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 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
175 "synchronize_%s() called too soon", rtp->name))
178 /* Wait for the grace period. */
179 wait_rcu_gp(rtp->call_func);
182 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
183 static int __noreturn rcu_tasks_kthread(void *arg)
186 struct rcu_head *list;
187 struct rcu_head *next;
188 struct rcu_tasks *rtp = arg;
190 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
191 housekeeping_affine(current, HK_FLAG_RCU);
192 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
195 * Each pass through the following loop makes one check for
196 * newly arrived callbacks, and, if there are some, waits for
197 * one RCU-tasks grace period and then invokes the callbacks.
198 * This loop is terminated by the system going down. ;-)
201 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
203 /* Pick up any new callbacks. */
204 raw_spin_lock_irqsave(&rtp->cbs_lock, flags);
205 smp_mb__after_spinlock(); // Order updates vs. GP.
206 list = rtp->cbs_head;
207 rtp->cbs_head = NULL;
208 rtp->cbs_tail = &rtp->cbs_head;
209 raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags);
211 /* If there were none, wait a bit and start over. */
213 wait_event_interruptible(rtp->cbs_wq,
214 READ_ONCE(rtp->cbs_head));
215 if (!rtp->cbs_head) {
216 WARN_ON(signal_pending(current));
217 set_tasks_gp_state(rtp, RTGS_WAIT_WAIT_CBS);
218 schedule_timeout_idle(HZ/10);
223 // Wait for one grace period.
224 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
225 rtp->gp_start = jiffies;
229 /* Invoke the callbacks. */
230 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
239 /* Paranoid sleep to keep this from entering a tight loop */
240 schedule_timeout_idle(rtp->gp_sleep);
244 /* Spawn RCU-tasks grace-period kthread. */
245 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
247 struct task_struct *t;
249 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
250 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
252 smp_mb(); /* Ensure others see full kthread. */
255 #ifndef CONFIG_TINY_RCU
258 * Print any non-default Tasks RCU settings.
260 static void __init rcu_tasks_bootup_oddness(void)
262 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
263 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
264 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
265 #endif /* #ifdef CONFIG_TASKS_RCU */
266 #ifdef CONFIG_TASKS_RCU
267 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
268 #endif /* #ifdef CONFIG_TASKS_RCU */
269 #ifdef CONFIG_TASKS_RUDE_RCU
270 pr_info("\tRude variant of Tasks RCU enabled.\n");
271 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
272 #ifdef CONFIG_TASKS_TRACE_RCU
273 pr_info("\tTracing variant of Tasks RCU enabled.\n");
274 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
277 #endif /* #ifndef CONFIG_TINY_RCU */
279 #ifndef CONFIG_TINY_RCU
280 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
281 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
283 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
285 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
286 jiffies - data_race(rtp->gp_jiffies),
287 data_race(rtp->n_gps),
288 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
289 ".k"[!!data_race(rtp->kthread_ptr)],
290 ".C"[!!data_race(rtp->cbs_head)],
293 #endif // #ifndef CONFIG_TINY_RCU
295 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
297 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
299 ////////////////////////////////////////////////////////////////////////
301 // Shared code between task-list-scanning variants of Tasks RCU.
303 /* Wait for one RCU-tasks grace period. */
304 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
306 struct task_struct *g, *t;
307 unsigned long lastreport;
311 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
315 * There were callbacks, so we need to wait for an RCU-tasks
316 * grace period. Start off by scanning the task list for tasks
317 * that are not already voluntarily blocked. Mark these tasks
318 * and make a list of them in holdouts.
320 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
322 for_each_process_thread(g, t)
323 rtp->pertask_func(t, &holdouts);
326 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
327 rtp->postscan_func(&holdouts);
330 * Each pass through the following loop scans the list of holdout
331 * tasks, removing any that are no longer holdouts. When the list
332 * is empty, we are done.
334 lastreport = jiffies;
336 // Start off with initial wait and slowly back off to 1 HZ wait.
337 fract = rtp->init_fract;
339 while (!list_empty(&holdouts)) {
344 /* Slowly back off waiting for holdouts */
345 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
346 schedule_timeout_idle(fract);
351 rtst = READ_ONCE(rcu_task_stall_timeout);
352 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
354 lastreport = jiffies;
356 WARN_ON(signal_pending(current));
357 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
358 rtp->holdouts_func(&holdouts, needreport, &firstreport);
361 set_tasks_gp_state(rtp, RTGS_POST_GP);
362 rtp->postgp_func(rtp);
365 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
367 #ifdef CONFIG_TASKS_RCU
369 ////////////////////////////////////////////////////////////////////////
371 // Simple variant of RCU whose quiescent states are voluntary context
372 // switch, cond_resched_rcu_qs(), user-space execution, and idle.
373 // As such, grace periods can take one good long time. There are no
374 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
375 // because this implementation is intended to get the system into a safe
376 // state for some of the manipulations involved in tracing and the like.
377 // Finally, this implementation does not support high call_rcu_tasks()
378 // rates from multiple CPUs. If this is required, per-CPU callback lists
381 // The implementation uses rcu_tasks_wait_gp(), which relies on function
382 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
383 // function sets these function pointers up so that rcu_tasks_wait_gp()
384 // invokes these functions in this order:
386 // rcu_tasks_pregp_step():
387 // Invokes synchronize_rcu() in order to wait for all in-flight
388 // t->on_rq and t->nvcsw transitions to complete. This works because
389 // all such transitions are carried out with interrupts disabled.
390 // rcu_tasks_pertask(), invoked on every non-idle task:
391 // For every runnable non-idle task other than the current one, use
392 // get_task_struct() to pin down that task, snapshot that task's
393 // number of voluntary context switches, and add that task to the
395 // rcu_tasks_postscan():
396 // Invoke synchronize_srcu() to ensure that all tasks that were
397 // in the process of exiting (and which thus might not know to
398 // synchronize with this RCU Tasks grace period) have completed
400 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
401 // Scans the holdout list, attempting to identify a quiescent state
402 // for each task on the list. If there is a quiescent state, the
403 // corresponding task is removed from the holdout list.
404 // rcu_tasks_postgp():
405 // Invokes synchronize_rcu() in order to ensure that all prior
406 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
407 // to have happened before the end of this RCU Tasks grace period.
408 // Again, this works because all such transitions are carried out
409 // with interrupts disabled.
411 // For each exiting task, the exit_tasks_rcu_start() and
412 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
413 // read-side critical sections waited for by rcu_tasks_postscan().
415 // Pre-grace-period update-side code is ordered before the grace via the
416 // ->cbs_lock and the smp_mb__after_spinlock(). Pre-grace-period read-side
417 // code is ordered before the grace period via synchronize_rcu() call
418 // in rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
421 /* Pre-grace-period preparation. */
422 static void rcu_tasks_pregp_step(void)
425 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
426 * to complete. Invoking synchronize_rcu() suffices because all
427 * these transitions occur with interrupts disabled. Without this
428 * synchronize_rcu(), a read-side critical section that started
429 * before the grace period might be incorrectly seen as having
430 * started after the grace period.
432 * This synchronize_rcu() also dispenses with the need for a
433 * memory barrier on the first store to t->rcu_tasks_holdout,
434 * as it forces the store to happen after the beginning of the
440 /* Per-task initial processing. */
441 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
443 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
445 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
446 WRITE_ONCE(t->rcu_tasks_holdout, true);
447 list_add(&t->rcu_tasks_holdout_list, hop);
451 /* Processing between scanning taskslist and draining the holdout list. */
452 static void rcu_tasks_postscan(struct list_head *hop)
455 * Exiting tasks may escape the tasklist scan. Those are vulnerable
456 * until their final schedule() with TASK_DEAD state. To cope with
457 * this, divide the fragile exit path part in two intersecting
458 * read side critical sections:
460 * 1) An _SRCU_ read side starting before calling exit_notify(),
461 * which may remove the task from the tasklist, and ending after
462 * the final preempt_disable() call in do_exit().
464 * 2) An _RCU_ read side starting with the final preempt_disable()
465 * call in do_exit() and ending with the final call to schedule()
466 * with TASK_DEAD state.
468 * This handles the part 1). And postgp will handle part 2) with a
469 * call to synchronize_rcu().
471 synchronize_srcu(&tasks_rcu_exit_srcu);
474 /* See if tasks are still holding out, complain if so. */
475 static void check_holdout_task(struct task_struct *t,
476 bool needreport, bool *firstreport)
480 if (!READ_ONCE(t->rcu_tasks_holdout) ||
481 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
482 !READ_ONCE(t->on_rq) ||
483 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
484 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
485 WRITE_ONCE(t->rcu_tasks_holdout, false);
486 list_del_init(&t->rcu_tasks_holdout_list);
490 rcu_request_urgent_qs_task(t);
494 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
495 *firstreport = false;
498 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
499 t, ".I"[is_idle_task(t)],
500 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
501 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
502 t->rcu_tasks_idle_cpu, cpu);
506 /* Scan the holdout lists for tasks no longer holding out. */
507 static void check_all_holdout_tasks(struct list_head *hop,
508 bool needreport, bool *firstreport)
510 struct task_struct *t, *t1;
512 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
513 check_holdout_task(t, needreport, firstreport);
518 /* Finish off the Tasks-RCU grace period. */
519 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
522 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
523 * memory barriers prior to them in the schedule() path, memory
524 * reordering on other CPUs could cause their RCU-tasks read-side
525 * critical sections to extend past the end of the grace period.
526 * However, because these ->nvcsw updates are carried out with
527 * interrupts disabled, we can use synchronize_rcu() to force the
528 * needed ordering on all such CPUs.
530 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
531 * accesses to be within the grace period, avoiding the need for
532 * memory barriers for ->rcu_tasks_holdout accesses.
534 * In addition, this synchronize_rcu() waits for exiting tasks
535 * to complete their final preempt_disable() region of execution,
536 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
537 * enforcing the whole region before tasklist removal until
538 * the final schedule() with TASK_DEAD state to be an RCU TASKS
539 * read side critical section.
544 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
545 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
548 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
549 * @rhp: structure to be used for queueing the RCU updates.
550 * @func: actual callback function to be invoked after the grace period
552 * The callback function will be invoked some time after a full grace
553 * period elapses, in other words after all currently executing RCU
554 * read-side critical sections have completed. call_rcu_tasks() assumes
555 * that the read-side critical sections end at a voluntary context
556 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
557 * or transition to usermode execution. As such, there are no read-side
558 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
559 * this primitive is intended to determine that all tasks have passed
560 * through a safe state, not so much for data-structure synchronization.
562 * See the description of call_rcu() for more detailed information on
563 * memory ordering guarantees.
565 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
567 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
569 EXPORT_SYMBOL_GPL(call_rcu_tasks);
572 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
574 * Control will return to the caller some time after a full rcu-tasks
575 * grace period has elapsed, in other words after all currently
576 * executing rcu-tasks read-side critical sections have elapsed. These
577 * read-side critical sections are delimited by calls to schedule(),
578 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
579 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
581 * This is a very specialized primitive, intended only for a few uses in
582 * tracing and other situations requiring manipulation of function
583 * preambles and profiling hooks. The synchronize_rcu_tasks() function
584 * is not (yet) intended for heavy use from multiple CPUs.
586 * See the description of synchronize_rcu() for more detailed information
587 * on memory ordering guarantees.
589 void synchronize_rcu_tasks(void)
591 synchronize_rcu_tasks_generic(&rcu_tasks);
593 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
596 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
598 * Although the current implementation is guaranteed to wait, it is not
599 * obligated to, for example, if there are no pending callbacks.
601 void rcu_barrier_tasks(void)
603 /* There is only one callback queue, so this is easy. ;-) */
604 synchronize_rcu_tasks();
606 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
608 static int __init rcu_spawn_tasks_kthread(void)
610 rcu_tasks.gp_sleep = HZ / 10;
611 rcu_tasks.init_fract = HZ / 10;
612 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
613 rcu_tasks.pertask_func = rcu_tasks_pertask;
614 rcu_tasks.postscan_func = rcu_tasks_postscan;
615 rcu_tasks.holdouts_func = check_all_holdout_tasks;
616 rcu_tasks.postgp_func = rcu_tasks_postgp;
617 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
621 #if !defined(CONFIG_TINY_RCU)
622 void show_rcu_tasks_classic_gp_kthread(void)
624 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
626 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
627 #endif // !defined(CONFIG_TINY_RCU)
630 * Contribute to protect against tasklist scan blind spot while the
631 * task is exiting and may be removed from the tasklist. See
632 * corresponding synchronize_srcu() for further details.
634 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
636 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
640 * Contribute to protect against tasklist scan blind spot while the
641 * task is exiting and may be removed from the tasklist. See
642 * corresponding synchronize_srcu() for further details.
644 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
646 struct task_struct *t = current;
648 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
652 * Contribute to protect against tasklist scan blind spot while the
653 * task is exiting and may be removed from the tasklist. See
654 * corresponding synchronize_srcu() for further details.
656 void exit_tasks_rcu_finish(void)
658 exit_tasks_rcu_stop();
659 exit_tasks_rcu_finish_trace(current);
662 #else /* #ifdef CONFIG_TASKS_RCU */
663 void exit_tasks_rcu_start(void) { }
664 void exit_tasks_rcu_stop(void) { }
665 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
666 #endif /* #else #ifdef CONFIG_TASKS_RCU */
668 #ifdef CONFIG_TASKS_RUDE_RCU
670 ////////////////////////////////////////////////////////////////////////
672 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
673 // passing an empty function to schedule_on_each_cpu(). This approach
674 // provides an asynchronous call_rcu_tasks_rude() API and batching of
675 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
676 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
677 // and induces otherwise unnecessary context switches on all online CPUs,
678 // whether idle or not.
680 // Callback handling is provided by the rcu_tasks_kthread() function.
682 // Ordering is provided by the scheduler's context-switch code.
684 // Empty function to allow workqueues to force a context switch.
685 static void rcu_tasks_be_rude(struct work_struct *work)
689 // Wait for one rude RCU-tasks grace period.
690 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
692 rtp->n_ipis += cpumask_weight(cpu_online_mask);
693 schedule_on_each_cpu(rcu_tasks_be_rude);
696 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
697 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
701 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
702 * @rhp: structure to be used for queueing the RCU updates.
703 * @func: actual callback function to be invoked after the grace period
705 * The callback function will be invoked some time after a full grace
706 * period elapses, in other words after all currently executing RCU
707 * read-side critical sections have completed. call_rcu_tasks_rude()
708 * assumes that the read-side critical sections end at context switch,
709 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
710 * there are no read-side primitives analogous to rcu_read_lock() and
711 * rcu_read_unlock() because this primitive is intended to determine
712 * that all tasks have passed through a safe state, not so much for
713 * data-structure synchronization.
715 * See the description of call_rcu() for more detailed information on
716 * memory ordering guarantees.
718 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
720 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
722 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
725 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
727 * Control will return to the caller some time after a rude rcu-tasks
728 * grace period has elapsed, in other words after all currently
729 * executing rcu-tasks read-side critical sections have elapsed. These
730 * read-side critical sections are delimited by calls to schedule(),
731 * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
732 * anyway) cond_resched().
734 * This is a very specialized primitive, intended only for a few uses in
735 * tracing and other situations requiring manipulation of function preambles
736 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
737 * (yet) intended for heavy use from multiple CPUs.
739 * See the description of synchronize_rcu() for more detailed information
740 * on memory ordering guarantees.
742 void synchronize_rcu_tasks_rude(void)
744 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
746 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
749 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
751 * Although the current implementation is guaranteed to wait, it is not
752 * obligated to, for example, if there are no pending callbacks.
754 void rcu_barrier_tasks_rude(void)
756 /* There is only one callback queue, so this is easy. ;-) */
757 synchronize_rcu_tasks_rude();
759 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
761 static int __init rcu_spawn_tasks_rude_kthread(void)
763 rcu_tasks_rude.gp_sleep = HZ / 10;
764 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
768 #if !defined(CONFIG_TINY_RCU)
769 void show_rcu_tasks_rude_gp_kthread(void)
771 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
773 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
774 #endif // !defined(CONFIG_TINY_RCU)
775 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
777 ////////////////////////////////////////////////////////////////////////
779 // Tracing variant of Tasks RCU. This variant is designed to be used
780 // to protect tracing hooks, including those of BPF. This variant
783 // 1. Has explicit read-side markers to allow finite grace periods
784 // in the face of in-kernel loops for PREEMPT=n builds.
786 // 2. Protects code in the idle loop, exception entry/exit, and
787 // CPU-hotplug code paths, similar to the capabilities of SRCU.
789 // 3. Avoids expensive read-side instruction, having overhead similar
790 // to that of Preemptible RCU.
792 // There are of course downsides. The grace-period code can send IPIs to
793 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
794 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
795 // is a single callback queue guarded by a single lock, again, much as for
796 // Tasks RCU. If needed, these downsides can be at least partially remedied.
798 // Perhaps most important, this variant of RCU does not affect the vanilla
799 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
800 // readers can operate from idle, offline, and exception entry/exit in no
801 // way allows rcu_preempt and rcu_sched readers to also do so.
803 // The implementation uses rcu_tasks_wait_gp(), which relies on function
804 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
805 // function sets these function pointers up so that rcu_tasks_wait_gp()
806 // invokes these functions in this order:
808 // rcu_tasks_trace_pregp_step():
809 // Initialize the count of readers and block CPU-hotplug operations.
810 // rcu_tasks_trace_pertask(), invoked on every non-idle task:
811 // Initialize per-task state and attempt to identify an immediate
812 // quiescent state for that task, or, failing that, attempt to
813 // set that task's .need_qs flag so that task's next outermost
814 // rcu_read_unlock_trace() will report the quiescent state (in which
815 // case the count of readers is incremented). If both attempts fail,
816 // the task is added to a "holdout" list. Note that IPIs are used
817 // to invoke trc_read_check_handler() in the context of running tasks
818 // in order to avoid ordering overhead on common-case shared-variable
820 // rcu_tasks_trace_postscan():
821 // Initialize state and attempt to identify an immediate quiescent
822 // state as above (but only for idle tasks), unblock CPU-hotplug
823 // operations, and wait for an RCU grace period to avoid races with
824 // tasks that are in the process of exiting.
825 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
826 // Scans the holdout list, attempting to identify a quiescent state
827 // for each task on the list. If there is a quiescent state, the
828 // corresponding task is removed from the holdout list.
829 // rcu_tasks_trace_postgp():
830 // Wait for the count of readers do drop to zero, reporting any stalls.
831 // Also execute full memory barriers to maintain ordering with code
832 // executing after the grace period.
834 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
836 // Pre-grace-period update-side code is ordered before the grace
837 // period via the ->cbs_lock and barriers in rcu_tasks_kthread().
838 // Pre-grace-period read-side code is ordered before the grace period by
839 // atomic_dec_and_test() of the count of readers (for IPIed readers) and by
840 // scheduler context-switch ordering (for locked-down non-running readers).
842 // The lockdep state must be outside of #ifdef to be useful.
843 #ifdef CONFIG_DEBUG_LOCK_ALLOC
844 static struct lock_class_key rcu_lock_trace_key;
845 struct lockdep_map rcu_trace_lock_map =
846 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
847 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
848 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
850 #ifdef CONFIG_TASKS_TRACE_RCU
852 static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
853 static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
855 // Record outstanding IPIs to each CPU. No point in sending two...
856 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
858 // The number of detections of task quiescent state relying on
859 // heavyweight readers executing explicit memory barriers.
860 static unsigned long n_heavy_reader_attempts;
861 static unsigned long n_heavy_reader_updates;
862 static unsigned long n_heavy_reader_ofl_updates;
864 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
865 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
869 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
870 * while the scheduler locks are held.
872 static void rcu_read_unlock_iw(struct irq_work *iwp)
876 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
878 /* If we are the last reader, wake up the grace-period kthread. */
879 void rcu_read_unlock_trace_special(struct task_struct *t, int nesting)
881 int nq = READ_ONCE(t->trc_reader_special.b.need_qs);
883 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
884 t->trc_reader_special.b.need_mb)
885 smp_mb(); // Pairs with update-side barriers.
886 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
888 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
889 WRITE_ONCE(t->trc_reader_nesting, nesting);
890 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
891 irq_work_queue(&rcu_tasks_trace_iw);
893 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
895 /* Add a task to the holdout list, if it is not already on the list. */
896 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
898 if (list_empty(&t->trc_holdout_list)) {
900 list_add(&t->trc_holdout_list, bhp);
904 /* Remove a task from the holdout list, if it is in fact present. */
905 static void trc_del_holdout(struct task_struct *t)
907 if (!list_empty(&t->trc_holdout_list)) {
908 list_del_init(&t->trc_holdout_list);
913 /* IPI handler to check task state. */
914 static void trc_read_check_handler(void *t_in)
916 struct task_struct *t = current;
917 struct task_struct *texp = t_in;
919 // If the task is no longer running on this CPU, leave.
920 if (unlikely(texp != t)) {
921 goto reset_ipi; // Already on holdout list, so will check later.
924 // If the task is not in a read-side critical section, and
925 // if this is the last reader, awaken the grace-period kthread.
926 if (likely(!READ_ONCE(t->trc_reader_nesting))) {
927 WRITE_ONCE(t->trc_reader_checked, true);
930 // If we are racing with an rcu_read_unlock_trace(), try again later.
931 if (unlikely(READ_ONCE(t->trc_reader_nesting) < 0))
933 WRITE_ONCE(t->trc_reader_checked, true);
935 // Get here if the task is in a read-side critical section. Set
936 // its state so that it will awaken the grace-period kthread upon
937 // exit from that critical section.
938 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
939 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
940 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
943 // Allow future IPIs to be sent on CPU and for task.
944 // Also order this IPI handler against any later manipulations of
945 // the intended task.
946 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
947 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
950 /* Callback function for scheduler to check locked-down task. */
951 static bool trc_inspect_reader(struct task_struct *t, void *arg)
953 int cpu = task_cpu(t);
955 bool ofl = cpu_is_offline(cpu);
958 WARN_ON_ONCE(ofl && !is_idle_task(t));
960 // If no chance of heavyweight readers, do it the hard way.
961 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
964 // If heavyweight readers are enabled on the remote task,
965 // we can inspect its state despite its currently running.
966 // However, we cannot safely change its state.
967 n_heavy_reader_attempts++;
968 if (!ofl && // Check for "running" idle tasks on offline CPUs.
969 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
970 return false; // No quiescent state, do it the hard way.
971 n_heavy_reader_updates++;
973 n_heavy_reader_ofl_updates++;
976 // The task is not running, so C-language access is safe.
977 nesting = t->trc_reader_nesting;
980 // If not exiting a read-side critical section, mark as checked
981 // so that the grace-period kthread will remove it from the
983 t->trc_reader_checked = nesting >= 0;
985 return !nesting; // If in QS, done, otherwise try again later.
987 // The task is in a read-side critical section, so set up its
988 // state so that it will awaken the grace-period kthread upon exit
989 // from that critical section.
990 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
991 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
992 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
996 /* Attempt to extract the state for the specified task. */
997 static void trc_wait_for_one_reader(struct task_struct *t,
998 struct list_head *bhp)
1002 // If a previous IPI is still in flight, let it complete.
1003 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1006 // The current task had better be in a quiescent state.
1008 t->trc_reader_checked = true;
1009 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1013 // Attempt to nail down the task for inspection.
1015 if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) {
1021 // If this task is not yet on the holdout list, then we are in
1022 // an RCU read-side critical section. Otherwise, the invocation of
1023 // rcu_add_holdout() that added it to the list did the necessary
1024 // get_task_struct(). Either way, the task cannot be freed out
1025 // from under this code.
1027 // If currently running, send an IPI, either way, add to list.
1028 trc_add_holdout(t, bhp);
1030 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1031 // The task is currently running, so try IPIing it.
1034 // If there is already an IPI outstanding, let it happen.
1035 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1038 per_cpu(trc_ipi_to_cpu, cpu) = true;
1039 t->trc_ipi_to_cpu = cpu;
1040 rcu_tasks_trace.n_ipis++;
1041 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1042 // Just in case there is some other reason for
1043 // failure than the target CPU being offline.
1044 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1046 rcu_tasks_trace.n_ipis_fails++;
1047 per_cpu(trc_ipi_to_cpu, cpu) = false;
1048 t->trc_ipi_to_cpu = -1;
1053 /* Initialize for a new RCU-tasks-trace grace period. */
1054 static void rcu_tasks_trace_pregp_step(void)
1058 // Allow for fast-acting IPIs.
1059 atomic_set(&trc_n_readers_need_end, 1);
1061 // There shouldn't be any old IPIs, but...
1062 for_each_possible_cpu(cpu)
1063 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1065 // Disable CPU hotplug across the tasklist scan.
1066 // This also waits for all readers in CPU-hotplug code paths.
1070 /* Do first-round processing for the specified task. */
1071 static void rcu_tasks_trace_pertask(struct task_struct *t,
1072 struct list_head *hop)
1074 // During early boot when there is only the one boot CPU, there
1075 // is no idle task for the other CPUs. Just return.
1076 if (unlikely(t == NULL))
1079 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1080 WRITE_ONCE(t->trc_reader_checked, false);
1081 t->trc_ipi_to_cpu = -1;
1082 trc_wait_for_one_reader(t, hop);
1086 * Do intermediate processing between task and holdout scans and
1087 * pick up the idle tasks.
1089 static void rcu_tasks_trace_postscan(struct list_head *hop)
1093 for_each_possible_cpu(cpu)
1094 rcu_tasks_trace_pertask(idle_task(cpu), hop);
1096 // Re-enable CPU hotplug now that the tasklist scan has completed.
1099 // Wait for late-stage exiting tasks to finish exiting.
1100 // These might have passed the call to exit_tasks_rcu_finish().
1102 // Any tasks that exit after this point will set ->trc_reader_checked.
1105 /* Show the state of a task stalling the current RCU tasks trace GP. */
1106 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1111 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1112 *firstreport = false;
1114 // FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1116 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1118 ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0],
1119 ".i"[is_idle_task(t)],
1120 ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)],
1121 READ_ONCE(t->trc_reader_nesting),
1122 " N"[!!READ_ONCE(t->trc_reader_special.b.need_qs)],
1127 /* List stalled IPIs for RCU tasks trace. */
1128 static void show_stalled_ipi_trace(void)
1132 for_each_possible_cpu(cpu)
1133 if (per_cpu(trc_ipi_to_cpu, cpu))
1134 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1137 /* Do one scan of the holdout list. */
1138 static void check_all_holdout_tasks_trace(struct list_head *hop,
1139 bool needreport, bool *firstreport)
1141 struct task_struct *g, *t;
1143 // Disable CPU hotplug across the holdout list scan.
1146 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1147 // If safe and needed, try to check the current task.
1148 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1149 !READ_ONCE(t->trc_reader_checked))
1150 trc_wait_for_one_reader(t, hop);
1152 // If check succeeded, remove this task from the list.
1153 if (READ_ONCE(t->trc_reader_checked))
1155 else if (needreport)
1156 show_stalled_task_trace(t, firstreport);
1159 // Re-enable CPU hotplug now that the holdout list scan has completed.
1164 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1165 show_stalled_ipi_trace();
1169 static void rcu_tasks_trace_empty_fn(void *unused)
1173 /* Wait for grace period to complete and provide ordering. */
1174 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1178 struct task_struct *g, *t;
1179 LIST_HEAD(holdouts);
1182 // Wait for any lingering IPI handlers to complete. Note that
1183 // if a CPU has gone offline or transitioned to userspace in the
1184 // meantime, all IPI handlers should have been drained beforehand.
1185 // Yes, this assumes that CPUs process IPIs in order. If that ever
1186 // changes, there will need to be a recheck and/or timed wait.
1187 for_each_online_cpu(cpu)
1188 if (smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu)))
1189 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1191 // Remove the safety count.
1192 smp_mb__before_atomic(); // Order vs. earlier atomics
1193 atomic_dec(&trc_n_readers_need_end);
1194 smp_mb__after_atomic(); // Order vs. later atomics
1196 // Wait for readers.
1197 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1199 ret = wait_event_idle_exclusive_timeout(
1201 atomic_read(&trc_n_readers_need_end) == 0,
1202 READ_ONCE(rcu_task_stall_timeout));
1204 break; // Count reached zero.
1205 // Stall warning time, so make a list of the offenders.
1207 for_each_process_thread(g, t)
1208 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1209 trc_add_holdout(t, &holdouts);
1212 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1213 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1214 show_stalled_task_trace(t, &firstreport);
1215 trc_del_holdout(t); // Release task_struct reference.
1218 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1219 show_stalled_ipi_trace();
1220 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1222 smp_mb(); // Caller's code must be ordered after wakeup.
1223 // Pairs with pretty much every ordering primitive.
1226 /* Report any needed quiescent state for this exiting task. */
1227 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1229 WRITE_ONCE(t->trc_reader_checked, true);
1230 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1231 WRITE_ONCE(t->trc_reader_nesting, 0);
1232 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1233 rcu_read_unlock_trace_special(t, 0);
1237 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1238 * @rhp: structure to be used for queueing the RCU updates.
1239 * @func: actual callback function to be invoked after the grace period
1241 * The callback function will be invoked some time after a full grace
1242 * period elapses, in other words after all currently executing RCU
1243 * read-side critical sections have completed. call_rcu_tasks_trace()
1244 * assumes that the read-side critical sections end at context switch,
1245 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
1246 * there are no read-side primitives analogous to rcu_read_lock() and
1247 * rcu_read_unlock() because this primitive is intended to determine
1248 * that all tasks have passed through a safe state, not so much for
1249 * data-structure synchronization.
1251 * See the description of call_rcu() for more detailed information on
1252 * memory ordering guarantees.
1254 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1256 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1258 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1261 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1263 * Control will return to the caller some time after a trace rcu-tasks
1264 * grace period has elapsed, in other words after all currently executing
1265 * rcu-tasks read-side critical sections have elapsed. These read-side
1266 * critical sections are delimited by calls to rcu_read_lock_trace()
1267 * and rcu_read_unlock_trace().
1269 * This is a very specialized primitive, intended only for a few uses in
1270 * tracing and other situations requiring manipulation of function preambles
1271 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1272 * (yet) intended for heavy use from multiple CPUs.
1274 * See the description of synchronize_rcu() for more detailed information
1275 * on memory ordering guarantees.
1277 void synchronize_rcu_tasks_trace(void)
1279 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1280 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1282 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1285 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1287 * Although the current implementation is guaranteed to wait, it is not
1288 * obligated to, for example, if there are no pending callbacks.
1290 void rcu_barrier_tasks_trace(void)
1292 /* There is only one callback queue, so this is easy. ;-) */
1293 synchronize_rcu_tasks_trace();
1295 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1297 static int __init rcu_spawn_tasks_trace_kthread(void)
1299 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1300 rcu_tasks_trace.gp_sleep = HZ / 10;
1301 rcu_tasks_trace.init_fract = HZ / 10;
1303 rcu_tasks_trace.gp_sleep = HZ / 200;
1304 if (rcu_tasks_trace.gp_sleep <= 0)
1305 rcu_tasks_trace.gp_sleep = 1;
1306 rcu_tasks_trace.init_fract = HZ / 200;
1307 if (rcu_tasks_trace.init_fract <= 0)
1308 rcu_tasks_trace.init_fract = 1;
1310 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1311 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1312 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1313 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1314 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1315 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1319 #if !defined(CONFIG_TINY_RCU)
1320 void show_rcu_tasks_trace_gp_kthread(void)
1324 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1325 data_race(n_heavy_reader_ofl_updates),
1326 data_race(n_heavy_reader_updates),
1327 data_race(n_heavy_reader_attempts));
1328 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1330 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1331 #endif // !defined(CONFIG_TINY_RCU)
1333 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1334 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1335 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1337 #ifndef CONFIG_TINY_RCU
1338 void show_rcu_tasks_gp_kthreads(void)
1340 show_rcu_tasks_classic_gp_kthread();
1341 show_rcu_tasks_rude_gp_kthread();
1342 show_rcu_tasks_trace_gp_kthread();
1344 #endif /* #ifndef CONFIG_TINY_RCU */
1346 #ifdef CONFIG_PROVE_RCU
1347 struct rcu_tasks_test_desc {
1353 static struct rcu_tasks_test_desc tests[] = {
1355 .name = "call_rcu_tasks()",
1356 /* If not defined, the test is skipped. */
1357 .notrun = !IS_ENABLED(CONFIG_TASKS_RCU),
1360 .name = "call_rcu_tasks_rude()",
1361 /* If not defined, the test is skipped. */
1362 .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1365 .name = "call_rcu_tasks_trace()",
1366 /* If not defined, the test is skipped. */
1367 .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1371 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1373 struct rcu_tasks_test_desc *rttd =
1374 container_of(rhp, struct rcu_tasks_test_desc, rh);
1376 pr_info("Callback from %s invoked.\n", rttd->name);
1378 rttd->notrun = true;
1381 static void rcu_tasks_initiate_self_tests(void)
1383 pr_info("Running RCU-tasks wait API self tests\n");
1384 #ifdef CONFIG_TASKS_RCU
1385 synchronize_rcu_tasks();
1386 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1389 #ifdef CONFIG_TASKS_RUDE_RCU
1390 synchronize_rcu_tasks_rude();
1391 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1394 #ifdef CONFIG_TASKS_TRACE_RCU
1395 synchronize_rcu_tasks_trace();
1396 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1400 static int rcu_tasks_verify_self_tests(void)
1405 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1406 if (!tests[i].notrun) { // still hanging.
1407 pr_err("%s has been failed.\n", tests[i].name);
1417 late_initcall(rcu_tasks_verify_self_tests);
1418 #else /* #ifdef CONFIG_PROVE_RCU */
1419 static void rcu_tasks_initiate_self_tests(void) { }
1420 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1422 void __init rcu_init_tasks_generic(void)
1424 #ifdef CONFIG_TASKS_RCU
1425 rcu_spawn_tasks_kthread();
1428 #ifdef CONFIG_TASKS_RUDE_RCU
1429 rcu_spawn_tasks_rude_kthread();
1432 #ifdef CONFIG_TASKS_TRACE_RCU
1433 rcu_spawn_tasks_trace_kthread();
1436 // Run the self-tests.
1437 rcu_tasks_initiate_self_tests();
1440 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1441 static inline void rcu_tasks_bootup_oddness(void) {}
1442 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */