu_tasks_holdout
) ||
470 t
->rcu_tasks_nvcsw
!= READ_ONCE(t
->nvcsw
) ||
471 !READ_ONCE(t
->on_rq
) ||
472 (IS_ENABLED(CONFIG_NO_HZ_FULL
) &&
473 !is_idle_task(t
) && t
->rcu_tasks_idle_cpu
>= 0)) {
474 WRITE_ONCE(t
->rcu_tasks_holdout
, false);
475 list_del_init(&t
->rcu_tasks_holdout_list
);
479 rcu_request_urgent_qs_task(t
);
483 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
484 *firstreport
= false;
487 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
488 t
, ".I"[is_idle_task(t
)],
489 "N."[cpu
< 0 || !tick_nohz_full_cpu(cpu
)],
490 t
->rcu_tasks_nvcsw
, t
->nvcsw
, t
->rcu_tasks_holdout
,
491 t
->rcu_tasks_idle_cpu
, cpu
);
495 /* Scan the holdout lists for tasks no longer holding out. */
496 static void check_all_holdout_tasks(struct list_head
*hop
,
497 bool needreport
, bool *firstreport
)
499 struct task_struct
*t
, *t1
;
501 list_for_each_entry_safe(t
, t1
, hop
, rcu_tasks_holdout_list
) {
502 check_holdout_task(t
, needreport
, firstreport
);
507 /* Finish off the Tasks-RCU grace period. */
508 static void rcu_tasks_postgp(struct rcu_tasks
*rtp
)
511 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
512 * memory barriers prior to them in the schedule() path, memory
513 * reordering on other CPUs could cause their RCU-tasks read-side
514 * critical sections to extend past the end of the grace period.
515 * However, because these ->nvcsw updates are carried out with
516 * interrupts disabled, we can use synchronize_rcu() to force the
517 * needed ordering on all such CPUs.
519 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
520 * accesses to be within the grace period, avoiding the need for
521 * memory barriers for ->rcu_tasks_holdout accesses.
523 * In addition, this synchronize_rcu() waits for exiting tasks
524 * to complete their final preempt_disable() region of execution,
525 * cleaning up after the synchronize_srcu() above.
530 void call_rcu_tasks(struct rcu_head
*rhp
, rcu_callback_t func
);
531 DEFINE_RCU_TASKS(rcu_tasks
, rcu_tasks_wait_gp
, call_rcu_tasks
, "RCU Tasks");
534 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
535 * @rhp: structure to be used for queueing the RCU updates.
536 * @func: actual callback function to be invoked after the grace period
538 * The callback function will be invoked some time after a full grace
539 * period elapses, in other words after all currently executing RCU
540 * read-side critical sections have completed. call_rcu_tasks() assumes
541 * that the read-side critical sections end at a voluntary context
542 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
543 * or transition to usermode execution. As such, there are no read-side
544 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
545 * this primitive is intended to determine that all tasks have passed
546 * through a safe state, not so much for data-structure synchronization.
548 * See the description of call_rcu() for more detailed information on
549 * memory ordering guarantees.
551 void call_rcu_tasks(struct rcu_head
*rhp
, rcu_callback_t func
)
553 call_rcu_tasks_generic(rhp
, func
, &rcu_tasks
);
555 EXPORT_SYMBOL_GPL(call_rcu_tasks
);
558 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
560 * Control will return to the caller some time after a full rcu-tasks
561 * grace period has elapsed, in other words after all currently
562 * executing rcu-tasks read-side critical sections have elapsed. These
563 * read-side critical sections are delimited by calls to schedule(),
564 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
565 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
567 * This is a very specialized primitive, intended only for a few uses in
568 * tracing and other situations requiring manipulation of function
569 * preambles and profiling hooks. The synchronize_rcu_tasks() function
570 * is not (yet) intended for heavy use from multiple CPUs.
572 * See the description of synchronize_rcu() for more detailed information
573 * on memory ordering guarantees.
575 void synchronize_rcu_tasks(void)
577 synchronize_rcu_tasks_generic(&rcu_tasks
);
579 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks
);
582 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
584 * Although the current implementation is guaranteed to wait, it is not
585 * obligated to, for example, if there are no pending callbacks.
587 void rcu_barrier_tasks(void)
589 /* There is only one callback queue, so this is easy. ;-) */
590 synchronize_rcu_tasks();
592 EXPORT_SYMBOL_GPL(rcu_barrier_tasks
);
594 static int __init
rcu_spawn_tasks_kthread(void)
596 rcu_tasks
.gp_sleep
= HZ
/ 10;
597 rcu_tasks
.init_fract
= HZ
/ 10;
598 rcu_tasks
.pregp_func
= rcu_tasks_pregp_step
;
599 rcu_tasks
.pertask_func
= rcu_tasks_pertask
;
600 rcu_tasks
.postscan_func
= rcu_tasks_postscan
;
601 rcu_tasks
.holdouts_func
= check_all_holdout_tasks
;
602 rcu_tasks
.postgp_func
= rcu_tasks_postgp
;
603 rcu_spawn_tasks_kthread_generic(&rcu_tasks
);
607 #if !defined(CONFIG_TINY_RCU)
608 void show_rcu_tasks_classic_gp_kthread(void)
610 show_rcu_tasks_generic_gp_kthread(&rcu_tasks
, "");
612 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread
);
613 #endif // !defined(CONFIG_TINY_RCU)
615 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
616 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu
)
619 current
->rcu_tasks_idx
= __srcu_read_lock(&tasks_rcu_exit_srcu
);
623 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
624 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu
)
626 struct task_struct
*t
= current
;
629 __srcu_read_unlock(&tasks_rcu_exit_srcu
, t
->rcu_tasks_idx
);
631 exit_tasks_rcu_finish_trace(t
);
634 #else /* #ifdef CONFIG_TASKS_RCU */
635 void exit_tasks_rcu_start(void) { }
636 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current
); }
637 #endif /* #else #ifdef CONFIG_TASKS_RCU */
639 #ifdef CONFIG_TASKS_RUDE_RCU
641 ////////////////////////////////////////////////////////////////////////
643 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
644 // passing an empty function to schedule_on_each_cpu(). This approach
645 // provides an asynchronous call_rcu_tasks_rude() API and batching of
646 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
647 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
648 // and induces otherwise unnecessary context switches on all online CPUs,
649 // whether idle or not.
651 // Callback handling is provided by the rcu_tasks_kthread() function.
653 // Ordering is provided by the scheduler's context-switch code.
655 // Empty function to allow workqueues to force a context switch.
656 static void rcu_tasks_be_rude(struct work_struct
*work
)
660 // Wait for one rude RCU-tasks grace period.
661 static void rcu_tasks_rude_wait_gp(struct rcu_tasks
*rtp
)
663 rtp
->n_ipis
+= cpumask_weight(cpu_online_mask
);
664 schedule_on_each_cpu(rcu_tasks_be_rude
);
667 void call_rcu_tasks_rude(struct rcu_head
*rhp
, rcu_callback_t func
);
668 DEFINE_RCU_TASKS(rcu_tasks_rude
, rcu_tasks_rude_wait_gp
, call_rcu_tasks_rude
,
672 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
673 * @rhp: structure to be used for queueing the RCU updates.
674 * @func: actual callback function to be invoked after the grace period
676 * The callback function will be invoked some time after a full grace
677 * period elapses, in other words after all currently executing RCU
678 * read-side critical sections have completed. call_rcu_tasks_rude()
679 * assumes that the read-side critical sections end at context switch,
680 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
681 * there are no read-side primitives analogous to rcu_read_lock() and
682 * rcu_read_unlock() because this primitive is intended to determine
683 * that all tasks have passed through a safe state, not so much for
684 * data-structure synchronization.
686 * See the description of call_rcu() for more detailed information on
687 * memory ordering guarantees.
689 void call_rcu_tasks_rude(struct rcu_head
*rhp
, rcu_callback_t func
)
691 call_rcu_tasks_generic(rhp
, func
, &rcu_tasks_rude
);
693 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude
);
696 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
698 * Control will return to the caller some time after a rude rcu-tasks
699 * grace period has elapsed, in other words after all currently
700 * executing rcu-tasks read-side critical sections have elapsed. These
701 * read-side critical sections are delimited by calls to schedule(),
702 * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
703 * anyway) cond_resched().
705 * This is a very specialized primitive, intended only for a few uses in
706 * tracing and other situations requiring manipulation of function preambles
707 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
708 * (yet) intended for heavy use from multiple CPUs.
710 * See the description of synchronize_rcu() for more detailed information
711 * on memory ordering guarantees.
713 void synchronize_rcu_tasks_rude(void)
715 synchronize_rcu_tasks_generic(&rcu_tasks_rude
);
717 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude
);
720 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
722 * Although the current implementation is guaranteed to wait, it is not
723 * obligated to, for example, if there are no pending callbacks.
725 void rcu_barrier_tasks_rude(void)
727 /* There is only one callback queue, so this is easy. ;-) */
728 synchronize_rcu_tasks_rude();
730 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude
);
732 static int __init
rcu_spawn_tasks_rude_kthread(void)
734 rcu_tasks_rude
.gp_sleep
= HZ
/ 10;
735 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude
);
739 #if !defined(CONFIG_TINY_RCU)
740 void show_rcu_tasks_rude_gp_kthread(void)
742 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude
, "");
744 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread
);
745 #endif // !defined(CONFIG_TINY_RCU)
746 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
748 ////////////////////////////////////////////////////////////////////////
750 // Tracing variant of Tasks RCU. This variant is designed to be used
751 // to protect tracing hooks, including those of BPF. This variant
754 // 1. Has explicit read-side markers to allow finite grace periods
755 // in the face of in-kernel loops for PREEMPT=n builds.
757 // 2. Protects code in the idle loop, exception entry/exit, and
758 // CPU-hotplug code paths, similar to the capabilities of SRCU.
760 // 3. Avoids expensive read-side instruction, having overhead similar
761 // to that of Preemptible RCU.
763 // There are of course downsides. The grace-period code can send IPIs to
764 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
765 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
766 // is a single callback queue guarded by a single lock, again, much as for
767 // Tasks RCU. If needed, these downsides can be at least partially remedied.
769 // Perhaps most important, this variant of RCU does not affect the vanilla
770 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
771 // readers can operate from idle, offline, and exception entry/exit in no
772 // way allows rcu_preempt and rcu_sched readers to also do so.
774 // The implementation uses rcu_tasks_wait_gp(), which relies on function
775 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
776 // function sets these function pointers up so that rcu_tasks_wait_gp()
777 // invokes these functions in this order:
779 // rcu_tasks_trace_pregp_step():
780 // Initialize the count of readers and block CPU-hotplug operations.
781 // rcu_tasks_trace_pertask(), invoked on every non-idle task:
782 // Initialize per-task state and attempt to identify an immediate
783 // quiescent state for that task, or, failing that, attempt to
784 // set that task's .need_qs flag so that task's next outermost
785 // rcu_read_unlock_trace() will report the quiescent state (in which
786 // case the count of readers is incremented). If both attempts fail,
787 // the task is added to a "holdout" list. Note that IPIs are used
788 // to invoke trc_read_check_handler() in the context of running tasks
789 // in order to avoid ordering overhead on common-case shared-variable
791 // rcu_tasks_trace_postscan():
792 // Initialize state and attempt to identify an immediate quiescent
793 // state as above (but only for idle tasks), unblock CPU-hotplug
794 // operations, and wait for an RCU grace period to avoid races with
795 // tasks that are in the process of exiting.
796 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
797 // Scans the holdout list, attempting to identify a quiescent state
798 // for each task on the list. If there is a quiescent state, the
799 // corresponding task is removed from the holdout list.
800 // rcu_tasks_trace_postgp():
801 // Wait for the count of readers do drop to zero, reporting any stalls.
802 // Also execute full memory barriers to maintain ordering with code
803 // executing after the grace period.
805 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
807 // Pre-grace-period update-side code is ordered before the grace
808 // period via the ->cbs_lock and barriers in rcu_tasks_kthread().
809 // Pre-grace-period read-side code is ordered before the grace period by
810 // atomic_dec_and_test() of the count of readers (for IPIed readers) and by
811 // scheduler context-switch ordering (for locked-down non-running readers).
813 // The lockdep state must be outside of #ifdef to be useful.
814 #ifdef CONFIG_DEBUG_LOCK_ALLOC
815 static struct lock_class_key rcu_lock_trace_key
;
816 struct lockdep_map rcu_trace_lock_map
=
817 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key
);
818 EXPORT_SYMBOL_GPL(rcu_trace_lock_map
);
819 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
821 #ifdef CONFIG_TASKS_TRACE_RCU
823 static atomic_t trc_n_readers_need_end
; // Number of waited-for readers.
824 static DECLARE_WAIT_QUEUE_HEAD(trc_wait
); // List of holdout tasks.
826 // Record outstanding IPIs to each CPU. No point in sending two...
827 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu
);
829 // The number of detections of task quiescent state relying on
830 // heavyweight readers executing explicit memory barriers.
831 static unsigned long n_heavy_reader_attempts
;
832 static unsigned long n_heavy_reader_updates
;
833 static unsigned long n_heavy_reader_ofl_updates
;
835 void call_rcu_tasks_trace(struct rcu_head
*rhp
, rcu_callback_t func
);
836 DEFINE_RCU_TASKS(rcu_tasks_trace
, rcu_tasks_wait_gp
, call_rcu_tasks_trace
,
840 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
841 * while the scheduler locks are held.
843 static void rcu_read_unlock_iw(struct irq_work
*iwp
)
847 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw
, rcu_read_unlock_iw
);
849 /* If we are the last reader, wake up the grace-period kthread. */
850 void rcu_read_unlock_trace_special(struct task_struct
*t
, int nesting
)
852 int nq
= READ_ONCE(t
->trc_reader_special
.b
.need_qs
);
854 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB
) &&
855 t
->trc_reader_special
.b
.need_mb
)
856 smp_mb(); // Pairs with update-side barriers.
857 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
859 WRITE_ONCE(t
->trc_reader_special
.b
.need_qs
, false);
860 WRITE_ONCE(t
->trc_reader_nesting
, nesting
);
861 if (nq
&& atomic_dec_and_test(&trc_n_readers_need_end
))
862 irq_work_queue(&rcu_tasks_trace_iw
);
864 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special
);
866 /* Add a task to the holdout list, if it is not already on the list. */
867 static void trc_add_holdout(struct task_struct
*t
, struct list_head
*bhp
)
869 if (list_empty(&t
->trc_holdout_list
)) {
871 list_add(&t
->trc_holdout_list
, bhp
);
875 /* Remove a task from the holdout list, if it is in fact present. */
876 static void trc_del_holdout(struct task_struct
*t
)
878 if (!list_empty(&t
->trc_holdout_list
)) {
879 list_del_init(&t
->trc_holdout_list
);
884 /* IPI handler to check task state. */
885 static void trc_read_check_handler(void *t_in
)
887 struct task_struct
*t
= current
;
888 struct task_struct
*texp
= t_in
;
890 // If the task is no longer running on this CPU, leave.
891 if (unlikely(texp
!= t
)) {
892 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end
)))
894 goto reset_ipi
; // Already on holdout list, so will check later.
897 // If the task is not in a read-side critical section, and
898 // if this is the last reader, awaken the grace-period kthread.
899 if (likely(!READ_ONCE(t
->trc_reader_nesting
))) {
900 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end
)))
902 // Mark as checked after decrement to avoid false
903 // positives on the above WARN_ON_ONCE().
904 WRITE_ONCE(t
->trc_reader_checked
, true);
907 // If we are racing with an rcu_read_unlock_trace(), try again later.
908 if (unlikely(READ_ONCE(t
->trc_reader_nesting
) < 0)) {
909 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end
)))
913 WRITE_ONCE(t
->trc_reader_checked
, true);
915 // Get here if the task is in a read-side critical section. Set
916 // its state so that it will awaken the grace-period kthread upon
917 // exit from that critical section.
918 WARN_ON_ONCE(READ_ONCE(t
->trc_reader_special
.b
.need_qs
));
919 WRITE_ONCE(t
->trc_reader_special
.b
.need_qs
, true);
922 // Allow future IPIs to be sent on CPU and for task.
923 // Also order this IPI handler against any later manipulations of
924 // the intended task.
925 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu
, smp_processor_id()), false); // ^^^
926 smp_store_release(&texp
->trc_ipi_to_cpu
, -1); // ^^^
929 /* Callback function for scheduler to check locked-down task. */
930 static bool trc_inspect_reader(struct task_struct
*t
, void *arg
)
932 int cpu
= task_cpu(t
);
934 bool ofl
= cpu_is_offline(cpu
);
937 WARN_ON_ONCE(ofl
&& !is_idle_task(t
));
939 // If no chance of heavyweight readers, do it the hard way.
940 if (!ofl
&& !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB
))
943 // If heavyweight readers are enabled on the remote task,
944 // we can inspect its state despite its currently running.
945 // However, we cannot safely change its state.
946 n_heavy_reader_attempts
++;
947 if (!ofl
&& // Check for "running" idle tasks on offline CPUs.
948 !rcu_dynticks_zero_in_eqs(cpu
, &t
->trc_reader_nesting
))
949 return false; // No quiescent state, do it the hard way.
950 n_heavy_reader_updates
++;
952 n_heavy_reader_ofl_updates
++;
955 // The task is not running, so C-language access is safe.
956 in_qs
= likely(!t
->trc_reader_nesting
);
959 // Mark as checked so that the grace-period kthread will
960 // remove it from the holdout list.
961 t
->trc_reader_checked
= true;
964 return true; // Already in quiescent state, done!!!
966 // The task is in a read-side critical section, so set up its
967 // state so that it will awaken the grace-period kthread upon exit
968 // from that critical section.
969 atomic_inc(&trc_n_readers_need_end
); // One more to wait on.
970 WARN_ON_ONCE(READ_ONCE(t
->trc_reader_special
.b
.need_qs
));
971 WRITE_ONCE(t
->trc_reader_special
.b
.need_qs
, true);
975 /* Attempt to extract the state for the specified task. */
976 static void trc_wait_for_one_reader(struct task_struct
*t
,
977 struct list_head
*bhp
)
981 // If a previous IPI is still in flight, let it complete.
982 if (smp_load_acquire(&t
->trc_ipi_to_cpu
) != -1) // Order IPI
985 // The current task had better be in a quiescent state.
987 t
->trc_reader_checked
= true;
988 WARN_ON_ONCE(READ_ONCE(t
->trc_reader_nesting
));
992 // Attempt to nail down the task for inspection.
994 if (try_invoke_on_locked_down_task(t
, trc_inspect_reader
, NULL
)) {
1000 // If this task is not yet on the holdout list, then we are in
1001 // an RCU read-side critical section. Otherwise, the invocation of
1002 // rcu_add_holdout() that added it to the list did the necessary
1003 // get_task_struct(). Either way, the task cannot be freed out
1004 // from under this code.
1006 // If currently running, send an IPI, either way, add to list.
1007 trc_add_holdout(t
, bhp
);
1009 time_after(jiffies
+ 1, rcu_tasks_trace
.gp_start
+ rcu_task_ipi_delay
)) {
1010 // The task is currently running, so try IPIing it.
1013 // If there is already an IPI outstanding, let it happen.
1014 if (per_cpu(trc_ipi_to_cpu
, cpu
) || t
->trc_ipi_to_cpu
>= 0)
1017 atomic_inc(&trc_n_readers_need_end
);
1018 per_cpu(trc_ipi_to_cpu
, cpu
) = true;
1019 t
->trc_ipi_to_cpu
= cpu
;
1020 rcu_tasks_trace
.n_ipis
++;
1021 if (smp_call_function_single(cpu
,
1022 trc_read_check_handler
, t
, 0)) {
1023 // Just in case there is some other reason for
1024 // failure than the target CPU being offline.
1025 rcu_tasks_trace
.n_ipis_fails
++;
1026 per_cpu(trc_ipi_to_cpu
, cpu
) = false;
1027 t
->trc_ipi_to_cpu
= cpu
;
1028 if (atomic_dec_and_test(&trc_n_readers_need_end
)) {
1036 /* Initialize for a new RCU-tasks-trace grace period. */
1037 static void rcu_tasks_trace_pregp_step(void)
1041 // Allow for fast-acting IPIs.
1042 atomic_set(&trc_n_readers_need_end
, 1);
1044 // There shouldn't be any old IPIs, but...
1045 for_each_possible_cpu(cpu
)
1046 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu
, cpu
));
1048 // Disable CPU hotplug across the tasklist scan.
1049 // This also waits for all readers in CPU-hotplug code paths.
1053 /* Do first-round processing for the specified task. */
1054 static void rcu_tasks_trace_pertask(struct task_struct
*t
,
1055 struct list_head
*hop
)
1057 // During early boot when there is only the one boot CPU, there
1058 // is no idle task for the other CPUs. Just return.
1059 if (unlikely(t
== NULL
))
1062 WRITE_ONCE(t
->trc_reader_special
.b
.need_qs
, false);
1063 WRITE_ONCE(t
->trc_reader_checked
, false);
1064 t
->trc_ipi_to_cpu
= -1;
1065 trc_wait_for_one_reader(t
, hop
);
1069 * Do intermediate processing between task and holdout scans and
1070 * pick up the idle tasks.
1072 static void rcu_tasks_trace_postscan(struct list_head
*hop
)
1076 for_each_possible_cpu(cpu
)
1077 rcu_tasks_trace_pertask(idle_task(cpu
), hop
);
1079 // Re-enable CPU hotplug now that the tasklist scan has completed.
1082 // Wait for late-stage exiting tasks to finish exiting.
1083 // These might have passed the call to exit_tasks_rcu_finish().
1085 // Any tasks that exit after this point will set ->trc_reader_checked.
1088 /* Show the state of a task stalling the current RCU tasks trace GP. */
1089 static void show_stalled_task_trace(struct task_struct
*t
, bool *firstreport
)
1094 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1095 *firstreport
= false;
1097 // FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1099 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1101 ".I"[READ_ONCE(t
->trc_ipi_to_cpu
) > 0],
1102 ".i"[is_idle_task(t
)],
1103 ".N"[cpu
> 0 && tick_nohz_full_cpu(cpu
)],
1104 READ_ONCE(t
->trc_reader_nesting
),
1105 " N"[!!READ_ONCE(t
->trc_reader_special
.b
.need_qs
)],
1110 /* List stalled IPIs for RCU tasks trace. */
1111 static void show_stalled_ipi_trace(void)
1115 for_each_possible_cpu(cpu
)
1116 if (per_cpu(trc_ipi_to_cpu
, cpu
))
1117 pr_alert("\tIPI outstanding to CPU %d\n", cpu
);
1120 /* Do one scan of the holdout list. */
1121 static void check_all_holdout_tasks_trace(struct list_head
*hop
,
1122 bool needreport
, bool *firstreport
)
1124 struct task_struct
*g
, *t
;
1126 // Disable CPU hotplug across the holdout list scan.
1129 list_for_each_entry_safe(t
, g
, hop
, trc_holdout_list
) {
1130 // If safe and needed, try to check the current task.
1131 if (READ_ONCE(t
->trc_ipi_to_cpu
) == -1 &&
1132 !READ_ONCE(t
->trc_reader_checked
))
1133 trc_wait_for_one_reader(t
, hop
);
1135 // If check succeeded, remove this task from the list.
1136 if (READ_ONCE(t
->trc_reader_checked
))
1138 else if (needreport
)
1139 show_stalled_task_trace(t
, firstreport
);
1142 // Re-enable CPU hotplug now that the holdout list scan has completed.
1147 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1148 show_stalled_ipi_trace();
1152 /* Wait for grace period to complete and provide ordering. */
1153 static void rcu_tasks_trace_postgp(struct rcu_tasks
*rtp
)
1156 struct task_struct
*g
, *t
;
1157 LIST_HEAD(holdouts
);
1160 // Remove the safety count.
1161 smp_mb__before_atomic(); // Order vs. earlier atomics
1162 atomic_dec(&trc_n_readers_need_end
);
1163 smp_mb__after_atomic(); // Order vs. later atomics
1165 // Wait for readers.
1166 set_tasks_gp_state(rtp
, RTGS_WAIT_READERS
);
1168 ret
= wait_event_idle_exclusive_timeout(
1170 atomic_read(&trc_n_readers_need_end
) == 0,
1171 READ_ONCE(rcu_task_stall_timeout
));
1173 break; // Count reached zero.
1174 // Stall warning time, so make a list of the offenders.
1176 for_each_process_thread(g
, t
)
1177 if (READ_ONCE(t
->trc_reader_special
.b
.need_qs
))
1178 trc_add_holdout(t
, &holdouts
);
1181 list_for_each_entry_safe(t
, g
, &holdouts
, trc_holdout_list
) {
1182 if (READ_ONCE(t
->trc_reader_special
.b
.need_qs
))
1183 show_stalled_task_trace(t
, &firstreport
);
1184 trc_del_holdout(t
); // Release task_struct reference.
1187 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1188 show_stalled_ipi_trace();
1189 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end
));
1191 smp_mb(); // Caller's code must be ordered after wakeup.
1192 // Pairs with pretty much every ordering primitive.
1195 /* Report any needed quiescent state for this exiting task. */
1196 static void exit_tasks_rcu_finish_trace(struct task_struct
*t
)
1198 WRITE_ONCE(t
->trc_reader_checked
, true);
1199 WARN_ON_ONCE(READ_ONCE(t
->trc_reader_nesting
));
1200 WRITE_ONCE(t
->trc_reader_nesting
, 0);
1201 if (WARN_ON_ONCE(READ_ONCE(t
->trc_reader_special
.b
.need_qs
)))
1202 rcu_read_unlock_trace_special(t
, 0);
1206 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1207 * @rhp: structure to be used for queueing the RCU updates.
1208 * @func: actual callback function to be invoked after the grace period
1210 * The callback function will be invoked some time after a full grace
1211 * period elapses, in other words after all currently executing RCU
1212 * read-side critical sections have completed. call_rcu_tasks_trace()
1213 * assumes that the read-side critical sections end at context switch,
1214 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
1215 * there are no read-side primitives analogous to rcu_read_lock() and
1216 * rcu_read_unlock() because this primitive is intended to determine
1217 * that all tasks have passed through a safe state, not so much for
1218 * data-structure synchronization.
1220 * See the description of call_rcu() for more detailed information on
1221 * memory ordering guarantees.
1223 void call_rcu_tasks_trace(struct rcu_head
*rhp
, rcu_callback_t func
)
1225 call_rcu_tasks_generic(rhp
, func
, &rcu_tasks_trace
);
1227 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace
);
1230 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1232 * Control will return to the caller some time after a trace rcu-tasks
1233 * grace period has elapsed, in other words after all currently executing
1234 * rcu-tasks read-side critical sections have elapsed. These read-side
1235 * critical sections are delimited by calls to rcu_read_lock_trace()
1236 * and rcu_read_unlock_trace().
1238 * This is a very specialized primitive, intended only for a few uses in
1239 * tracing and other situations requiring manipulation of function preambles
1240 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1241 * (yet) intended for heavy use from multiple CPUs.
1243 * See the description of synchronize_rcu() for more detailed information
1244 * on memory ordering guarantees.
1246 void synchronize_rcu_tasks_trace(void)
1248 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map
), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1249 synchronize_rcu_tasks_generic(&rcu_tasks_trace
);
1251 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace
);
1254 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1256 * Although the current implementation is guaranteed to wait, it is not
1257 * obligated to, for example, if there are no pending callbacks.
1259 void rcu_barrier_tasks_trace(void)
1261 /* There is only one callback queue, so this is easy. ;-) */
1262 synchronize_rcu_tasks_trace();
1264 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace
);
1266 static int __init
rcu_spawn_tasks_trace_kthread(void)
1268 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB
)) {
1269 rcu_tasks_trace
.gp_sleep
= HZ
/ 10;
1270 rcu_tasks_trace
.init_fract
= HZ
/ 10;
1272 rcu_tasks_trace
.gp_sleep
= HZ
/ 200;
1273 if (rcu_tasks_trace
.gp_sleep
<= 0)
1274 rcu_tasks_trace
.gp_sleep
= 1;
1275 rcu_tasks_trace
.init_fract
= HZ
/ 200;
1276 if (rcu_tasks_trace
.init_fract
<= 0)
1277 rcu_tasks_trace
.init_fract
= 1;
1279 rcu_tasks_trace
.pregp_func
= rcu_tasks_trace_pregp_step
;
1280 rcu_tasks_trace
.pertask_func
= rcu_tasks_trace_pertask
;
1281 rcu_tasks_trace
.postscan_func
= rcu_tasks_trace_postscan
;
1282 rcu_tasks_trace
.holdouts_func
= check_all_holdout_tasks_trace
;
1283 rcu_tasks_trace
.postgp_func
= rcu_tasks_trace_postgp
;
1284 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace
);
1288 #if !defined(CONFIG_TINY_RCU)
1289 void show_rcu_tasks_trace_gp_kthread(void)
1293 sprintf(buf
, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end
),
1294 data_race(n_heavy_reader_ofl_updates
),
1295 data_race(n_heavy_reader_updates
),
1296 data_race(n_heavy_reader_attempts
));
1297 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace
, buf
);
1299 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread
);
1300 #endif // !defined(CONFIG_TINY_RCU)
1302 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1303 static void exit_tasks_rcu_finish_trace(struct task_struct
*t
) { }
1304 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1306 #ifndef CONFIG_TINY_RCU
1307 void show_rcu_tasks_gp_kthreads(void)
1309 show_rcu_tasks_classic_gp_kthread();
1310 show_rcu_tasks_rude_gp_kthread();
1311 show_rcu_tasks_trace_gp_kthread();
1313 #endif /* #ifndef CONFIG_TINY_RCU */
1315 #ifdef CONFIG_PROVE_RCU
1316 struct rcu_tasks_test_desc
{
1322 static struct rcu_tasks_test_desc tests
[] = {
1324 .name
= "call_rcu_tasks()",
1325 /* If not defined, the test is skipped. */
1326 .notrun
= !IS_ENABLED(CONFIG_TASKS_RCU
),
1329 .name
= "call_rcu_tasks_rude()",
1330 /* If not defined, the test is skipped. */
1331 .notrun
= !IS_ENABLED(CONFIG_TASKS_RUDE_RCU
),
1334 .name
= "call_rcu_tasks_trace()",
1335 /* If not defined, the test is skipped. */
1336 .notrun
= !IS_ENABLED(CONFIG_TASKS_TRACE_RCU
)
1340 static void test_rcu_tasks_callback(struct rcu_head
*rhp
)
1342 struct rcu_tasks_test_desc
*rttd
=
1343 container_of(rhp
, struct rcu_tasks_test_desc
, rh
);
1345 pr_info("Callback from %s invoked.\n", rttd
->name
);
1347 rttd
->notrun
= true;
1350 static void rcu_tasks_initiate_self_tests(void)
1352 pr_info("Running RCU-tasks wait API self tests\n");
1353 #ifdef CONFIG_TASKS_RCU
1354 synchronize_rcu_tasks();
1355 call_rcu_tasks(&tests
[0].rh
, test_rcu_tasks_callback
);
1358 #ifdef CONFIG_TASKS_RUDE_RCU
1359 synchronize_rcu_tasks_rude();
1360 call_rcu_tasks_rude(&tests
[1].rh
, test_rcu_tasks_callback
);
1363 #ifdef CONFIG_TASKS_TRACE_RCU
1364 synchronize_rcu_tasks_trace();
1365 call_rcu_tasks_trace(&tests
[2].rh
, test_rcu_tasks_callback
);
1369 static int rcu_tasks_verify_self_tests(void)
1374 for (i
= 0; i
< ARRAY_SIZE(tests
); i
++) {
1375 if (!tests
[i
].notrun
) { // still hanging.
1376 pr_err("%s has been failed.\n", tests
[i
].name
);
1386 late_initcall(rcu_tasks_verify_self_tests
);
1387 #else /* #ifdef CONFIG_PROVE_RCU */
1388 static void rcu_tasks_initiate_self_tests(void) { }
1389 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1391 void __init
rcu_init_tasks_generic(void)
1393 #ifdef CONFIG_TASKS_RCU
1394 rcu_spawn_tasks_kthread();
1397 #ifdef CONFIG_TASKS_RUDE_RCU
1398 rcu_spawn_tasks_rude_kthread();
1401 #ifdef CONFIG_TASKS_TRACE_RCU
1402 rcu_spawn_tasks_trace_kthread();
1405 // Run the self-tests.
1406 rcu_tasks_initiate_self_tests();
1409 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1410 static inline void rcu_tasks_bootup_oddness(void) {}
1411 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */