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 * 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 allowning 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 * @holdout_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, e.g., at core_initcall() time. */
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 /* Pre-grace-period preparation. */
382 static void rcu_tasks_pregp_step(void)
385 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
386 * to complete. Invoking synchronize_rcu() suffices because all
387 * these transitions occur with interrupts disabled. Without this
388 * synchronize_rcu(), a read-side critical section that started
389 * before the grace period might be incorrectly seen as having
390 * started after the grace period.
392 * This synchronize_rcu() also dispenses with the need for a
393 * memory barrier on the first store to t->rcu_tasks_holdout,
394 * as it forces the store to happen after the beginning of the
400 /* Per-task initial processing. */
401 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
403 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
405 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
406 WRITE_ONCE(t->rcu_tasks_holdout, true);
407 list_add(&t->rcu_tasks_holdout_list, hop);
411 /* Processing between scanning taskslist and draining the holdout list. */
412 static void rcu_tasks_postscan(struct list_head *hop)
415 * Wait for tasks that are in the process of exiting. This
416 * does only part of the job, ensuring that all tasks that were
417 * previously exiting reach the point where they have disabled
418 * preemption, allowing the later synchronize_rcu() to finish
421 synchronize_srcu(&tasks_rcu_exit_srcu);
424 /* See if tasks are still holding out, complain if so. */
425 static void check_holdout_task(struct task_struct *t,
426 bool needreport, bool *firstreport)
430 if (!READ_ONCE(t->rcu_tasks_holdout) ||
431 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
432 !READ_ONCE(t->on_rq) ||
433 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
434 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
435 WRITE_ONCE(t->rcu_tasks_holdout, false);
436 list_del_init(&t->rcu_tasks_holdout_list);
440 rcu_request_urgent_qs_task(t);
444 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
445 *firstreport = false;
448 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
449 t, ".I"[is_idle_task(t)],
450 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
451 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
452 t->rcu_tasks_idle_cpu, cpu);
456 /* Scan the holdout lists for tasks no longer holding out. */
457 static void check_all_holdout_tasks(struct list_head *hop,
458 bool needreport, bool *firstreport)
460 struct task_struct *t, *t1;
462 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
463 check_holdout_task(t, needreport, firstreport);
468 /* Finish off the Tasks-RCU grace period. */
469 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
472 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
473 * memory barriers prior to them in the schedule() path, memory
474 * reordering on other CPUs could cause their RCU-tasks read-side
475 * critical sections to extend past the end of the grace period.
476 * However, because these ->nvcsw updates are carried out with
477 * interrupts disabled, we can use synchronize_rcu() to force the
478 * needed ordering on all such CPUs.
480 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
481 * accesses to be within the grace period, avoiding the need for
482 * memory barriers for ->rcu_tasks_holdout accesses.
484 * In addition, this synchronize_rcu() waits for exiting tasks
485 * to complete their final preempt_disable() region of execution,
486 * cleaning up after the synchronize_srcu() above.
491 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
492 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
495 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
496 * @rhp: structure to be used for queueing the RCU updates.
497 * @func: actual callback function to be invoked after the grace period
499 * The callback function will be invoked some time after a full grace
500 * period elapses, in other words after all currently executing RCU
501 * read-side critical sections have completed. call_rcu_tasks() assumes
502 * that the read-side critical sections end at a voluntary context
503 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
504 * or transition to usermode execution. As such, there are no read-side
505 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
506 * this primitive is intended to determine that all tasks have passed
507 * through a safe state, not so much for data-strcuture synchronization.
509 * See the description of call_rcu() for more detailed information on
510 * memory ordering guarantees.
512 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
514 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
516 EXPORT_SYMBOL_GPL(call_rcu_tasks);
519 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
521 * Control will return to the caller some time after a full rcu-tasks
522 * grace period has elapsed, in other words after all currently
523 * executing rcu-tasks read-side critical sections have elapsed. These
524 * read-side critical sections are delimited by calls to schedule(),
525 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
526 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
528 * This is a very specialized primitive, intended only for a few uses in
529 * tracing and other situations requiring manipulation of function
530 * preambles and profiling hooks. The synchronize_rcu_tasks() function
531 * is not (yet) intended for heavy use from multiple CPUs.
533 * See the description of synchronize_rcu() for more detailed information
534 * on memory ordering guarantees.
536 void synchronize_rcu_tasks(void)
538 synchronize_rcu_tasks_generic(&rcu_tasks);
540 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
543 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
545 * Although the current implementation is guaranteed to wait, it is not
546 * obligated to, for example, if there are no pending callbacks.
548 void rcu_barrier_tasks(void)
550 /* There is only one callback queue, so this is easy. ;-) */
551 synchronize_rcu_tasks();
553 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
555 static int __init rcu_spawn_tasks_kthread(void)
557 rcu_tasks.gp_sleep = HZ / 10;
558 rcu_tasks.init_fract = HZ / 10;
559 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
560 rcu_tasks.pertask_func = rcu_tasks_pertask;
561 rcu_tasks.postscan_func = rcu_tasks_postscan;
562 rcu_tasks.holdouts_func = check_all_holdout_tasks;
563 rcu_tasks.postgp_func = rcu_tasks_postgp;
564 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
567 core_initcall(rcu_spawn_tasks_kthread);
569 #if !defined(CONFIG_TINY_RCU)
570 void show_rcu_tasks_classic_gp_kthread(void)
572 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
574 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
575 #endif // !defined(CONFIG_TINY_RCU)
577 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
578 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
581 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
585 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
586 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
588 struct task_struct *t = current;
591 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
593 exit_tasks_rcu_finish_trace(t);
596 #else /* #ifdef CONFIG_TASKS_RCU */
597 void exit_tasks_rcu_start(void) { }
598 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
599 #endif /* #else #ifdef CONFIG_TASKS_RCU */
601 #ifdef CONFIG_TASKS_RUDE_RCU
603 ////////////////////////////////////////////////////////////////////////
605 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
606 // passing an empty function to schedule_on_each_cpu(). This approach
607 // provides an asynchronous call_rcu_tasks_rude() API and batching
608 // of concurrent calls to the synchronous synchronize_rcu_rude() API.
609 // This sends IPIs far and wide and induces otherwise unnecessary context
610 // switches on all online CPUs, whether idle or not.
612 // Empty function to allow workqueues to force a context switch.
613 static void rcu_tasks_be_rude(struct work_struct *work)
617 // Wait for one rude RCU-tasks grace period.
618 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
620 rtp->n_ipis += cpumask_weight(cpu_online_mask);
621 schedule_on_each_cpu(rcu_tasks_be_rude);
624 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
625 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
629 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
630 * @rhp: structure to be used for queueing the RCU updates.
631 * @func: actual callback function to be invoked after the grace period
633 * The callback function will be invoked some time after a full grace
634 * period elapses, in other words after all currently executing RCU
635 * read-side critical sections have completed. call_rcu_tasks_rude()
636 * assumes that the read-side critical sections end at context switch,
637 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
638 * there are no read-side primitives analogous to rcu_read_lock() and
639 * rcu_read_unlock() because this primitive is intended to determine
640 * that all tasks have passed through a safe state, not so much for
641 * data-strcuture synchronization.
643 * See the description of call_rcu() for more detailed information on
644 * memory ordering guarantees.
646 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
648 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
650 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
653 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
655 * Control will return to the caller some time after a rude rcu-tasks
656 * grace period has elapsed, in other words after all currently
657 * executing rcu-tasks read-side critical sections have elapsed. These
658 * read-side critical sections are delimited by calls to schedule(),
659 * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
660 * anyway) cond_resched().
662 * This is a very specialized primitive, intended only for a few uses in
663 * tracing and other situations requiring manipulation of function preambles
664 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
665 * (yet) intended for heavy use from multiple CPUs.
667 * See the description of synchronize_rcu() for more detailed information
668 * on memory ordering guarantees.
670 void synchronize_rcu_tasks_rude(void)
672 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
674 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
677 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
679 * Although the current implementation is guaranteed to wait, it is not
680 * obligated to, for example, if there are no pending callbacks.
682 void rcu_barrier_tasks_rude(void)
684 /* There is only one callback queue, so this is easy. ;-) */
685 synchronize_rcu_tasks_rude();
687 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
689 static int __init rcu_spawn_tasks_rude_kthread(void)
691 rcu_tasks_rude.gp_sleep = HZ / 10;
692 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
695 core_initcall(rcu_spawn_tasks_rude_kthread);
697 #if !defined(CONFIG_TINY_RCU)
698 void show_rcu_tasks_rude_gp_kthread(void)
700 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
702 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
703 #endif // !defined(CONFIG_TINY_RCU)
704 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
706 ////////////////////////////////////////////////////////////////////////
708 // Tracing variant of Tasks RCU. This variant is designed to be used
709 // to protect tracing hooks, including those of BPF. This variant
712 // 1. Has explicit read-side markers to allow finite grace periods
713 // in the face of in-kernel loops for PREEMPT=n builds.
715 // 2. Protects code in the idle loop, exception entry/exit, and
716 // CPU-hotplug code paths, similar to the capabilities of SRCU.
718 // 3. Avoids expensive read-side instruction, having overhead similar
719 // to that of Preemptible RCU.
721 // There are of course downsides. The grace-period code can send IPIs to
722 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
723 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
724 // is a single callback queue guarded by a single lock, again, much as for
725 // Tasks RCU. If needed, these downsides can be at least partially remedied.
727 // Perhaps most important, this variant of RCU does not affect the vanilla
728 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
729 // readers can operate from idle, offline, and exception entry/exit in no
730 // way allows rcu_preempt and rcu_sched readers to also do so.
732 // The lockdep state must be outside of #ifdef to be useful.
733 #ifdef CONFIG_DEBUG_LOCK_ALLOC
734 static struct lock_class_key rcu_lock_trace_key;
735 struct lockdep_map rcu_trace_lock_map =
736 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
737 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
738 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
740 #ifdef CONFIG_TASKS_TRACE_RCU
742 static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
743 static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
745 // Record outstanding IPIs to each CPU. No point in sending two...
746 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
748 // The number of detections of task quiescent state relying on
749 // heavyweight readers executing explicit memory barriers.
750 static unsigned long n_heavy_reader_attempts;
751 static unsigned long n_heavy_reader_updates;
752 static unsigned long n_heavy_reader_ofl_updates;
754 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
755 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
759 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
760 * while the scheduler locks are held.
762 static void rcu_read_unlock_iw(struct irq_work *iwp)
766 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
768 /* If we are the last reader, wake up the grace-period kthread. */
769 void rcu_read_unlock_trace_special(struct task_struct *t, int nesting)
771 int nq = t->trc_reader_special.b.need_qs;
773 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
774 t->trc_reader_special.b.need_mb)
775 smp_mb(); // Pairs with update-side barriers.
776 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
778 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
779 WRITE_ONCE(t->trc_reader_nesting, nesting);
780 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
781 irq_work_queue(&rcu_tasks_trace_iw);
783 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
785 /* Add a task to the holdout list, if it is not already on the list. */
786 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
788 if (list_empty(&t->trc_holdout_list)) {
790 list_add(&t->trc_holdout_list, bhp);
794 /* Remove a task from the holdout list, if it is in fact present. */
795 static void trc_del_holdout(struct task_struct *t)
797 if (!list_empty(&t->trc_holdout_list)) {
798 list_del_init(&t->trc_holdout_list);
803 /* IPI handler to check task state. */
804 static void trc_read_check_handler(void *t_in)
806 struct task_struct *t = current;
807 struct task_struct *texp = t_in;
809 // If the task is no longer running on this CPU, leave.
810 if (unlikely(texp != t)) {
811 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
813 goto reset_ipi; // Already on holdout list, so will check later.
816 // If the task is not in a read-side critical section, and
817 // if this is the last reader, awaken the grace-period kthread.
818 if (likely(!t->trc_reader_nesting)) {
819 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
821 // Mark as checked after decrement to avoid false
822 // positives on the above WARN_ON_ONCE().
823 WRITE_ONCE(t->trc_reader_checked, true);
826 // If we are racing with an rcu_read_unlock_trace(), try again later.
827 if (unlikely(t->trc_reader_nesting < 0)) {
828 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
832 WRITE_ONCE(t->trc_reader_checked, true);
834 // Get here if the task is in a read-side critical section. Set
835 // its state so that it will awaken the grace-period kthread upon
836 // exit from that critical section.
837 WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
838 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
841 // Allow future IPIs to be sent on CPU and for task.
842 // Also order this IPI handler against any later manipulations of
843 // the intended task.
844 smp_store_release(&per_cpu(trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
845 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
848 /* Callback function for scheduler to check locked-down task. */
849 static bool trc_inspect_reader(struct task_struct *t, void *arg)
851 int cpu = task_cpu(t);
853 bool ofl = cpu_is_offline(cpu);
856 WARN_ON_ONCE(ofl && !is_idle_task(t));
858 // If no chance of heavyweight readers, do it the hard way.
859 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
862 // If heavyweight readers are enabled on the remote task,
863 // we can inspect its state despite its currently running.
864 // However, we cannot safely change its state.
865 n_heavy_reader_attempts++;
866 if (!ofl && // Check for "running" idle tasks on offline CPUs.
867 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
868 return false; // No quiescent state, do it the hard way.
869 n_heavy_reader_updates++;
871 n_heavy_reader_ofl_updates++;
874 in_qs = likely(!t->trc_reader_nesting);
877 // Mark as checked. Because this is called from the grace-period
878 // kthread, also remove the task from the holdout list.
879 t->trc_reader_checked = true;
883 return true; // Already in quiescent state, done!!!
885 // The task is in a read-side critical section, so set up its
886 // state so that it will awaken the grace-period kthread upon exit
887 // from that critical section.
888 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
889 WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
890 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
894 /* Attempt to extract the state for the specified task. */
895 static void trc_wait_for_one_reader(struct task_struct *t,
896 struct list_head *bhp)
900 // If a previous IPI is still in flight, let it complete.
901 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
904 // The current task had better be in a quiescent state.
906 t->trc_reader_checked = true;
908 WARN_ON_ONCE(t->trc_reader_nesting);
912 // Attempt to nail down the task for inspection.
914 if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) {
920 // If currently running, send an IPI, either way, add to list.
921 trc_add_holdout(t, bhp);
923 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
924 // The task is currently running, so try IPIing it.
927 // If there is already an IPI outstanding, let it happen.
928 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
931 atomic_inc(&trc_n_readers_need_end);
932 per_cpu(trc_ipi_to_cpu, cpu) = true;
933 t->trc_ipi_to_cpu = cpu;
934 rcu_tasks_trace.n_ipis++;
935 if (smp_call_function_single(cpu,
936 trc_read_check_handler, t, 0)) {
937 // Just in case there is some other reason for
938 // failure than the target CPU being offline.
939 rcu_tasks_trace.n_ipis_fails++;
940 per_cpu(trc_ipi_to_cpu, cpu) = false;
941 t->trc_ipi_to_cpu = cpu;
942 if (atomic_dec_and_test(&trc_n_readers_need_end)) {
950 /* Initialize for a new RCU-tasks-trace grace period. */
951 static void rcu_tasks_trace_pregp_step(void)
955 // Allow for fast-acting IPIs.
956 atomic_set(&trc_n_readers_need_end, 1);
958 // There shouldn't be any old IPIs, but...
959 for_each_possible_cpu(cpu)
960 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
962 // Disable CPU hotplug across the tasklist scan.
963 // This also waits for all readers in CPU-hotplug code paths.
967 /* Do first-round processing for the specified task. */
968 static void rcu_tasks_trace_pertask(struct task_struct *t,
969 struct list_head *hop)
971 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
972 WRITE_ONCE(t->trc_reader_checked, false);
973 t->trc_ipi_to_cpu = -1;
974 trc_wait_for_one_reader(t, hop);
978 * Do intermediate processing between task and holdout scans and
979 * pick up the idle tasks.
981 static void rcu_tasks_trace_postscan(struct list_head *hop)
985 for_each_possible_cpu(cpu)
986 rcu_tasks_trace_pertask(idle_task(cpu), hop);
988 // Re-enable CPU hotplug now that the tasklist scan has completed.
991 // Wait for late-stage exiting tasks to finish exiting.
992 // These might have passed the call to exit_tasks_rcu_finish().
994 // Any tasks that exit after this point will set ->trc_reader_checked.
997 /* Show the state of a task stalling the current RCU tasks trace GP. */
998 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1003 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1004 *firstreport = false;
1006 // FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1008 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1010 ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0],
1011 ".i"[is_idle_task(t)],
1012 ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)],
1013 t->trc_reader_nesting,
1014 " N"[!!t->trc_reader_special.b.need_qs],
1019 /* List stalled IPIs for RCU tasks trace. */
1020 static void show_stalled_ipi_trace(void)
1024 for_each_possible_cpu(cpu)
1025 if (per_cpu(trc_ipi_to_cpu, cpu))
1026 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1029 /* Do one scan of the holdout list. */
1030 static void check_all_holdout_tasks_trace(struct list_head *hop,
1031 bool needreport, bool *firstreport)
1033 struct task_struct *g, *t;
1035 // Disable CPU hotplug across the holdout list scan.
1038 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1039 // If safe and needed, try to check the current task.
1040 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1041 !READ_ONCE(t->trc_reader_checked))
1042 trc_wait_for_one_reader(t, hop);
1044 // If check succeeded, remove this task from the list.
1045 if (READ_ONCE(t->trc_reader_checked))
1047 else if (needreport)
1048 show_stalled_task_trace(t, firstreport);
1051 // Re-enable CPU hotplug now that the holdout list scan has completed.
1056 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1057 show_stalled_ipi_trace();
1061 /* Wait for grace period to complete and provide ordering. */
1062 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1065 struct task_struct *g, *t;
1066 LIST_HEAD(holdouts);
1069 // Remove the safety count.
1070 smp_mb__before_atomic(); // Order vs. earlier atomics
1071 atomic_dec(&trc_n_readers_need_end);
1072 smp_mb__after_atomic(); // Order vs. later atomics
1074 // Wait for readers.
1075 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1077 ret = wait_event_idle_exclusive_timeout(
1079 atomic_read(&trc_n_readers_need_end) == 0,
1080 READ_ONCE(rcu_task_stall_timeout));
1082 break; // Count reached zero.
1083 // Stall warning time, so make a list of the offenders.
1085 for_each_process_thread(g, t)
1086 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1087 trc_add_holdout(t, &holdouts);
1090 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1091 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1092 show_stalled_task_trace(t, &firstreport);
1093 trc_del_holdout(t); // Release task_struct reference.
1096 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1097 show_stalled_ipi_trace();
1098 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1100 smp_mb(); // Caller's code must be ordered after wakeup.
1101 // Pairs with pretty much every ordering primitive.
1104 /* Report any needed quiescent state for this exiting task. */
1105 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1107 WRITE_ONCE(t->trc_reader_checked, true);
1108 WARN_ON_ONCE(t->trc_reader_nesting);
1109 WRITE_ONCE(t->trc_reader_nesting, 0);
1110 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1111 rcu_read_unlock_trace_special(t, 0);
1115 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1116 * @rhp: structure to be used for queueing the RCU updates.
1117 * @func: actual callback function to be invoked after the grace period
1119 * The callback function will be invoked some time after a full grace
1120 * period elapses, in other words after all currently executing RCU
1121 * read-side critical sections have completed. call_rcu_tasks_trace()
1122 * assumes that the read-side critical sections end at context switch,
1123 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
1124 * there are no read-side primitives analogous to rcu_read_lock() and
1125 * rcu_read_unlock() because this primitive is intended to determine
1126 * that all tasks have passed through a safe state, not so much for
1127 * data-strcuture synchronization.
1129 * See the description of call_rcu() for more detailed information on
1130 * memory ordering guarantees.
1132 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1134 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1136 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1139 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1141 * Control will return to the caller some time after a trace rcu-tasks
1142 * grace period has elapsed, in other words after all currently executing
1143 * rcu-tasks read-side critical sections have elapsed. These read-side
1144 * critical sections are delimited by calls to rcu_read_lock_trace()
1145 * and rcu_read_unlock_trace().
1147 * This is a very specialized primitive, intended only for a few uses in
1148 * tracing and other situations requiring manipulation of function preambles
1149 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1150 * (yet) intended for heavy use from multiple CPUs.
1152 * See the description of synchronize_rcu() for more detailed information
1153 * on memory ordering guarantees.
1155 void synchronize_rcu_tasks_trace(void)
1157 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1158 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1160 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1163 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1165 * Although the current implementation is guaranteed to wait, it is not
1166 * obligated to, for example, if there are no pending callbacks.
1168 void rcu_barrier_tasks_trace(void)
1170 /* There is only one callback queue, so this is easy. ;-) */
1171 synchronize_rcu_tasks_trace();
1173 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1175 static int __init rcu_spawn_tasks_trace_kthread(void)
1177 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1178 rcu_tasks_trace.gp_sleep = HZ / 10;
1179 rcu_tasks_trace.init_fract = HZ / 10;
1181 rcu_tasks_trace.gp_sleep = HZ / 200;
1182 if (rcu_tasks_trace.gp_sleep <= 0)
1183 rcu_tasks_trace.gp_sleep = 1;
1184 rcu_tasks_trace.init_fract = HZ / 200;
1185 if (rcu_tasks_trace.init_fract <= 0)
1186 rcu_tasks_trace.init_fract = 1;
1188 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1189 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1190 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1191 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1192 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1193 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1196 core_initcall(rcu_spawn_tasks_trace_kthread);
1198 #if !defined(CONFIG_TINY_RCU)
1199 void show_rcu_tasks_trace_gp_kthread(void)
1203 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1204 data_race(n_heavy_reader_ofl_updates),
1205 data_race(n_heavy_reader_updates),
1206 data_race(n_heavy_reader_attempts));
1207 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1209 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1210 #endif // !defined(CONFIG_TINY_RCU)
1212 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1213 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1214 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1216 #ifndef CONFIG_TINY_RCU
1217 void show_rcu_tasks_gp_kthreads(void)
1219 show_rcu_tasks_classic_gp_kthread();
1220 show_rcu_tasks_rude_gp_kthread();
1221 show_rcu_tasks_trace_gp_kthread();
1223 #endif /* #ifndef CONFIG_TINY_RCU */
1225 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1226 static inline void rcu_tasks_bootup_oddness(void) {}
1227 void show_rcu_tasks_gp_kthreads(void) {}
1228 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */