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. ;-)
201 /* Pick up any new callbacks. */
202 raw_spin_lock_irqsave(&rtp->cbs_lock, flags);
203 smp_mb__after_spinlock(); // Order updates vs. GP.
204 list = rtp->cbs_head;
205 rtp->cbs_head = NULL;
206 rtp->cbs_tail = &rtp->cbs_head;
207 raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags);
209 /* If there were none, wait a bit and start over. */
211 wait_event_interruptible(rtp->cbs_wq,
212 READ_ONCE(rtp->cbs_head));
213 if (!rtp->cbs_head) {
214 WARN_ON(signal_pending(current));
215 set_tasks_gp_state(rtp, RTGS_WAIT_WAIT_CBS);
216 schedule_timeout_idle(HZ/10);
221 // Wait for one grace period.
222 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
223 rtp->gp_start = jiffies;
227 /* Invoke the callbacks. */
228 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
237 /* Paranoid sleep to keep this from entering a tight loop */
238 schedule_timeout_idle(rtp->gp_sleep);
240 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
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 * Wait for tasks that are in the process of exiting. This
456 * does only part of the job, ensuring that all tasks that were
457 * previously exiting reach the point where they have disabled
458 * preemption, allowing the later synchronize_rcu() to finish
461 synchronize_srcu(&tasks_rcu_exit_srcu);
464 /* See if tasks are still holding out, complain if so. */
465 static void check_holdout_task(struct task_struct *t,
466 bool needreport, bool *firstreport)
470 if (!READ_ONCE(t->rcu_tasks_holdout) ||
471 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
472 !READ_ONCE(t->on_rq) ||
473 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
474 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
475 WRITE_ONCE(t->rcu_tasks_holdout, false);
476 list_del_init(&t->rcu_tasks_holdout_list);
480 rcu_request_urgent_qs_task(t);
484 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
485 *firstreport = false;
488 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
489 t, ".I"[is_idle_task(t)],
490 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
491 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
492 t->rcu_tasks_idle_cpu, cpu);
496 /* Scan the holdout lists for tasks no longer holding out. */
497 static void check_all_holdout_tasks(struct list_head *hop,
498 bool needreport, bool *firstreport)
500 struct task_struct *t, *t1;
502 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
503 check_holdout_task(t, needreport, firstreport);
508 /* Finish off the Tasks-RCU grace period. */
509 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
512 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
513 * memory barriers prior to them in the schedule() path, memory
514 * reordering on other CPUs could cause their RCU-tasks read-side
515 * critical sections to extend past the end of the grace period.
516 * However, because these ->nvcsw updates are carried out with
517 * interrupts disabled, we can use synchronize_rcu() to force the
518 * needed ordering on all such CPUs.
520 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
521 * accesses to be within the grace period, avoiding the need for
522 * memory barriers for ->rcu_tasks_holdout accesses.
524 * In addition, this synchronize_rcu() waits for exiting tasks
525 * to complete their final preempt_disable() region of execution,
526 * cleaning up after the synchronize_srcu() above.
531 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
532 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
535 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
536 * @rhp: structure to be used for queueing the RCU updates.
537 * @func: actual callback function to be invoked after the grace period
539 * The callback function will be invoked some time after a full grace
540 * period elapses, in other words after all currently executing RCU
541 * read-side critical sections have completed. call_rcu_tasks() assumes
542 * that the read-side critical sections end at a voluntary context
543 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
544 * or transition to usermode execution. As such, there are no read-side
545 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
546 * this primitive is intended to determine that all tasks have passed
547 * through a safe state, not so much for data-structure synchronization.
549 * See the description of call_rcu() for more detailed information on
550 * memory ordering guarantees.
552 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
554 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
556 EXPORT_SYMBOL_GPL(call_rcu_tasks);
559 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
561 * Control will return to the caller some time after a full rcu-tasks
562 * grace period has elapsed, in other words after all currently
563 * executing rcu-tasks read-side critical sections have elapsed. These
564 * read-side critical sections are delimited by calls to schedule(),
565 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
566 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
568 * This is a very specialized primitive, intended only for a few uses in
569 * tracing and other situations requiring manipulation of function
570 * preambles and profiling hooks. The synchronize_rcu_tasks() function
571 * is not (yet) intended for heavy use from multiple CPUs.
573 * See the description of synchronize_rcu() for more detailed information
574 * on memory ordering guarantees.
576 void synchronize_rcu_tasks(void)
578 synchronize_rcu_tasks_generic(&rcu_tasks);
580 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
583 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
585 * Although the current implementation is guaranteed to wait, it is not
586 * obligated to, for example, if there are no pending callbacks.
588 void rcu_barrier_tasks(void)
590 /* There is only one callback queue, so this is easy. ;-) */
591 synchronize_rcu_tasks();
593 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
595 static int __init rcu_spawn_tasks_kthread(void)
597 rcu_tasks.gp_sleep = HZ / 10;
598 rcu_tasks.init_fract = HZ / 10;
599 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
600 rcu_tasks.pertask_func = rcu_tasks_pertask;
601 rcu_tasks.postscan_func = rcu_tasks_postscan;
602 rcu_tasks.holdouts_func = check_all_holdout_tasks;
603 rcu_tasks.postgp_func = rcu_tasks_postgp;
604 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
608 #if !defined(CONFIG_TINY_RCU)
609 void show_rcu_tasks_classic_gp_kthread(void)
611 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
613 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
614 #endif // !defined(CONFIG_TINY_RCU)
616 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
617 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
620 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
624 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
625 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
627 struct task_struct *t = current;
630 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
632 exit_tasks_rcu_finish_trace(t);
635 #else /* #ifdef CONFIG_TASKS_RCU */
636 void exit_tasks_rcu_start(void) { }
637 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
638 #endif /* #else #ifdef CONFIG_TASKS_RCU */
640 #ifdef CONFIG_TASKS_RUDE_RCU
642 ////////////////////////////////////////////////////////////////////////
644 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
645 // passing an empty function to schedule_on_each_cpu(). This approach
646 // provides an asynchronous call_rcu_tasks_rude() API and batching of
647 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
648 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
649 // and induces otherwise unnecessary context switches on all online CPUs,
650 // whether idle or not.
652 // Callback handling is provided by the rcu_tasks_kthread() function.
654 // Ordering is provided by the scheduler's context-switch code.
656 // Empty function to allow workqueues to force a context switch.
657 static void rcu_tasks_be_rude(struct work_struct *work)
661 // Wait for one rude RCU-tasks grace period.
662 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
664 rtp->n_ipis += cpumask_weight(cpu_online_mask);
665 schedule_on_each_cpu(rcu_tasks_be_rude);
668 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
669 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
673 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
674 * @rhp: structure to be used for queueing the RCU updates.
675 * @func: actual callback function to be invoked after the grace period
677 * The callback function will be invoked some time after a full grace
678 * period elapses, in other words after all currently executing RCU
679 * read-side critical sections have completed. call_rcu_tasks_rude()
680 * assumes that the read-side critical sections end at context switch,
681 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
682 * there are no read-side primitives analogous to rcu_read_lock() and
683 * rcu_read_unlock() because this primitive is intended to determine
684 * that all tasks have passed through a safe state, not so much for
685 * data-structure synchronization.
687 * See the description of call_rcu() for more detailed information on
688 * memory ordering guarantees.
690 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
692 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
694 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
697 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
699 * Control will return to the caller some time after a rude rcu-tasks
700 * grace period has elapsed, in other words after all currently
701 * executing rcu-tasks read-side critical sections have elapsed. These
702 * read-side critical sections are delimited by calls to schedule(),
703 * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
704 * anyway) cond_resched().
706 * This is a very specialized primitive, intended only for a few uses in
707 * tracing and other situations requiring manipulation of function preambles
708 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
709 * (yet) intended for heavy use from multiple CPUs.
711 * See the description of synchronize_rcu() for more detailed information
712 * on memory ordering guarantees.
714 void synchronize_rcu_tasks_rude(void)
716 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
718 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
721 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
723 * Although the current implementation is guaranteed to wait, it is not
724 * obligated to, for example, if there are no pending callbacks.
726 void rcu_barrier_tasks_rude(void)
728 /* There is only one callback queue, so this is easy. ;-) */
729 synchronize_rcu_tasks_rude();
731 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
733 static int __init rcu_spawn_tasks_rude_kthread(void)
735 rcu_tasks_rude.gp_sleep = HZ / 10;
736 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
740 #if !defined(CONFIG_TINY_RCU)
741 void show_rcu_tasks_rude_gp_kthread(void)
743 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
745 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
746 #endif // !defined(CONFIG_TINY_RCU)
747 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
749 ////////////////////////////////////////////////////////////////////////
751 // Tracing variant of Tasks RCU. This variant is designed to be used
752 // to protect tracing hooks, including those of BPF. This variant
755 // 1. Has explicit read-side markers to allow finite grace periods
756 // in the face of in-kernel loops for PREEMPT=n builds.
758 // 2. Protects code in the idle loop, exception entry/exit, and
759 // CPU-hotplug code paths, similar to the capabilities of SRCU.
761 // 3. Avoids expensive read-side instruction, having overhead similar
762 // to that of Preemptible RCU.
764 // There are of course downsides. The grace-period code can send IPIs to
765 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
766 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
767 // is a single callback queue guarded by a single lock, again, much as for
768 // Tasks RCU. If needed, these downsides can be at least partially remedied.
770 // Perhaps most important, this variant of RCU does not affect the vanilla
771 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
772 // readers can operate from idle, offline, and exception entry/exit in no
773 // way allows rcu_preempt and rcu_sched readers to also do so.
775 // The implementation uses rcu_tasks_wait_gp(), which relies on function
776 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
777 // function sets these function pointers up so that rcu_tasks_wait_gp()
778 // invokes these functions in this order:
780 // rcu_tasks_trace_pregp_step():
781 // Initialize the count of readers and block CPU-hotplug operations.
782 // rcu_tasks_trace_pertask(), invoked on every non-idle task:
783 // Initialize per-task state and attempt to identify an immediate
784 // quiescent state for that task, or, failing that, attempt to
785 // set that task's .need_qs flag so that task's next outermost
786 // rcu_read_unlock_trace() will report the quiescent state (in which
787 // case the count of readers is incremented). If both attempts fail,
788 // the task is added to a "holdout" list. Note that IPIs are used
789 // to invoke trc_read_check_handler() in the context of running tasks
790 // in order to avoid ordering overhead on common-case shared-variable
792 // rcu_tasks_trace_postscan():
793 // Initialize state and attempt to identify an immediate quiescent
794 // state as above (but only for idle tasks), unblock CPU-hotplug
795 // operations, and wait for an RCU grace period to avoid races with
796 // tasks that are in the process of exiting.
797 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
798 // Scans the holdout list, attempting to identify a quiescent state
799 // for each task on the list. If there is a quiescent state, the
800 // corresponding task is removed from the holdout list.
801 // rcu_tasks_trace_postgp():
802 // Wait for the count of readers do drop to zero, reporting any stalls.
803 // Also execute full memory barriers to maintain ordering with code
804 // executing after the grace period.
806 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
808 // Pre-grace-period update-side code is ordered before the grace
809 // period via the ->cbs_lock and barriers in rcu_tasks_kthread().
810 // Pre-grace-period read-side code is ordered before the grace period by
811 // atomic_dec_and_test() of the count of readers (for IPIed readers) and by
812 // scheduler context-switch ordering (for locked-down non-running readers).
814 // The lockdep state must be outside of #ifdef to be useful.
815 #ifdef CONFIG_DEBUG_LOCK_ALLOC
816 static struct lock_class_key rcu_lock_trace_key;
817 struct lockdep_map rcu_trace_lock_map =
818 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
819 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
820 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
822 #ifdef CONFIG_TASKS_TRACE_RCU
824 static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
825 static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
827 // Record outstanding IPIs to each CPU. No point in sending two...
828 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
830 // The number of detections of task quiescent state relying on
831 // heavyweight readers executing explicit memory barriers.
832 static unsigned long n_heavy_reader_attempts;
833 static unsigned long n_heavy_reader_updates;
834 static unsigned long n_heavy_reader_ofl_updates;
836 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
837 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
841 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
842 * while the scheduler locks are held.
844 static void rcu_read_unlock_iw(struct irq_work *iwp)
848 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
850 /* If we are the last reader, wake up the grace-period kthread. */
851 void rcu_read_unlock_trace_special(struct task_struct *t, int nesting)
853 int nq = READ_ONCE(t->trc_reader_special.b.need_qs);
855 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
856 t->trc_reader_special.b.need_mb)
857 smp_mb(); // Pairs with update-side barriers.
858 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
860 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
861 WRITE_ONCE(t->trc_reader_nesting, nesting);
862 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
863 irq_work_queue(&rcu_tasks_trace_iw);
865 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
867 /* Add a task to the holdout list, if it is not already on the list. */
868 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
870 if (list_empty(&t->trc_holdout_list)) {
872 list_add(&t->trc_holdout_list, bhp);
876 /* Remove a task from the holdout list, if it is in fact present. */
877 static void trc_del_holdout(struct task_struct *t)
879 if (!list_empty(&t->trc_holdout_list)) {
880 list_del_init(&t->trc_holdout_list);
885 /* IPI handler to check task state. */
886 static void trc_read_check_handler(void *t_in)
888 struct task_struct *t = current;
889 struct task_struct *texp = t_in;
891 // If the task is no longer running on this CPU, leave.
892 if (unlikely(texp != t)) {
893 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
895 goto reset_ipi; // Already on holdout list, so will check later.
898 // If the task is not in a read-side critical section, and
899 // if this is the last reader, awaken the grace-period kthread.
900 if (likely(!READ_ONCE(t->trc_reader_nesting))) {
901 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
903 // Mark as checked after decrement to avoid false
904 // positives on the above WARN_ON_ONCE().
905 WRITE_ONCE(t->trc_reader_checked, true);
908 // If we are racing with an rcu_read_unlock_trace(), try again later.
909 if (unlikely(READ_ONCE(t->trc_reader_nesting) < 0)) {
910 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
914 WRITE_ONCE(t->trc_reader_checked, true);
916 // Get here if the task is in a read-side critical section. Set
917 // its state so that it will awaken the grace-period kthread upon
918 // exit from that critical section.
919 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
920 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
923 // Allow future IPIs to be sent on CPU and for task.
924 // Also order this IPI handler against any later manipulations of
925 // the intended task.
926 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
927 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
930 /* Callback function for scheduler to check locked-down task. */
931 static bool trc_inspect_reader(struct task_struct *t, void *arg)
933 int cpu = task_cpu(t);
935 bool ofl = cpu_is_offline(cpu);
938 WARN_ON_ONCE(ofl && !is_idle_task(t));
940 // If no chance of heavyweight readers, do it the hard way.
941 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
944 // If heavyweight readers are enabled on the remote task,
945 // we can inspect its state despite its currently running.
946 // However, we cannot safely change its state.
947 n_heavy_reader_attempts++;
948 if (!ofl && // Check for "running" idle tasks on offline CPUs.
949 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
950 return false; // No quiescent state, do it the hard way.
951 n_heavy_reader_updates++;
953 n_heavy_reader_ofl_updates++;
956 // The task is not running, so C-language access is safe.
957 in_qs = likely(!t->trc_reader_nesting);
960 // Mark as checked so that the grace-period kthread will
961 // remove it from the holdout list.
962 t->trc_reader_checked = true;
965 return true; // Already in quiescent state, done!!!
967 // The task is in a read-side critical section, so set up its
968 // state so that it will awaken the grace-period kthread upon exit
969 // from that critical section.
970 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
971 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
972 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
976 /* Attempt to extract the state for the specified task. */
977 static void trc_wait_for_one_reader(struct task_struct *t,
978 struct list_head *bhp)
982 // If a previous IPI is still in flight, let it complete.
983 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
986 // The current task had better be in a quiescent state.
988 t->trc_reader_checked = true;
989 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
993 // Attempt to nail down the task for inspection.
995 if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) {
1001 // If this task is not yet on the holdout list, then we are in
1002 // an RCU read-side critical section. Otherwise, the invocation of
1003 // rcu_add_holdout() that added it to the list did the necessary
1004 // get_task_struct(). Either way, the task cannot be freed out
1005 // from under this code.
1007 // If currently running, send an IPI, either way, add to list.
1008 trc_add_holdout(t, bhp);
1010 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1011 // The task is currently running, so try IPIing it.
1014 // If there is already an IPI outstanding, let it happen.
1015 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1018 atomic_inc(&trc_n_readers_need_end);
1019 per_cpu(trc_ipi_to_cpu, cpu) = true;
1020 t->trc_ipi_to_cpu = cpu;
1021 rcu_tasks_trace.n_ipis++;
1022 if (smp_call_function_single(cpu,
1023 trc_read_check_handler, t, 0)) {
1024 // Just in case there is some other reason for
1025 // failure than the target CPU being offline.
1026 rcu_tasks_trace.n_ipis_fails++;
1027 per_cpu(trc_ipi_to_cpu, cpu) = false;
1028 t->trc_ipi_to_cpu = cpu;
1029 if (atomic_dec_and_test(&trc_n_readers_need_end)) {
1037 /* Initialize for a new RCU-tasks-trace grace period. */
1038 static void rcu_tasks_trace_pregp_step(void)
1042 // Allow for fast-acting IPIs.
1043 atomic_set(&trc_n_readers_need_end, 1);
1045 // There shouldn't be any old IPIs, but...
1046 for_each_possible_cpu(cpu)
1047 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1049 // Disable CPU hotplug across the tasklist scan.
1050 // This also waits for all readers in CPU-hotplug code paths.
1054 /* Do first-round processing for the specified task. */
1055 static void rcu_tasks_trace_pertask(struct task_struct *t,
1056 struct list_head *hop)
1058 // During early boot when there is only the one boot CPU, there
1059 // is no idle task for the other CPUs. Just return.
1060 if (unlikely(t == NULL))
1063 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1064 WRITE_ONCE(t->trc_reader_checked, false);
1065 t->trc_ipi_to_cpu = -1;
1066 trc_wait_for_one_reader(t, hop);
1070 * Do intermediate processing between task and holdout scans and
1071 * pick up the idle tasks.
1073 static void rcu_tasks_trace_postscan(struct list_head *hop)
1077 for_each_possible_cpu(cpu)
1078 rcu_tasks_trace_pertask(idle_task(cpu), hop);
1080 // Re-enable CPU hotplug now that the tasklist scan has completed.
1083 // Wait for late-stage exiting tasks to finish exiting.
1084 // These might have passed the call to exit_tasks_rcu_finish().
1086 // Any tasks that exit after this point will set ->trc_reader_checked.
1089 /* Show the state of a task stalling the current RCU tasks trace GP. */
1090 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1095 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1096 *firstreport = false;
1098 // FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1100 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1102 ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0],
1103 ".i"[is_idle_task(t)],
1104 ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)],
1105 READ_ONCE(t->trc_reader_nesting),
1106 " N"[!!READ_ONCE(t->trc_reader_special.b.need_qs)],
1111 /* List stalled IPIs for RCU tasks trace. */
1112 static void show_stalled_ipi_trace(void)
1116 for_each_possible_cpu(cpu)
1117 if (per_cpu(trc_ipi_to_cpu, cpu))
1118 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1121 /* Do one scan of the holdout list. */
1122 static void check_all_holdout_tasks_trace(struct list_head *hop,
1123 bool needreport, bool *firstreport)
1125 struct task_struct *g, *t;
1127 // Disable CPU hotplug across the holdout list scan.
1130 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1131 // If safe and needed, try to check the current task.
1132 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1133 !READ_ONCE(t->trc_reader_checked))
1134 trc_wait_for_one_reader(t, hop);
1136 // If check succeeded, remove this task from the list.
1137 if (READ_ONCE(t->trc_reader_checked))
1139 else if (needreport)
1140 show_stalled_task_trace(t, firstreport);
1143 // Re-enable CPU hotplug now that the holdout list scan has completed.
1148 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1149 show_stalled_ipi_trace();
1153 /* Wait for grace period to complete and provide ordering. */
1154 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1157 struct task_struct *g, *t;
1158 LIST_HEAD(holdouts);
1161 // Remove the safety count.
1162 smp_mb__before_atomic(); // Order vs. earlier atomics
1163 atomic_dec(&trc_n_readers_need_end);
1164 smp_mb__after_atomic(); // Order vs. later atomics
1166 // Wait for readers.
1167 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1169 ret = wait_event_idle_exclusive_timeout(
1171 atomic_read(&trc_n_readers_need_end) == 0,
1172 READ_ONCE(rcu_task_stall_timeout));
1174 break; // Count reached zero.
1175 // Stall warning time, so make a list of the offenders.
1177 for_each_process_thread(g, t)
1178 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1179 trc_add_holdout(t, &holdouts);
1182 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1183 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1184 show_stalled_task_trace(t, &firstreport);
1185 trc_del_holdout(t); // Release task_struct reference.
1188 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1189 show_stalled_ipi_trace();
1190 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1192 smp_mb(); // Caller's code must be ordered after wakeup.
1193 // Pairs with pretty much every ordering primitive.
1196 /* Report any needed quiescent state for this exiting task. */
1197 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1199 WRITE_ONCE(t->trc_reader_checked, true);
1200 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1201 WRITE_ONCE(t->trc_reader_nesting, 0);
1202 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1203 rcu_read_unlock_trace_special(t, 0);
1207 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1208 * @rhp: structure to be used for queueing the RCU updates.
1209 * @func: actual callback function to be invoked after the grace period
1211 * The callback function will be invoked some time after a full grace
1212 * period elapses, in other words after all currently executing RCU
1213 * read-side critical sections have completed. call_rcu_tasks_trace()
1214 * assumes that the read-side critical sections end at context switch,
1215 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
1216 * there are no read-side primitives analogous to rcu_read_lock() and
1217 * rcu_read_unlock() because this primitive is intended to determine
1218 * that all tasks have passed through a safe state, not so much for
1219 * data-structure synchronization.
1221 * See the description of call_rcu() for more detailed information on
1222 * memory ordering guarantees.
1224 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1226 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1228 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1231 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1233 * Control will return to the caller some time after a trace rcu-tasks
1234 * grace period has elapsed, in other words after all currently executing
1235 * rcu-tasks read-side critical sections have elapsed. These read-side
1236 * critical sections are delimited by calls to rcu_read_lock_trace()
1237 * and rcu_read_unlock_trace().
1239 * This is a very specialized primitive, intended only for a few uses in
1240 * tracing and other situations requiring manipulation of function preambles
1241 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1242 * (yet) intended for heavy use from multiple CPUs.
1244 * See the description of synchronize_rcu() for more detailed information
1245 * on memory ordering guarantees.
1247 void synchronize_rcu_tasks_trace(void)
1249 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1250 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1252 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1255 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1257 * Although the current implementation is guaranteed to wait, it is not
1258 * obligated to, for example, if there are no pending callbacks.
1260 void rcu_barrier_tasks_trace(void)
1262 /* There is only one callback queue, so this is easy. ;-) */
1263 synchronize_rcu_tasks_trace();
1265 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1267 static int __init rcu_spawn_tasks_trace_kthread(void)
1269 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1270 rcu_tasks_trace.gp_sleep = HZ / 10;
1271 rcu_tasks_trace.init_fract = HZ / 10;
1273 rcu_tasks_trace.gp_sleep = HZ / 200;
1274 if (rcu_tasks_trace.gp_sleep <= 0)
1275 rcu_tasks_trace.gp_sleep = 1;
1276 rcu_tasks_trace.init_fract = HZ / 200;
1277 if (rcu_tasks_trace.init_fract <= 0)
1278 rcu_tasks_trace.init_fract = 1;
1280 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1281 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1282 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1283 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1284 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1285 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1289 #if !defined(CONFIG_TINY_RCU)
1290 void show_rcu_tasks_trace_gp_kthread(void)
1294 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1295 data_race(n_heavy_reader_ofl_updates),
1296 data_race(n_heavy_reader_updates),
1297 data_race(n_heavy_reader_attempts));
1298 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1300 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1301 #endif // !defined(CONFIG_TINY_RCU)
1303 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1304 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1305 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1307 #ifndef CONFIG_TINY_RCU
1308 void show_rcu_tasks_gp_kthreads(void)
1310 show_rcu_tasks_classic_gp_kthread();
1311 show_rcu_tasks_rude_gp_kthread();
1312 show_rcu_tasks_trace_gp_kthread();
1314 #endif /* #ifndef CONFIG_TINY_RCU */
1316 #ifdef CONFIG_PROVE_RCU
1317 struct rcu_tasks_test_desc {
1323 static struct rcu_tasks_test_desc tests[] = {
1325 .name = "call_rcu_tasks()",
1326 /* If not defined, the test is skipped. */
1327 .notrun = !IS_ENABLED(CONFIG_TASKS_RCU),
1330 .name = "call_rcu_tasks_rude()",
1331 /* If not defined, the test is skipped. */
1332 .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1335 .name = "call_rcu_tasks_trace()",
1336 /* If not defined, the test is skipped. */
1337 .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1341 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1343 struct rcu_tasks_test_desc *rttd =
1344 container_of(rhp, struct rcu_tasks_test_desc, rh);
1346 pr_info("Callback from %s invoked.\n", rttd->name);
1348 rttd->notrun = true;
1351 static void rcu_tasks_initiate_self_tests(void)
1353 pr_info("Running RCU-tasks wait API self tests\n");
1354 #ifdef CONFIG_TASKS_RCU
1355 synchronize_rcu_tasks();
1356 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1359 #ifdef CONFIG_TASKS_RUDE_RCU
1360 synchronize_rcu_tasks_rude();
1361 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1364 #ifdef CONFIG_TASKS_TRACE_RCU
1365 synchronize_rcu_tasks_trace();
1366 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1370 static int rcu_tasks_verify_self_tests(void)
1375 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1376 if (!tests[i].notrun) { // still hanging.
1377 pr_err("%s has been failed.\n", tests[i].name);
1387 late_initcall(rcu_tasks_verify_self_tests);
1388 #else /* #ifdef CONFIG_PROVE_RCU */
1389 static void rcu_tasks_initiate_self_tests(void) { }
1390 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1392 void __init rcu_init_tasks_generic(void)
1394 #ifdef CONFIG_TASKS_RCU
1395 rcu_spawn_tasks_kthread();
1398 #ifdef CONFIG_TASKS_RUDE_RCU
1399 rcu_spawn_tasks_rude_kthread();
1402 #ifdef CONFIG_TASKS_TRACE_RCU
1403 rcu_spawn_tasks_trace_kthread();
1406 // Run the self-tests.
1407 rcu_tasks_initiate_self_tests();
1410 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1411 static inline void rcu_tasks_bootup_oddness(void) {}
1412 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */